stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/stat.h>
   5 #include <math.h>
   6
   7 #include "fio.h"
   8 #include "diskutil.h"
   9 #include "lib/ieee754.h"
  10 #include "json.h"
  11 #include "lib/getrusage.h"
  12 #include "idletime.h"
  13 #include "lib/pow2.h"
  14 #include "lib/output_buffer.h"
  15 #include "helper_thread.h"
  16 #include "smalloc.h"
  17 #include "zbd.h"
  18 #include "oslib/asprintf.h"
  19
  20 #define LOG_MSEC_SLACK  1
  21
  22 struct fio_sem *stat_sem;
  23
  24 void clear_rusage_stat(struct thread_data *td)
  25 {
  26         struct thread_stat *ts = &td->ts;
  27
  28         fio_getrusage(&td->ru_start);
  29         ts->usr_time = ts->sys_time = 0;
  30         ts->ctx = 0;
  31         ts->minf = ts->majf = 0;
  32 }
  33
  34 void update_rusage_stat(struct thread_data *td)
  35 {
  36         struct thread_stat *ts = &td->ts;
  37
  38         fio_getrusage(&td->ru_end);
  39         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
  40                                         &td->ru_end.ru_utime);
  41         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
  42                                         &td->ru_end.ru_stime);
  43         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  44                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  45         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  46         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  47
  48         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  49 }
  50
  51 /*
  52  * Given a latency, return the index of the corresponding bucket in
  53  * the structure tracking percentiles.
  54  *
  55  * (1) find the group (and error bits) that the value (latency)
  56  * belongs to by looking at its MSB. (2) find the bucket number in the
  57  * group by looking at the index bits.
  58  *
  59  */
  60 static unsigned int plat_val_to_idx(unsigned long long val)
  61 {
  62         unsigned int msb, error_bits, base, offset, idx;
  63
  64         /* Find MSB starting from bit 0 */
  65         if (val == 0)
  66                 msb = 0;
  67         else
  68                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
  69
  70         /*
  71          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  72          * all bits of the sample as index
  73          */
  74         if (msb <= FIO_IO_U_PLAT_BITS)
  75                 return val;
  76
  77         /* Compute the number of error bits to discard*/
  78         error_bits = msb - FIO_IO_U_PLAT_BITS;
  79
  80         /* Compute the number of buckets before the group */
  81         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  82
  83         /*
  84          * Discard the error bits and apply the mask to find the
  85          * index for the buckets in the group
  86          */
  87         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  88
  89         /* Make sure the index does not exceed (array size - 1) */
  90         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  91                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  92
  93         return idx;
  94 }
  95
  96 /*
  97  * Convert the given index of the bucket array to the value
  98  * represented by the bucket
  99  */
 100 static unsigned long long plat_idx_to_val(unsigned int idx)
 101 {
 102         unsigned int error_bits;
 103         unsigned long long 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 = ((unsigned long long) 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(uint64_t *io_u_plat, unsigned long long nr,
 138                                    fio_fp64_t *plist, unsigned long long **output,
 139                                    unsigned long long *maxv, unsigned long long *minv)
 140 {
 141         unsigned long long sum = 0;
 142         unsigned int len, i, j = 0;
 143         unsigned long long *ovals = NULL;
 144         bool is_last;
 145
 146         *minv = -1ULL;
 147         *maxv = 0;
 148
 149         len = 0;
 150         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 151                 len++;
 152
 153         if (!len)
 154                 return 0;
 155
 156         /*
 157          * Sort the percentile list. Note that it may already be sorted if
 158          * we are using the default values, but since it's a short list this
 159          * isn't a worry. Also note that this does not work for NaN values.
 160          */
 161         if (len > 1)
 162                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 163
 164         ovals = malloc(len * sizeof(*ovals));
 165         if (!ovals)
 166                 return 0;
 167
 168         /*
 169          * Calculate bucket values, note down max and min values
 170          */
 171         is_last = false;
 172         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 173                 sum += io_u_plat[i];
 174                 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
 175                         assert(plist[j].u.f <= 100.0);
 176
 177                         ovals[j] = plat_idx_to_val(i);
 178                         if (ovals[j] < *minv)
 179                                 *minv = ovals[j];
 180                         if (ovals[j] > *maxv)
 181                                 *maxv = ovals[j];
 182
 183                         is_last = (j == len - 1) != 0;
 184                         if (is_last)
 185                                 break;
 186
 187                         j++;
 188                 }
 189         }
 190
 191         if (!is_last)
 192                 log_err("fio: error calculating latency percentiles\n");
 193
 194         *output = ovals;
 195         return len;
 196 }
 197
 198 /*
 199  * Find and display the p-th percentile of clat
 200  */
 201 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 202                                   fio_fp64_t *plist, unsigned int precision,
 203                                   const char *pre, struct buf_output *out)
 204 {
 205         unsigned int divisor, len, i, j = 0;
 206         unsigned long long minv, maxv;
 207         unsigned long long *ovals;
 208         int per_line, scale_down, time_width;
 209         bool is_last;
 210         char fmt[32];
 211
 212         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 213         if (!len || !ovals)
 214                 goto out;
 215
 216         /*
 217          * We default to nsecs, but if the value range is such that we
 218          * should scale down to usecs or msecs, do that.
 219          */
 220         if (minv > 2000000 && maxv > 99999999ULL) {
 221                 scale_down = 2;
 222                 divisor = 1000000;
 223                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 224         } else if (minv > 2000 && maxv > 99999) {
 225                 scale_down = 1;
 226                 divisor = 1000;
 227                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 228         } else {
 229                 scale_down = 0;
 230                 divisor = 1;
 231                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 232         }
 233
 234
 235         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 236         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 237                         precision, time_width);
 238         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 239         per_line = (80 - 7) / (precision + 10 + time_width);
 240
 241         for (j = 0; j < len; j++) {
 242                 /* for formatting */
 243                 if (j != 0 && (j % per_line) == 0)
 244                         log_buf(out, "     |");
 245
 246                 /* end of the list */
 247                 is_last = (j == len - 1) != 0;
 248
 249                 for (i = 0; i < scale_down; i++)
 250                         ovals[j] = (ovals[j] + 999) / 1000;
 251
 252                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 253
 254                 if (is_last)
 255                         break;
 256
 257                 if ((j % per_line) == per_line - 1)     /* for formatting */
 258                         log_buf(out, "\n");
 259         }
 260
 261 out:
 262         free(ovals);
 263 }
 264
 265 bool calc_lat(struct io_stat *is, unsigned long long *min,
 266               unsigned long long *max, double *mean, double *dev)
 267 {
 268         double n = (double) is->samples;
 269
 270         if (n == 0)
 271                 return false;
 272
 273         *min = is->min_val;
 274         *max = is->max_val;
 275         *mean = is->mean.u.f;
 276
 277         if (n > 1.0)
 278                 *dev = sqrt(is->S.u.f / (n - 1.0));
 279         else
 280                 *dev = 0;
 281
 282         return true;
 283 }
 284
 285 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 286 {
 287         char *io, *agg, *min, *max;
 288         char *ioalt, *aggalt, *minalt, *maxalt;
 289         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 290         int i;
 291
 292         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 293
 294         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 295                 const int i2p = is_power_of_2(rs->kb_base);
 296
 297                 if (!rs->max_run[i])
 298                         continue;
 299
 300                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
 301                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
 302                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
 303                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 304                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 305                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 306                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 307                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 308                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 309                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 310                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 311                                 (unsigned long long) rs->min_run[i],
 312                                 (unsigned long long) rs->max_run[i]);
 313
 314                 free(io);
 315                 free(agg);
 316                 free(min);
 317                 free(max);
 318                 free(ioalt);
 319                 free(aggalt);
 320                 free(minalt);
 321                 free(maxalt);
 322         }
 323 }
 324
 325 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
 326 {
 327         int i;
 328
 329         /*
 330          * Do depth distribution calculations
 331          */
 332         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 333                 if (total) {
 334                         io_u_dist[i] = (double) map[i] / (double) total;
 335                         io_u_dist[i] *= 100.0;
 336                         if (io_u_dist[i] < 0.1 && map[i])
 337                                 io_u_dist[i] = 0.1;
 338                 } else
 339                         io_u_dist[i] = 0.0;
 340         }
 341 }
 342
 343 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 344                           uint64_t *src, int nr)
 345 {
 346         unsigned long total = ddir_rw_sum(ts->total_io_u);
 347         int i;
 348
 349         /*
 350          * Do latency distribution calculations
 351          */
 352         for (i = 0; i < nr; i++) {
 353                 if (total) {
 354                         dst[i] = (double) src[i] / (double) total;
 355                         dst[i] *= 100.0;
 356                         if (dst[i] < 0.01 && src[i])
 357                                 dst[i] = 0.01;
 358                 } else
 359                         dst[i] = 0.0;
 360         }
 361 }
 362
 363 /*
 364  * To keep the terse format unaltered, add all of the ns latency
 365  * buckets to the first us latency bucket
 366  */
 367 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 368 {
 369         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 370         int i;
 371
 372         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 373
 374         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 375                 ntotal += ts->io_u_lat_n[i];
 376
 377         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 378 }
 379
 380 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 381 {
 382         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 383 }
 384
 385 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 386 {
 387         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 388 }
 389
 390 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 391 {
 392         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 393 }
 394
 395 static void display_lat(const char *name, unsigned long long min,
 396                         unsigned long long max, double mean, double dev,
 397                         struct buf_output *out)
 398 {
 399         const char *base = "(nsec)";
 400         char *minp, *maxp;
 401
 402         if (nsec_to_msec(&min, &max, &mean, &dev))
 403                 base = "(msec)";
 404         else if (nsec_to_usec(&min, &max, &mean, &dev))
 405                 base = "(usec)";
 406
 407         minp = num2str(min, 6, 1, 0, N2S_NONE);
 408         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 409
 410         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 411                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 412
 413         free(minp);
 414         free(maxp);
 415 }
 416
 417 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 418                              int ddir, struct buf_output *out)
 419 {
 420         unsigned long runt;
 421         unsigned long long min, max, bw, iops;
 422         double mean, dev;
 423         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
 424         int i2p;
 425
 426         if (ddir_sync(ddir)) {
 427                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
 428                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
 429                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
 430                         show_clat_percentiles(ts->io_u_sync_plat,
 431                                                 ts->sync_stat.samples,
 432                                                 ts->percentile_list,
 433                                                 ts->percentile_precision,
 434                                                 io_ddir_name(ddir), out);
 435                 }
 436                 return;
 437         }
 438
 439         assert(ddir_rw(ddir));
 440
 441         if (!ts->runtime[ddir])
 442                 return;
 443
 444         i2p = is_power_of_2(rs->kb_base);
 445         runt = ts->runtime[ddir];
 446
 447         bw = (1000 * ts->io_bytes[ddir]) / runt;
 448         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
 449         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
 450         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
 451
 452         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 453         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
 454         if (ddir == DDIR_WRITE)
 455                 post_st = zbd_write_status(ts);
 456         else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
 457                 uint64_t total;
 458                 double hit;
 459
 460                 total = ts->cachehit + ts->cachemiss;
 461                 hit = (double) ts->cachehit / (double) total;
 462                 hit *= 100.0;
 463                 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
 464                         post_st = NULL;
 465         }
 466
 467         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
 468                         rs->unified_rw_rep ? "mixed" : io_ddir_name(ddir),
 469                         iops_p, bw_p, bw_p_alt, io_p,
 470                         (unsigned long long) ts->runtime[ddir],
 471                         post_st ? : "");
 472
 473         free(post_st);
 474         free(io_p);
 475         free(bw_p);
 476         free(bw_p_alt);
 477         free(iops_p);
 478
 479         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 480                 display_lat("slat", min, max, mean, dev, out);
 481         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 482                 display_lat("clat", min, max, mean, dev, out);
 483         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 484                 display_lat(" lat", min, max, mean, dev, out);
 485         if (calc_lat(&ts->clat_high_prio_stat[ddir], &min, &max, &mean, &dev))
 486                 display_lat("prio_clat", min, max, mean, dev, out);
 487
 488         if (ts->clat_percentiles || ts->lat_percentiles) {
 489                 const char *name = ts->clat_percentiles ? "clat" : " lat";
 490                 char prio_name[32];
 491                 uint64_t samples;
 492
 493                 if (ts->clat_percentiles)
 494                         samples = ts->clat_stat[ddir].samples;
 495                 else
 496                         samples = ts->lat_stat[ddir].samples;
 497
 498                 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
 499                                         samples,
 500                                         ts->percentile_list,
 501                                         ts->percentile_precision, name, out);
 502
 503                 /* Only print this if some high and low priority stats were collected */
 504                 if (ts->clat_high_prio_stat[ddir].samples > 0 &&
 505                         ts->clat_prio_stat[ddir].samples > 0)
 506                 {
 507                         sprintf(prio_name, "high prio (%.2f%%) %s",
 508                                         100. * (double) ts->clat_high_prio_stat[ddir].samples / (double) samples,
 509                                         name);
 510                         show_clat_percentiles(ts->io_u_plat_high_prio[ddir],
 511                                                 ts->clat_high_prio_stat[ddir].samples,
 512                                                 ts->percentile_list,
 513                                                 ts->percentile_precision, prio_name, out);
 514
 515                         sprintf(prio_name, "low prio (%.2f%%) %s",
 516                                         100. * (double) ts->clat_prio_stat[ddir].samples / (double) samples,
 517                                         name);
 518                         show_clat_percentiles(ts->io_u_plat_prio[ddir],
 519                                                 ts->clat_prio_stat[ddir].samples,
 520                                                 ts->percentile_list,
 521                                                 ts->percentile_precision, prio_name, out);
 522                 }
 523         }
 524         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 525                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 526                 const char *bw_str;
 527
 528                 if ((rs->unit_base == 1) && i2p)
 529                         bw_str = "Kibit";
 530                 else if (rs->unit_base == 1)
 531                         bw_str = "kbit";
 532                 else if (i2p)
 533                         bw_str = "KiB";
 534                 else
 535                         bw_str = "kB";
 536
 537                 if (rs->agg[ddir]) {
 538                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 539                         if (p_of_agg > 100.0)
 540                                 p_of_agg = 100.0;
 541                 }
 542
 543                 if (rs->unit_base == 1) {
 544                         min *= 8.0;
 545                         max *= 8.0;
 546                         mean *= 8.0;
 547                         dev *= 8.0;
 548                 }
 549
 550                 if (mean > fkb_base * fkb_base) {
 551                         min /= fkb_base;
 552                         max /= fkb_base;
 553                         mean /= fkb_base;
 554                         dev /= fkb_base;
 555                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 556                 }
 557
 558                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 559                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 560                         bw_str, min, max, p_of_agg, mean, dev,
 561                         (&ts->bw_stat[ddir])->samples);
 562         }
 563         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 564                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 565                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 566                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 567         }
 568 }
 569
 570 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
 571                      const char *msg, struct buf_output *out)
 572 {
 573         bool new_line = true, shown = false;
 574         int i, line = 0;
 575
 576         for (i = 0; i < nr; i++) {
 577                 if (io_u_lat[i] <= 0.0)
 578                         continue;
 579                 shown = true;
 580                 if (new_line) {
 581                         if (line)
 582                                 log_buf(out, "\n");
 583                         log_buf(out, "  lat (%s)   : ", msg);
 584                         new_line = false;
 585                         line = 0;
 586                 }
 587                 if (line)
 588                         log_buf(out, ", ");
 589                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 590                 line++;
 591                 if (line == 5)
 592                         new_line = true;
 593         }
 594
 595         if (shown)
 596                 log_buf(out, "\n");
 597
 598         return true;
 599 }
 600
 601 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 602 {
 603         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 604                                  "250=", "500=", "750=", "1000=", };
 605
 606         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 607 }
 608
 609 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 610 {
 611         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 612                                  "250=", "500=", "750=", "1000=", };
 613
 614         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 615 }
 616
 617 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 618 {
 619         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 620                                  "250=", "500=", "750=", "1000=", "2000=",
 621                                  ">=2000=", };
 622
 623         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 624 }
 625
 626 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 627 {
 628         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 629         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 630         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 631
 632         stat_calc_lat_n(ts, io_u_lat_n);
 633         stat_calc_lat_u(ts, io_u_lat_u);
 634         stat_calc_lat_m(ts, io_u_lat_m);
 635
 636         show_lat_n(io_u_lat_n, out);
 637         show_lat_u(io_u_lat_u, out);
 638         show_lat_m(io_u_lat_m, out);
 639 }
 640
 641 static int block_state_category(int block_state)
 642 {
 643         switch (block_state) {
 644         case BLOCK_STATE_UNINIT:
 645                 return 0;
 646         case BLOCK_STATE_TRIMMED:
 647         case BLOCK_STATE_WRITTEN:
 648                 return 1;
 649         case BLOCK_STATE_WRITE_FAILURE:
 650         case BLOCK_STATE_TRIM_FAILURE:
 651                 return 2;
 652         default:
 653                 /* Silence compile warning on some BSDs and have a return */
 654                 assert(0);
 655                 return -1;
 656         }
 657 }
 658
 659 static int compare_block_infos(const void *bs1, const void *bs2)
 660 {
 661         uint64_t block1 = *(uint64_t *)bs1;
 662         uint64_t block2 = *(uint64_t *)bs2;
 663         int state1 = BLOCK_INFO_STATE(block1);
 664         int state2 = BLOCK_INFO_STATE(block2);
 665         int bscat1 = block_state_category(state1);
 666         int bscat2 = block_state_category(state2);
 667         int cycles1 = BLOCK_INFO_TRIMS(block1);
 668         int cycles2 = BLOCK_INFO_TRIMS(block2);
 669
 670         if (bscat1 < bscat2)
 671                 return -1;
 672         if (bscat1 > bscat2)
 673                 return 1;
 674
 675         if (cycles1 < cycles2)
 676                 return -1;
 677         if (cycles1 > cycles2)
 678                 return 1;
 679
 680         if (state1 < state2)
 681                 return -1;
 682         if (state1 > state2)
 683                 return 1;
 684
 685         assert(block1 == block2);
 686         return 0;
 687 }
 688
 689 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 690                                   fio_fp64_t *plist, unsigned int **percentiles,
 691                                   unsigned int *types)
 692 {
 693         int len = 0;
 694         int i, nr_uninit;
 695
 696         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 697
 698         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 699                 len++;
 700
 701         if (!len)
 702                 return 0;
 703
 704         /*
 705          * Sort the percentile list. Note that it may already be sorted if
 706          * we are using the default values, but since it's a short list this
 707          * isn't a worry. Also note that this does not work for NaN values.
 708          */
 709         if (len > 1)
 710                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 711
 712         /* Start only after the uninit entries end */
 713         for (nr_uninit = 0;
 714              nr_uninit < nr_block_infos
 715                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 716              nr_uninit ++)
 717                 ;
 718
 719         if (nr_uninit == nr_block_infos)
 720                 return 0;
 721
 722         *percentiles = calloc(len, sizeof(**percentiles));
 723
 724         for (i = 0; i < len; i++) {
 725                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 726                                 + nr_uninit;
 727                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 728         }
 729
 730         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 731         for (i = 0; i < nr_block_infos; i++)
 732                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 733
 734         return len;
 735 }
 736
 737 static const char *block_state_names[] = {
 738         [BLOCK_STATE_UNINIT] = "unwritten",
 739         [BLOCK_STATE_TRIMMED] = "trimmed",
 740         [BLOCK_STATE_WRITTEN] = "written",
 741         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 742         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 743 };
 744
 745 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 746                              fio_fp64_t *plist, struct buf_output *out)
 747 {
 748         int len, pos, i;
 749         unsigned int *percentiles = NULL;
 750         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 751
 752         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 753                                      &percentiles, block_state_counts);
 754
 755         log_buf(out, "  block lifetime percentiles :\n   |");
 756         pos = 0;
 757         for (i = 0; i < len; i++) {
 758                 uint32_t block_info = percentiles[i];
 759 #define LINE_LENGTH     75
 760                 char str[LINE_LENGTH];
 761                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 762                                      plist[i].u.f, block_info,
 763                                      i == len - 1 ? '\n' : ',');
 764                 assert(strln < LINE_LENGTH);
 765                 if (pos + strln > LINE_LENGTH) {
 766                         pos = 0;
 767                         log_buf(out, "\n   |");
 768                 }
 769                 log_buf(out, "%s", str);
 770                 pos += strln;
 771 #undef LINE_LENGTH
 772         }
 773         if (percentiles)
 774                 free(percentiles);
 775
 776         log_buf(out, "        states               :");
 777         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 778                 log_buf(out, " %s=%u%c",
 779                          block_state_names[i], block_state_counts[i],
 780                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 781 }
 782
 783 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 784 {
 785         char *p1, *p1alt, *p2;
 786         unsigned long long bw_mean, iops_mean;
 787         const int i2p = is_power_of_2(ts->kb_base);
 788
 789         if (!ts->ss_dur)
 790                 return;
 791
 792         bw_mean = steadystate_bw_mean(ts);
 793         iops_mean = steadystate_iops_mean(ts);
 794
 795         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
 796         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
 797         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
 798
 799         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 800                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
 801                 p1, p1alt, p2,
 802                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
 803                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
 804                 ts->ss_criterion.u.f,
 805                 ts->ss_state & FIO_SS_PCT ? "%" : "");
 806
 807         free(p1);
 808         free(p1alt);
 809         free(p2);
 810 }
 811
 812 static void show_agg_stats(struct disk_util_agg *agg, int terse,
 813                            struct buf_output *out)
 814 {
 815         if (!agg->slavecount)
 816                 return;
 817
 818         if (!terse) {
 819                 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
 820                          "aggrticks=%llu/%llu, aggrin_queue=%llu, "
 821                          "aggrutil=%3.2f%%",
 822                         (unsigned long long) agg->ios[0] / agg->slavecount,
 823                         (unsigned long long) agg->ios[1] / agg->slavecount,
 824                         (unsigned long long) agg->merges[0] / agg->slavecount,
 825                         (unsigned long long) agg->merges[1] / agg->slavecount,
 826                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 827                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 828                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 829                         agg->max_util.u.f);
 830         } else {
 831                 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
 832                         (unsigned long long) agg->ios[0] / agg->slavecount,
 833                         (unsigned long long) agg->ios[1] / agg->slavecount,
 834                         (unsigned long long) agg->merges[0] / agg->slavecount,
 835                         (unsigned long long) agg->merges[1] / agg->slavecount,
 836                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 837                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 838                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 839                         agg->max_util.u.f);
 840         }
 841 }
 842
 843 static void aggregate_slaves_stats(struct disk_util *masterdu)
 844 {
 845         struct disk_util_agg *agg = &masterdu->agg;
 846         struct disk_util_stat *dus;
 847         struct flist_head *entry;
 848         struct disk_util *slavedu;
 849         double util;
 850
 851         flist_for_each(entry, &masterdu->slaves) {
 852                 slavedu = flist_entry(entry, struct disk_util, slavelist);
 853                 dus = &slavedu->dus;
 854                 agg->ios[0] += dus->s.ios[0];
 855                 agg->ios[1] += dus->s.ios[1];
 856                 agg->merges[0] += dus->s.merges[0];
 857                 agg->merges[1] += dus->s.merges[1];
 858                 agg->sectors[0] += dus->s.sectors[0];
 859                 agg->sectors[1] += dus->s.sectors[1];
 860                 agg->ticks[0] += dus->s.ticks[0];
 861                 agg->ticks[1] += dus->s.ticks[1];
 862                 agg->time_in_queue += dus->s.time_in_queue;
 863                 agg->slavecount++;
 864
 865                 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
 866                 /* System utilization is the utilization of the
 867                  * component with the highest utilization.
 868                  */
 869                 if (util > agg->max_util.u.f)
 870                         agg->max_util.u.f = util;
 871
 872         }
 873
 874         if (agg->max_util.u.f > 100.0)
 875                 agg->max_util.u.f = 100.0;
 876 }
 877
 878 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
 879                      int terse, struct buf_output *out)
 880 {
 881         double util = 0;
 882
 883         if (dus->s.msec)
 884                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
 885         if (util > 100.0)
 886                 util = 100.0;
 887
 888         if (!terse) {
 889                 if (agg->slavecount)
 890                         log_buf(out, "  ");
 891
 892                 log_buf(out, "  %s: ios=%llu/%llu, merge=%llu/%llu, "
 893                          "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
 894                                 dus->name,
 895                                 (unsigned long long) dus->s.ios[0],
 896                                 (unsigned long long) dus->s.ios[1],
 897                                 (unsigned long long) dus->s.merges[0],
 898                                 (unsigned long long) dus->s.merges[1],
 899                                 (unsigned long long) dus->s.ticks[0],
 900                                 (unsigned long long) dus->s.ticks[1],
 901                                 (unsigned long long) dus->s.time_in_queue,
 902                                 util);
 903         } else {
 904                 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
 905                                 dus->name,
 906                                 (unsigned long long) dus->s.ios[0],
 907                                 (unsigned long long) dus->s.ios[1],
 908                                 (unsigned long long) dus->s.merges[0],
 909                                 (unsigned long long) dus->s.merges[1],
 910                                 (unsigned long long) dus->s.ticks[0],
 911                                 (unsigned long long) dus->s.ticks[1],
 912                                 (unsigned long long) dus->s.time_in_queue,
 913                                 util);
 914         }
 915
 916         /*
 917          * If the device has slaves, aggregate the stats for
 918          * those slave devices also.
 919          */
 920         show_agg_stats(agg, terse, out);
 921
 922         if (!terse)
 923                 log_buf(out, "\n");
 924 }
 925
 926 void json_array_add_disk_util(struct disk_util_stat *dus,
 927                 struct disk_util_agg *agg, struct json_array *array)
 928 {
 929         struct json_object *obj;
 930         double util = 0;
 931
 932         if (dus->s.msec)
 933                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
 934         if (util > 100.0)
 935                 util = 100.0;
 936
 937         obj = json_create_object();
 938         json_array_add_value_object(array, obj);
 939
 940         json_object_add_value_string(obj, "name", dus->name);
 941         json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
 942         json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
 943         json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
 944         json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
 945         json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
 946         json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
 947         json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
 948         json_object_add_value_float(obj, "util", util);
 949
 950         /*
 951          * If the device has slaves, aggregate the stats for
 952          * those slave devices also.
 953          */
 954         if (!agg->slavecount)
 955                 return;
 956         json_object_add_value_int(obj, "aggr_read_ios",
 957                                 agg->ios[0] / agg->slavecount);
 958         json_object_add_value_int(obj, "aggr_write_ios",
 959                                 agg->ios[1] / agg->slavecount);
 960         json_object_add_value_int(obj, "aggr_read_merges",
 961                                 agg->merges[0] / agg->slavecount);
 962         json_object_add_value_int(obj, "aggr_write_merge",
 963                                 agg->merges[1] / agg->slavecount);
 964         json_object_add_value_int(obj, "aggr_read_ticks",
 965                                 agg->ticks[0] / agg->slavecount);
 966         json_object_add_value_int(obj, "aggr_write_ticks",
 967                                 agg->ticks[1] / agg->slavecount);
 968         json_object_add_value_int(obj, "aggr_in_queue",
 969                                 agg->time_in_queue / agg->slavecount);
 970         json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
 971 }
 972
 973 static void json_object_add_disk_utils(struct json_object *obj,
 974                                        struct flist_head *head)
 975 {
 976         struct json_array *array = json_create_array();
 977         struct flist_head *entry;
 978         struct disk_util *du;
 979
 980         json_object_add_value_array(obj, "disk_util", array);
 981
 982         flist_for_each(entry, head) {
 983                 du = flist_entry(entry, struct disk_util, list);
 984
 985                 aggregate_slaves_stats(du);
 986                 json_array_add_disk_util(&du->dus, &du->agg, array);
 987         }
 988 }
 989
 990 void show_disk_util(int terse, struct json_object *parent,
 991                     struct buf_output *out)
 992 {
 993         struct flist_head *entry;
 994         struct disk_util *du;
 995         bool do_json;
 996
 997         if (!is_running_backend())
 998                 return;
 999
1000         if (flist_empty(&disk_list)) {
1001                 return;
1002         }
1003
1004         if ((output_format & FIO_OUTPUT_JSON) && parent)
1005                 do_json = true;
1006         else
1007                 do_json = false;
1008
1009         if (!terse && !do_json)
1010                 log_buf(out, "\nDisk stats (read/write):\n");
1011
1012         if (do_json)
1013                 json_object_add_disk_utils(parent, &disk_list);
1014         else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1015                 flist_for_each(entry, &disk_list) {
1016                         du = flist_entry(entry, struct disk_util, list);
1017
1018                         aggregate_slaves_stats(du);
1019                         print_disk_util(&du->dus, &du->agg, terse, out);
1020                 }
1021         }
1022 }
1023
1024 static void show_thread_status_normal(struct thread_stat *ts,
1025                                       struct group_run_stats *rs,
1026                                       struct buf_output *out)
1027 {
1028         double usr_cpu, sys_cpu;
1029         unsigned long runtime;
1030         double io_u_dist[FIO_IO_U_MAP_NR];
1031         time_t time_p;
1032         char time_buf[32];
1033
1034         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1035                 return;
1036
1037         memset(time_buf, 0, sizeof(time_buf));
1038
1039         time(&time_p);
1040         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1041
1042         if (!ts->error) {
1043                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1044                                         ts->name, ts->groupid, ts->members,
1045                                         ts->error, (int) ts->pid, time_buf);
1046         } else {
1047                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1048                                         ts->name, ts->groupid, ts->members,
1049                                         ts->error, ts->verror, (int) ts->pid,
1050                                         time_buf);
1051         }
1052
1053         if (strlen(ts->description))
1054                 log_buf(out, "  Description  : [%s]\n", ts->description);
1055
1056         if (ts->io_bytes[DDIR_READ])
1057                 show_ddir_status(rs, ts, DDIR_READ, out);
1058         if (ts->io_bytes[DDIR_WRITE])
1059                 show_ddir_status(rs, ts, DDIR_WRITE, out);
1060         if (ts->io_bytes[DDIR_TRIM])
1061                 show_ddir_status(rs, ts, DDIR_TRIM, out);
1062
1063         show_latencies(ts, out);
1064
1065         if (ts->sync_stat.samples)
1066                 show_ddir_status(rs, ts, DDIR_SYNC, out);
1067
1068         runtime = ts->total_run_time;
1069         if (runtime) {
1070                 double runt = (double) runtime;
1071
1072                 usr_cpu = (double) ts->usr_time * 100 / runt;
1073                 sys_cpu = (double) ts->sys_time * 100 / runt;
1074         } else {
1075                 usr_cpu = 0;
1076                 sys_cpu = 0;
1077         }
1078
1079         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1080                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1081                         (unsigned long long) ts->ctx,
1082                         (unsigned long long) ts->majf,
1083                         (unsigned long long) ts->minf);
1084
1085         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1086         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1087                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1088                                         io_u_dist[1], io_u_dist[2],
1089                                         io_u_dist[3], io_u_dist[4],
1090                                         io_u_dist[5], io_u_dist[6]);
1091
1092         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1093         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1094                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1095                                         io_u_dist[1], io_u_dist[2],
1096                                         io_u_dist[3], io_u_dist[4],
1097                                         io_u_dist[5], io_u_dist[6]);
1098         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1099         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1100                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1101                                         io_u_dist[1], io_u_dist[2],
1102                                         io_u_dist[3], io_u_dist[4],
1103                                         io_u_dist[5], io_u_dist[6]);
1104         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
1105                                  " short=%llu,%llu,%llu,0"
1106                                  " dropped=%llu,%llu,%llu,0\n",
1107                                         (unsigned long long) ts->total_io_u[0],
1108                                         (unsigned long long) ts->total_io_u[1],
1109                                         (unsigned long long) ts->total_io_u[2],
1110                                         (unsigned long long) ts->total_io_u[3],
1111                                         (unsigned long long) ts->short_io_u[0],
1112                                         (unsigned long long) ts->short_io_u[1],
1113                                         (unsigned long long) ts->short_io_u[2],
1114                                         (unsigned long long) ts->drop_io_u[0],
1115                                         (unsigned long long) ts->drop_io_u[1],
1116                                         (unsigned long long) ts->drop_io_u[2]);
1117         if (ts->continue_on_error) {
1118                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
1119                                         (unsigned long long)ts->total_err_count,
1120                                         ts->first_error,
1121                                         strerror(ts->first_error));
1122         }
1123         if (ts->latency_depth) {
1124                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1125                                         (unsigned long long)ts->latency_target,
1126                                         (unsigned long long)ts->latency_window,
1127                                         ts->latency_percentile.u.f,
1128                                         ts->latency_depth);
1129         }
1130
1131         if (ts->nr_block_infos)
1132                 show_block_infos(ts->nr_block_infos, ts->block_infos,
1133                                   ts->percentile_list, out);
1134
1135         if (ts->ss_dur)
1136                 show_ss_normal(ts, out);
1137 }
1138
1139 static void show_ddir_status_terse(struct thread_stat *ts,
1140                                    struct group_run_stats *rs, int ddir,
1141                                    int ver, struct buf_output *out)
1142 {
1143         unsigned long long min, max, minv, maxv, bw, iops;
1144         unsigned long long *ovals = NULL;
1145         double mean, dev;
1146         unsigned int len;
1147         int i, bw_stat;
1148
1149         assert(ddir_rw(ddir));
1150
1151         iops = bw = 0;
1152         if (ts->runtime[ddir]) {
1153                 uint64_t runt = ts->runtime[ddir];
1154
1155                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1156                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1157         }
1158
1159         log_buf(out, ";%llu;%llu;%llu;%llu",
1160                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1161                                         (unsigned long long) ts->runtime[ddir]);
1162
1163         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1164                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1165         else
1166                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1167
1168         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1169                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1170         else
1171                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1172
1173         if (ts->clat_percentiles || ts->lat_percentiles) {
1174                 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1175                                         ts->clat_stat[ddir].samples,
1176                                         ts->percentile_list, &ovals, &maxv,
1177                                         &minv);
1178         } else
1179                 len = 0;
1180
1181         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1182                 if (i >= len) {
1183                         log_buf(out, ";0%%=0");
1184                         continue;
1185                 }
1186                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1187         }
1188
1189         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1190                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1191         else
1192                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1193
1194         free(ovals);
1195
1196         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1197         if (bw_stat) {
1198                 double p_of_agg = 100.0;
1199
1200                 if (rs->agg[ddir]) {
1201                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1202                         if (p_of_agg > 100.0)
1203                                 p_of_agg = 100.0;
1204                 }
1205
1206                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1207         } else
1208                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1209
1210         if (ver == 5) {
1211                 if (bw_stat)
1212                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1213                 else
1214                         log_buf(out, ";%lu", 0UL);
1215
1216                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1217                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1218                                 mean, dev, (&ts->iops_stat[ddir])->samples);
1219                 else
1220                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1221         }
1222 }
1223
1224 static void add_ddir_status_json(struct thread_stat *ts,
1225                 struct group_run_stats *rs, int ddir, struct json_object *parent)
1226 {
1227         unsigned long long min, max, minv, maxv;
1228         unsigned long long bw_bytes, bw;
1229         unsigned long long *ovals = NULL;
1230         double mean, dev, iops;
1231         unsigned int len;
1232         int i;
1233         struct json_object *dir_object, *tmp_object, *percentile_object = NULL,
1234                 *clat_bins_object = NULL;
1235         char buf[120];
1236         double p_of_agg = 100.0;
1237
1238         assert(ddir_rw(ddir) || ddir_sync(ddir));
1239
1240         if (ts->unified_rw_rep && ddir != DDIR_READ)
1241                 return;
1242
1243         dir_object = json_create_object();
1244         json_object_add_value_object(parent,
1245                 ts->unified_rw_rep ? "mixed" : io_ddir_name(ddir), dir_object);
1246
1247         if (ddir_rw(ddir)) {
1248                 bw_bytes = 0;
1249                 bw = 0;
1250                 iops = 0.0;
1251                 if (ts->runtime[ddir]) {
1252                         uint64_t runt = ts->runtime[ddir];
1253
1254                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1255                         bw = bw_bytes / 1024; /* KiB/s */
1256                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1257                 }
1258
1259                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1260                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1261                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1262                 json_object_add_value_int(dir_object, "bw", bw);
1263                 json_object_add_value_float(dir_object, "iops", iops);
1264                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1265                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1266                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1267                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1268
1269                 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
1270                         min = max = 0;
1271                         mean = dev = 0.0;
1272                 }
1273                 tmp_object = json_create_object();
1274                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1275                 json_object_add_value_int(tmp_object, "min", min);
1276                 json_object_add_value_int(tmp_object, "max", max);
1277                 json_object_add_value_float(tmp_object, "mean", mean);
1278                 json_object_add_value_float(tmp_object, "stddev", dev);
1279
1280                 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1281                         min = max = 0;
1282                         mean = dev = 0.0;
1283                 }
1284                 tmp_object = json_create_object();
1285                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1286                 json_object_add_value_int(tmp_object, "min", min);
1287                 json_object_add_value_int(tmp_object, "max", max);
1288                 json_object_add_value_float(tmp_object, "mean", mean);
1289                 json_object_add_value_float(tmp_object, "stddev", dev);
1290         } else {
1291                 if (!calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
1292                         min = max = 0;
1293                         mean = dev = 0.0;
1294                 }
1295
1296                 tmp_object = json_create_object();
1297                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1298                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1299                 json_object_add_value_int(tmp_object, "min", min);
1300                 json_object_add_value_int(tmp_object, "max", max);
1301                 json_object_add_value_float(tmp_object, "mean", mean);
1302                 json_object_add_value_float(tmp_object, "stddev", dev);
1303         }
1304
1305         if (ts->clat_percentiles || ts->lat_percentiles) {
1306                 if (ddir_rw(ddir)) {
1307                         uint64_t samples;
1308
1309                         if (ts->clat_percentiles)
1310                                 samples = ts->clat_stat[ddir].samples;
1311                         else
1312                                 samples = ts->lat_stat[ddir].samples;
1313
1314                         len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1315                                         samples, ts->percentile_list, &ovals,
1316                                         &maxv, &minv);
1317                 } else {
1318                         len = calc_clat_percentiles(ts->io_u_sync_plat,
1319                                         ts->sync_stat.samples,
1320                                         ts->percentile_list, &ovals, &maxv,
1321                                         &minv);
1322                 }
1323
1324                 if (len > FIO_IO_U_LIST_MAX_LEN)
1325                         len = FIO_IO_U_LIST_MAX_LEN;
1326         } else
1327                 len = 0;
1328
1329         if (ts->clat_percentiles) {
1330                 percentile_object = json_create_object();
1331                 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1332                 for (i = 0; i < len; i++) {
1333                         snprintf(buf, sizeof(buf), "%f",
1334                                  ts->percentile_list[i].u.f);
1335                         json_object_add_value_int(percentile_object, buf,
1336                                                   ovals[i]);
1337                 }
1338         }
1339
1340         free(ovals);
1341
1342         if (output_format & FIO_OUTPUT_JSON_PLUS && ts->clat_percentiles) {
1343                 clat_bins_object = json_create_object();
1344                 json_object_add_value_object(tmp_object, "bins",
1345                                              clat_bins_object);
1346
1347                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1348                         if (ddir_rw(ddir)) {
1349                                 if (ts->io_u_plat[FIO_CLAT][ddir][i]) {
1350                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1351                                         json_object_add_value_int(clat_bins_object, buf, ts->io_u_plat[FIO_CLAT][ddir][i]);
1352                                 }
1353                         } else {
1354                                 if (ts->io_u_sync_plat[i]) {
1355                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1356                                         json_object_add_value_int(clat_bins_object, buf, ts->io_u_sync_plat[i]);
1357                                 }
1358                         }
1359                 }
1360         }
1361
1362
1363         /* Only print PRIO latencies if some high priority samples were gathered */
1364         if (ts->clat_high_prio_stat[ddir].samples > 0) {
1365                 /* START OF HIGH PRIO CLAT */
1366             if (!calc_lat(&ts->clat_high_prio_stat[ddir], &min, &max, &mean, &dev)) {
1367                         min = max = 0;
1368                         mean = dev = 0.0;
1369                 }
1370                 tmp_object = json_create_object();
1371                 json_object_add_value_object(dir_object, "clat_prio",
1372                                 tmp_object);
1373                 json_object_add_value_int(tmp_object, "samples",
1374                                 ts->clat_high_prio_stat[ddir].samples);
1375                 json_object_add_value_int(tmp_object, "min", min);
1376                 json_object_add_value_int(tmp_object, "max", max);
1377                 json_object_add_value_float(tmp_object, "mean", mean);
1378                 json_object_add_value_float(tmp_object, "stddev", dev);
1379
1380                 if (ts->clat_percentiles) {
1381                         len = calc_clat_percentiles(ts->io_u_plat_high_prio[ddir],
1382                                                 ts->clat_high_prio_stat[ddir].samples,
1383                                                 ts->percentile_list, &ovals, &maxv,
1384                                                 &minv);
1385                 } else
1386                         len = 0;
1387
1388                 percentile_object = json_create_object();
1389                 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1390                 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1391                         if (i >= len) {
1392                                 json_object_add_value_int(percentile_object, "0.00", 0);
1393                                 continue;
1394                         }
1395                         snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1396                         json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1397                 }
1398
1399                 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1400                         clat_bins_object = json_create_object();
1401                         json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1402                         for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1403                                 snprintf(buf, sizeof(buf), "%d", i);
1404                                 json_object_add_value_int(clat_bins_object, (const char *)buf,
1405                                                 ts->io_u_plat_high_prio[ddir][i]);
1406                         }
1407                         json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS",
1408                                         FIO_IO_U_PLAT_BITS);
1409                         json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL",
1410                                         FIO_IO_U_PLAT_VAL);
1411                         json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR",
1412                                         FIO_IO_U_PLAT_NR);
1413                 }
1414                 /* END OF HIGH PRIO CLAT */
1415
1416                 /* START OF PRIO CLAT */
1417             if (!calc_lat(&ts->clat_prio_stat[ddir], &min, &max, &mean, &dev)) {
1418                         min = max = 0;
1419                         mean = dev = 0.0;
1420                 }
1421                 tmp_object = json_create_object();
1422                 json_object_add_value_object(dir_object, "clat_low_prio",
1423                                 tmp_object);
1424                 json_object_add_value_int(tmp_object, "samples",
1425                                 ts->clat_prio_stat[ddir].samples);
1426                 json_object_add_value_int(tmp_object, "min", min);
1427                 json_object_add_value_int(tmp_object, "max", max);
1428                 json_object_add_value_float(tmp_object, "mean", mean);
1429                 json_object_add_value_float(tmp_object, "stddev", dev);
1430
1431                 if (ts->clat_percentiles) {
1432                         len = calc_clat_percentiles(ts->io_u_plat_prio[ddir],
1433                                                 ts->clat_prio_stat[ddir].samples,
1434                                                 ts->percentile_list, &ovals, &maxv,
1435                                                 &minv);
1436                 } else
1437                         len = 0;
1438
1439                 percentile_object = json_create_object();
1440                 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1441                 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1442                         if (i >= len) {
1443                                 json_object_add_value_int(percentile_object, "0.00", 0);
1444                                 continue;
1445                         }
1446                         snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1447                         json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1448                 }
1449
1450                 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1451                         clat_bins_object = json_create_object();
1452                         json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1453                         for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1454                                 snprintf(buf, sizeof(buf), "%d", i);
1455                                 json_object_add_value_int(clat_bins_object, (const char *)buf,
1456                                                 ts->io_u_plat_prio[ddir][i]);
1457                         }
1458                         json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS",
1459                                         FIO_IO_U_PLAT_BITS);
1460                         json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL",
1461                                         FIO_IO_U_PLAT_VAL);
1462                         json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR",
1463                                         FIO_IO_U_PLAT_NR);
1464                 }
1465                 /* END OF PRIO CLAT */
1466         }
1467
1468         if (!ddir_rw(ddir))
1469                 return;
1470
1471         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1472                 min = max = 0;
1473                 mean = dev = 0.0;
1474         }
1475         tmp_object = json_create_object();
1476         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1477         json_object_add_value_int(tmp_object, "min", min);
1478         json_object_add_value_int(tmp_object, "max", max);
1479         json_object_add_value_float(tmp_object, "mean", mean);
1480         json_object_add_value_float(tmp_object, "stddev", dev);
1481         if (ts->lat_percentiles)
1482                 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1483         if (output_format & FIO_OUTPUT_JSON_PLUS && ts->lat_percentiles)
1484                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1485
1486         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1487                 if (rs->agg[ddir]) {
1488                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1489                         if (p_of_agg > 100.0)
1490                                 p_of_agg = 100.0;
1491                 }
1492         } else {
1493                 min = max = 0;
1494                 p_of_agg = mean = dev = 0.0;
1495         }
1496         json_object_add_value_int(dir_object, "bw_min", min);
1497         json_object_add_value_int(dir_object, "bw_max", max);
1498         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1499         json_object_add_value_float(dir_object, "bw_mean", mean);
1500         json_object_add_value_float(dir_object, "bw_dev", dev);
1501         json_object_add_value_int(dir_object, "bw_samples",
1502                                 (&ts->bw_stat[ddir])->samples);
1503
1504         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1505                 min = max = 0;
1506                 mean = dev = 0.0;
1507         }
1508         json_object_add_value_int(dir_object, "iops_min", min);
1509         json_object_add_value_int(dir_object, "iops_max", max);
1510         json_object_add_value_float(dir_object, "iops_mean", mean);
1511         json_object_add_value_float(dir_object, "iops_stddev", dev);
1512         json_object_add_value_int(dir_object, "iops_samples",
1513                                 (&ts->iops_stat[ddir])->samples);
1514
1515         if (ts->cachehit + ts->cachemiss) {
1516                 uint64_t total;
1517                 double hit;
1518
1519                 total = ts->cachehit + ts->cachemiss;
1520                 hit = (double) ts->cachehit / (double) total;
1521                 hit *= 100.0;
1522                 json_object_add_value_float(dir_object, "cachehit", hit);
1523         }
1524 }
1525
1526 static void show_thread_status_terse_all(struct thread_stat *ts,
1527                                          struct group_run_stats *rs, int ver,
1528                                          struct buf_output *out)
1529 {
1530         double io_u_dist[FIO_IO_U_MAP_NR];
1531         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1532         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1533         double usr_cpu, sys_cpu;
1534         int i;
1535
1536         /* General Info */
1537         if (ver == 2)
1538                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1539         else
1540                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1541                         ts->name, ts->groupid, ts->error);
1542
1543         /* Log Read Status */
1544         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1545         /* Log Write Status */
1546         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1547         /* Log Trim Status */
1548         if (ver == 2 || ver == 4 || ver == 5)
1549                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1550
1551         /* CPU Usage */
1552         if (ts->total_run_time) {
1553                 double runt = (double) ts->total_run_time;
1554
1555                 usr_cpu = (double) ts->usr_time * 100 / runt;
1556                 sys_cpu = (double) ts->sys_time * 100 / runt;
1557         } else {
1558                 usr_cpu = 0;
1559                 sys_cpu = 0;
1560         }
1561
1562         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1563                                                 (unsigned long long) ts->ctx,
1564                                                 (unsigned long long) ts->majf,
1565                                                 (unsigned long long) ts->minf);
1566
1567         /* Calc % distribution of IO depths, usecond, msecond latency */
1568         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1569         stat_calc_lat_nu(ts, io_u_lat_u);
1570         stat_calc_lat_m(ts, io_u_lat_m);
1571
1572         /* Only show fixed 7 I/O depth levels*/
1573         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1574                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1575                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1576
1577         /* Microsecond latency */
1578         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1579                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1580         /* Millisecond latency */
1581         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1582                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1583
1584         /* disk util stats, if any */
1585         if (ver >= 3 && is_running_backend())
1586                 show_disk_util(1, NULL, out);
1587
1588         /* Additional output if continue_on_error set - default off*/
1589         if (ts->continue_on_error)
1590                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1591
1592         /* Additional output if description is set */
1593         if (strlen(ts->description)) {
1594                 if (ver == 2)
1595                         log_buf(out, "\n");
1596                 log_buf(out, ";%s", ts->description);
1597         }
1598
1599         log_buf(out, "\n");
1600 }
1601
1602 static void json_add_job_opts(struct json_object *root, const char *name,
1603                               struct flist_head *opt_list)
1604 {
1605         struct json_object *dir_object;
1606         struct flist_head *entry;
1607         struct print_option *p;
1608
1609         if (flist_empty(opt_list))
1610                 return;
1611
1612         dir_object = json_create_object();
1613         json_object_add_value_object(root, name, dir_object);
1614
1615         flist_for_each(entry, opt_list) {
1616                 const char *pos = "";
1617
1618                 p = flist_entry(entry, struct print_option, list);
1619                 if (p->value)
1620                         pos = p->value;
1621                 json_object_add_value_string(dir_object, p->name, pos);
1622         }
1623 }
1624
1625 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1626                                                    struct group_run_stats *rs,
1627                                                    struct flist_head *opt_list)
1628 {
1629         struct json_object *root, *tmp;
1630         struct jobs_eta *je;
1631         double io_u_dist[FIO_IO_U_MAP_NR];
1632         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1633         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1634         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1635         double usr_cpu, sys_cpu;
1636         int i;
1637         size_t size;
1638
1639         root = json_create_object();
1640         json_object_add_value_string(root, "jobname", ts->name);
1641         json_object_add_value_int(root, "groupid", ts->groupid);
1642         json_object_add_value_int(root, "error", ts->error);
1643
1644         /* ETA Info */
1645         je = get_jobs_eta(true, &size);
1646         if (je) {
1647                 json_object_add_value_int(root, "eta", je->eta_sec);
1648                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1649         }
1650
1651         if (opt_list)
1652                 json_add_job_opts(root, "job options", opt_list);
1653
1654         add_ddir_status_json(ts, rs, DDIR_READ, root);
1655         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1656         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1657         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1658
1659         /* CPU Usage */
1660         if (ts->total_run_time) {
1661                 double runt = (double) ts->total_run_time;
1662
1663                 usr_cpu = (double) ts->usr_time * 100 / runt;
1664                 sys_cpu = (double) ts->sys_time * 100 / runt;
1665         } else {
1666                 usr_cpu = 0;
1667                 sys_cpu = 0;
1668         }
1669         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1670         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1671         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1672         json_object_add_value_int(root, "ctx", ts->ctx);
1673         json_object_add_value_int(root, "majf", ts->majf);
1674         json_object_add_value_int(root, "minf", ts->minf);
1675
1676         /* Calc % distribution of IO depths */
1677         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1678         tmp = json_create_object();
1679         json_object_add_value_object(root, "iodepth_level", tmp);
1680         /* Only show fixed 7 I/O depth levels*/
1681         for (i = 0; i < 7; i++) {
1682                 char name[20];
1683                 if (i < 6)
1684                         snprintf(name, 20, "%d", 1 << i);
1685                 else
1686                         snprintf(name, 20, ">=%d", 1 << i);
1687                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1688         }
1689
1690         /* Calc % distribution of submit IO depths */
1691         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1692         tmp = json_create_object();
1693         json_object_add_value_object(root, "iodepth_submit", tmp);
1694         /* Only show fixed 7 I/O depth levels*/
1695         for (i = 0; i < 7; i++) {
1696                 char name[20];
1697                 if (i == 0)
1698                         snprintf(name, 20, "0");
1699                 else if (i < 6)
1700                         snprintf(name, 20, "%d", 1 << (i+1));
1701                 else
1702                         snprintf(name, 20, ">=%d", 1 << i);
1703                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1704         }
1705
1706         /* Calc % distribution of completion IO depths */
1707         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1708         tmp = json_create_object();
1709         json_object_add_value_object(root, "iodepth_complete", tmp);
1710         /* Only show fixed 7 I/O depth levels*/
1711         for (i = 0; i < 7; i++) {
1712                 char name[20];
1713                 if (i == 0)
1714                         snprintf(name, 20, "0");
1715                 else if (i < 6)
1716                         snprintf(name, 20, "%d", 1 << (i+1));
1717                 else
1718                         snprintf(name, 20, ">=%d", 1 << i);
1719                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1720         }
1721
1722         /* Calc % distribution of nsecond, usecond, msecond latency */
1723         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1724         stat_calc_lat_n(ts, io_u_lat_n);
1725         stat_calc_lat_u(ts, io_u_lat_u);
1726         stat_calc_lat_m(ts, io_u_lat_m);
1727
1728         /* Nanosecond latency */
1729         tmp = json_create_object();
1730         json_object_add_value_object(root, "latency_ns", tmp);
1731         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1732                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1733                                  "250", "500", "750", "1000", };
1734                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1735         }
1736         /* Microsecond latency */
1737         tmp = json_create_object();
1738         json_object_add_value_object(root, "latency_us", tmp);
1739         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1740                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1741                                  "250", "500", "750", "1000", };
1742                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1743         }
1744         /* Millisecond latency */
1745         tmp = json_create_object();
1746         json_object_add_value_object(root, "latency_ms", tmp);
1747         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1748                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1749                                  "250", "500", "750", "1000", "2000",
1750                                  ">=2000", };
1751                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1752         }
1753
1754         /* Additional output if continue_on_error set - default off*/
1755         if (ts->continue_on_error) {
1756                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1757                 json_object_add_value_int(root, "first_error", ts->first_error);
1758         }
1759
1760         if (ts->latency_depth) {
1761                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1762                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1763                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1764                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1765         }
1766
1767         /* Additional output if description is set */
1768         if (strlen(ts->description))
1769                 json_object_add_value_string(root, "desc", ts->description);
1770
1771         if (ts->nr_block_infos) {
1772                 /* Block error histogram and types */
1773                 int len;
1774                 unsigned int *percentiles = NULL;
1775                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1776
1777                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1778                                              ts->percentile_list,
1779                                              &percentiles, block_state_counts);
1780
1781                 if (len) {
1782                         struct json_object *block, *percentile_object, *states;
1783                         int state;
1784                         block = json_create_object();
1785                         json_object_add_value_object(root, "block", block);
1786
1787                         percentile_object = json_create_object();
1788                         json_object_add_value_object(block, "percentiles",
1789                                                      percentile_object);
1790                         for (i = 0; i < len; i++) {
1791                                 char buf[20];
1792                                 snprintf(buf, sizeof(buf), "%f",
1793                                          ts->percentile_list[i].u.f);
1794                                 json_object_add_value_int(percentile_object,
1795                                                           buf,
1796                                                           percentiles[i]);
1797                         }
1798
1799                         states = json_create_object();
1800                         json_object_add_value_object(block, "states", states);
1801                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1802                                 json_object_add_value_int(states,
1803                                         block_state_names[state],
1804                                         block_state_counts[state]);
1805                         }
1806                         free(percentiles);
1807                 }
1808         }
1809
1810         if (ts->ss_dur) {
1811                 struct json_object *data;
1812                 struct json_array *iops, *bw;
1813                 int j, k, l;
1814                 char ss_buf[64];
1815
1816                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1817                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1818                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1819                         (float) ts->ss_limit.u.f,
1820                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1821
1822                 tmp = json_create_object();
1823                 json_object_add_value_object(root, "steadystate", tmp);
1824                 json_object_add_value_string(tmp, "ss", ss_buf);
1825                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1826                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1827
1828                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1829                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1830                 json_object_add_value_string(tmp, "criterion", ss_buf);
1831                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1832                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1833
1834                 data = json_create_object();
1835                 json_object_add_value_object(tmp, "data", data);
1836                 bw = json_create_array();
1837                 iops = json_create_array();
1838
1839                 /*
1840                 ** if ss was attained or the buffer is not full,
1841                 ** ss->head points to the first element in the list.
1842                 ** otherwise it actually points to the second element
1843                 ** in the list
1844                 */
1845                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1846                         j = ts->ss_head;
1847                 else
1848                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1849                 for (l = 0; l < ts->ss_dur; l++) {
1850                         k = (j + l) % ts->ss_dur;
1851                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1852                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1853                 }
1854                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1855                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1856                 json_object_add_value_array(data, "iops", iops);
1857                 json_object_add_value_array(data, "bw", bw);
1858         }
1859
1860         return root;
1861 }
1862
1863 static void show_thread_status_terse(struct thread_stat *ts,
1864                                      struct group_run_stats *rs,
1865                                      struct buf_output *out)
1866 {
1867         if (terse_version >= 2 && terse_version <= 5)
1868                 show_thread_status_terse_all(ts, rs, terse_version, out);
1869         else
1870                 log_err("fio: bad terse version!? %d\n", terse_version);
1871 }
1872
1873 struct json_object *show_thread_status(struct thread_stat *ts,
1874                                        struct group_run_stats *rs,
1875                                        struct flist_head *opt_list,
1876                                        struct buf_output *out)
1877 {
1878         struct json_object *ret = NULL;
1879
1880         if (output_format & FIO_OUTPUT_TERSE)
1881                 show_thread_status_terse(ts, rs,  out);
1882         if (output_format & FIO_OUTPUT_JSON)
1883                 ret = show_thread_status_json(ts, rs, opt_list);
1884         if (output_format & FIO_OUTPUT_NORMAL)
1885                 show_thread_status_normal(ts, rs,  out);
1886
1887         return ret;
1888 }
1889
1890 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1891 {
1892         double mean, S;
1893
1894         dst->min_val = min(dst->min_val, src->min_val);
1895         dst->max_val = max(dst->max_val, src->max_val);
1896
1897         /*
1898          * Compute new mean and S after the merge
1899          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1900          *  #Parallel_algorithm>
1901          */
1902         if (first) {
1903                 mean = src->mean.u.f;
1904                 S = src->S.u.f;
1905         } else {
1906                 double delta = src->mean.u.f - dst->mean.u.f;
1907
1908                 mean = ((src->mean.u.f * src->samples) +
1909                         (dst->mean.u.f * dst->samples)) /
1910                         (dst->samples + src->samples);
1911
1912                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1913                         (dst->samples * src->samples) /
1914                         (dst->samples + src->samples);
1915         }
1916
1917         dst->samples += src->samples;
1918         dst->mean.u.f = mean;
1919         dst->S.u.f = S;
1920
1921 }
1922
1923 /*
1924  * We sum two kinds of stats - one that is time based, in which case we
1925  * apply the proper summing technique, and then one that is iops/bw
1926  * numbers. For group_reporting, we should just add those up, not make
1927  * them the mean of everything.
1928  */
1929 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first,
1930                      bool pure_sum)
1931 {
1932         if (src->samples == 0)
1933                 return;
1934
1935         if (!pure_sum) {
1936                 __sum_stat(dst, src, first);
1937                 return;
1938         }
1939
1940         if (first) {
1941                 dst->min_val = src->min_val;
1942                 dst->max_val = src->max_val;
1943                 dst->samples = src->samples;
1944                 dst->mean.u.f = src->mean.u.f;
1945                 dst->S.u.f = src->S.u.f;
1946         } else {
1947                 dst->min_val += src->min_val;
1948                 dst->max_val += src->max_val;
1949                 dst->samples += src->samples;
1950                 dst->mean.u.f += src->mean.u.f;
1951                 dst->S.u.f += src->S.u.f;
1952         }
1953 }
1954
1955 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1956 {
1957         int i;
1958
1959         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1960                 if (dst->max_run[i] < src->max_run[i])
1961                         dst->max_run[i] = src->max_run[i];
1962                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1963                         dst->min_run[i] = src->min_run[i];
1964                 if (dst->max_bw[i] < src->max_bw[i])
1965                         dst->max_bw[i] = src->max_bw[i];
1966                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1967                         dst->min_bw[i] = src->min_bw[i];
1968
1969                 dst->iobytes[i] += src->iobytes[i];
1970                 dst->agg[i] += src->agg[i];
1971         }
1972
1973         if (!dst->kb_base)
1974                 dst->kb_base = src->kb_base;
1975         if (!dst->unit_base)
1976                 dst->unit_base = src->unit_base;
1977         if (!dst->sig_figs)
1978                 dst->sig_figs = src->sig_figs;
1979 }
1980
1981 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1982                       bool first)
1983 {
1984         int k, l, m;
1985
1986         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1987                 if (!dst->unified_rw_rep) {
1988                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first, false);
1989                         sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], first, false);
1990                         sum_stat(&dst->clat_prio_stat[l], &src->clat_prio_stat[l], first, false);
1991                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first, false);
1992                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first, false);
1993                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first, true);
1994                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first, true);
1995
1996                         dst->io_bytes[l] += src->io_bytes[l];
1997
1998                         if (dst->runtime[l] < src->runtime[l])
1999                                 dst->runtime[l] = src->runtime[l];
2000                 } else {
2001                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first, false);
2002                         sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], first, false);
2003                         sum_stat(&dst->clat_prio_stat[l], &src->clat_prio_stat[l], first, false);
2004                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first, false);
2005                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first, false);
2006                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first, true);
2007                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first, true);
2008
2009                         dst->io_bytes[0] += src->io_bytes[l];
2010
2011                         if (dst->runtime[0] < src->runtime[l])
2012                                 dst->runtime[0] = src->runtime[l];
2013
2014                         /*
2015                          * We're summing to the same destination, so override
2016                          * 'first' after the first iteration of the loop
2017                          */
2018                         first = false;
2019                 }
2020         }
2021
2022         sum_stat(&dst->sync_stat, &src->sync_stat, first, false);
2023         dst->usr_time += src->usr_time;
2024         dst->sys_time += src->sys_time;
2025         dst->ctx += src->ctx;
2026         dst->majf += src->majf;
2027         dst->minf += src->minf;
2028
2029         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2030                 dst->io_u_map[k] += src->io_u_map[k];
2031                 dst->io_u_submit[k] += src->io_u_submit[k];
2032                 dst->io_u_complete[k] += src->io_u_complete[k];
2033         }
2034
2035         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2036                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2037         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2038                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2039         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2040                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2041
2042         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2043                 if (!dst->unified_rw_rep) {
2044                         dst->total_io_u[k] += src->total_io_u[k];
2045                         dst->short_io_u[k] += src->short_io_u[k];
2046                         dst->drop_io_u[k] += src->drop_io_u[k];
2047                 } else {
2048                         dst->total_io_u[0] += src->total_io_u[k];
2049                         dst->short_io_u[0] += src->short_io_u[k];
2050                         dst->drop_io_u[0] += src->drop_io_u[k];
2051                 }
2052         }
2053
2054         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2055
2056         for (k = 0; k < FIO_LAT_CNT; k++)
2057                 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2058                         for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2059                                 if (!dst->unified_rw_rep)
2060                                         dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2061                                 else
2062                                         dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2063
2064         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2065                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2066
2067         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2068                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
2069                         if (!dst->unified_rw_rep) {
2070                                 dst->io_u_plat_high_prio[k][m] += src->io_u_plat_high_prio[k][m];
2071                                 dst->io_u_plat_prio[k][m] += src->io_u_plat_prio[k][m];
2072                         } else {
2073                                 dst->io_u_plat_high_prio[0][m] += src->io_u_plat_high_prio[k][m];
2074                                 dst->io_u_plat_prio[0][m] += src->io_u_plat_prio[k][m];
2075                         }
2076
2077                 }
2078         }
2079
2080         dst->total_run_time += src->total_run_time;
2081         dst->total_submit += src->total_submit;
2082         dst->total_complete += src->total_complete;
2083         dst->nr_zone_resets += src->nr_zone_resets;
2084         dst->cachehit += src->cachehit;
2085         dst->cachemiss += src->cachemiss;
2086 }
2087
2088 void init_group_run_stat(struct group_run_stats *gs)
2089 {
2090         int i;
2091         memset(gs, 0, sizeof(*gs));
2092
2093         for (i = 0; i < DDIR_RWDIR_CNT; i++)
2094                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2095 }
2096
2097 void init_thread_stat(struct thread_stat *ts)
2098 {
2099         int j;
2100
2101         memset(ts, 0, sizeof(*ts));
2102
2103         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2104                 ts->lat_stat[j].min_val = -1UL;
2105                 ts->clat_stat[j].min_val = -1UL;
2106                 ts->slat_stat[j].min_val = -1UL;
2107                 ts->bw_stat[j].min_val = -1UL;
2108                 ts->iops_stat[j].min_val = -1UL;
2109                 ts->clat_high_prio_stat[j].min_val = -1UL;
2110                 ts->clat_prio_stat[j].min_val = -1UL;
2111         }
2112         ts->sync_stat.min_val = -1UL;
2113         ts->groupid = -1;
2114 }
2115
2116 void __show_run_stats(void)
2117 {
2118         struct group_run_stats *runstats, *rs;
2119         struct thread_data *td;
2120         struct thread_stat *threadstats, *ts;
2121         int i, j, k, nr_ts, last_ts, idx;
2122         bool kb_base_warned = false;
2123         bool unit_base_warned = false;
2124         struct json_object *root = NULL;
2125         struct json_array *array = NULL;
2126         struct buf_output output[FIO_OUTPUT_NR];
2127         struct flist_head **opt_lists;
2128
2129         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2130
2131         for (i = 0; i < groupid + 1; i++)
2132                 init_group_run_stat(&runstats[i]);
2133
2134         /*
2135          * find out how many threads stats we need. if group reporting isn't
2136          * enabled, it's one-per-td.
2137          */
2138         nr_ts = 0;
2139         last_ts = -1;
2140         for_each_td(td, i) {
2141                 if (!td->o.group_reporting) {
2142                         nr_ts++;
2143                         continue;
2144                 }
2145                 if (last_ts == td->groupid)
2146                         continue;
2147                 if (!td->o.stats)
2148                         continue;
2149
2150                 last_ts = td->groupid;
2151                 nr_ts++;
2152         }
2153
2154         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2155         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2156
2157         for (i = 0; i < nr_ts; i++) {
2158                 init_thread_stat(&threadstats[i]);
2159                 opt_lists[i] = NULL;
2160         }
2161
2162         j = 0;
2163         last_ts = -1;
2164         idx = 0;
2165         for_each_td(td, i) {
2166                 if (!td->o.stats)
2167                         continue;
2168                 if (idx && (!td->o.group_reporting ||
2169                     (td->o.group_reporting && last_ts != td->groupid))) {
2170                         idx = 0;
2171                         j++;
2172                 }
2173
2174                 last_ts = td->groupid;
2175
2176                 ts = &threadstats[j];
2177
2178                 ts->clat_percentiles = td->o.clat_percentiles;
2179                 ts->lat_percentiles = td->o.lat_percentiles;
2180                 ts->percentile_precision = td->o.percentile_precision;
2181                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2182                 opt_lists[j] = &td->opt_list;
2183
2184                 idx++;
2185                 ts->members++;
2186
2187                 if (ts->groupid == -1) {
2188                         /*
2189                          * These are per-group shared already
2190                          */
2191                         snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2192                         if (td->o.description)
2193                                 snprintf(ts->description,
2194                                          sizeof(ts->description), "%s",
2195                                          td->o.description);
2196                         else
2197                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2198
2199                         /*
2200                          * If multiple entries in this group, this is
2201                          * the first member.
2202                          */
2203                         ts->thread_number = td->thread_number;
2204                         ts->groupid = td->groupid;
2205
2206                         /*
2207                          * first pid in group, not very useful...
2208                          */
2209                         ts->pid = td->pid;
2210
2211                         ts->kb_base = td->o.kb_base;
2212                         ts->unit_base = td->o.unit_base;
2213                         ts->sig_figs = td->o.sig_figs;
2214                         ts->unified_rw_rep = td->o.unified_rw_rep;
2215                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2216                         log_info("fio: kb_base differs for jobs in group, using"
2217                                  " %u as the base\n", ts->kb_base);
2218                         kb_base_warned = true;
2219                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2220                         log_info("fio: unit_base differs for jobs in group, using"
2221                                  " %u as the base\n", ts->unit_base);
2222                         unit_base_warned = true;
2223                 }
2224
2225                 ts->continue_on_error = td->o.continue_on_error;
2226                 ts->total_err_count += td->total_err_count;
2227                 ts->first_error = td->first_error;
2228                 if (!ts->error) {
2229                         if (!td->error && td->o.continue_on_error &&
2230                             td->first_error) {
2231                                 ts->error = td->first_error;
2232                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2233                                          td->verror);
2234                         } else  if (td->error) {
2235                                 ts->error = td->error;
2236                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2237                                          td->verror);
2238                         }
2239                 }
2240
2241                 ts->latency_depth = td->latency_qd;
2242                 ts->latency_target = td->o.latency_target;
2243                 ts->latency_percentile = td->o.latency_percentile;
2244                 ts->latency_window = td->o.latency_window;
2245
2246                 ts->nr_block_infos = td->ts.nr_block_infos;
2247                 for (k = 0; k < ts->nr_block_infos; k++)
2248                         ts->block_infos[k] = td->ts.block_infos[k];
2249
2250                 sum_thread_stats(ts, &td->ts, idx == 1);
2251
2252                 if (td->o.ss_dur) {
2253                         ts->ss_state = td->ss.state;
2254                         ts->ss_dur = td->ss.dur;
2255                         ts->ss_head = td->ss.head;
2256                         ts->ss_bw_data = td->ss.bw_data;
2257                         ts->ss_iops_data = td->ss.iops_data;
2258                         ts->ss_limit.u.f = td->ss.limit;
2259                         ts->ss_slope.u.f = td->ss.slope;
2260                         ts->ss_deviation.u.f = td->ss.deviation;
2261                         ts->ss_criterion.u.f = td->ss.criterion;
2262                 }
2263                 else
2264                         ts->ss_dur = ts->ss_state = 0;
2265         }
2266
2267         for (i = 0; i < nr_ts; i++) {
2268                 unsigned long long bw;
2269
2270                 ts = &threadstats[i];
2271                 if (ts->groupid == -1)
2272                         continue;
2273                 rs = &runstats[ts->groupid];
2274                 rs->kb_base = ts->kb_base;
2275                 rs->unit_base = ts->unit_base;
2276                 rs->sig_figs = ts->sig_figs;
2277                 rs->unified_rw_rep += ts->unified_rw_rep;
2278
2279                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2280                         if (!ts->runtime[j])
2281                                 continue;
2282                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2283                                 rs->min_run[j] = ts->runtime[j];
2284                         if (ts->runtime[j] > rs->max_run[j])
2285                                 rs->max_run[j] = ts->runtime[j];
2286
2287                         bw = 0;
2288                         if (ts->runtime[j])
2289                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2290                         if (bw < rs->min_bw[j])
2291                                 rs->min_bw[j] = bw;
2292                         if (bw > rs->max_bw[j])
2293                                 rs->max_bw[j] = bw;
2294
2295                         rs->iobytes[j] += ts->io_bytes[j];
2296                 }
2297         }
2298
2299         for (i = 0; i < groupid + 1; i++) {
2300                 int ddir;
2301
2302                 rs = &runstats[i];
2303
2304                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2305                         if (rs->max_run[ddir])
2306                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2307                                                 rs->max_run[ddir];
2308                 }
2309         }
2310
2311         for (i = 0; i < FIO_OUTPUT_NR; i++)
2312                 buf_output_init(&output[i]);
2313
2314         /*
2315          * don't overwrite last signal output
2316          */
2317         if (output_format & FIO_OUTPUT_NORMAL)
2318                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2319         if (output_format & FIO_OUTPUT_JSON) {
2320                 struct thread_data *global;
2321                 char time_buf[32];
2322                 struct timeval now;
2323                 unsigned long long ms_since_epoch;
2324                 time_t tv_sec;
2325
2326                 gettimeofday(&now, NULL);
2327                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2328                                  (unsigned long long)(now.tv_usec) / 1000;
2329
2330                 tv_sec = now.tv_sec;
2331                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2332                 if (time_buf[strlen(time_buf) - 1] == '\n')
2333                         time_buf[strlen(time_buf) - 1] = '\0';
2334
2335                 root = json_create_object();
2336                 json_object_add_value_string(root, "fio version", fio_version_string);
2337                 json_object_add_value_int(root, "timestamp", now.tv_sec);
2338                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2339                 json_object_add_value_string(root, "time", time_buf);
2340                 global = get_global_options();
2341                 json_add_job_opts(root, "global options", &global->opt_list);
2342                 array = json_create_array();
2343                 json_object_add_value_array(root, "jobs", array);
2344         }
2345
2346         if (is_backend)
2347                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2348
2349         for (i = 0; i < nr_ts; i++) {
2350                 ts = &threadstats[i];
2351                 rs = &runstats[ts->groupid];
2352
2353                 if (is_backend) {
2354                         fio_server_send_job_options(opt_lists[i], i);
2355                         fio_server_send_ts(ts, rs);
2356                 } else {
2357                         if (output_format & FIO_OUTPUT_TERSE)
2358                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2359                         if (output_format & FIO_OUTPUT_JSON) {
2360                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2361                                 json_array_add_value_object(array, tmp);
2362                         }
2363                         if (output_format & FIO_OUTPUT_NORMAL)
2364                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2365                 }
2366         }
2367         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2368                 /* disk util stats, if any */
2369                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2370
2371                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2372
2373                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2374                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2375                 json_free_object(root);
2376         }
2377
2378         for (i = 0; i < groupid + 1; i++) {
2379                 rs = &runstats[i];
2380
2381                 rs->groupid = i;
2382                 if (is_backend)
2383                         fio_server_send_gs(rs);
2384                 else if (output_format & FIO_OUTPUT_NORMAL)
2385                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2386         }
2387
2388         if (is_backend)
2389                 fio_server_send_du();
2390         else if (output_format & FIO_OUTPUT_NORMAL) {
2391                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2392                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2393         }
2394
2395         for (i = 0; i < FIO_OUTPUT_NR; i++) {
2396                 struct buf_output *out = &output[i];
2397
2398                 log_info_buf(out->buf, out->buflen);
2399                 buf_output_free(out);
2400         }
2401
2402         fio_idle_prof_cleanup();
2403
2404         log_info_flush();
2405         free(runstats);
2406         free(threadstats);
2407         free(opt_lists);
2408 }
2409
2410 void __show_running_run_stats(void)
2411 {
2412         struct thread_data *td;
2413         unsigned long long *rt;
2414         struct timespec ts;
2415         int i;
2416
2417         fio_sem_down(stat_sem);
2418
2419         rt = malloc(thread_number * sizeof(unsigned long long));
2420         fio_gettime(&ts, NULL);
2421
2422         for_each_td(td, i) {
2423                 td->update_rusage = 1;
2424                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
2425                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
2426                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2427                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2428
2429                 rt[i] = mtime_since(&td->start, &ts);
2430                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2431                         td->ts.runtime[DDIR_READ] += rt[i];
2432                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2433                         td->ts.runtime[DDIR_WRITE] += rt[i];
2434                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2435                         td->ts.runtime[DDIR_TRIM] += rt[i];
2436         }
2437
2438         for_each_td(td, i) {
2439                 if (td->runstate >= TD_EXITED)
2440                         continue;
2441                 if (td->rusage_sem) {
2442                         td->update_rusage = 1;
2443                         fio_sem_down(td->rusage_sem);
2444                 }
2445                 td->update_rusage = 0;
2446         }
2447
2448         __show_run_stats();
2449
2450         for_each_td(td, i) {
2451                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2452                         td->ts.runtime[DDIR_READ] -= rt[i];
2453                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2454                         td->ts.runtime[DDIR_WRITE] -= rt[i];
2455                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2456                         td->ts.runtime[DDIR_TRIM] -= rt[i];
2457         }
2458
2459         free(rt);
2460         fio_sem_up(stat_sem);
2461 }
2462
2463 static bool status_interval_init;
2464 static struct timespec status_time;
2465 static bool status_file_disabled;
2466
2467 #define FIO_STATUS_FILE         "fio-dump-status"
2468
2469 static int check_status_file(void)
2470 {
2471         struct stat sb;
2472         const char *temp_dir;
2473         char fio_status_file_path[PATH_MAX];
2474
2475         if (status_file_disabled)
2476                 return 0;
2477
2478         temp_dir = getenv("TMPDIR");
2479         if (temp_dir == NULL) {
2480                 temp_dir = getenv("TEMP");
2481                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2482                         temp_dir = NULL;
2483         }
2484         if (temp_dir == NULL)
2485                 temp_dir = "/tmp";
2486 #ifdef __COVERITY__
2487         __coverity_tainted_data_sanitize__(temp_dir);
2488 #endif
2489
2490         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2491
2492         if (stat(fio_status_file_path, &sb))
2493                 return 0;
2494
2495         if (unlink(fio_status_file_path) < 0) {
2496                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2497                                                         strerror(errno));
2498                 log_err("fio: disabling status file updates\n");
2499                 status_file_disabled = true;
2500         }
2501
2502         return 1;
2503 }
2504
2505 void check_for_running_stats(void)
2506 {
2507         if (status_interval) {
2508                 if (!status_interval_init) {
2509                         fio_gettime(&status_time, NULL);
2510                         status_interval_init = true;
2511                 } else if (mtime_since_now(&status_time) >= status_interval) {
2512                         show_running_run_stats();
2513                         fio_gettime(&status_time, NULL);
2514                         return;
2515                 }
2516         }
2517         if (check_status_file()) {
2518                 show_running_run_stats();
2519                 return;
2520         }
2521 }
2522
2523 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2524 {
2525         double val = data;
2526         double delta;
2527
2528         if (data > is->max_val)
2529                 is->max_val = data;
2530         if (data < is->min_val)
2531                 is->min_val = data;
2532
2533         delta = val - is->mean.u.f;
2534         if (delta) {
2535                 is->mean.u.f += delta / (is->samples + 1.0);
2536                 is->S.u.f += delta * (val - is->mean.u.f);
2537         }
2538
2539         is->samples++;
2540 }
2541
2542 /*
2543  * Return a struct io_logs, which is added to the tail of the log
2544  * list for 'iolog'.
2545  */
2546 static struct io_logs *get_new_log(struct io_log *iolog)
2547 {
2548         size_t new_size, new_samples;
2549         struct io_logs *cur_log;
2550
2551         /*
2552          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2553          * forever
2554          */
2555         if (!iolog->cur_log_max)
2556                 new_samples = DEF_LOG_ENTRIES;
2557         else {
2558                 new_samples = iolog->cur_log_max * 2;
2559                 if (new_samples > MAX_LOG_ENTRIES)
2560                         new_samples = MAX_LOG_ENTRIES;
2561         }
2562
2563         new_size = new_samples * log_entry_sz(iolog);
2564
2565         cur_log = smalloc(sizeof(*cur_log));
2566         if (cur_log) {
2567                 INIT_FLIST_HEAD(&cur_log->list);
2568                 cur_log->log = malloc(new_size);
2569                 if (cur_log->log) {
2570                         cur_log->nr_samples = 0;
2571                         cur_log->max_samples = new_samples;
2572                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2573                         iolog->cur_log_max = new_samples;
2574                         return cur_log;
2575                 }
2576                 sfree(cur_log);
2577         }
2578
2579         return NULL;
2580 }
2581
2582 /*
2583  * Add and return a new log chunk, or return current log if big enough
2584  */
2585 static struct io_logs *regrow_log(struct io_log *iolog)
2586 {
2587         struct io_logs *cur_log;
2588         int i;
2589
2590         if (!iolog || iolog->disabled)
2591                 goto disable;
2592
2593         cur_log = iolog_cur_log(iolog);
2594         if (!cur_log) {
2595                 cur_log = get_new_log(iolog);
2596                 if (!cur_log)
2597                         return NULL;
2598         }
2599
2600         if (cur_log->nr_samples < cur_log->max_samples)
2601                 return cur_log;
2602
2603         /*
2604          * No room for a new sample. If we're compressing on the fly, flush
2605          * out the current chunk
2606          */
2607         if (iolog->log_gz) {
2608                 if (iolog_cur_flush(iolog, cur_log)) {
2609                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2610                         return NULL;
2611                 }
2612         }
2613
2614         /*
2615          * Get a new log array, and add to our list
2616          */
2617         cur_log = get_new_log(iolog);
2618         if (!cur_log) {
2619                 log_err("fio: failed extending iolog! Will stop logging.\n");
2620                 return NULL;
2621         }
2622
2623         if (!iolog->pending || !iolog->pending->nr_samples)
2624                 return cur_log;
2625
2626         /*
2627          * Flush pending items to new log
2628          */
2629         for (i = 0; i < iolog->pending->nr_samples; i++) {
2630                 struct io_sample *src, *dst;
2631
2632                 src = get_sample(iolog, iolog->pending, i);
2633                 dst = get_sample(iolog, cur_log, i);
2634                 memcpy(dst, src, log_entry_sz(iolog));
2635         }
2636         cur_log->nr_samples = iolog->pending->nr_samples;
2637
2638         iolog->pending->nr_samples = 0;
2639         return cur_log;
2640 disable:
2641         if (iolog)
2642                 iolog->disabled = true;
2643         return NULL;
2644 }
2645
2646 void regrow_logs(struct thread_data *td)
2647 {
2648         regrow_log(td->slat_log);
2649         regrow_log(td->clat_log);
2650         regrow_log(td->clat_hist_log);
2651         regrow_log(td->lat_log);
2652         regrow_log(td->bw_log);
2653         regrow_log(td->iops_log);
2654         td->flags &= ~TD_F_REGROW_LOGS;
2655 }
2656
2657 static struct io_logs *get_cur_log(struct io_log *iolog)
2658 {
2659         struct io_logs *cur_log;
2660
2661         cur_log = iolog_cur_log(iolog);
2662         if (!cur_log) {
2663                 cur_log = get_new_log(iolog);
2664                 if (!cur_log)
2665                         return NULL;
2666         }
2667
2668         if (cur_log->nr_samples < cur_log->max_samples)
2669                 return cur_log;
2670
2671         /*
2672          * Out of space. If we're in IO offload mode, or we're not doing
2673          * per unit logging (hence logging happens outside of the IO thread
2674          * as well), add a new log chunk inline. If we're doing inline
2675          * submissions, flag 'td' as needing a log regrow and we'll take
2676          * care of it on the submission side.
2677          */
2678         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2679             !per_unit_log(iolog))
2680                 return regrow_log(iolog);
2681
2682         if (iolog->td)
2683                 iolog->td->flags |= TD_F_REGROW_LOGS;
2684         if (iolog->pending)
2685                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2686         return iolog->pending;
2687 }
2688
2689 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2690                              enum fio_ddir ddir, unsigned long long bs,
2691                              unsigned long t, uint64_t offset, uint8_t priority_bit)
2692 {
2693         struct io_logs *cur_log;
2694
2695         if (iolog->disabled)
2696                 return;
2697         if (flist_empty(&iolog->io_logs))
2698                 iolog->avg_last[ddir] = t;
2699
2700         cur_log = get_cur_log(iolog);
2701         if (cur_log) {
2702                 struct io_sample *s;
2703
2704                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2705
2706                 s->data = data;
2707                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2708                 io_sample_set_ddir(iolog, s, ddir);
2709                 s->bs = bs;
2710                 s->priority_bit = priority_bit;
2711
2712                 if (iolog->log_offset) {
2713                         struct io_sample_offset *so = (void *) s;
2714
2715                         so->offset = offset;
2716                 }
2717
2718                 cur_log->nr_samples++;
2719                 return;
2720         }
2721
2722         iolog->disabled = true;
2723 }
2724
2725 static inline void reset_io_stat(struct io_stat *ios)
2726 {
2727         ios->min_val = -1ULL;
2728         ios->max_val = ios->samples = 0;
2729         ios->mean.u.f = ios->S.u.f = 0;
2730 }
2731
2732 void reset_io_stats(struct thread_data *td)
2733 {
2734         struct thread_stat *ts = &td->ts;
2735         int i, j, k;
2736
2737         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2738                 reset_io_stat(&ts->clat_high_prio_stat[i]);
2739                 reset_io_stat(&ts->clat_prio_stat[i]);
2740                 reset_io_stat(&ts->clat_stat[i]);
2741                 reset_io_stat(&ts->slat_stat[i]);
2742                 reset_io_stat(&ts->lat_stat[i]);
2743                 reset_io_stat(&ts->bw_stat[i]);
2744                 reset_io_stat(&ts->iops_stat[i]);
2745
2746                 ts->io_bytes[i] = 0;
2747                 ts->runtime[i] = 0;
2748                 ts->total_io_u[i] = 0;
2749                 ts->short_io_u[i] = 0;
2750                 ts->drop_io_u[i] = 0;
2751
2752                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2753                         ts->io_u_plat_high_prio[i][j] = 0;
2754                         ts->io_u_plat_prio[i][j] = 0;
2755                         if (!i)
2756                                 ts->io_u_sync_plat[j] = 0;
2757                 }
2758         }
2759
2760         for (i = 0; i < FIO_LAT_CNT; i++)
2761                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
2762                         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2763                                 ts->io_u_plat[i][j][k] = 0;
2764
2765         ts->total_io_u[DDIR_SYNC] = 0;
2766
2767         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2768                 ts->io_u_map[i] = 0;
2769                 ts->io_u_submit[i] = 0;
2770                 ts->io_u_complete[i] = 0;
2771         }
2772
2773         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2774                 ts->io_u_lat_n[i] = 0;
2775         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2776                 ts->io_u_lat_u[i] = 0;
2777         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2778                 ts->io_u_lat_m[i] = 0;
2779
2780         ts->total_submit = 0;
2781         ts->total_complete = 0;
2782         ts->nr_zone_resets = 0;
2783         ts->cachehit = ts->cachemiss = 0;
2784 }
2785
2786 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2787                               unsigned long elapsed, bool log_max, uint8_t priority_bit)
2788 {
2789         /*
2790          * Note an entry in the log. Use the mean from the logged samples,
2791          * making sure to properly round up. Only write a log entry if we
2792          * had actual samples done.
2793          */
2794         if (iolog->avg_window[ddir].samples) {
2795                 union io_sample_data data;
2796
2797                 if (log_max)
2798                         data.val = iolog->avg_window[ddir].max_val;
2799                 else
2800                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2801
2802                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, priority_bit);
2803         }
2804
2805         reset_io_stat(&iolog->avg_window[ddir]);
2806 }
2807
2808 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2809                              bool log_max, uint8_t priority_bit)
2810 {
2811         int ddir;
2812
2813         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2814                 __add_stat_to_log(iolog, ddir, elapsed, log_max, priority_bit);
2815 }
2816
2817 static unsigned long add_log_sample(struct thread_data *td,
2818                                     struct io_log *iolog,
2819                                     union io_sample_data data,
2820                                     enum fio_ddir ddir, unsigned long long bs,
2821                                     uint64_t offset, uint8_t priority_bit)
2822 {
2823         unsigned long elapsed, this_window;
2824
2825         if (!ddir_rw(ddir))
2826                 return 0;
2827
2828         elapsed = mtime_since_now(&td->epoch);
2829
2830         /*
2831          * If no time averaging, just add the log sample.
2832          */
2833         if (!iolog->avg_msec) {
2834                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset, priority_bit);
2835                 return 0;
2836         }
2837
2838         /*
2839          * Add the sample. If the time period has passed, then
2840          * add that entry to the log and clear.
2841          */
2842         add_stat_sample(&iolog->avg_window[ddir], data.val);
2843
2844         /*
2845          * If period hasn't passed, adding the above sample is all we
2846          * need to do.
2847          */
2848         this_window = elapsed - iolog->avg_last[ddir];
2849         if (elapsed < iolog->avg_last[ddir])
2850                 return iolog->avg_last[ddir] - elapsed;
2851         else if (this_window < iolog->avg_msec) {
2852                 unsigned long diff = iolog->avg_msec - this_window;
2853
2854                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2855                         return diff;
2856         }
2857
2858         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0, priority_bit);
2859
2860         iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2861         return iolog->avg_msec;
2862 }
2863
2864 void finalize_logs(struct thread_data *td, bool unit_logs)
2865 {
2866         unsigned long elapsed;
2867
2868         elapsed = mtime_since_now(&td->epoch);
2869
2870         if (td->clat_log && unit_logs)
2871                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0, 0);
2872         if (td->slat_log && unit_logs)
2873                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0, 0);
2874         if (td->lat_log && unit_logs)
2875                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0, 0);
2876         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2877                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0, 0);
2878         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2879                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0, 0);
2880 }
2881
2882 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned long long bs,
2883                                         uint8_t priority_bit)
2884 {
2885         struct io_log *iolog;
2886
2887         if (!ddir_rw(ddir))
2888                 return;
2889
2890         iolog = agg_io_log[ddir];
2891         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, priority_bit);
2892 }
2893
2894 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2895 {
2896         unsigned int idx = plat_val_to_idx(nsec);
2897         assert(idx < FIO_IO_U_PLAT_NR);
2898
2899         ts->io_u_sync_plat[idx]++;
2900         add_stat_sample(&ts->sync_stat, nsec);
2901 }
2902
2903 static void add_clat_percentile_sample(struct thread_stat *ts,
2904                                 unsigned long long nsec, enum fio_ddir ddir, uint8_t priority_bit)
2905 {
2906         unsigned int idx = plat_val_to_idx(nsec);
2907         assert(idx < FIO_IO_U_PLAT_NR);
2908
2909         ts->io_u_plat[FIO_CLAT][ddir][idx]++;
2910
2911         if (!priority_bit) {
2912                 ts->io_u_plat_prio[ddir][idx]++;
2913         } else {
2914                 ts->io_u_plat_high_prio[ddir][idx]++;
2915         }
2916 }
2917
2918 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2919                      unsigned long long nsec, unsigned long long bs,
2920                      uint64_t offset, uint8_t priority_bit)
2921 {
2922         const bool needs_lock = td_async_processing(td);
2923         unsigned long elapsed, this_window;
2924         struct thread_stat *ts = &td->ts;
2925         struct io_log *iolog = td->clat_hist_log;
2926
2927         if (needs_lock)
2928                 __td_io_u_lock(td);
2929
2930         add_stat_sample(&ts->clat_stat[ddir], nsec);
2931
2932         if (priority_bit) {
2933                 add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
2934         } else {
2935                 add_stat_sample(&ts->clat_prio_stat[ddir], nsec);
2936         }
2937
2938         if (td->clat_log)
2939                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2940                                offset, priority_bit);
2941
2942         if (ts->clat_percentiles) {
2943                 add_clat_percentile_sample(ts, nsec, ddir, priority_bit);
2944         }
2945
2946         if (iolog && iolog->hist_msec) {
2947                 struct io_hist *hw = &iolog->hist_window[ddir];
2948
2949                 hw->samples++;
2950                 elapsed = mtime_since_now(&td->epoch);
2951                 if (!hw->hist_last)
2952                         hw->hist_last = elapsed;
2953                 this_window = elapsed - hw->hist_last;
2954
2955                 if (this_window >= iolog->hist_msec) {
2956                         uint64_t *io_u_plat;
2957                         struct io_u_plat_entry *dst;
2958
2959                         /*
2960                          * Make a byte-for-byte copy of the latency histogram
2961                          * stored in td->ts.io_u_plat[ddir], recording it in a
2962                          * log sample. Note that the matching call to free() is
2963                          * located in iolog.c after printing this sample to the
2964                          * log file.
2965                          */
2966                         io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
2967                         dst = malloc(sizeof(struct io_u_plat_entry));
2968                         memcpy(&(dst->io_u_plat), io_u_plat,
2969                                 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
2970                         flist_add(&dst->list, &hw->list);
2971                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2972                                                 elapsed, offset, priority_bit);
2973
2974                         /*
2975                          * Update the last time we recorded as being now, minus
2976                          * any drift in time we encountered before actually
2977                          * making the record.
2978                          */
2979                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2980                         hw->samples = 0;
2981                 }
2982         }
2983
2984         if (needs_lock)
2985                 __td_io_u_unlock(td);
2986 }
2987
2988 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2989                         unsigned long usec, unsigned long long bs, uint64_t offset,
2990                         uint8_t priority_bit)
2991 {
2992         const bool needs_lock = td_async_processing(td);
2993         struct thread_stat *ts = &td->ts;
2994
2995         if (!ddir_rw(ddir))
2996                 return;
2997
2998         if (needs_lock)
2999                 __td_io_u_lock(td);
3000
3001         add_stat_sample(&ts->slat_stat[ddir], usec);
3002
3003         if (td->slat_log)
3004                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset,
3005                         priority_bit);
3006
3007         if (needs_lock)
3008                 __td_io_u_unlock(td);
3009 }
3010
3011 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3012                     unsigned long long nsec, unsigned long long bs,
3013                     uint64_t offset, uint8_t priority_bit)
3014 {
3015         const bool needs_lock = td_async_processing(td);
3016         struct thread_stat *ts = &td->ts;
3017
3018         if (!ddir_rw(ddir))
3019                 return;
3020
3021         if (needs_lock)
3022                 __td_io_u_lock(td);
3023
3024         add_stat_sample(&ts->lat_stat[ddir], nsec);
3025
3026         if (td->lat_log)
3027                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3028                                offset, priority_bit);
3029
3030         if (ts->lat_percentiles)
3031                 add_clat_percentile_sample(ts, nsec, ddir, priority_bit);
3032
3033         if (needs_lock)
3034                 __td_io_u_unlock(td);
3035 }
3036
3037 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3038                    unsigned int bytes, unsigned long long spent)
3039 {
3040         const bool needs_lock = td_async_processing(td);
3041         struct thread_stat *ts = &td->ts;
3042         unsigned long rate;
3043
3044         if (spent)
3045                 rate = (unsigned long) (bytes * 1000000ULL / spent);
3046         else
3047                 rate = 0;
3048
3049         if (needs_lock)
3050                 __td_io_u_lock(td);
3051
3052         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3053
3054         if (td->bw_log)
3055                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3056                                bytes, io_u->offset, io_u_is_prio(io_u));
3057
3058         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3059
3060         if (needs_lock)
3061                 __td_io_u_unlock(td);
3062 }
3063
3064 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3065                          struct timespec *t, unsigned int avg_time,
3066                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3067                          struct io_stat *stat, struct io_log *log,
3068                          bool is_kb)
3069 {
3070         const bool needs_lock = td_async_processing(td);
3071         unsigned long spent, rate;
3072         enum fio_ddir ddir;
3073         unsigned long next, next_log;
3074
3075         next_log = avg_time;
3076
3077         spent = mtime_since(parent_tv, t);
3078         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
3079                 return avg_time - spent;
3080
3081         if (needs_lock)
3082                 __td_io_u_lock(td);
3083
3084         /*
3085          * Compute both read and write rates for the interval.
3086          */
3087         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3088                 uint64_t delta;
3089
3090                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3091                 if (!delta)
3092                         continue; /* No entries for interval */
3093
3094                 if (spent) {
3095                         if (is_kb)
3096                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
3097                         else
3098                                 rate = (delta * 1000) / spent;
3099                 } else
3100                         rate = 0;
3101
3102                 add_stat_sample(&stat[ddir], rate);
3103
3104                 if (log) {
3105                         unsigned long long bs = 0;
3106
3107                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3108                                 bs = td->o.min_bs[ddir];
3109
3110                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0, 0);
3111                         next_log = min(next_log, next);
3112                 }
3113
3114                 stat_io_bytes[ddir] = this_io_bytes[ddir];
3115         }
3116
3117         *parent_tv = *t;
3118
3119         if (needs_lock)
3120                 __td_io_u_unlock(td);
3121
3122         if (spent <= avg_time)
3123                 next = avg_time;
3124         else
3125                 next = avg_time - (1 + spent - avg_time);
3126
3127         return min(next, next_log);
3128 }
3129
3130 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3131 {
3132         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3133                                 td->this_io_bytes, td->stat_io_bytes,
3134                                 td->ts.bw_stat, td->bw_log, true);
3135 }
3136
3137 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3138                      unsigned int bytes)
3139 {
3140         const bool needs_lock = td_async_processing(td);
3141         struct thread_stat *ts = &td->ts;
3142
3143         if (needs_lock)
3144                 __td_io_u_lock(td);
3145
3146         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3147
3148         if (td->iops_log)
3149                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3150                                bytes, io_u->offset, io_u_is_prio(io_u));
3151
3152         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3153
3154         if (needs_lock)
3155                 __td_io_u_unlock(td);
3156 }
3157
3158 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3159 {
3160         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3161                                 td->this_io_blocks, td->stat_io_blocks,
3162                                 td->ts.iops_stat, td->iops_log, false);
3163 }
3164
3165 /*
3166  * Returns msecs to next event
3167  */
3168 int calc_log_samples(void)
3169 {
3170         struct thread_data *td;
3171         unsigned int next = ~0U, tmp;
3172         struct timespec now;
3173         int i;
3174
3175         fio_gettime(&now, NULL);
3176
3177         for_each_td(td, i) {
3178                 if (!td->o.stats)
3179                         continue;
3180                 if (in_ramp_time(td) ||
3181                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3182                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3183                         continue;
3184                 }
3185                 if (!td->bw_log ||
3186                         (td->bw_log && !per_unit_log(td->bw_log))) {
3187                         tmp = add_bw_samples(td, &now);
3188                         if (tmp < next)
3189                                 next = tmp;
3190                 }
3191                 if (!td->iops_log ||
3192                         (td->iops_log && !per_unit_log(td->iops_log))) {
3193                         tmp = add_iops_samples(td, &now);
3194                         if (tmp < next)
3195                                 next = tmp;
3196                 }
3197         }
3198
3199         return next == ~0U ? 0 : next;
3200 }
3201
3202 void stat_init(void)
3203 {
3204         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3205 }
3206
3207 void stat_exit(void)
3208 {
3209         /*
3210          * When we have the mutex, we know out-of-band access to it
3211          * have ended.
3212          */
3213         fio_sem_down(stat_sem);
3214         fio_sem_remove(stat_sem);
3215 }
3216
3217 /*
3218  * Called from signal handler. Wake up status thread.
3219  */
3220 void show_running_run_stats(void)
3221 {
3222         helper_do_stat();
3223 }
3224
3225 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3226 {
3227         /* Ignore io_u's which span multiple blocks--they will just get
3228          * inaccurate counts. */
3229         int idx = (io_u->offset - io_u->file->file_offset)
3230                         / td->o.bs[DDIR_TRIM];
3231         uint32_t *info = &td->ts.block_infos[idx];
3232         assert(idx < td->ts.nr_block_infos);
3233         return info;
3234 }