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 int 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 int 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 0;
 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 1;
 278 }
 279
 280 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 281 {
 282         char *p1, *p2, *p3, *p4;
 283         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 284         int i;
 285
 286         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 287
 288         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 289                 const int i2p = is_power_of_2(rs->kb_base);
 290
 291                 if (!rs->max_run[i])
 292                         continue;
 293
 294                 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
 295                 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
 296                 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 297                 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 298
 299                 log_buf(out, "%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
 300                          " mint=%llumsec, maxt=%llumsec\n",
 301                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 302                                 p1, p2, p3, p4,
 303                                 (unsigned long long) rs->min_run[i],
 304                                 (unsigned long long) rs->max_run[i]);
 305
 306                 free(p1);
 307                 free(p2);
 308                 free(p3);
 309                 free(p4);
 310         }
 311 }
 312
 313 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 314 {
 315         int i;
 316
 317         /*
 318          * Do depth distribution calculations
 319          */
 320         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 321                 if (total) {
 322                         io_u_dist[i] = (double) map[i] / (double) total;
 323                         io_u_dist[i] *= 100.0;
 324                         if (io_u_dist[i] < 0.1 && map[i])
 325                                 io_u_dist[i] = 0.1;
 326                 } else
 327                         io_u_dist[i] = 0.0;
 328         }
 329 }
 330
 331 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 332                           unsigned int *src, int nr)
 333 {
 334         unsigned long total = ddir_rw_sum(ts->total_io_u);
 335         int i;
 336
 337         /*
 338          * Do latency distribution calculations
 339          */
 340         for (i = 0; i < nr; i++) {
 341                 if (total) {
 342                         dst[i] = (double) src[i] / (double) total;
 343                         dst[i] *= 100.0;
 344                         if (dst[i] < 0.01 && src[i])
 345                                 dst[i] = 0.01;
 346                 } else
 347                         dst[i] = 0.0;
 348         }
 349 }
 350
 351 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 352 {
 353         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 354 }
 355
 356 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 357 {
 358         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 359 }
 360
 361 static void display_lat(const char *name, unsigned long min, unsigned long max,
 362                         double mean, double dev, struct buf_output *out)
 363 {
 364         const char *base = "(usec)";
 365         char *minp, *maxp;
 366
 367         if (!usec_to_msec(&min, &max, &mean, &dev))
 368                 base = "(msec)";
 369
 370         minp = num2str(min, 6, 1, 0, 0);
 371         maxp = num2str(max, 6, 1, 0, 0);
 372
 373         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 374                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 375
 376         free(minp);
 377         free(maxp);
 378 }
 379
 380 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 381                              int ddir, struct buf_output *out)
 382 {
 383         const char *str[] = { "read ", "write", "trim" };
 384         unsigned long min, max, runt;
 385         unsigned long long bw, iops;
 386         double mean, dev;
 387         char *io_p, *bw_p, *iops_p;
 388         int i2p;
 389
 390         assert(ddir_rw(ddir));
 391
 392         if (!ts->runtime[ddir])
 393                 return;
 394
 395         i2p = is_power_of_2(rs->kb_base);
 396         runt = ts->runtime[ddir];
 397
 398         bw = (1000 * ts->io_bytes[ddir]) / runt;
 399         io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
 400         bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
 401
 402         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 403         iops_p = num2str(iops, 6, 1, 0, 0);
 404
 405         log_buf(out, "  %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
 406                                 rs->unified_rw_rep ? "mixed" : str[ddir],
 407                                 io_p, bw_p, iops_p,
 408                                 (unsigned long long) ts->runtime[ddir]);
 409
 410         free(io_p);
 411         free(bw_p);
 412         free(iops_p);
 413
 414         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 415                 display_lat("slat", min, max, mean, dev, out);
 416         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 417                 display_lat("clat", min, max, mean, dev, out);
 418         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 419                 display_lat(" lat", min, max, mean, dev, out);
 420
 421         if (ts->clat_percentiles) {
 422                 show_clat_percentiles(ts->io_u_plat[ddir],
 423                                         ts->clat_stat[ddir].samples,
 424                                         ts->percentile_list,
 425                                         ts->percentile_precision, out);
 426         }
 427         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 428                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 429                 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
 430
 431                 if (rs->unit_base == 1) {
 432                         min *= 8.0;
 433                         max *= 8.0;
 434                         mean *= 8.0;
 435                         dev *= 8.0;
 436                 }
 437
 438                 if (rs->agg[ddir]) {
 439                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 440                         if (p_of_agg > 100.0)
 441                                 p_of_agg = 100.0;
 442                 }
 443
 444                 if (mean > fkb_base * fkb_base) {
 445                         min /= fkb_base;
 446                         max /= fkb_base;
 447                         mean /= fkb_base;
 448                         dev /= fkb_base;
 449                         bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
 450                 }
 451
 452                 log_buf(out, "    bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
 453                          " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
 454                                                         p_of_agg, mean, dev);
 455         }
 456 }
 457
 458 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 459                     const char *msg, struct buf_output *out)
 460 {
 461         int new_line = 1, i, line = 0, shown = 0;
 462
 463         for (i = 0; i < nr; i++) {
 464                 if (io_u_lat[i] <= 0.0)
 465                         continue;
 466                 shown = 1;
 467                 if (new_line) {
 468                         if (line)
 469                                 log_buf(out, "\n");
 470                         log_buf(out, "    lat (%s) : ", msg);
 471                         new_line = 0;
 472                         line = 0;
 473                 }
 474                 if (line)
 475                         log_buf(out, ", ");
 476                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 477                 line++;
 478                 if (line == 5)
 479                         new_line = 1;
 480         }
 481
 482         if (shown)
 483                 log_buf(out, "\n");
 484
 485         return shown;
 486 }
 487
 488 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 489 {
 490         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 491                                  "250=", "500=", "750=", "1000=", };
 492
 493         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 494 }
 495
 496 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 497 {
 498         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 499                                  "250=", "500=", "750=", "1000=", "2000=",
 500                                  ">=2000=", };
 501
 502         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 503 }
 504
 505 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 506 {
 507         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 508         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 509
 510         stat_calc_lat_u(ts, io_u_lat_u);
 511         stat_calc_lat_m(ts, io_u_lat_m);
 512
 513         show_lat_u(io_u_lat_u, out);
 514         show_lat_m(io_u_lat_m, out);
 515 }
 516
 517 static int block_state_category(int block_state)
 518 {
 519         switch (block_state) {
 520         case BLOCK_STATE_UNINIT:
 521                 return 0;
 522         case BLOCK_STATE_TRIMMED:
 523         case BLOCK_STATE_WRITTEN:
 524                 return 1;
 525         case BLOCK_STATE_WRITE_FAILURE:
 526         case BLOCK_STATE_TRIM_FAILURE:
 527                 return 2;
 528         default:
 529                 /* Silence compile warning on some BSDs and have a return */
 530                 assert(0);
 531                 return -1;
 532         }
 533 }
 534
 535 static int compare_block_infos(const void *bs1, const void *bs2)
 536 {
 537         uint32_t block1 = *(uint32_t *)bs1;
 538         uint32_t block2 = *(uint32_t *)bs2;
 539         int state1 = BLOCK_INFO_STATE(block1);
 540         int state2 = BLOCK_INFO_STATE(block2);
 541         int bscat1 = block_state_category(state1);
 542         int bscat2 = block_state_category(state2);
 543         int cycles1 = BLOCK_INFO_TRIMS(block1);
 544         int cycles2 = BLOCK_INFO_TRIMS(block2);
 545
 546         if (bscat1 < bscat2)
 547                 return -1;
 548         if (bscat1 > bscat2)
 549                 return 1;
 550
 551         if (cycles1 < cycles2)
 552                 return -1;
 553         if (cycles1 > cycles2)
 554                 return 1;
 555
 556         if (state1 < state2)
 557                 return -1;
 558         if (state1 > state2)
 559                 return 1;
 560
 561         assert(block1 == block2);
 562         return 0;
 563 }
 564
 565 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 566                                   fio_fp64_t *plist, unsigned int **percentiles,
 567                                   unsigned int *types)
 568 {
 569         int len = 0;
 570         int i, nr_uninit;
 571
 572         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 573
 574         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 575                 len++;
 576
 577         if (!len)
 578                 return 0;
 579
 580         /*
 581          * Sort the percentile list. Note that it may already be sorted if
 582          * we are using the default values, but since it's a short list this
 583          * isn't a worry. Also note that this does not work for NaN values.
 584          */
 585         if (len > 1)
 586                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 587
 588         nr_uninit = 0;
 589         /* Start only after the uninit entries end */
 590         for (nr_uninit = 0;
 591              nr_uninit < nr_block_infos
 592                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 593              nr_uninit ++)
 594                 ;
 595
 596         if (nr_uninit == nr_block_infos)
 597                 return 0;
 598
 599         *percentiles = calloc(len, sizeof(**percentiles));
 600
 601         for (i = 0; i < len; i++) {
 602                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 603                                 + nr_uninit;
 604                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 605         }
 606
 607         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 608         for (i = 0; i < nr_block_infos; i++)
 609                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 610
 611         return len;
 612 }
 613
 614 static const char *block_state_names[] = {
 615         [BLOCK_STATE_UNINIT] = "unwritten",
 616         [BLOCK_STATE_TRIMMED] = "trimmed",
 617         [BLOCK_STATE_WRITTEN] = "written",
 618         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 619         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 620 };
 621
 622 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 623                              fio_fp64_t *plist, struct buf_output *out)
 624 {
 625         int len, pos, i;
 626         unsigned int *percentiles = NULL;
 627         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 628
 629         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 630                                      &percentiles, block_state_counts);
 631
 632         log_buf(out, "  block lifetime percentiles :\n   |");
 633         pos = 0;
 634         for (i = 0; i < len; i++) {
 635                 uint32_t block_info = percentiles[i];
 636 #define LINE_LENGTH     75
 637                 char str[LINE_LENGTH];
 638                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 639                                      plist[i].u.f, block_info,
 640                                      i == len - 1 ? '\n' : ',');
 641                 assert(strln < LINE_LENGTH);
 642                 if (pos + strln > LINE_LENGTH) {
 643                         pos = 0;
 644                         log_buf(out, "\n   |");
 645                 }
 646                 log_buf(out, "%s", str);
 647                 pos += strln;
 648 #undef LINE_LENGTH
 649         }
 650         if (percentiles)
 651                 free(percentiles);
 652
 653         log_buf(out, "        states               :");
 654         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 655                 log_buf(out, " %s=%u%c",
 656                          block_state_names[i], block_state_counts[i],
 657                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 658 }
 659
 660 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 661 {
 662         char *p1, *p2;
 663         struct steadystate_data *ss = ts->ss;
 664         unsigned long long bw_mean, iops_mean;
 665         const int i2p = is_power_of_2(ts->kb_base);
 666
 667         if (!ss->state)
 668                 return;
 669
 670         bw_mean = steadystate_bw_mean(ss);
 671         iops_mean = steadystate_iops_mean(ss);
 672
 673         p1 = num2str(bw_mean / ts->kb_base, 6, ts->kb_base, i2p, ts->unit_base);
 674         p2 = num2str(iops_mean, 6, 1, 0, 0);
 675
 676         log_buf(out, "  steadystate  : attained=%s, bw=%s/s, iops=%s, %s%s=%.3f%s\n",
 677                 ss->state & __FIO_SS_ATTAINED ? "yes" : "no",
 678                 p1, p2,
 679                 ss->state & __FIO_SS_IOPS ? "iops" : "bw",
 680                 ss->state & __FIO_SS_SLOPE ? " slope": " mean dev",
 681                 ss->criterion,
 682                 ss->state & __FIO_SS_PCT ? "%" : "");
 683
 684         free(p1);
 685         free(p2);
 686 }
 687
 688 static void show_thread_status_normal(struct thread_stat *ts,
 689                                       struct group_run_stats *rs,
 690                                       struct buf_output *out)
 691 {
 692         double usr_cpu, sys_cpu;
 693         unsigned long runtime;
 694         double io_u_dist[FIO_IO_U_MAP_NR];
 695         time_t time_p;
 696         char time_buf[32];
 697
 698         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 699                 return;
 700
 701         time(&time_p);
 702         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 703
 704         if (!ts->error) {
 705                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 706                                         ts->name, ts->groupid, ts->members,
 707                                         ts->error, (int) ts->pid, time_buf);
 708         } else {
 709                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 710                                         ts->name, ts->groupid, ts->members,
 711                                         ts->error, ts->verror, (int) ts->pid,
 712                                         time_buf);
 713         }
 714
 715         if (strlen(ts->description))
 716                 log_buf(out, "  Description  : [%s]\n", ts->description);
 717
 718         if (ts->io_bytes[DDIR_READ])
 719                 show_ddir_status(rs, ts, DDIR_READ, out);
 720         if (ts->io_bytes[DDIR_WRITE])
 721                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 722         if (ts->io_bytes[DDIR_TRIM])
 723                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 724
 725         show_latencies(ts, out);
 726
 727         runtime = ts->total_run_time;
 728         if (runtime) {
 729                 double runt = (double) runtime;
 730
 731                 usr_cpu = (double) ts->usr_time * 100 / runt;
 732                 sys_cpu = (double) ts->sys_time * 100 / runt;
 733         } else {
 734                 usr_cpu = 0;
 735                 sys_cpu = 0;
 736         }
 737
 738         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 739                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 740                         (unsigned long long) ts->ctx,
 741                         (unsigned long long) ts->majf,
 742                         (unsigned long long) ts->minf);
 743
 744         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 745         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 746                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 747                                         io_u_dist[1], io_u_dist[2],
 748                                         io_u_dist[3], io_u_dist[4],
 749                                         io_u_dist[5], io_u_dist[6]);
 750
 751         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 752         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 753                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 754                                         io_u_dist[1], io_u_dist[2],
 755                                         io_u_dist[3], io_u_dist[4],
 756                                         io_u_dist[5], io_u_dist[6]);
 757         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 758         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 759                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 760                                         io_u_dist[1], io_u_dist[2],
 761                                         io_u_dist[3], io_u_dist[4],
 762                                         io_u_dist[5], io_u_dist[6]);
 763         log_buf(out, "     issued    : total=r=%llu/w=%llu/d=%llu,"
 764                                  " short=r=%llu/w=%llu/d=%llu,"
 765                                  " drop=r=%llu/w=%llu/d=%llu\n",
 766                                         (unsigned long long) ts->total_io_u[0],
 767                                         (unsigned long long) ts->total_io_u[1],
 768                                         (unsigned long long) ts->total_io_u[2],
 769                                         (unsigned long long) ts->short_io_u[0],
 770                                         (unsigned long long) ts->short_io_u[1],
 771                                         (unsigned long long) ts->short_io_u[2],
 772                                         (unsigned long long) ts->drop_io_u[0],
 773                                         (unsigned long long) ts->drop_io_u[1],
 774                                         (unsigned long long) ts->drop_io_u[2]);
 775         if (ts->continue_on_error) {
 776                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 777                                         (unsigned long long)ts->total_err_count,
 778                                         ts->first_error,
 779                                         strerror(ts->first_error));
 780         }
 781         if (ts->latency_depth) {
 782                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 783                                         (unsigned long long)ts->latency_target,
 784                                         (unsigned long long)ts->latency_window,
 785                                         ts->latency_percentile.u.f,
 786                                         ts->latency_depth);
 787         }
 788
 789         if (ts->nr_block_infos)
 790                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 791                                   ts->percentile_list, out);
 792
 793         if (ts->ss)
 794                 show_ss_normal(ts, out);
 795 }
 796
 797 static void show_ddir_status_terse(struct thread_stat *ts,
 798                                    struct group_run_stats *rs, int ddir,
 799                                    struct buf_output *out)
 800 {
 801         unsigned long min, max;
 802         unsigned long long bw, iops;
 803         unsigned int *ovals = NULL;
 804         double mean, dev;
 805         unsigned int len, minv, maxv;
 806         int i;
 807
 808         assert(ddir_rw(ddir));
 809
 810         iops = bw = 0;
 811         if (ts->runtime[ddir]) {
 812                 uint64_t runt = ts->runtime[ddir];
 813
 814                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 815                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 816         }
 817
 818         log_buf(out, ";%llu;%llu;%llu;%llu",
 819                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 820                                         (unsigned long long) ts->runtime[ddir]);
 821
 822         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 823                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 824         else
 825                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 826
 827         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 828                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 829         else
 830                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 831
 832         if (ts->clat_percentiles) {
 833                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 834                                         ts->clat_stat[ddir].samples,
 835                                         ts->percentile_list, &ovals, &maxv,
 836                                         &minv);
 837         } else
 838                 len = 0;
 839
 840         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 841                 if (i >= len) {
 842                         log_buf(out, ";0%%=0");
 843                         continue;
 844                 }
 845                 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
 846         }
 847
 848         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 849                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 850         else
 851                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 852
 853         if (ovals)
 854                 free(ovals);
 855
 856         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 857                 double p_of_agg = 100.0;
 858
 859                 if (rs->agg[ddir]) {
 860                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 861                         if (p_of_agg > 100.0)
 862                                 p_of_agg = 100.0;
 863                 }
 864
 865                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 866         } else
 867                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 868 }
 869
 870 static void add_ddir_status_json(struct thread_stat *ts,
 871                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 872 {
 873         unsigned long min, max;
 874         unsigned long long bw;
 875         unsigned int *ovals = NULL;
 876         double mean, dev, iops;
 877         unsigned int len, minv, maxv;
 878         int i;
 879         const char *ddirname[] = {"read", "write", "trim"};
 880         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 881         char buf[120];
 882         double p_of_agg = 100.0;
 883
 884         assert(ddir_rw(ddir));
 885
 886         if (ts->unified_rw_rep && ddir != DDIR_READ)
 887                 return;
 888
 889         dir_object = json_create_object();
 890         json_object_add_value_object(parent,
 891                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 892
 893         bw = 0;
 894         iops = 0.0;
 895         if (ts->runtime[ddir]) {
 896                 uint64_t runt = ts->runtime[ddir];
 897
 898                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 899                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 900         }
 901
 902         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
 903         json_object_add_value_int(dir_object, "bw", bw);
 904         json_object_add_value_float(dir_object, "iops", iops);
 905         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 906         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 907         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 908         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 909
 910         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 911                 min = max = 0;
 912                 mean = dev = 0.0;
 913         }
 914         tmp_object = json_create_object();
 915         json_object_add_value_object(dir_object, "slat", tmp_object);
 916         json_object_add_value_int(tmp_object, "min", min);
 917         json_object_add_value_int(tmp_object, "max", max);
 918         json_object_add_value_float(tmp_object, "mean", mean);
 919         json_object_add_value_float(tmp_object, "stddev", dev);
 920
 921         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 922                 min = max = 0;
 923                 mean = dev = 0.0;
 924         }
 925         tmp_object = json_create_object();
 926         json_object_add_value_object(dir_object, "clat", tmp_object);
 927         json_object_add_value_int(tmp_object, "min", min);
 928         json_object_add_value_int(tmp_object, "max", max);
 929         json_object_add_value_float(tmp_object, "mean", mean);
 930         json_object_add_value_float(tmp_object, "stddev", dev);
 931
 932         if (ts->clat_percentiles) {
 933                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 934                                         ts->clat_stat[ddir].samples,
 935                                         ts->percentile_list, &ovals, &maxv,
 936                                         &minv);
 937         } else
 938                 len = 0;
 939
 940         percentile_object = json_create_object();
 941         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 942         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 943                 if (i >= len) {
 944                         json_object_add_value_int(percentile_object, "0.00", 0);
 945                         continue;
 946                 }
 947                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 948                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 949         }
 950
 951         if (output_format & FIO_OUTPUT_JSON_PLUS) {
 952                 clat_bins_object = json_create_object();
 953                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
 954                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
 955                         snprintf(buf, sizeof(buf), "%d", i);
 956                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
 957                 }
 958                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS);
 959                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL);
 960                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR);
 961         }
 962
 963         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
 964                 min = max = 0;
 965                 mean = dev = 0.0;
 966         }
 967         tmp_object = json_create_object();
 968         json_object_add_value_object(dir_object, "lat", tmp_object);
 969         json_object_add_value_int(tmp_object, "min", min);
 970         json_object_add_value_int(tmp_object, "max", max);
 971         json_object_add_value_float(tmp_object, "mean", mean);
 972         json_object_add_value_float(tmp_object, "stddev", dev);
 973         if (ovals)
 974                 free(ovals);
 975
 976         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 977                 if (rs->agg[ddir]) {
 978                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 979                         if (p_of_agg > 100.0)
 980                                 p_of_agg = 100.0;
 981                 }
 982         } else {
 983                 min = max = 0;
 984                 p_of_agg = mean = dev = 0.0;
 985         }
 986         json_object_add_value_int(dir_object, "bw_min", min);
 987         json_object_add_value_int(dir_object, "bw_max", max);
 988         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
 989         json_object_add_value_float(dir_object, "bw_mean", mean);
 990         json_object_add_value_float(dir_object, "bw_dev", dev);
 991 }
 992
 993 static void show_thread_status_terse_v2(struct thread_stat *ts,
 994                                         struct group_run_stats *rs,
 995                                         struct buf_output *out)
 996 {
 997         double io_u_dist[FIO_IO_U_MAP_NR];
 998         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 999         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1000         double usr_cpu, sys_cpu;
1001         int i;
1002
1003         /* General Info */
1004         log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1005         /* Log Read Status */
1006         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1007         /* Log Write Status */
1008         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1009         /* Log Trim Status */
1010         show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1011
1012         /* CPU Usage */
1013         if (ts->total_run_time) {
1014                 double runt = (double) ts->total_run_time;
1015
1016                 usr_cpu = (double) ts->usr_time * 100 / runt;
1017                 sys_cpu = (double) ts->sys_time * 100 / runt;
1018         } else {
1019                 usr_cpu = 0;
1020                 sys_cpu = 0;
1021         }
1022
1023         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1024                                                 (unsigned long long) ts->ctx,
1025                                                 (unsigned long long) ts->majf,
1026                                                 (unsigned long long) ts->minf);
1027
1028         /* Calc % distribution of IO depths, usecond, msecond latency */
1029         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1030         stat_calc_lat_u(ts, io_u_lat_u);
1031         stat_calc_lat_m(ts, io_u_lat_m);
1032
1033         /* Only show fixed 7 I/O depth levels*/
1034         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1035                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1036                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1037
1038         /* Microsecond latency */
1039         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1040                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1041         /* Millisecond latency */
1042         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1043                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1044         /* Additional output if continue_on_error set - default off*/
1045         if (ts->continue_on_error)
1046                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1047         log_buf(out, "\n");
1048
1049         /* Additional output if description is set */
1050         if (strlen(ts->description))
1051                 log_buf(out, ";%s", ts->description);
1052
1053         log_buf(out, "\n");
1054 }
1055
1056 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1057                                            struct group_run_stats *rs, int ver,
1058                                            struct buf_output *out)
1059 {
1060         double io_u_dist[FIO_IO_U_MAP_NR];
1061         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1062         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1063         double usr_cpu, sys_cpu;
1064         int i;
1065
1066         /* General Info */
1067         log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1068                                         ts->name, ts->groupid, ts->error);
1069         /* Log Read Status */
1070         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1071         /* Log Write Status */
1072         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1073         /* Log Trim Status */
1074         if (ver == 4)
1075                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1076
1077         /* CPU Usage */
1078         if (ts->total_run_time) {
1079                 double runt = (double) ts->total_run_time;
1080
1081                 usr_cpu = (double) ts->usr_time * 100 / runt;
1082                 sys_cpu = (double) ts->sys_time * 100 / runt;
1083         } else {
1084                 usr_cpu = 0;
1085                 sys_cpu = 0;
1086         }
1087
1088         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1089                                                 (unsigned long long) ts->ctx,
1090                                                 (unsigned long long) ts->majf,
1091                                                 (unsigned long long) ts->minf);
1092
1093         /* Calc % distribution of IO depths, usecond, msecond latency */
1094         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1095         stat_calc_lat_u(ts, io_u_lat_u);
1096         stat_calc_lat_m(ts, io_u_lat_m);
1097
1098         /* Only show fixed 7 I/O depth levels*/
1099         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1100                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1101                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1102
1103         /* Microsecond latency */
1104         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1105                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1106         /* Millisecond latency */
1107         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1108                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1109
1110         /* disk util stats, if any */
1111         show_disk_util(1, NULL, out);
1112
1113         /* Additional output if continue_on_error set - default off*/
1114         if (ts->continue_on_error)
1115                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1116
1117         /* Additional output if description is set */
1118         if (strlen(ts->description))
1119                 log_buf(out, ";%s", ts->description);
1120
1121         log_buf(out, "\n");
1122 }
1123
1124 void json_add_job_opts(struct json_object *root, const char *name,
1125                        struct flist_head *opt_list, bool num_jobs)
1126 {
1127         struct json_object *dir_object;
1128         struct flist_head *entry;
1129         struct print_option *p;
1130
1131         if (flist_empty(opt_list))
1132                 return;
1133
1134         dir_object = json_create_object();
1135         json_object_add_value_object(root, name, dir_object);
1136
1137         flist_for_each(entry, opt_list) {
1138                 const char *pos = "";
1139
1140                 p = flist_entry(entry, struct print_option, list);
1141                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1142                         continue;
1143                 if (p->value)
1144                         pos = p->value;
1145                 json_object_add_value_string(dir_object, p->name, pos);
1146         }
1147 }
1148
1149 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1150                                                    struct group_run_stats *rs,
1151                                                    struct flist_head *opt_list)
1152 {
1153         struct json_object *root, *tmp;
1154         struct jobs_eta *je;
1155         double io_u_dist[FIO_IO_U_MAP_NR];
1156         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1157         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1158         double usr_cpu, sys_cpu;
1159         int i;
1160         size_t size;
1161
1162         root = json_create_object();
1163         json_object_add_value_string(root, "jobname", ts->name);
1164         json_object_add_value_int(root, "groupid", ts->groupid);
1165         json_object_add_value_int(root, "error", ts->error);
1166
1167         /* ETA Info */
1168         je = get_jobs_eta(true, &size);
1169         if (je) {
1170                 json_object_add_value_int(root, "eta", je->eta_sec);
1171                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1172         }
1173
1174         if (opt_list)
1175                 json_add_job_opts(root, "job options", opt_list, true);
1176
1177         add_ddir_status_json(ts, rs, DDIR_READ, root);
1178         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1179         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1180
1181         /* CPU Usage */
1182         if (ts->total_run_time) {
1183                 double runt = (double) ts->total_run_time;
1184
1185                 usr_cpu = (double) ts->usr_time * 100 / runt;
1186                 sys_cpu = (double) ts->sys_time * 100 / runt;
1187         } else {
1188                 usr_cpu = 0;
1189                 sys_cpu = 0;
1190         }
1191         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1192         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1193         json_object_add_value_int(root, "ctx", ts->ctx);
1194         json_object_add_value_int(root, "majf", ts->majf);
1195         json_object_add_value_int(root, "minf", ts->minf);
1196
1197
1198         /* Calc % distribution of IO depths, usecond, msecond latency */
1199         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1200         stat_calc_lat_u(ts, io_u_lat_u);
1201         stat_calc_lat_m(ts, io_u_lat_m);
1202
1203         tmp = json_create_object();
1204         json_object_add_value_object(root, "iodepth_level", tmp);
1205         /* Only show fixed 7 I/O depth levels*/
1206         for (i = 0; i < 7; i++) {
1207                 char name[20];
1208                 if (i < 6)
1209                         snprintf(name, 20, "%d", 1 << i);
1210                 else
1211                         snprintf(name, 20, ">=%d", 1 << i);
1212                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1213         }
1214
1215         tmp = json_create_object();
1216         json_object_add_value_object(root, "latency_us", tmp);
1217         /* Microsecond latency */
1218         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1219                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1220                                  "250", "500", "750", "1000", };
1221                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1222         }
1223         /* Millisecond latency */
1224         tmp = json_create_object();
1225         json_object_add_value_object(root, "latency_ms", tmp);
1226         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1227                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1228                                  "250", "500", "750", "1000", "2000",
1229                                  ">=2000", };
1230                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1231         }
1232
1233         /* Additional output if continue_on_error set - default off*/
1234         if (ts->continue_on_error) {
1235                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1236                 json_object_add_value_int(root, "first_error", ts->first_error);
1237         }
1238
1239         if (ts->latency_depth) {
1240                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1241                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1242                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1243                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1244         }
1245
1246         /* Additional output if description is set */
1247         if (strlen(ts->description))
1248                 json_object_add_value_string(root, "desc", ts->description);
1249
1250         if (ts->nr_block_infos) {
1251                 /* Block error histogram and types */
1252                 int len;
1253                 unsigned int *percentiles = NULL;
1254                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1255
1256                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1257                                              ts->percentile_list,
1258                                              &percentiles, block_state_counts);
1259
1260                 if (len) {
1261                         struct json_object *block, *percentile_object, *states;
1262                         int state;
1263                         block = json_create_object();
1264                         json_object_add_value_object(root, "block", block);
1265
1266                         percentile_object = json_create_object();
1267                         json_object_add_value_object(block, "percentiles",
1268                                                      percentile_object);
1269                         for (i = 0; i < len; i++) {
1270                                 char buf[20];
1271                                 snprintf(buf, sizeof(buf), "%f",
1272                                          ts->percentile_list[i].u.f);
1273                                 json_object_add_value_int(percentile_object,
1274                                                           (const char *)buf,
1275                                                           percentiles[i]);
1276                         }
1277
1278                         states = json_create_object();
1279                         json_object_add_value_object(block, "states", states);
1280                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1281                                 json_object_add_value_int(states,
1282                                         block_state_names[state],
1283                                         block_state_counts[state]);
1284                         }
1285                         free(percentiles);
1286                 }
1287         }
1288
1289         if (ts->ss) {
1290                 struct json_object *data;
1291                 struct json_array *iops, *bw;
1292                 struct steadystate_data *ss = ts->ss;
1293                 int i, j, k;
1294                 char ss_buf[64];
1295
1296                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1297                         ss->state & __FIO_SS_IOPS ? "iops" : "bw",
1298                         ss->state & __FIO_SS_SLOPE ? "_slope" : "",
1299                         (float) ss->limit,
1300                         ss->state & __FIO_SS_PCT ? "%" : "");
1301
1302                 tmp = json_create_object();
1303                 json_object_add_value_object(root, "steadystate", tmp);
1304                 json_object_add_value_string(tmp, "ss", ss_buf);
1305                 json_object_add_value_int(tmp, "duration", (int)ss->dur);
1306                 json_object_add_value_int(tmp, "steadystate_ramptime", ss->ramp_time / 1000000L);
1307                 json_object_add_value_int(tmp, "attained", (ss->state & __FIO_SS_ATTAINED) > 0);
1308
1309                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ss->criterion,
1310                         ss->state & __FIO_SS_PCT ? "%" : "");
1311                 json_object_add_value_string(tmp, "criterion", ss_buf);
1312                 json_object_add_value_float(tmp, "max_deviation", ss->deviation);
1313                 json_object_add_value_float(tmp, "slope", ss->slope);
1314
1315                 data = json_create_object();
1316                 json_object_add_value_object(tmp, "data", data);
1317                 bw = json_create_array();
1318                 iops = json_create_array();
1319
1320                 /*
1321                 ** if ss was attained or the buffer is not full,
1322                 ** ss->head points to the first element in the list.
1323                 ** otherwise it actually points to the second element
1324                 ** in the list
1325                 */
1326                 if ((ss->state & __FIO_SS_ATTAINED) || ss->sum_y == 0)
1327                         j = ss->head;
1328                 else
1329                         j = ss->head == 0 ? ss->dur - 1 : ss->head - 1;
1330                 for (i = 0; i < ss->dur; i++) {
1331                         k = (j + i) % ss->dur;
1332                         json_array_add_value_int(bw, ss->bw_data[k]);
1333                         json_array_add_value_int(iops, ss->iops_data[k]);
1334                 }
1335                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ss));
1336                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ss));
1337                 json_object_add_value_array(data, "iops", iops);
1338                 json_object_add_value_array(data, "bw", bw);
1339         }
1340
1341         return root;
1342 }
1343
1344 static void show_thread_status_terse(struct thread_stat *ts,
1345                                      struct group_run_stats *rs,
1346                                      struct buf_output *out)
1347 {
1348         if (terse_version == 2)
1349                 show_thread_status_terse_v2(ts, rs, out);
1350         else if (terse_version == 3 || terse_version == 4)
1351                 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1352         else
1353                 log_err("fio: bad terse version!? %d\n", terse_version);
1354 }
1355
1356 struct json_object *show_thread_status(struct thread_stat *ts,
1357                                        struct group_run_stats *rs,
1358                                        struct flist_head *opt_list,
1359                                        struct buf_output *out)
1360 {
1361         struct json_object *ret = NULL;
1362
1363         if (output_format & FIO_OUTPUT_TERSE)
1364                 show_thread_status_terse(ts, rs,  out);
1365         if (output_format & FIO_OUTPUT_JSON)
1366                 ret = show_thread_status_json(ts, rs, opt_list);
1367         if (output_format & FIO_OUTPUT_NORMAL)
1368                 show_thread_status_normal(ts, rs,  out);
1369
1370         return ret;
1371 }
1372
1373 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1374 {
1375         double mean, S;
1376
1377         if (src->samples == 0)
1378                 return;
1379
1380         dst->min_val = min(dst->min_val, src->min_val);
1381         dst->max_val = max(dst->max_val, src->max_val);
1382
1383         /*
1384          * Compute new mean and S after the merge
1385          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1386          *  #Parallel_algorithm>
1387          */
1388         if (first) {
1389                 mean = src->mean.u.f;
1390                 S = src->S.u.f;
1391         } else {
1392                 double delta = src->mean.u.f - dst->mean.u.f;
1393
1394                 mean = ((src->mean.u.f * src->samples) +
1395                         (dst->mean.u.f * dst->samples)) /
1396                         (dst->samples + src->samples);
1397
1398                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1399                         (dst->samples * src->samples) /
1400                         (dst->samples + src->samples);
1401         }
1402
1403         dst->samples += src->samples;
1404         dst->mean.u.f = mean;
1405         dst->S.u.f = S;
1406 }
1407
1408 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1409 {
1410         int i;
1411
1412         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1413                 if (dst->max_run[i] < src->max_run[i])
1414                         dst->max_run[i] = src->max_run[i];
1415                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1416                         dst->min_run[i] = src->min_run[i];
1417                 if (dst->max_bw[i] < src->max_bw[i])
1418                         dst->max_bw[i] = src->max_bw[i];
1419                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1420                         dst->min_bw[i] = src->min_bw[i];
1421
1422                 dst->io_kb[i] += src->io_kb[i];
1423                 dst->agg[i] += src->agg[i];
1424         }
1425
1426         if (!dst->kb_base)
1427                 dst->kb_base = src->kb_base;
1428         if (!dst->unit_base)
1429                 dst->unit_base = src->unit_base;
1430 }
1431
1432 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1433                       bool first)
1434 {
1435         int l, k;
1436
1437         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1438                 if (!dst->unified_rw_rep) {
1439                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1440                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1441                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1442                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1443
1444                         dst->io_bytes[l] += src->io_bytes[l];
1445
1446                         if (dst->runtime[l] < src->runtime[l])
1447                                 dst->runtime[l] = src->runtime[l];
1448                 } else {
1449                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1450                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1451                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1452                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1453
1454                         dst->io_bytes[0] += src->io_bytes[l];
1455
1456                         if (dst->runtime[0] < src->runtime[l])
1457                                 dst->runtime[0] = src->runtime[l];
1458
1459                         /*
1460                          * We're summing to the same destination, so override
1461                          * 'first' after the first iteration of the loop
1462                          */
1463                         first = false;
1464                 }
1465         }
1466
1467         dst->usr_time += src->usr_time;
1468         dst->sys_time += src->sys_time;
1469         dst->ctx += src->ctx;
1470         dst->majf += src->majf;
1471         dst->minf += src->minf;
1472
1473         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1474                 dst->io_u_map[k] += src->io_u_map[k];
1475         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1476                 dst->io_u_submit[k] += src->io_u_submit[k];
1477         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1478                 dst->io_u_complete[k] += src->io_u_complete[k];
1479         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1480                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1481         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1482                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1483
1484         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1485                 if (!dst->unified_rw_rep) {
1486                         dst->total_io_u[k] += src->total_io_u[k];
1487                         dst->short_io_u[k] += src->short_io_u[k];
1488                         dst->drop_io_u[k] += src->drop_io_u[k];
1489                 } else {
1490                         dst->total_io_u[0] += src->total_io_u[k];
1491                         dst->short_io_u[0] += src->short_io_u[k];
1492                         dst->drop_io_u[0] += src->drop_io_u[k];
1493                 }
1494         }
1495
1496         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1497                 int m;
1498
1499                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1500                         if (!dst->unified_rw_rep)
1501                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1502                         else
1503                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1504                 }
1505         }
1506
1507         dst->total_run_time += src->total_run_time;
1508         dst->total_submit += src->total_submit;
1509         dst->total_complete += src->total_complete;
1510 }
1511
1512 void init_group_run_stat(struct group_run_stats *gs)
1513 {
1514         int i;
1515         memset(gs, 0, sizeof(*gs));
1516
1517         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1518                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1519 }
1520
1521 void init_thread_stat(struct thread_stat *ts)
1522 {
1523         int j;
1524
1525         memset(ts, 0, sizeof(*ts));
1526
1527         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1528                 ts->lat_stat[j].min_val = -1UL;
1529                 ts->clat_stat[j].min_val = -1UL;
1530                 ts->slat_stat[j].min_val = -1UL;
1531                 ts->bw_stat[j].min_val = -1UL;
1532         }
1533         ts->groupid = -1;
1534 }
1535
1536 void __show_run_stats(void)
1537 {
1538         struct group_run_stats *runstats, *rs;
1539         struct thread_data *td;
1540         struct thread_stat *threadstats, *ts;
1541         int i, j, k, nr_ts, last_ts, idx;
1542         int kb_base_warned = 0;
1543         int unit_base_warned = 0;
1544         struct json_object *root = NULL;
1545         struct json_array *array = NULL;
1546         struct buf_output output[FIO_OUTPUT_NR];
1547         struct flist_head **opt_lists;
1548
1549         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1550
1551         for (i = 0; i < groupid + 1; i++)
1552                 init_group_run_stat(&runstats[i]);
1553
1554         /*
1555          * find out how many threads stats we need. if group reporting isn't
1556          * enabled, it's one-per-td.
1557          */
1558         nr_ts = 0;
1559         last_ts = -1;
1560         for_each_td(td, i) {
1561                 if (!td->o.group_reporting) {
1562                         nr_ts++;
1563                         continue;
1564                 }
1565                 if (last_ts == td->groupid)
1566                         continue;
1567
1568                 last_ts = td->groupid;
1569                 nr_ts++;
1570         }
1571
1572         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1573         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1574
1575         for (i = 0; i < nr_ts; i++) {
1576                 init_thread_stat(&threadstats[i]);
1577                 opt_lists[i] = NULL;
1578         }
1579
1580         j = 0;
1581         last_ts = -1;
1582         idx = 0;
1583         for_each_td(td, i) {
1584                 if (idx && (!td->o.group_reporting ||
1585                     (td->o.group_reporting && last_ts != td->groupid))) {
1586                         idx = 0;
1587                         j++;
1588                 }
1589
1590                 last_ts = td->groupid;
1591
1592                 ts = &threadstats[j];
1593
1594                 ts->clat_percentiles = td->o.clat_percentiles;
1595                 ts->percentile_precision = td->o.percentile_precision;
1596                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1597                 opt_lists[j] = &td->opt_list;
1598
1599                 idx++;
1600                 ts->members++;
1601
1602                 if (ts->groupid == -1) {
1603                         /*
1604                          * These are per-group shared already
1605                          */
1606                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1607                         if (td->o.description)
1608                                 strncpy(ts->description, td->o.description,
1609                                                 FIO_JOBDESC_SIZE - 1);
1610                         else
1611                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1612
1613                         /*
1614                          * If multiple entries in this group, this is
1615                          * the first member.
1616                          */
1617                         ts->thread_number = td->thread_number;
1618                         ts->groupid = td->groupid;
1619
1620                         /*
1621                          * first pid in group, not very useful...
1622                          */
1623                         ts->pid = td->pid;
1624
1625                         ts->kb_base = td->o.kb_base;
1626                         ts->unit_base = td->o.unit_base;
1627                         ts->unified_rw_rep = td->o.unified_rw_rep;
1628                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1629                         log_info("fio: kb_base differs for jobs in group, using"
1630                                  " %u as the base\n", ts->kb_base);
1631                         kb_base_warned = 1;
1632                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1633                         log_info("fio: unit_base differs for jobs in group, using"
1634                                  " %u as the base\n", ts->unit_base);
1635                         unit_base_warned = 1;
1636                 }
1637
1638                 ts->continue_on_error = td->o.continue_on_error;
1639                 ts->total_err_count += td->total_err_count;
1640                 ts->first_error = td->first_error;
1641                 if (!ts->error) {
1642                         if (!td->error && td->o.continue_on_error &&
1643                             td->first_error) {
1644                                 ts->error = td->first_error;
1645                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1646                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1647                         } else  if (td->error) {
1648                                 ts->error = td->error;
1649                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1650                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1651                         }
1652                 }
1653
1654                 ts->latency_depth = td->latency_qd;
1655                 ts->latency_target = td->o.latency_target;
1656                 ts->latency_percentile = td->o.latency_percentile;
1657                 ts->latency_window = td->o.latency_window;
1658
1659                 ts->nr_block_infos = td->ts.nr_block_infos;
1660                 for (k = 0; k < ts->nr_block_infos; k++)
1661                         ts->block_infos[k] = td->ts.block_infos[k];
1662
1663                 sum_thread_stats(ts, &td->ts, idx == 1);
1664
1665                 if (td->o.ss_dur)
1666                         ts->ss = &td->ss;
1667                 else
1668                         ts->ss = NULL;
1669         }
1670
1671         for (i = 0; i < nr_ts; i++) {
1672                 unsigned long long bw;
1673
1674                 ts = &threadstats[i];
1675                 if (ts->groupid == -1)
1676                         continue;
1677                 rs = &runstats[ts->groupid];
1678                 rs->kb_base = ts->kb_base;
1679                 rs->unit_base = ts->unit_base;
1680                 rs->unified_rw_rep += ts->unified_rw_rep;
1681
1682                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1683                         if (!ts->runtime[j])
1684                                 continue;
1685                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1686                                 rs->min_run[j] = ts->runtime[j];
1687                         if (ts->runtime[j] > rs->max_run[j])
1688                                 rs->max_run[j] = ts->runtime[j];
1689
1690                         bw = 0;
1691                         if (ts->runtime[j]) {
1692                                 unsigned long runt = ts->runtime[j];
1693                                 unsigned long long kb;
1694
1695                                 kb = ts->io_bytes[j] / rs->kb_base;
1696                                 bw = kb * 1000 / runt;
1697                         }
1698                         if (bw < rs->min_bw[j])
1699                                 rs->min_bw[j] = bw;
1700                         if (bw > rs->max_bw[j])
1701                                 rs->max_bw[j] = bw;
1702
1703                         rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1704                 }
1705         }
1706
1707         for (i = 0; i < groupid + 1; i++) {
1708                 int ddir;
1709
1710                 rs = &runstats[i];
1711
1712                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1713                         if (rs->max_run[ddir])
1714                                 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1715                                                 rs->max_run[ddir];
1716                 }
1717         }
1718
1719         for (i = 0; i < FIO_OUTPUT_NR; i++)
1720                 buf_output_init(&output[i]);
1721
1722         /*
1723          * don't overwrite last signal output
1724          */
1725         if (output_format & FIO_OUTPUT_NORMAL)
1726                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1727         if (output_format & FIO_OUTPUT_JSON) {
1728                 struct thread_data *global;
1729                 char time_buf[32];
1730                 struct timeval now;
1731                 unsigned long long ms_since_epoch;
1732
1733                 gettimeofday(&now, NULL);
1734                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1735                                  (unsigned long long)(now.tv_usec) / 1000;
1736
1737                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1738                                 sizeof(time_buf));
1739                 time_buf[strlen(time_buf) - 1] = '\0';
1740
1741                 root = json_create_object();
1742                 json_object_add_value_string(root, "fio version", fio_version_string);
1743                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1744                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1745                 json_object_add_value_string(root, "time", time_buf);
1746                 global = get_global_options();
1747                 json_add_job_opts(root, "global options", &global->opt_list, false);
1748                 array = json_create_array();
1749                 json_object_add_value_array(root, "jobs", array);
1750         }
1751
1752         if (is_backend)
1753                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1754
1755         for (i = 0; i < nr_ts; i++) {
1756                 ts = &threadstats[i];
1757                 rs = &runstats[ts->groupid];
1758
1759                 if (is_backend) {
1760                         fio_server_send_job_options(opt_lists[i], i);
1761                         fio_server_send_ts(ts, rs);
1762                 } else {
1763                         if (output_format & FIO_OUTPUT_TERSE)
1764                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1765                         if (output_format & FIO_OUTPUT_JSON) {
1766                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1767                                 json_array_add_value_object(array, tmp);
1768                         }
1769                         if (output_format & FIO_OUTPUT_NORMAL)
1770                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1771                 }
1772         }
1773         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1774                 /* disk util stats, if any */
1775                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1776
1777                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1778
1779                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1780                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1781                 json_free_object(root);
1782         }
1783
1784         for (i = 0; i < groupid + 1; i++) {
1785                 rs = &runstats[i];
1786
1787                 rs->groupid = i;
1788                 if (is_backend)
1789                         fio_server_send_gs(rs);
1790                 else if (output_format & FIO_OUTPUT_NORMAL)
1791                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1792         }
1793
1794         if (is_backend)
1795                 fio_server_send_du();
1796         else if (output_format & FIO_OUTPUT_NORMAL) {
1797                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1798                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1799         }
1800
1801         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1802                 buf_output_flush(&output[i]);
1803                 buf_output_free(&output[i]);
1804         }
1805
1806         log_info_flush();
1807         free(runstats);
1808         free(threadstats);
1809         free(opt_lists);
1810 }
1811
1812 void show_run_stats(void)
1813 {
1814         fio_mutex_down(stat_mutex);
1815         __show_run_stats();
1816         fio_mutex_up(stat_mutex);
1817 }
1818
1819 void __show_running_run_stats(void)
1820 {
1821         struct thread_data *td;
1822         unsigned long long *rt;
1823         struct timeval tv;
1824         int i;
1825
1826         fio_mutex_down(stat_mutex);
1827
1828         rt = malloc(thread_number * sizeof(unsigned long long));
1829         fio_gettime(&tv, NULL);
1830
1831         for_each_td(td, i) {
1832                 td->update_rusage = 1;
1833                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1834                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1835                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1836                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1837
1838                 rt[i] = mtime_since(&td->start, &tv);
1839                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1840                         td->ts.runtime[DDIR_READ] += rt[i];
1841                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1842                         td->ts.runtime[DDIR_WRITE] += rt[i];
1843                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1844                         td->ts.runtime[DDIR_TRIM] += rt[i];
1845         }
1846
1847         for_each_td(td, i) {
1848                 if (td->runstate >= TD_EXITED)
1849                         continue;
1850                 if (td->rusage_sem) {
1851                         td->update_rusage = 1;
1852                         fio_mutex_down(td->rusage_sem);
1853                 }
1854                 td->update_rusage = 0;
1855         }
1856
1857         __show_run_stats();
1858
1859         for_each_td(td, i) {
1860                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1861                         td->ts.runtime[DDIR_READ] -= rt[i];
1862                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1863                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1864                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1865                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1866         }
1867
1868         free(rt);
1869         fio_mutex_up(stat_mutex);
1870 }
1871
1872 static int status_interval_init;
1873 static struct timeval status_time;
1874 static int status_file_disabled;
1875
1876 #define FIO_STATUS_FILE         "fio-dump-status"
1877
1878 static int check_status_file(void)
1879 {
1880         struct stat sb;
1881         const char *temp_dir;
1882         char fio_status_file_path[PATH_MAX];
1883
1884         if (status_file_disabled)
1885                 return 0;
1886
1887         temp_dir = getenv("TMPDIR");
1888         if (temp_dir == NULL) {
1889                 temp_dir = getenv("TEMP");
1890                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1891                         temp_dir = NULL;
1892         }
1893         if (temp_dir == NULL)
1894                 temp_dir = "/tmp";
1895
1896         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1897
1898         if (stat(fio_status_file_path, &sb))
1899                 return 0;
1900
1901         if (unlink(fio_status_file_path) < 0) {
1902                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1903                                                         strerror(errno));
1904                 log_err("fio: disabling status file updates\n");
1905                 status_file_disabled = 1;
1906         }
1907
1908         return 1;
1909 }
1910
1911 void check_for_running_stats(void)
1912 {
1913         if (status_interval) {
1914                 if (!status_interval_init) {
1915                         fio_gettime(&status_time, NULL);
1916                         status_interval_init = 1;
1917                 } else if (mtime_since_now(&status_time) >= status_interval) {
1918                         show_running_run_stats();
1919                         fio_gettime(&status_time, NULL);
1920                         return;
1921                 }
1922         }
1923         if (check_status_file()) {
1924                 show_running_run_stats();
1925                 return;
1926         }
1927 }
1928
1929 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1930 {
1931         double val = data;
1932         double delta;
1933
1934         if (data > is->max_val)
1935                 is->max_val = data;
1936         if (data < is->min_val)
1937                 is->min_val = data;
1938
1939         delta = val - is->mean.u.f;
1940         if (delta) {
1941                 is->mean.u.f += delta / (is->samples + 1.0);
1942                 is->S.u.f += delta * (val - is->mean.u.f);
1943         }
1944
1945         is->samples++;
1946 }
1947
1948 /*
1949  * Return a struct io_logs, which is added to the tail of the log
1950  * list for 'iolog'.
1951  */
1952 static struct io_logs *get_new_log(struct io_log *iolog)
1953 {
1954         size_t new_size, new_samples;
1955         struct io_logs *cur_log;
1956
1957         /*
1958          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1959          * forever
1960          */
1961         if (!iolog->cur_log_max)
1962                 new_samples = DEF_LOG_ENTRIES;
1963         else {
1964                 new_samples = iolog->cur_log_max * 2;
1965                 if (new_samples > MAX_LOG_ENTRIES)
1966                         new_samples = MAX_LOG_ENTRIES;
1967         }
1968
1969         new_size = new_samples * log_entry_sz(iolog);
1970
1971         cur_log = smalloc(sizeof(*cur_log));
1972         if (cur_log) {
1973                 INIT_FLIST_HEAD(&cur_log->list);
1974                 cur_log->log = malloc(new_size);
1975                 if (cur_log->log) {
1976                         cur_log->nr_samples = 0;
1977                         cur_log->max_samples = new_samples;
1978                         flist_add_tail(&cur_log->list, &iolog->io_logs);
1979                         iolog->cur_log_max = new_samples;
1980                         return cur_log;
1981                 }
1982                 sfree(cur_log);
1983         }
1984
1985         return NULL;
1986 }
1987
1988 /*
1989  * Add and return a new log chunk, or return current log if big enough
1990  */
1991 static struct io_logs *regrow_log(struct io_log *iolog)
1992 {
1993         struct io_logs *cur_log;
1994         int i;
1995
1996         if (!iolog || iolog->disabled)
1997                 goto disable;
1998
1999         cur_log = iolog_cur_log(iolog);
2000         if (!cur_log) {
2001                 cur_log = get_new_log(iolog);
2002                 if (!cur_log)
2003                         return NULL;
2004         }
2005
2006         if (cur_log->nr_samples < cur_log->max_samples)
2007                 return cur_log;
2008
2009         /*
2010          * No room for a new sample. If we're compressing on the fly, flush
2011          * out the current chunk
2012          */
2013         if (iolog->log_gz) {
2014                 if (iolog_cur_flush(iolog, cur_log)) {
2015                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2016                         return NULL;
2017                 }
2018         }
2019
2020         /*
2021          * Get a new log array, and add to our list
2022          */
2023         cur_log = get_new_log(iolog);
2024         if (!cur_log) {
2025                 log_err("fio: failed extending iolog! Will stop logging.\n");
2026                 return NULL;
2027         }
2028
2029         if (!iolog->pending || !iolog->pending->nr_samples)
2030                 return cur_log;
2031
2032         /*
2033          * Flush pending items to new log
2034          */
2035         for (i = 0; i < iolog->pending->nr_samples; i++) {
2036                 struct io_sample *src, *dst;
2037
2038                 src = get_sample(iolog, iolog->pending, i);
2039                 dst = get_sample(iolog, cur_log, i);
2040                 memcpy(dst, src, log_entry_sz(iolog));
2041         }
2042         cur_log->nr_samples = iolog->pending->nr_samples;
2043
2044         iolog->pending->nr_samples = 0;
2045         return cur_log;
2046 disable:
2047         if (iolog)
2048                 iolog->disabled = true;
2049         return NULL;
2050 }
2051
2052 void regrow_logs(struct thread_data *td)
2053 {
2054         regrow_log(td->slat_log);
2055         regrow_log(td->clat_log);
2056         regrow_log(td->clat_hist_log);
2057         regrow_log(td->lat_log);
2058         regrow_log(td->bw_log);
2059         regrow_log(td->iops_log);
2060         td->flags &= ~TD_F_REGROW_LOGS;
2061 }
2062
2063 static struct io_logs *get_cur_log(struct io_log *iolog)
2064 {
2065         struct io_logs *cur_log;
2066
2067         cur_log = iolog_cur_log(iolog);
2068         if (!cur_log) {
2069                 cur_log = get_new_log(iolog);
2070                 if (!cur_log)
2071                         return NULL;
2072         }
2073
2074         if (cur_log->nr_samples < cur_log->max_samples)
2075                 return cur_log;
2076
2077         /*
2078          * Out of space. If we're in IO offload mode, or we're not doing
2079          * per unit logging (hence logging happens outside of the IO thread
2080          * as well), add a new log chunk inline. If we're doing inline
2081          * submissions, flag 'td' as needing a log regrow and we'll take
2082          * care of it on the submission side.
2083          */
2084         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2085             !per_unit_log(iolog))
2086                 return regrow_log(iolog);
2087
2088         iolog->td->flags |= TD_F_REGROW_LOGS;
2089         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2090         return iolog->pending;
2091 }
2092
2093 static void __add_log_sample(struct io_log *iolog, unsigned long val,
2094                              enum fio_ddir ddir, unsigned int bs,
2095                              unsigned long t, uint64_t offset)
2096 {
2097         struct io_logs *cur_log;
2098
2099         if (iolog->disabled)
2100                 return;
2101         if (flist_empty(&iolog->io_logs))
2102                 iolog->avg_last = t;
2103
2104         cur_log = get_cur_log(iolog);
2105         if (cur_log) {
2106                 struct io_sample *s;
2107
2108                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2109
2110                 s->val = val;
2111                 s->time = t;
2112                 io_sample_set_ddir(iolog, s, ddir);
2113                 s->bs = bs;
2114
2115                 if (iolog->log_offset) {
2116                         struct io_sample_offset *so = (void *) s;
2117
2118                         so->offset = offset;
2119                 }
2120
2121                 cur_log->nr_samples++;
2122                 return;
2123         }
2124
2125         iolog->disabled = true;
2126 }
2127
2128 static inline void reset_io_stat(struct io_stat *ios)
2129 {
2130         ios->max_val = ios->min_val = ios->samples = 0;
2131         ios->mean.u.f = ios->S.u.f = 0;
2132 }
2133
2134 void reset_io_stats(struct thread_data *td)
2135 {
2136         struct thread_stat *ts = &td->ts;
2137         int i, j;
2138
2139         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2140                 reset_io_stat(&ts->clat_stat[i]);
2141                 reset_io_stat(&ts->slat_stat[i]);
2142                 reset_io_stat(&ts->lat_stat[i]);
2143                 reset_io_stat(&ts->bw_stat[i]);
2144                 reset_io_stat(&ts->iops_stat[i]);
2145
2146                 ts->io_bytes[i] = 0;
2147                 ts->runtime[i] = 0;
2148
2149                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2150                         ts->io_u_plat[i][j] = 0;
2151         }
2152
2153         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2154                 ts->io_u_map[i] = 0;
2155                 ts->io_u_submit[i] = 0;
2156                 ts->io_u_complete[i] = 0;
2157                 ts->io_u_lat_u[i] = 0;
2158                 ts->io_u_lat_m[i] = 0;
2159                 ts->total_submit = 0;
2160                 ts->total_complete = 0;
2161         }
2162
2163         for (i = 0; i < 3; i++) {
2164                 ts->total_io_u[i] = 0;
2165                 ts->short_io_u[i] = 0;
2166                 ts->drop_io_u[i] = 0;
2167         }
2168 }
2169
2170 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2171                               unsigned long elapsed, bool log_max)
2172 {
2173         /*
2174          * Note an entry in the log. Use the mean from the logged samples,
2175          * making sure to properly round up. Only write a log entry if we
2176          * had actual samples done.
2177          */
2178         if (iolog->avg_window[ddir].samples) {
2179                 unsigned long val;
2180
2181                 if (log_max)
2182                         val = iolog->avg_window[ddir].max_val;
2183                 else
2184                         val = iolog->avg_window[ddir].mean.u.f + 0.50;
2185
2186                 __add_log_sample(iolog, val, ddir, 0, elapsed, 0);
2187         }
2188
2189         reset_io_stat(&iolog->avg_window[ddir]);
2190 }
2191
2192 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2193                              bool log_max)
2194 {
2195         int ddir;
2196
2197         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2198                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2199 }
2200
2201 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2202                            unsigned long val, enum fio_ddir ddir,
2203                            unsigned int bs, uint64_t offset)
2204 {
2205         unsigned long elapsed, this_window;
2206
2207         if (!ddir_rw(ddir))
2208                 return 0;
2209
2210         elapsed = mtime_since_now(&td->epoch);
2211
2212         /*
2213          * If no time averaging, just add the log sample.
2214          */
2215         if (!iolog->avg_msec) {
2216                 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
2217                 return 0;
2218         }
2219
2220         /*
2221          * Add the sample. If the time period has passed, then
2222          * add that entry to the log and clear.
2223          */
2224         add_stat_sample(&iolog->avg_window[ddir], val);
2225
2226         /*
2227          * If period hasn't passed, adding the above sample is all we
2228          * need to do.
2229          */
2230         this_window = elapsed - iolog->avg_last;
2231         if (elapsed < iolog->avg_last)
2232                 return iolog->avg_last - elapsed;
2233         else if (this_window < iolog->avg_msec) {
2234                 int diff = iolog->avg_msec - this_window;
2235
2236                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2237                         return diff;
2238         }
2239
2240         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2241
2242         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2243         return iolog->avg_msec;
2244 }
2245
2246 void finalize_logs(struct thread_data *td, bool unit_logs)
2247 {
2248         unsigned long elapsed;
2249
2250         elapsed = mtime_since_now(&td->epoch);
2251
2252         if (td->clat_log && unit_logs)
2253                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2254         if (td->slat_log && unit_logs)
2255                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2256         if (td->lat_log && unit_logs)
2257                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2258         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2259                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2260         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2261                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2262 }
2263
2264 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
2265 {
2266         struct io_log *iolog;
2267
2268         if (!ddir_rw(ddir))
2269                 return;
2270
2271         iolog = agg_io_log[ddir];
2272         __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
2273 }
2274
2275 static void add_clat_percentile_sample(struct thread_stat *ts,
2276                                 unsigned long usec, enum fio_ddir ddir)
2277 {
2278         unsigned int idx = plat_val_to_idx(usec);
2279         assert(idx < FIO_IO_U_PLAT_NR);
2280
2281         ts->io_u_plat[ddir][idx]++;
2282 }
2283
2284 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2285                      unsigned long usec, unsigned int bs, uint64_t offset)
2286 {
2287         unsigned long elapsed, this_window;
2288         struct thread_stat *ts = &td->ts;
2289         struct io_log *iolog = td->clat_hist_log;
2290
2291         td_io_u_lock(td);
2292
2293         add_stat_sample(&ts->clat_stat[ddir], usec);
2294
2295         if (td->clat_log)
2296                 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
2297
2298         if (ts->clat_percentiles)
2299                 add_clat_percentile_sample(ts, usec, ddir);
2300
2301         if (iolog && iolog->hist_msec) {
2302                 struct io_hist *hw = &iolog->hist_window[ddir];
2303
2304                 hw->samples++;
2305                 elapsed = mtime_since_now(&td->epoch);
2306                 if (!hw->hist_last)
2307                         hw->hist_last = elapsed;
2308                 this_window = elapsed - hw->hist_last;
2309
2310                 if (this_window >= iolog->hist_msec) {
2311                         unsigned int *io_u_plat;
2312                         unsigned int *dst;
2313
2314                         /*
2315                          * Make a byte-for-byte copy of the latency histogram
2316                          * stored in td->ts.io_u_plat[ddir], recording it in a
2317                          * log sample. Note that the matching call to free() is
2318                          * located in iolog.c after printing this sample to the
2319                          * log file.
2320                          */
2321                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2322                         dst = malloc(FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2323                         memcpy(dst, io_u_plat,
2324                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2325                         __add_log_sample(iolog, (unsigned long )dst, ddir, bs,
2326                                                 elapsed, offset);
2327
2328                         /*
2329                          * Update the last time we recorded as being now, minus
2330                          * any drift in time we encountered before actually
2331                          * making the record.
2332                          */
2333                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2334                         hw->samples = 0;
2335                 }
2336         }
2337
2338         td_io_u_unlock(td);
2339 }
2340
2341 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2342                      unsigned long usec, unsigned int bs, uint64_t offset)
2343 {
2344         struct thread_stat *ts = &td->ts;
2345
2346         if (!ddir_rw(ddir))
2347                 return;
2348
2349         td_io_u_lock(td);
2350
2351         add_stat_sample(&ts->slat_stat[ddir], usec);
2352
2353         if (td->slat_log)
2354                 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
2355
2356         td_io_u_unlock(td);
2357 }
2358
2359 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2360                     unsigned long usec, unsigned int bs, uint64_t offset)
2361 {
2362         struct thread_stat *ts = &td->ts;
2363
2364         if (!ddir_rw(ddir))
2365                 return;
2366
2367         td_io_u_lock(td);
2368
2369         add_stat_sample(&ts->lat_stat[ddir], usec);
2370
2371         if (td->lat_log)
2372                 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2373
2374         td_io_u_unlock(td);
2375 }
2376
2377 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2378                    unsigned int bytes, unsigned long spent)
2379 {
2380         struct thread_stat *ts = &td->ts;
2381         unsigned long rate;
2382
2383         if (spent)
2384                 rate = bytes * 1000 / spent;
2385         else
2386                 rate = 0;
2387
2388         td_io_u_lock(td);
2389
2390         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2391
2392         if (td->bw_log)
2393                 add_log_sample(td, td->bw_log, rate, io_u->ddir, bytes, io_u->offset);
2394
2395         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2396         td_io_u_unlock(td);
2397 }
2398
2399 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2400 {
2401         struct thread_stat *ts = &td->ts;
2402         unsigned long spent, rate;
2403         enum fio_ddir ddir;
2404         unsigned int next, next_log;
2405
2406         next_log = td->o.bw_avg_time;
2407
2408         spent = mtime_since(&td->bw_sample_time, t);
2409         if (spent < td->o.bw_avg_time &&
2410             td->o.bw_avg_time - spent >= LOG_MSEC_SLACK)
2411                 return td->o.bw_avg_time - spent;
2412
2413         td_io_u_lock(td);
2414
2415         /*
2416          * Compute both read and write rates for the interval.
2417          */
2418         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2419                 uint64_t delta;
2420
2421                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2422                 if (!delta)
2423                         continue; /* No entries for interval */
2424
2425                 if (spent)
2426                         rate = delta * 1000 / spent / 1024;
2427                 else
2428                         rate = 0;
2429
2430                 add_stat_sample(&ts->bw_stat[ddir], rate);
2431
2432                 if (td->bw_log) {
2433                         unsigned int bs = 0;
2434
2435                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2436                                 bs = td->o.min_bs[ddir];
2437
2438                         next = add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2439                         next_log = min(next_log, next);
2440                 }
2441
2442                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2443         }
2444
2445         timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time);
2446
2447         td_io_u_unlock(td);
2448
2449         if (spent <= td->o.bw_avg_time)
2450                 return min(next_log, td->o.bw_avg_time);
2451
2452         next = td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time);
2453         return min(next, next_log);
2454 }
2455
2456 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2457                      unsigned int bytes)
2458 {
2459         struct thread_stat *ts = &td->ts;
2460
2461         td_io_u_lock(td);
2462
2463         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2464
2465         if (td->iops_log)
2466                 add_log_sample(td, td->iops_log, 1, io_u->ddir, bytes, io_u->offset);
2467
2468         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2469         td_io_u_unlock(td);
2470 }
2471
2472 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2473 {
2474         struct thread_stat *ts = &td->ts;
2475         unsigned long spent, iops;
2476         enum fio_ddir ddir;
2477         unsigned int next, next_log;
2478
2479         next_log = td->o.iops_avg_time;
2480
2481         spent = mtime_since(&td->iops_sample_time, t);
2482         if (spent < td->o.iops_avg_time &&
2483             td->o.iops_avg_time - spent >= LOG_MSEC_SLACK)
2484                 return td->o.iops_avg_time - spent;
2485
2486         td_io_u_lock(td);
2487
2488         /*
2489          * Compute both read and write rates for the interval.
2490          */
2491         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2492                 uint64_t delta;
2493
2494                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2495                 if (!delta)
2496                         continue; /* No entries for interval */
2497
2498                 if (spent)
2499                         iops = (delta * 1000) / spent;
2500                 else
2501                         iops = 0;
2502
2503                 add_stat_sample(&ts->iops_stat[ddir], iops);
2504
2505                 if (td->iops_log) {
2506                         unsigned int bs = 0;
2507
2508                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2509                                 bs = td->o.min_bs[ddir];
2510
2511                         next = add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2512                         next_log = min(next_log, next);
2513                 }
2514
2515                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2516         }
2517
2518         timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time);
2519
2520         td_io_u_unlock(td);
2521
2522         if (spent <= td->o.iops_avg_time)
2523                 return min(next_log, td->o.iops_avg_time);
2524
2525         next = td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
2526         return min(next, next_log);
2527 }
2528
2529 /*
2530  * Returns msecs to next event
2531  */
2532 int calc_log_samples(void)
2533 {
2534         struct thread_data *td;
2535         unsigned int next = ~0U, tmp;
2536         struct timeval now;
2537         int i;
2538
2539         fio_gettime(&now, NULL);
2540
2541         for_each_td(td, i) {
2542                 if (in_ramp_time(td) ||
2543                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2544                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2545                         continue;
2546                 }
2547                 if (!per_unit_log(td->bw_log)) {
2548                         tmp = add_bw_samples(td, &now);
2549                         if (tmp < next)
2550                                 next = tmp;
2551                 }
2552                 if (!per_unit_log(td->iops_log)) {
2553                         tmp = add_iops_samples(td, &now);
2554                         if (tmp < next)
2555                                 next = tmp;
2556                 }
2557         }
2558
2559         return next == ~0U ? 0 : next;
2560 }
2561
2562 void stat_init(void)
2563 {
2564         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2565 }
2566
2567 void stat_exit(void)
2568 {
2569         /*
2570          * When we have the mutex, we know out-of-band access to it
2571          * have ended.
2572          */
2573         fio_mutex_down(stat_mutex);
2574         fio_mutex_remove(stat_mutex);
2575 }
2576
2577 /*
2578  * Called from signal handler. Wake up status thread.
2579  */
2580 void show_running_run_stats(void)
2581 {
2582         helper_do_stat();
2583 }
2584
2585 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2586 {
2587         /* Ignore io_u's which span multiple blocks--they will just get
2588          * inaccurate counts. */
2589         int idx = (io_u->offset - io_u->file->file_offset)
2590                         / td->o.bs[DDIR_TRIM];
2591         uint32_t *info = &td->ts.block_infos[idx];
2592         assert(idx < td->ts.nr_block_infos);
2593         return info;
2594 }