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 #ifdef WIN32
  21 #define LOG_MSEC_SLACK  2
  22 #else
  23 #define LOG_MSEC_SLACK  1
  24 #endif
  25
  26 struct fio_sem *stat_sem;
  27
  28 void clear_rusage_stat(struct thread_data *td)
  29 {
  30         struct thread_stat *ts = &td->ts;
  31
  32         fio_getrusage(&td->ru_start);
  33         ts->usr_time = ts->sys_time = 0;
  34         ts->ctx = 0;
  35         ts->minf = ts->majf = 0;
  36 }
  37
  38 void update_rusage_stat(struct thread_data *td)
  39 {
  40         struct thread_stat *ts = &td->ts;
  41
  42         fio_getrusage(&td->ru_end);
  43         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
  44                                         &td->ru_end.ru_utime);
  45         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
  46                                         &td->ru_end.ru_stime);
  47         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  48                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  49         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  50         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  51
  52         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  53 }
  54
  55 /*
  56  * Given a latency, return the index of the corresponding bucket in
  57  * the structure tracking percentiles.
  58  *
  59  * (1) find the group (and error bits) that the value (latency)
  60  * belongs to by looking at its MSB. (2) find the bucket number in the
  61  * group by looking at the index bits.
  62  *
  63  */
  64 static unsigned int plat_val_to_idx(unsigned long long val)
  65 {
  66         unsigned int msb, error_bits, base, offset, idx;
  67
  68         /* Find MSB starting from bit 0 */
  69         if (val == 0)
  70                 msb = 0;
  71         else
  72                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
  73
  74         /*
  75          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  76          * all bits of the sample as index
  77          */
  78         if (msb <= FIO_IO_U_PLAT_BITS)
  79                 return val;
  80
  81         /* Compute the number of error bits to discard*/
  82         error_bits = msb - FIO_IO_U_PLAT_BITS;
  83
  84         /* Compute the number of buckets before the group */
  85         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  86
  87         /*
  88          * Discard the error bits and apply the mask to find the
  89          * index for the buckets in the group
  90          */
  91         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  92
  93         /* Make sure the index does not exceed (array size - 1) */
  94         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  95                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  96
  97         return idx;
  98 }
  99
 100 /*
 101  * Convert the given index of the bucket array to the value
 102  * represented by the bucket
 103  */
 104 static unsigned long long plat_idx_to_val(unsigned int idx)
 105 {
 106         unsigned int error_bits;
 107         unsigned long long k, base;
 108
 109         assert(idx < FIO_IO_U_PLAT_NR);
 110
 111         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 112          * all bits of the sample as index */
 113         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 114                 return idx;
 115
 116         /* Find the group and compute the minimum value of that group */
 117         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 118         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
 119
 120         /* Find its bucket number of the group */
 121         k = idx % FIO_IO_U_PLAT_VAL;
 122
 123         /* Return the mean of the range of the bucket */
 124         return base + ((k + 0.5) * (1 << error_bits));
 125 }
 126
 127 static int double_cmp(const void *a, const void *b)
 128 {
 129         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 130         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 131         int cmp = 0;
 132
 133         if (fa.u.f > fb.u.f)
 134                 cmp = 1;
 135         else if (fa.u.f < fb.u.f)
 136                 cmp = -1;
 137
 138         return cmp;
 139 }
 140
 141 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 142                                    fio_fp64_t *plist, unsigned long long **output,
 143                                    unsigned long long *maxv, unsigned long long *minv)
 144 {
 145         unsigned long long sum = 0;
 146         unsigned int len, i, j = 0;
 147         unsigned long long *ovals = NULL;
 148         bool is_last;
 149
 150         *minv = -1ULL;
 151         *maxv = 0;
 152
 153         len = 0;
 154         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 155                 len++;
 156
 157         if (!len)
 158                 return 0;
 159
 160         /*
 161          * Sort the percentile list. Note that it may already be sorted if
 162          * we are using the default values, but since it's a short list this
 163          * isn't a worry. Also note that this does not work for NaN values.
 164          */
 165         if (len > 1)
 166                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 167
 168         ovals = malloc(len * sizeof(*ovals));
 169         if (!ovals)
 170                 return 0;
 171
 172         /*
 173          * Calculate bucket values, note down max and min values
 174          */
 175         is_last = false;
 176         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 177                 sum += io_u_plat[i];
 178                 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
 179                         assert(plist[j].u.f <= 100.0);
 180
 181                         ovals[j] = plat_idx_to_val(i);
 182                         if (ovals[j] < *minv)
 183                                 *minv = ovals[j];
 184                         if (ovals[j] > *maxv)
 185                                 *maxv = ovals[j];
 186
 187                         is_last = (j == len - 1) != 0;
 188                         if (is_last)
 189                                 break;
 190
 191                         j++;
 192                 }
 193         }
 194
 195         if (!is_last)
 196                 log_err("fio: error calculating latency percentiles\n");
 197
 198         *output = ovals;
 199         return len;
 200 }
 201
 202 /*
 203  * Find and display the p-th percentile of clat
 204  */
 205 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 206                                   fio_fp64_t *plist, unsigned int precision,
 207                                   const char *pre, struct buf_output *out)
 208 {
 209         unsigned int divisor, len, i, j = 0;
 210         unsigned long long minv, maxv;
 211         unsigned long long *ovals;
 212         int per_line, scale_down, time_width;
 213         bool is_last;
 214         char fmt[32];
 215
 216         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 217         if (!len || !ovals)
 218                 return;
 219
 220         /*
 221          * We default to nsecs, but if the value range is such that we
 222          * should scale down to usecs or msecs, do that.
 223          */
 224         if (minv > 2000000 && maxv > 99999999ULL) {
 225                 scale_down = 2;
 226                 divisor = 1000000;
 227                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 228         } else if (minv > 2000 && maxv > 99999) {
 229                 scale_down = 1;
 230                 divisor = 1000;
 231                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 232         } else {
 233                 scale_down = 0;
 234                 divisor = 1;
 235                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 236         }
 237
 238
 239         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 240         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 241                         precision, time_width);
 242         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 243         per_line = (80 - 7) / (precision + 10 + time_width);
 244
 245         for (j = 0; j < len; j++) {
 246                 /* for formatting */
 247                 if (j != 0 && (j % per_line) == 0)
 248                         log_buf(out, "     |");
 249
 250                 /* end of the list */
 251                 is_last = (j == len - 1) != 0;
 252
 253                 for (i = 0; i < scale_down; i++)
 254                         ovals[j] = (ovals[j] + 999) / 1000;
 255
 256                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 257
 258                 if (is_last)
 259                         break;
 260
 261                 if ((j % per_line) == per_line - 1)     /* for formatting */
 262                         log_buf(out, "\n");
 263         }
 264
 265         free(ovals);
 266 }
 267
 268 static int get_nr_prios_with_samples(struct thread_stat *ts, enum fio_ddir ddir)
 269 {
 270         int i, nr_prios_with_samples = 0;
 271
 272         for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
 273                 if (ts->clat_prio[ddir][i].clat_stat.samples)
 274                         nr_prios_with_samples++;
 275         }
 276
 277         return nr_prios_with_samples;
 278 }
 279
 280 bool calc_lat(struct io_stat *is, unsigned long long *min,
 281               unsigned long long *max, double *mean, double *dev)
 282 {
 283         double n = (double) is->samples;
 284
 285         if (n == 0)
 286                 return false;
 287
 288         *min = is->min_val;
 289         *max = is->max_val;
 290         *mean = is->mean.u.f;
 291
 292         if (n > 1.0)
 293                 *dev = sqrt(is->S.u.f / (n - 1.0));
 294         else
 295                 *dev = 0;
 296
 297         return true;
 298 }
 299
 300 void show_mixed_group_stats(struct group_run_stats *rs, struct buf_output *out)
 301 {
 302         char *io, *agg, *min, *max;
 303         char *ioalt, *aggalt, *minalt, *maxalt;
 304         uint64_t io_mix = 0, agg_mix = 0, min_mix = -1, max_mix = 0;
 305         uint64_t min_run = -1, max_run = 0;
 306         const int i2p = is_power_of_2(rs->kb_base);
 307         int i;
 308
 309         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 310                 if (!rs->max_run[i])
 311                         continue;
 312                 io_mix += rs->iobytes[i];
 313                 agg_mix += rs->agg[i];
 314                 min_mix = min_mix < rs->min_bw[i] ? min_mix : rs->min_bw[i];
 315                 max_mix = max_mix > rs->max_bw[i] ? max_mix : rs->max_bw[i];
 316                 min_run = min_run < rs->min_run[i] ? min_run : rs->min_run[i];
 317                 max_run = max_run > rs->max_run[i] ? max_run : rs->max_run[i];
 318         }
 319         io = num2str(io_mix, rs->sig_figs, 1, i2p, N2S_BYTE);
 320         ioalt = num2str(io_mix, rs->sig_figs, 1, !i2p, N2S_BYTE);
 321         agg = num2str(agg_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 322         aggalt = num2str(agg_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 323         min = num2str(min_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 324         minalt = num2str(min_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 325         max = num2str(max_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 326         maxalt = num2str(max_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 327         log_buf(out, "  MIXED: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 328                         agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 329                         (unsigned long long) min_run,
 330                         (unsigned long long) max_run);
 331         free(io);
 332         free(agg);
 333         free(min);
 334         free(max);
 335         free(ioalt);
 336         free(aggalt);
 337         free(minalt);
 338         free(maxalt);
 339 }
 340
 341 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 342 {
 343         char *io, *agg, *min, *max;
 344         char *ioalt, *aggalt, *minalt, *maxalt;
 345         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 346         int i;
 347
 348         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 349
 350         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 351                 const int i2p = is_power_of_2(rs->kb_base);
 352
 353                 if (!rs->max_run[i])
 354                         continue;
 355
 356                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
 357                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
 358                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
 359                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 360                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 361                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 362                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 363                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 364                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 365                                 (rs->unified_rw_rep == UNIFIED_MIXED) ? "  MIXED" : str[i],
 366                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 367                                 (unsigned long long) rs->min_run[i],
 368                                 (unsigned long long) rs->max_run[i]);
 369
 370                 free(io);
 371                 free(agg);
 372                 free(min);
 373                 free(max);
 374                 free(ioalt);
 375                 free(aggalt);
 376                 free(minalt);
 377                 free(maxalt);
 378         }
 379
 380         /* Need to aggregate statistics to show mixed values */
 381         if (rs->unified_rw_rep == UNIFIED_BOTH)
 382                 show_mixed_group_stats(rs, out);
 383 }
 384
 385 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
 386 {
 387         int i;
 388
 389         /*
 390          * Do depth distribution calculations
 391          */
 392         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 393                 if (total) {
 394                         io_u_dist[i] = (double) map[i] / (double) total;
 395                         io_u_dist[i] *= 100.0;
 396                         if (io_u_dist[i] < 0.1 && map[i])
 397                                 io_u_dist[i] = 0.1;
 398                 } else
 399                         io_u_dist[i] = 0.0;
 400         }
 401 }
 402
 403 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 404                           uint64_t *src, int nr)
 405 {
 406         unsigned long total = ddir_rw_sum(ts->total_io_u);
 407         int i;
 408
 409         /*
 410          * Do latency distribution calculations
 411          */
 412         for (i = 0; i < nr; i++) {
 413                 if (total) {
 414                         dst[i] = (double) src[i] / (double) total;
 415                         dst[i] *= 100.0;
 416                         if (dst[i] < 0.01 && src[i])
 417                                 dst[i] = 0.01;
 418                 } else
 419                         dst[i] = 0.0;
 420         }
 421 }
 422
 423 /*
 424  * To keep the terse format unaltered, add all of the ns latency
 425  * buckets to the first us latency bucket
 426  */
 427 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 428 {
 429         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 430         int i;
 431
 432         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 433
 434         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 435                 ntotal += ts->io_u_lat_n[i];
 436
 437         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 438 }
 439
 440 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 441 {
 442         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 443 }
 444
 445 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 446 {
 447         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 448 }
 449
 450 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 451 {
 452         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 453 }
 454
 455 static void display_lat(const char *name, unsigned long long min,
 456                         unsigned long long max, double mean, double dev,
 457                         struct buf_output *out)
 458 {
 459         const char *base = "(nsec)";
 460         char *minp, *maxp;
 461
 462         if (nsec_to_msec(&min, &max, &mean, &dev))
 463                 base = "(msec)";
 464         else if (nsec_to_usec(&min, &max, &mean, &dev))
 465                 base = "(usec)";
 466
 467         minp = num2str(min, 6, 1, 0, N2S_NONE);
 468         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 469
 470         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 471                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 472
 473         free(minp);
 474         free(maxp);
 475 }
 476
 477 static struct thread_stat *gen_mixed_ddir_stats_from_ts(struct thread_stat *ts)
 478 {
 479         struct thread_stat *ts_lcl;
 480
 481         /*
 482          * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
 483          * Trims (ddir = 2)
 484          */
 485         ts_lcl = malloc(sizeof(struct thread_stat));
 486         if (!ts_lcl) {
 487                 log_err("fio: failed to allocate local thread stat\n");
 488                 return NULL;
 489         }
 490
 491         init_thread_stat(ts_lcl);
 492
 493         /* calculate mixed stats  */
 494         ts_lcl->unified_rw_rep = UNIFIED_MIXED;
 495         ts_lcl->lat_percentiles = ts->lat_percentiles;
 496         ts_lcl->clat_percentiles = ts->clat_percentiles;
 497         ts_lcl->slat_percentiles = ts->slat_percentiles;
 498         ts_lcl->percentile_precision = ts->percentile_precision;
 499         memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
 500
 501         sum_thread_stats(ts_lcl, ts);
 502
 503         return ts_lcl;
 504 }
 505
 506 static double convert_agg_kbytes_percent(struct group_run_stats *rs,
 507                                          enum fio_ddir ddir, int mean)
 508 {
 509         double p_of_agg = 100.0;
 510         if (rs && rs->agg[ddir] > 1024) {
 511                 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
 512
 513                 if (p_of_agg > 100.0)
 514                         p_of_agg = 100.0;
 515         }
 516         return p_of_agg;
 517 }
 518
 519 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 520                              enum fio_ddir ddir, struct buf_output *out)
 521 {
 522         unsigned long runt;
 523         unsigned long long min, max, bw, iops;
 524         double mean, dev;
 525         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
 526         int i2p, i;
 527         const char *clat_type = ts->lat_percentiles ? "lat" : "clat";
 528
 529         if (ddir_sync(ddir)) {
 530                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
 531                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
 532                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
 533                         show_clat_percentiles(ts->io_u_sync_plat,
 534                                                 ts->sync_stat.samples,
 535                                                 ts->percentile_list,
 536                                                 ts->percentile_precision,
 537                                                 io_ddir_name(ddir), out);
 538                 }
 539                 return;
 540         }
 541
 542         assert(ddir_rw(ddir));
 543
 544         if (!ts->runtime[ddir])
 545                 return;
 546
 547         i2p = is_power_of_2(rs->kb_base);
 548         runt = ts->runtime[ddir];
 549
 550         bw = (1000 * ts->io_bytes[ddir]) / runt;
 551         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
 552         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
 553         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
 554
 555         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 556         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
 557         if (ddir == DDIR_WRITE)
 558                 post_st = zbd_write_status(ts);
 559         else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
 560                 uint64_t total;
 561                 double hit;
 562
 563                 total = ts->cachehit + ts->cachemiss;
 564                 hit = (double) ts->cachehit / (double) total;
 565                 hit *= 100.0;
 566                 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
 567                         post_st = NULL;
 568         }
 569
 570         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
 571                         (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
 572                         iops_p, bw_p, bw_p_alt, io_p,
 573                         (unsigned long long) ts->runtime[ddir],
 574                         post_st ? : "");
 575
 576         free(post_st);
 577         free(io_p);
 578         free(bw_p);
 579         free(bw_p_alt);
 580         free(iops_p);
 581
 582         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 583                 display_lat("slat", min, max, mean, dev, out);
 584         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 585                 display_lat("clat", min, max, mean, dev, out);
 586         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 587                 display_lat(" lat", min, max, mean, dev, out);
 588
 589         /* Only print per prio stats if there are >= 2 prios with samples */
 590         if (get_nr_prios_with_samples(ts, ddir) >= 2) {
 591                 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
 592                         if (calc_lat(&ts->clat_prio[ddir][i].clat_stat, &min,
 593                                      &max, &mean, &dev)) {
 594                                 char buf[64];
 595
 596                                 snprintf(buf, sizeof(buf),
 597                                          "%s prio %u/%u",
 598                                          clat_type,
 599                                          ts->clat_prio[ddir][i].ioprio >> 13,
 600                                          ts->clat_prio[ddir][i].ioprio & 7);
 601                                 display_lat(buf, min, max, mean, dev, out);
 602                         }
 603                 }
 604         }
 605
 606         if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
 607                 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
 608                                         ts->slat_stat[ddir].samples,
 609                                         ts->percentile_list,
 610                                         ts->percentile_precision, "slat", out);
 611         if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
 612                 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
 613                                         ts->clat_stat[ddir].samples,
 614                                         ts->percentile_list,
 615                                         ts->percentile_precision, "clat", out);
 616         if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
 617                 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
 618                                         ts->lat_stat[ddir].samples,
 619                                         ts->percentile_list,
 620                                         ts->percentile_precision, "lat", out);
 621
 622         if (ts->clat_percentiles || ts->lat_percentiles) {
 623                 char prio_name[64];
 624                 uint64_t samples;
 625
 626                 if (ts->lat_percentiles)
 627                         samples = ts->lat_stat[ddir].samples;
 628                 else
 629                         samples = ts->clat_stat[ddir].samples;
 630
 631                 /* Only print per prio stats if there are >= 2 prios with samples */
 632                 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
 633                         for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
 634                                 uint64_t prio_samples = ts->clat_prio[ddir][i].clat_stat.samples;
 635
 636                                 if (prio_samples > 0) {
 637                                         snprintf(prio_name, sizeof(prio_name),
 638                                                  "%s prio %u/%u (%.2f%% of IOs)",
 639                                                  clat_type,
 640                                                  ts->clat_prio[ddir][i].ioprio >> 13,
 641                                                  ts->clat_prio[ddir][i].ioprio & 7,
 642                                                  100. * (double) prio_samples / (double) samples);
 643                                         show_clat_percentiles(ts->clat_prio[ddir][i].io_u_plat,
 644                                                               prio_samples, ts->percentile_list,
 645                                                               ts->percentile_precision,
 646                                                               prio_name, out);
 647                                 }
 648                         }
 649                 }
 650         }
 651
 652         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 653                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 654                 const char *bw_str;
 655
 656                 if ((rs->unit_base == 1) && i2p)
 657                         bw_str = "Kibit";
 658                 else if (rs->unit_base == 1)
 659                         bw_str = "kbit";
 660                 else if (i2p)
 661                         bw_str = "KiB";
 662                 else
 663                         bw_str = "kB";
 664
 665                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
 666
 667                 if (rs->unit_base == 1) {
 668                         min *= 8.0;
 669                         max *= 8.0;
 670                         mean *= 8.0;
 671                         dev *= 8.0;
 672                 }
 673
 674                 if (mean > fkb_base * fkb_base) {
 675                         min /= fkb_base;
 676                         max /= fkb_base;
 677                         mean /= fkb_base;
 678                         dev /= fkb_base;
 679                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 680                 }
 681
 682                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 683                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 684                         bw_str, min, max, p_of_agg, mean, dev,
 685                         (&ts->bw_stat[ddir])->samples);
 686         }
 687         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 688                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 689                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 690                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 691         }
 692 }
 693
 694 static void show_mixed_ddir_status(struct group_run_stats *rs,
 695                                    struct thread_stat *ts,
 696                                    struct buf_output *out)
 697 {
 698         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
 699
 700         if (ts_lcl)
 701                 show_ddir_status(rs, ts_lcl, DDIR_READ, out);
 702
 703         free_clat_prio_stats(ts_lcl);
 704         free(ts_lcl);
 705 }
 706
 707 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
 708                      const char *msg, struct buf_output *out)
 709 {
 710         bool new_line = true, shown = false;
 711         int i, line = 0;
 712
 713         for (i = 0; i < nr; i++) {
 714                 if (io_u_lat[i] <= 0.0)
 715                         continue;
 716                 shown = true;
 717                 if (new_line) {
 718                         if (line)
 719                                 log_buf(out, "\n");
 720                         log_buf(out, "  lat (%s)   : ", msg);
 721                         new_line = false;
 722                         line = 0;
 723                 }
 724                 if (line)
 725                         log_buf(out, ", ");
 726                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 727                 line++;
 728                 if (line == 5)
 729                         new_line = true;
 730         }
 731
 732         if (shown)
 733                 log_buf(out, "\n");
 734
 735         return true;
 736 }
 737
 738 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 739 {
 740         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 741                                  "250=", "500=", "750=", "1000=", };
 742
 743         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 744 }
 745
 746 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 747 {
 748         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 749                                  "250=", "500=", "750=", "1000=", };
 750
 751         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 752 }
 753
 754 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 755 {
 756         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 757                                  "250=", "500=", "750=", "1000=", "2000=",
 758                                  ">=2000=", };
 759
 760         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 761 }
 762
 763 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 764 {
 765         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 766         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 767         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 768
 769         stat_calc_lat_n(ts, io_u_lat_n);
 770         stat_calc_lat_u(ts, io_u_lat_u);
 771         stat_calc_lat_m(ts, io_u_lat_m);
 772
 773         show_lat_n(io_u_lat_n, out);
 774         show_lat_u(io_u_lat_u, out);
 775         show_lat_m(io_u_lat_m, out);
 776 }
 777
 778 static int block_state_category(int block_state)
 779 {
 780         switch (block_state) {
 781         case BLOCK_STATE_UNINIT:
 782                 return 0;
 783         case BLOCK_STATE_TRIMMED:
 784         case BLOCK_STATE_WRITTEN:
 785                 return 1;
 786         case BLOCK_STATE_WRITE_FAILURE:
 787         case BLOCK_STATE_TRIM_FAILURE:
 788                 return 2;
 789         default:
 790                 /* Silence compile warning on some BSDs and have a return */
 791                 assert(0);
 792                 return -1;
 793         }
 794 }
 795
 796 static int compare_block_infos(const void *bs1, const void *bs2)
 797 {
 798         uint64_t block1 = *(uint64_t *)bs1;
 799         uint64_t block2 = *(uint64_t *)bs2;
 800         int state1 = BLOCK_INFO_STATE(block1);
 801         int state2 = BLOCK_INFO_STATE(block2);
 802         int bscat1 = block_state_category(state1);
 803         int bscat2 = block_state_category(state2);
 804         int cycles1 = BLOCK_INFO_TRIMS(block1);
 805         int cycles2 = BLOCK_INFO_TRIMS(block2);
 806
 807         if (bscat1 < bscat2)
 808                 return -1;
 809         if (bscat1 > bscat2)
 810                 return 1;
 811
 812         if (cycles1 < cycles2)
 813                 return -1;
 814         if (cycles1 > cycles2)
 815                 return 1;
 816
 817         if (state1 < state2)
 818                 return -1;
 819         if (state1 > state2)
 820                 return 1;
 821
 822         assert(block1 == block2);
 823         return 0;
 824 }
 825
 826 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 827                                   fio_fp64_t *plist, unsigned int **percentiles,
 828                                   unsigned int *types)
 829 {
 830         int len = 0;
 831         int i, nr_uninit;
 832
 833         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 834
 835         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 836                 len++;
 837
 838         if (!len)
 839                 return 0;
 840
 841         /*
 842          * Sort the percentile list. Note that it may already be sorted if
 843          * we are using the default values, but since it's a short list this
 844          * isn't a worry. Also note that this does not work for NaN values.
 845          */
 846         if (len > 1)
 847                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 848
 849         /* Start only after the uninit entries end */
 850         for (nr_uninit = 0;
 851              nr_uninit < nr_block_infos
 852                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 853              nr_uninit ++)
 854                 ;
 855
 856         if (nr_uninit == nr_block_infos)
 857                 return 0;
 858
 859         *percentiles = calloc(len, sizeof(**percentiles));
 860
 861         for (i = 0; i < len; i++) {
 862                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 863                                 + nr_uninit;
 864                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 865         }
 866
 867         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 868         for (i = 0; i < nr_block_infos; i++)
 869                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 870
 871         return len;
 872 }
 873
 874 static const char *block_state_names[] = {
 875         [BLOCK_STATE_UNINIT] = "unwritten",
 876         [BLOCK_STATE_TRIMMED] = "trimmed",
 877         [BLOCK_STATE_WRITTEN] = "written",
 878         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 879         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 880 };
 881
 882 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 883                              fio_fp64_t *plist, struct buf_output *out)
 884 {
 885         int len, pos, i;
 886         unsigned int *percentiles = NULL;
 887         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 888
 889         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 890                                      &percentiles, block_state_counts);
 891
 892         log_buf(out, "  block lifetime percentiles :\n   |");
 893         pos = 0;
 894         for (i = 0; i < len; i++) {
 895                 uint32_t block_info = percentiles[i];
 896 #define LINE_LENGTH     75
 897                 char str[LINE_LENGTH];
 898                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 899                                      plist[i].u.f, block_info,
 900                                      i == len - 1 ? '\n' : ',');
 901                 assert(strln < LINE_LENGTH);
 902                 if (pos + strln > LINE_LENGTH) {
 903                         pos = 0;
 904                         log_buf(out, "\n   |");
 905                 }
 906                 log_buf(out, "%s", str);
 907                 pos += strln;
 908 #undef LINE_LENGTH
 909         }
 910         if (percentiles)
 911                 free(percentiles);
 912
 913         log_buf(out, "        states               :");
 914         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 915                 log_buf(out, " %s=%u%c",
 916                          block_state_names[i], block_state_counts[i],
 917                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 918 }
 919
 920 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 921 {
 922         char *p1, *p1alt, *p2;
 923         unsigned long long bw_mean, iops_mean;
 924         const int i2p = is_power_of_2(ts->kb_base);
 925
 926         if (!ts->ss_dur)
 927                 return;
 928
 929         bw_mean = steadystate_bw_mean(ts);
 930         iops_mean = steadystate_iops_mean(ts);
 931
 932         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
 933         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
 934         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
 935
 936         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 937                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
 938                 p1, p1alt, p2,
 939                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
 940                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
 941                 ts->ss_criterion.u.f,
 942                 ts->ss_state & FIO_SS_PCT ? "%" : "");
 943
 944         free(p1);
 945         free(p1alt);
 946         free(p2);
 947 }
 948
 949 static void show_agg_stats(struct disk_util_agg *agg, int terse,
 950                            struct buf_output *out)
 951 {
 952         if (!agg->slavecount)
 953                 return;
 954
 955         if (!terse) {
 956                 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
 957                          "aggrticks=%llu/%llu, aggrin_queue=%llu, "
 958                          "aggrutil=%3.2f%%",
 959                         (unsigned long long) agg->ios[0] / agg->slavecount,
 960                         (unsigned long long) agg->ios[1] / agg->slavecount,
 961                         (unsigned long long) agg->merges[0] / agg->slavecount,
 962                         (unsigned long long) agg->merges[1] / agg->slavecount,
 963                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 964                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 965                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 966                         agg->max_util.u.f);
 967         } else {
 968                 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
 969                         (unsigned long long) agg->ios[0] / agg->slavecount,
 970                         (unsigned long long) agg->ios[1] / agg->slavecount,
 971                         (unsigned long long) agg->merges[0] / agg->slavecount,
 972                         (unsigned long long) agg->merges[1] / agg->slavecount,
 973                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 974                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 975                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 976                         agg->max_util.u.f);
 977         }
 978 }
 979
 980 static void aggregate_slaves_stats(struct disk_util *masterdu)
 981 {
 982         struct disk_util_agg *agg = &masterdu->agg;
 983         struct disk_util_stat *dus;
 984         struct flist_head *entry;
 985         struct disk_util *slavedu;
 986         double util;
 987
 988         flist_for_each(entry, &masterdu->slaves) {
 989                 slavedu = flist_entry(entry, struct disk_util, slavelist);
 990                 dus = &slavedu->dus;
 991                 agg->ios[0] += dus->s.ios[0];
 992                 agg->ios[1] += dus->s.ios[1];
 993                 agg->merges[0] += dus->s.merges[0];
 994                 agg->merges[1] += dus->s.merges[1];
 995                 agg->sectors[0] += dus->s.sectors[0];
 996                 agg->sectors[1] += dus->s.sectors[1];
 997                 agg->ticks[0] += dus->s.ticks[0];
 998                 agg->ticks[1] += dus->s.ticks[1];
 999                 agg->time_in_queue += dus->s.time_in_queue;
1000                 agg->slavecount++;
1001
1002                 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
1003                 /* System utilization is the utilization of the
1004                  * component with the highest utilization.
1005                  */
1006                 if (util > agg->max_util.u.f)
1007                         agg->max_util.u.f = util;
1008
1009         }
1010
1011         if (agg->max_util.u.f > 100.0)
1012                 agg->max_util.u.f = 100.0;
1013 }
1014
1015 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
1016                      int terse, struct buf_output *out)
1017 {
1018         double util = 0;
1019
1020         if (dus->s.msec)
1021                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1022         if (util > 100.0)
1023                 util = 100.0;
1024
1025         if (!terse) {
1026                 if (agg->slavecount)
1027                         log_buf(out, "  ");
1028
1029                 log_buf(out, "  %s: ios=%llu/%llu, merge=%llu/%llu, "
1030                          "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
1031                                 dus->name,
1032                                 (unsigned long long) dus->s.ios[0],
1033                                 (unsigned long long) dus->s.ios[1],
1034                                 (unsigned long long) dus->s.merges[0],
1035                                 (unsigned long long) dus->s.merges[1],
1036                                 (unsigned long long) dus->s.ticks[0],
1037                                 (unsigned long long) dus->s.ticks[1],
1038                                 (unsigned long long) dus->s.time_in_queue,
1039                                 util);
1040         } else {
1041                 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1042                                 dus->name,
1043                                 (unsigned long long) dus->s.ios[0],
1044                                 (unsigned long long) dus->s.ios[1],
1045                                 (unsigned long long) dus->s.merges[0],
1046                                 (unsigned long long) dus->s.merges[1],
1047                                 (unsigned long long) dus->s.ticks[0],
1048                                 (unsigned long long) dus->s.ticks[1],
1049                                 (unsigned long long) dus->s.time_in_queue,
1050                                 util);
1051         }
1052
1053         /*
1054          * If the device has slaves, aggregate the stats for
1055          * those slave devices also.
1056          */
1057         show_agg_stats(agg, terse, out);
1058
1059         if (!terse)
1060                 log_buf(out, "\n");
1061 }
1062
1063 void json_array_add_disk_util(struct disk_util_stat *dus,
1064                 struct disk_util_agg *agg, struct json_array *array)
1065 {
1066         struct json_object *obj;
1067         double util = 0;
1068
1069         if (dus->s.msec)
1070                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1071         if (util > 100.0)
1072                 util = 100.0;
1073
1074         obj = json_create_object();
1075         json_array_add_value_object(array, obj);
1076
1077         json_object_add_value_string(obj, "name", (const char *)dus->name);
1078         json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1079         json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1080         json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1081         json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1082         json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1083         json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1084         json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1085         json_object_add_value_float(obj, "util", util);
1086
1087         /*
1088          * If the device has slaves, aggregate the stats for
1089          * those slave devices also.
1090          */
1091         if (!agg->slavecount)
1092                 return;
1093         json_object_add_value_int(obj, "aggr_read_ios",
1094                                 agg->ios[0] / agg->slavecount);
1095         json_object_add_value_int(obj, "aggr_write_ios",
1096                                 agg->ios[1] / agg->slavecount);
1097         json_object_add_value_int(obj, "aggr_read_merges",
1098                                 agg->merges[0] / agg->slavecount);
1099         json_object_add_value_int(obj, "aggr_write_merge",
1100                                 agg->merges[1] / agg->slavecount);
1101         json_object_add_value_int(obj, "aggr_read_ticks",
1102                                 agg->ticks[0] / agg->slavecount);
1103         json_object_add_value_int(obj, "aggr_write_ticks",
1104                                 agg->ticks[1] / agg->slavecount);
1105         json_object_add_value_int(obj, "aggr_in_queue",
1106                                 agg->time_in_queue / agg->slavecount);
1107         json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1108 }
1109
1110 static void json_object_add_disk_utils(struct json_object *obj,
1111                                        struct flist_head *head)
1112 {
1113         struct json_array *array = json_create_array();
1114         struct flist_head *entry;
1115         struct disk_util *du;
1116
1117         json_object_add_value_array(obj, "disk_util", array);
1118
1119         flist_for_each(entry, head) {
1120                 du = flist_entry(entry, struct disk_util, list);
1121
1122                 aggregate_slaves_stats(du);
1123                 json_array_add_disk_util(&du->dus, &du->agg, array);
1124         }
1125 }
1126
1127 void show_disk_util(int terse, struct json_object *parent,
1128                     struct buf_output *out)
1129 {
1130         struct flist_head *entry;
1131         struct disk_util *du;
1132         bool do_json;
1133
1134         if (!is_running_backend())
1135                 return;
1136
1137         if (flist_empty(&disk_list))
1138                 return;
1139
1140         if ((output_format & FIO_OUTPUT_JSON) && parent)
1141                 do_json = true;
1142         else
1143                 do_json = false;
1144
1145         if (!terse && !do_json)
1146                 log_buf(out, "\nDisk stats (read/write):\n");
1147
1148         if (do_json) {
1149                 json_object_add_disk_utils(parent, &disk_list);
1150         } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1151                 flist_for_each(entry, &disk_list) {
1152                         du = flist_entry(entry, struct disk_util, list);
1153
1154                         aggregate_slaves_stats(du);
1155                         print_disk_util(&du->dus, &du->agg, terse, out);
1156                 }
1157         }
1158 }
1159
1160 static void show_thread_status_normal(struct thread_stat *ts,
1161                                       struct group_run_stats *rs,
1162                                       struct buf_output *out)
1163 {
1164         double usr_cpu, sys_cpu;
1165         unsigned long runtime;
1166         double io_u_dist[FIO_IO_U_MAP_NR];
1167         time_t time_p;
1168         char time_buf[32];
1169
1170         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1171                 return;
1172
1173         memset(time_buf, 0, sizeof(time_buf));
1174
1175         time(&time_p);
1176         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1177
1178         if (!ts->error) {
1179                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1180                                         ts->name, ts->groupid, ts->members,
1181                                         ts->error, (int) ts->pid, time_buf);
1182         } else {
1183                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1184                                         ts->name, ts->groupid, ts->members,
1185                                         ts->error, ts->verror, (int) ts->pid,
1186                                         time_buf);
1187         }
1188
1189         if (strlen(ts->description))
1190                 log_buf(out, "  Description  : [%s]\n", ts->description);
1191
1192         for_each_rw_ddir(ddir) {
1193                 if (ts->io_bytes[ddir])
1194                         show_ddir_status(rs, ts, ddir, out);
1195         }
1196
1197         if (ts->unified_rw_rep == UNIFIED_BOTH)
1198                 show_mixed_ddir_status(rs, ts, out);
1199
1200         show_latencies(ts, out);
1201
1202         if (ts->sync_stat.samples)
1203                 show_ddir_status(rs, ts, DDIR_SYNC, out);
1204
1205         runtime = ts->total_run_time;
1206         if (runtime) {
1207                 double runt = (double) runtime;
1208
1209                 usr_cpu = (double) ts->usr_time * 100 / runt;
1210                 sys_cpu = (double) ts->sys_time * 100 / runt;
1211         } else {
1212                 usr_cpu = 0;
1213                 sys_cpu = 0;
1214         }
1215
1216         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1217                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1218                         (unsigned long long) ts->ctx,
1219                         (unsigned long long) ts->majf,
1220                         (unsigned long long) ts->minf);
1221
1222         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1223         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1224                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1225                                         io_u_dist[1], io_u_dist[2],
1226                                         io_u_dist[3], io_u_dist[4],
1227                                         io_u_dist[5], io_u_dist[6]);
1228
1229         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1230         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1231                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1232                                         io_u_dist[1], io_u_dist[2],
1233                                         io_u_dist[3], io_u_dist[4],
1234                                         io_u_dist[5], io_u_dist[6]);
1235         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1236         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1237                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1238                                         io_u_dist[1], io_u_dist[2],
1239                                         io_u_dist[3], io_u_dist[4],
1240                                         io_u_dist[5], io_u_dist[6]);
1241         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
1242                                  " short=%llu,%llu,%llu,0"
1243                                  " dropped=%llu,%llu,%llu,0\n",
1244                                         (unsigned long long) ts->total_io_u[0],
1245                                         (unsigned long long) ts->total_io_u[1],
1246                                         (unsigned long long) ts->total_io_u[2],
1247                                         (unsigned long long) ts->total_io_u[3],
1248                                         (unsigned long long) ts->short_io_u[0],
1249                                         (unsigned long long) ts->short_io_u[1],
1250                                         (unsigned long long) ts->short_io_u[2],
1251                                         (unsigned long long) ts->drop_io_u[0],
1252                                         (unsigned long long) ts->drop_io_u[1],
1253                                         (unsigned long long) ts->drop_io_u[2]);
1254         if (ts->continue_on_error) {
1255                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
1256                                         (unsigned long long)ts->total_err_count,
1257                                         ts->first_error,
1258                                         strerror(ts->first_error));
1259         }
1260         if (ts->latency_depth) {
1261                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1262                                         (unsigned long long)ts->latency_target,
1263                                         (unsigned long long)ts->latency_window,
1264                                         ts->latency_percentile.u.f,
1265                                         ts->latency_depth);
1266         }
1267
1268         if (ts->nr_block_infos)
1269                 show_block_infos(ts->nr_block_infos, ts->block_infos,
1270                                   ts->percentile_list, out);
1271
1272         if (ts->ss_dur)
1273                 show_ss_normal(ts, out);
1274 }
1275
1276 static void show_ddir_status_terse(struct thread_stat *ts,
1277                                    struct group_run_stats *rs,
1278                                    enum fio_ddir ddir, int ver,
1279                                    struct buf_output *out)
1280 {
1281         unsigned long long min, max, minv, maxv, bw, iops;
1282         unsigned long long *ovals = NULL;
1283         double mean, dev;
1284         unsigned int len;
1285         int i, bw_stat;
1286
1287         assert(ddir_rw(ddir));
1288
1289         iops = bw = 0;
1290         if (ts->runtime[ddir]) {
1291                 uint64_t runt = ts->runtime[ddir];
1292
1293                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1294                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1295         }
1296
1297         log_buf(out, ";%llu;%llu;%llu;%llu",
1298                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1299                                         (unsigned long long) ts->runtime[ddir]);
1300
1301         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1302                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1303         else
1304                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1305
1306         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1307                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1308         else
1309                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1310
1311         if (ts->lat_percentiles) {
1312                 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1313                                         ts->lat_stat[ddir].samples,
1314                                         ts->percentile_list, &ovals, &maxv,
1315                                         &minv);
1316         } else if (ts->clat_percentiles) {
1317                 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1318                                         ts->clat_stat[ddir].samples,
1319                                         ts->percentile_list, &ovals, &maxv,
1320                                         &minv);
1321         } else {
1322                 len = 0;
1323         }
1324
1325         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1326                 if (i >= len) {
1327                         log_buf(out, ";0%%=0");
1328                         continue;
1329                 }
1330                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1331         }
1332
1333         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1334                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1335         else
1336                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1337
1338         free(ovals);
1339
1340         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1341         if (bw_stat) {
1342                 double p_of_agg = 100.0;
1343
1344                 if (rs->agg[ddir]) {
1345                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1346                         if (p_of_agg > 100.0)
1347                                 p_of_agg = 100.0;
1348                 }
1349
1350                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1351         } else {
1352                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1353         }
1354
1355         if (ver == 5) {
1356                 if (bw_stat)
1357                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1358                 else
1359                         log_buf(out, ";%lu", 0UL);
1360
1361                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1362                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1363                                 mean, dev, (&ts->iops_stat[ddir])->samples);
1364                 else
1365                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1366         }
1367 }
1368
1369 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1370                                    struct group_run_stats *rs,
1371                                    int ver, struct buf_output *out)
1372 {
1373         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1374
1375         if (ts_lcl)
1376                 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1377
1378         free_clat_prio_stats(ts_lcl);
1379         free(ts_lcl);
1380 }
1381
1382 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1383                                              uint32_t percentiles,
1384                                              struct io_stat *lat_stat,
1385                                              uint64_t *io_u_plat)
1386 {
1387         char buf[120];
1388         double mean, dev;
1389         unsigned int i, len;
1390         struct json_object *lat_object, *percentile_object, *clat_bins_object;
1391         unsigned long long min, max, maxv, minv, *ovals = NULL;
1392
1393         if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1394                 min = max = 0;
1395                 mean = dev = 0.0;
1396         }
1397         lat_object = json_create_object();
1398         json_object_add_value_int(lat_object, "min", min);
1399         json_object_add_value_int(lat_object, "max", max);
1400         json_object_add_value_float(lat_object, "mean", mean);
1401         json_object_add_value_float(lat_object, "stddev", dev);
1402         json_object_add_value_int(lat_object, "N", lat_stat->samples);
1403
1404         if (percentiles && lat_stat->samples) {
1405                 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1406                                 ts->percentile_list, &ovals, &maxv, &minv);
1407
1408                 if (len > FIO_IO_U_LIST_MAX_LEN)
1409                         len = FIO_IO_U_LIST_MAX_LEN;
1410
1411                 percentile_object = json_create_object();
1412                 json_object_add_value_object(lat_object, "percentile", percentile_object);
1413                 for (i = 0; i < len; i++) {
1414                         snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1415                         json_object_add_value_int(percentile_object, buf, ovals[i]);
1416                 }
1417                 free(ovals);
1418
1419                 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1420                         clat_bins_object = json_create_object();
1421                         json_object_add_value_object(lat_object, "bins", clat_bins_object);
1422
1423                         for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1424                                 if (io_u_plat[i]) {
1425                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1426                                         json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1427                                 }
1428                 }
1429         }
1430
1431         return lat_object;
1432 }
1433
1434 static void add_ddir_status_json(struct thread_stat *ts,
1435                                  struct group_run_stats *rs, enum fio_ddir ddir,
1436                                  struct json_object *parent)
1437 {
1438         unsigned long long min, max;
1439         unsigned long long bw_bytes, bw;
1440         double mean, dev, iops;
1441         struct json_object *dir_object, *tmp_object;
1442         double p_of_agg = 100.0;
1443
1444         assert(ddir_rw(ddir) || ddir_sync(ddir));
1445
1446         if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1447                 return;
1448
1449         dir_object = json_create_object();
1450         json_object_add_value_object(parent,
1451                 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1452
1453         if (ddir_rw(ddir)) {
1454                 bw_bytes = 0;
1455                 bw = 0;
1456                 iops = 0.0;
1457                 if (ts->runtime[ddir]) {
1458                         uint64_t runt = ts->runtime[ddir];
1459
1460                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1461                         bw = bw_bytes / 1024; /* KiB/s */
1462                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1463                 }
1464
1465                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1466                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1467                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1468                 json_object_add_value_int(dir_object, "bw", bw);
1469                 json_object_add_value_float(dir_object, "iops", iops);
1470                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1471                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1472                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1473                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1474
1475                 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1476                                 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1477                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1478
1479                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1480                                 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1481                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1482
1483                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1484                                 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1485                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1486         } else {
1487                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1488                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1489                                 &ts->sync_stat, ts->io_u_sync_plat);
1490                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1491         }
1492
1493         if (!ddir_rw(ddir))
1494                 return;
1495
1496         /* Only include per prio stats if there are >= 2 prios with samples */
1497         if (get_nr_prios_with_samples(ts, ddir) >= 2) {
1498                 struct json_array *array = json_create_array();
1499                 const char *obj_name;
1500                 int i;
1501
1502                 if (ts->lat_percentiles)
1503                         obj_name = "lat_ns";
1504                 else
1505                         obj_name = "clat_ns";
1506
1507                 json_object_add_value_array(dir_object, "prios", array);
1508
1509                 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
1510                         if (ts->clat_prio[ddir][i].clat_stat.samples > 0) {
1511                                 struct json_object *obj = json_create_object();
1512                                 unsigned long long class, level;
1513
1514                                 class = ts->clat_prio[ddir][i].ioprio >> 13;
1515                                 json_object_add_value_int(obj, "prioclass", class);
1516                                 level = ts->clat_prio[ddir][i].ioprio & 7;
1517                                 json_object_add_value_int(obj, "prio", level);
1518
1519                                 tmp_object = add_ddir_lat_json(ts,
1520                                                                ts->clat_percentiles | ts->lat_percentiles,
1521                                                                &ts->clat_prio[ddir][i].clat_stat,
1522                                                                ts->clat_prio[ddir][i].io_u_plat);
1523                                 json_object_add_value_object(obj, obj_name, tmp_object);
1524                                 json_array_add_value_object(array, obj);
1525                         }
1526                 }
1527         }
1528
1529         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1530                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1531         } else {
1532                 min = max = 0;
1533                 p_of_agg = mean = dev = 0.0;
1534         }
1535
1536         json_object_add_value_int(dir_object, "bw_min", min);
1537         json_object_add_value_int(dir_object, "bw_max", max);
1538         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1539         json_object_add_value_float(dir_object, "bw_mean", mean);
1540         json_object_add_value_float(dir_object, "bw_dev", dev);
1541         json_object_add_value_int(dir_object, "bw_samples",
1542                                 (&ts->bw_stat[ddir])->samples);
1543
1544         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1545                 min = max = 0;
1546                 mean = dev = 0.0;
1547         }
1548         json_object_add_value_int(dir_object, "iops_min", min);
1549         json_object_add_value_int(dir_object, "iops_max", max);
1550         json_object_add_value_float(dir_object, "iops_mean", mean);
1551         json_object_add_value_float(dir_object, "iops_stddev", dev);
1552         json_object_add_value_int(dir_object, "iops_samples",
1553                                 (&ts->iops_stat[ddir])->samples);
1554
1555         if (ts->cachehit + ts->cachemiss) {
1556                 uint64_t total;
1557                 double hit;
1558
1559                 total = ts->cachehit + ts->cachemiss;
1560                 hit = (double) ts->cachehit / (double) total;
1561                 hit *= 100.0;
1562                 json_object_add_value_float(dir_object, "cachehit", hit);
1563         }
1564 }
1565
1566 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1567                 struct group_run_stats *rs, struct json_object *parent)
1568 {
1569         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1570
1571         /* add the aggregated stats to json parent */
1572         if (ts_lcl)
1573                 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1574
1575         free_clat_prio_stats(ts_lcl);
1576         free(ts_lcl);
1577 }
1578
1579 static void show_thread_status_terse_all(struct thread_stat *ts,
1580                                          struct group_run_stats *rs, int ver,
1581                                          struct buf_output *out)
1582 {
1583         double io_u_dist[FIO_IO_U_MAP_NR];
1584         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1585         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1586         double usr_cpu, sys_cpu;
1587         int i;
1588
1589         /* General Info */
1590         if (ver == 2)
1591                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1592         else
1593                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1594                         ts->name, ts->groupid, ts->error);
1595
1596         /* Log Read Status, or mixed if unified_rw_rep = 1 */
1597         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1598         if (ts->unified_rw_rep != UNIFIED_MIXED) {
1599                 /* Log Write Status */
1600                 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1601                 /* Log Trim Status */
1602                 if (ver == 2 || ver == 4 || ver == 5)
1603                         show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1604         }
1605         if (ts->unified_rw_rep == UNIFIED_BOTH)
1606                 show_mixed_ddir_status_terse(ts, rs, ver, out);
1607         /* CPU Usage */
1608         if (ts->total_run_time) {
1609                 double runt = (double) ts->total_run_time;
1610
1611                 usr_cpu = (double) ts->usr_time * 100 / runt;
1612                 sys_cpu = (double) ts->sys_time * 100 / runt;
1613         } else {
1614                 usr_cpu = 0;
1615                 sys_cpu = 0;
1616         }
1617
1618         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1619                                                 (unsigned long long) ts->ctx,
1620                                                 (unsigned long long) ts->majf,
1621                                                 (unsigned long long) ts->minf);
1622
1623         /* Calc % distribution of IO depths, usecond, msecond latency */
1624         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1625         stat_calc_lat_nu(ts, io_u_lat_u);
1626         stat_calc_lat_m(ts, io_u_lat_m);
1627
1628         /* Only show fixed 7 I/O depth levels*/
1629         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1630                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1631                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1632
1633         /* Microsecond latency */
1634         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1635                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1636         /* Millisecond latency */
1637         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1638                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1639
1640         /* disk util stats, if any */
1641         if (ver >= 3 && is_running_backend())
1642                 show_disk_util(1, NULL, out);
1643
1644         /* Additional output if continue_on_error set - default off*/
1645         if (ts->continue_on_error)
1646                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1647
1648         /* Additional output if description is set */
1649         if (strlen(ts->description)) {
1650                 if (ver == 2)
1651                         log_buf(out, "\n");
1652                 log_buf(out, ";%s", ts->description);
1653         }
1654
1655         log_buf(out, "\n");
1656 }
1657
1658 static void json_add_job_opts(struct json_object *root, const char *name,
1659                               struct flist_head *opt_list)
1660 {
1661         struct json_object *dir_object;
1662         struct flist_head *entry;
1663         struct print_option *p;
1664
1665         if (flist_empty(opt_list))
1666                 return;
1667
1668         dir_object = json_create_object();
1669         json_object_add_value_object(root, name, dir_object);
1670
1671         flist_for_each(entry, opt_list) {
1672                 p = flist_entry(entry, struct print_option, list);
1673                 json_object_add_value_string(dir_object, p->name, p->value);
1674         }
1675 }
1676
1677 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1678                                                    struct group_run_stats *rs,
1679                                                    struct flist_head *opt_list)
1680 {
1681         struct json_object *root, *tmp;
1682         struct jobs_eta *je;
1683         double io_u_dist[FIO_IO_U_MAP_NR];
1684         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1685         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1686         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1687         double usr_cpu, sys_cpu;
1688         int i;
1689         size_t size;
1690
1691         root = json_create_object();
1692         json_object_add_value_string(root, "jobname", ts->name);
1693         json_object_add_value_int(root, "groupid", ts->groupid);
1694         json_object_add_value_int(root, "error", ts->error);
1695
1696         /* ETA Info */
1697         je = get_jobs_eta(true, &size);
1698         if (je) {
1699                 json_object_add_value_int(root, "eta", je->eta_sec);
1700                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1701         }
1702
1703         if (opt_list)
1704                 json_add_job_opts(root, "job options", opt_list);
1705
1706         add_ddir_status_json(ts, rs, DDIR_READ, root);
1707         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1708         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1709         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1710
1711         if (ts->unified_rw_rep == UNIFIED_BOTH)
1712                 add_mixed_ddir_status_json(ts, rs, root);
1713
1714         /* CPU Usage */
1715         if (ts->total_run_time) {
1716                 double runt = (double) ts->total_run_time;
1717
1718                 usr_cpu = (double) ts->usr_time * 100 / runt;
1719                 sys_cpu = (double) ts->sys_time * 100 / runt;
1720         } else {
1721                 usr_cpu = 0;
1722                 sys_cpu = 0;
1723         }
1724         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1725         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1726         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1727         json_object_add_value_int(root, "ctx", ts->ctx);
1728         json_object_add_value_int(root, "majf", ts->majf);
1729         json_object_add_value_int(root, "minf", ts->minf);
1730
1731         /* Calc % distribution of IO depths */
1732         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1733         tmp = json_create_object();
1734         json_object_add_value_object(root, "iodepth_level", tmp);
1735         /* Only show fixed 7 I/O depth levels*/
1736         for (i = 0; i < 7; i++) {
1737                 char name[20];
1738                 if (i < 6)
1739                         snprintf(name, 20, "%d", 1 << i);
1740                 else
1741                         snprintf(name, 20, ">=%d", 1 << i);
1742                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1743         }
1744
1745         /* Calc % distribution of submit IO depths */
1746         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1747         tmp = json_create_object();
1748         json_object_add_value_object(root, "iodepth_submit", tmp);
1749         /* Only show fixed 7 I/O depth levels*/
1750         for (i = 0; i < 7; i++) {
1751                 char name[20];
1752                 if (i == 0)
1753                         snprintf(name, 20, "0");
1754                 else if (i < 6)
1755                         snprintf(name, 20, "%d", 1 << (i+1));
1756                 else
1757                         snprintf(name, 20, ">=%d", 1 << i);
1758                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1759         }
1760
1761         /* Calc % distribution of completion IO depths */
1762         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1763         tmp = json_create_object();
1764         json_object_add_value_object(root, "iodepth_complete", tmp);
1765         /* Only show fixed 7 I/O depth levels*/
1766         for (i = 0; i < 7; i++) {
1767                 char name[20];
1768                 if (i == 0)
1769                         snprintf(name, 20, "0");
1770                 else if (i < 6)
1771                         snprintf(name, 20, "%d", 1 << (i+1));
1772                 else
1773                         snprintf(name, 20, ">=%d", 1 << i);
1774                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1775         }
1776
1777         /* Calc % distribution of nsecond, usecond, msecond latency */
1778         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1779         stat_calc_lat_n(ts, io_u_lat_n);
1780         stat_calc_lat_u(ts, io_u_lat_u);
1781         stat_calc_lat_m(ts, io_u_lat_m);
1782
1783         /* Nanosecond latency */
1784         tmp = json_create_object();
1785         json_object_add_value_object(root, "latency_ns", tmp);
1786         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1787                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1788                                  "250", "500", "750", "1000", };
1789                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1790         }
1791         /* Microsecond latency */
1792         tmp = json_create_object();
1793         json_object_add_value_object(root, "latency_us", tmp);
1794         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1795                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1796                                  "250", "500", "750", "1000", };
1797                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1798         }
1799         /* Millisecond latency */
1800         tmp = json_create_object();
1801         json_object_add_value_object(root, "latency_ms", tmp);
1802         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1803                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1804                                  "250", "500", "750", "1000", "2000",
1805                                  ">=2000", };
1806                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1807         }
1808
1809         /* Additional output if continue_on_error set - default off*/
1810         if (ts->continue_on_error) {
1811                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1812                 json_object_add_value_int(root, "first_error", ts->first_error);
1813         }
1814
1815         if (ts->latency_depth) {
1816                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1817                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1818                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1819                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1820         }
1821
1822         /* Additional output if description is set */
1823         if (strlen(ts->description))
1824                 json_object_add_value_string(root, "desc", ts->description);
1825
1826         if (ts->nr_block_infos) {
1827                 /* Block error histogram and types */
1828                 int len;
1829                 unsigned int *percentiles = NULL;
1830                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1831
1832                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1833                                              ts->percentile_list,
1834                                              &percentiles, block_state_counts);
1835
1836                 if (len) {
1837                         struct json_object *block, *percentile_object, *states;
1838                         int state;
1839                         block = json_create_object();
1840                         json_object_add_value_object(root, "block", block);
1841
1842                         percentile_object = json_create_object();
1843                         json_object_add_value_object(block, "percentiles",
1844                                                      percentile_object);
1845                         for (i = 0; i < len; i++) {
1846                                 char buf[20];
1847                                 snprintf(buf, sizeof(buf), "%f",
1848                                          ts->percentile_list[i].u.f);
1849                                 json_object_add_value_int(percentile_object,
1850                                                           buf,
1851                                                           percentiles[i]);
1852                         }
1853
1854                         states = json_create_object();
1855                         json_object_add_value_object(block, "states", states);
1856                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1857                                 json_object_add_value_int(states,
1858                                         block_state_names[state],
1859                                         block_state_counts[state]);
1860                         }
1861                         free(percentiles);
1862                 }
1863         }
1864
1865         if (ts->ss_dur) {
1866                 struct json_object *data;
1867                 struct json_array *iops, *bw;
1868                 int j, k, l;
1869                 char ss_buf[64];
1870
1871                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1872                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1873                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1874                         (float) ts->ss_limit.u.f,
1875                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1876
1877                 tmp = json_create_object();
1878                 json_object_add_value_object(root, "steadystate", tmp);
1879                 json_object_add_value_string(tmp, "ss", ss_buf);
1880                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1881                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1882
1883                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1884                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1885                 json_object_add_value_string(tmp, "criterion", ss_buf);
1886                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1887                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1888
1889                 data = json_create_object();
1890                 json_object_add_value_object(tmp, "data", data);
1891                 bw = json_create_array();
1892                 iops = json_create_array();
1893
1894                 /*
1895                 ** if ss was attained or the buffer is not full,
1896                 ** ss->head points to the first element in the list.
1897                 ** otherwise it actually points to the second element
1898                 ** in the list
1899                 */
1900                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1901                         j = ts->ss_head;
1902                 else
1903                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1904                 for (l = 0; l < ts->ss_dur; l++) {
1905                         k = (j + l) % ts->ss_dur;
1906                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1907                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1908                 }
1909                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1910                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1911                 json_object_add_value_array(data, "iops", iops);
1912                 json_object_add_value_array(data, "bw", bw);
1913         }
1914
1915         return root;
1916 }
1917
1918 static void show_thread_status_terse(struct thread_stat *ts,
1919                                      struct group_run_stats *rs,
1920                                      struct buf_output *out)
1921 {
1922         if (terse_version >= 2 && terse_version <= 5)
1923                 show_thread_status_terse_all(ts, rs, terse_version, out);
1924         else
1925                 log_err("fio: bad terse version!? %d\n", terse_version);
1926 }
1927
1928 struct json_object *show_thread_status(struct thread_stat *ts,
1929                                        struct group_run_stats *rs,
1930                                        struct flist_head *opt_list,
1931                                        struct buf_output *out)
1932 {
1933         struct json_object *ret = NULL;
1934
1935         if (output_format & FIO_OUTPUT_TERSE)
1936                 show_thread_status_terse(ts, rs,  out);
1937         if (output_format & FIO_OUTPUT_JSON)
1938                 ret = show_thread_status_json(ts, rs, opt_list);
1939         if (output_format & FIO_OUTPUT_NORMAL)
1940                 show_thread_status_normal(ts, rs,  out);
1941
1942         return ret;
1943 }
1944
1945 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1946 {
1947         double mean, S;
1948
1949         dst->min_val = min(dst->min_val, src->min_val);
1950         dst->max_val = max(dst->max_val, src->max_val);
1951
1952         /*
1953          * Compute new mean and S after the merge
1954          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1955          *  #Parallel_algorithm>
1956          */
1957         if (first) {
1958                 mean = src->mean.u.f;
1959                 S = src->S.u.f;
1960         } else {
1961                 double delta = src->mean.u.f - dst->mean.u.f;
1962
1963                 mean = ((src->mean.u.f * src->samples) +
1964                         (dst->mean.u.f * dst->samples)) /
1965                         (dst->samples + src->samples);
1966
1967                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1968                         (dst->samples * src->samples) /
1969                         (dst->samples + src->samples);
1970         }
1971
1972         dst->samples += src->samples;
1973         dst->mean.u.f = mean;
1974         dst->S.u.f = S;
1975
1976 }
1977
1978 /*
1979  * We sum two kinds of stats - one that is time based, in which case we
1980  * apply the proper summing technique, and then one that is iops/bw
1981  * numbers. For group_reporting, we should just add those up, not make
1982  * them the mean of everything.
1983  */
1984 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
1985 {
1986         bool first = dst->samples == 0;
1987
1988         if (src->samples == 0)
1989                 return;
1990
1991         if (!pure_sum) {
1992                 __sum_stat(dst, src, first);
1993                 return;
1994         }
1995
1996         if (first) {
1997                 dst->min_val = src->min_val;
1998                 dst->max_val = src->max_val;
1999                 dst->samples = src->samples;
2000                 dst->mean.u.f = src->mean.u.f;
2001                 dst->S.u.f = src->S.u.f;
2002         } else {
2003                 dst->min_val += src->min_val;
2004                 dst->max_val += src->max_val;
2005                 dst->samples += src->samples;
2006                 dst->mean.u.f += src->mean.u.f;
2007                 dst->S.u.f += src->S.u.f;
2008         }
2009 }
2010
2011 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
2012 {
2013         int i;
2014
2015         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2016                 if (dst->max_run[i] < src->max_run[i])
2017                         dst->max_run[i] = src->max_run[i];
2018                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
2019                         dst->min_run[i] = src->min_run[i];
2020                 if (dst->max_bw[i] < src->max_bw[i])
2021                         dst->max_bw[i] = src->max_bw[i];
2022                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
2023                         dst->min_bw[i] = src->min_bw[i];
2024
2025                 dst->iobytes[i] += src->iobytes[i];
2026                 dst->agg[i] += src->agg[i];
2027         }
2028
2029         if (!dst->kb_base)
2030                 dst->kb_base = src->kb_base;
2031         if (!dst->unit_base)
2032                 dst->unit_base = src->unit_base;
2033         if (!dst->sig_figs)
2034                 dst->sig_figs = src->sig_figs;
2035 }
2036
2037 /*
2038  * Free the clat_prio_stat arrays allocated by alloc_clat_prio_stat_ddir().
2039  */
2040 void free_clat_prio_stats(struct thread_stat *ts)
2041 {
2042         enum fio_ddir ddir;
2043
2044         if (!ts)
2045                 return;
2046
2047         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2048                 sfree(ts->clat_prio[ddir]);
2049                 ts->clat_prio[ddir] = NULL;
2050                 ts->nr_clat_prio[ddir] = 0;
2051         }
2052 }
2053
2054 /*
2055  * Allocate a clat_prio_stat array. The array has to be allocated/freed using
2056  * smalloc/sfree, so that it is accessible by the process/thread summing the
2057  * thread_stats.
2058  */
2059 int alloc_clat_prio_stat_ddir(struct thread_stat *ts, enum fio_ddir ddir,
2060                               int nr_prios)
2061 {
2062         struct clat_prio_stat *clat_prio;
2063         int i;
2064
2065         clat_prio = scalloc(nr_prios, sizeof(*ts->clat_prio[ddir]));
2066         if (!clat_prio) {
2067                 log_err("fio: failed to allocate ts clat data\n");
2068                 return 1;
2069         }
2070
2071         for (i = 0; i < nr_prios; i++)
2072                 clat_prio[i].clat_stat.min_val = ULONG_MAX;
2073
2074         ts->clat_prio[ddir] = clat_prio;
2075         ts->nr_clat_prio[ddir] = nr_prios;
2076
2077         return 0;
2078 }
2079
2080 static int grow_clat_prio_stat(struct thread_stat *dst, enum fio_ddir ddir)
2081 {
2082         int curr_len = dst->nr_clat_prio[ddir];
2083         void *new_arr;
2084
2085         new_arr = scalloc(curr_len + 1, sizeof(*dst->clat_prio[ddir]));
2086         if (!new_arr) {
2087                 log_err("fio: failed to grow clat prio array\n");
2088                 return 1;
2089         }
2090
2091         memcpy(new_arr, dst->clat_prio[ddir],
2092                curr_len * sizeof(*dst->clat_prio[ddir]));
2093         sfree(dst->clat_prio[ddir]);
2094
2095         dst->clat_prio[ddir] = new_arr;
2096         dst->clat_prio[ddir][curr_len].clat_stat.min_val = ULONG_MAX;
2097         dst->nr_clat_prio[ddir]++;
2098
2099         return 0;
2100 }
2101
2102 static int find_clat_prio_index(struct thread_stat *dst, enum fio_ddir ddir,
2103                                 uint32_t ioprio)
2104 {
2105         int i, nr_prios = dst->nr_clat_prio[ddir];
2106
2107         for (i = 0; i < nr_prios; i++) {
2108                 if (dst->clat_prio[ddir][i].ioprio == ioprio)
2109                         return i;
2110         }
2111
2112         return -1;
2113 }
2114
2115 static int alloc_or_get_clat_prio_index(struct thread_stat *dst,
2116                                         enum fio_ddir ddir, uint32_t ioprio,
2117                                         int *idx)
2118 {
2119         int index = find_clat_prio_index(dst, ddir, ioprio);
2120
2121         if (index == -1) {
2122                 index = dst->nr_clat_prio[ddir];
2123
2124                 if (grow_clat_prio_stat(dst, ddir))
2125                         return 1;
2126
2127                 dst->clat_prio[ddir][index].ioprio = ioprio;
2128         }
2129
2130         *idx = index;
2131
2132         return 0;
2133 }
2134
2135 static int clat_prio_stats_copy(struct thread_stat *dst, struct thread_stat *src,
2136                                 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2137 {
2138         size_t sz = sizeof(*src->clat_prio[src_ddir]) *
2139                 src->nr_clat_prio[src_ddir];
2140
2141         dst->clat_prio[dst_ddir] = smalloc(sz);
2142         if (!dst->clat_prio[dst_ddir]) {
2143                 log_err("fio: failed to alloc clat prio array\n");
2144                 return 1;
2145         }
2146
2147         memcpy(dst->clat_prio[dst_ddir], src->clat_prio[src_ddir], sz);
2148         dst->nr_clat_prio[dst_ddir] = src->nr_clat_prio[src_ddir];
2149
2150         return 0;
2151 }
2152
2153 static int clat_prio_stat_add_samples(struct thread_stat *dst,
2154                                       enum fio_ddir dst_ddir, uint32_t ioprio,
2155                                       struct io_stat *io_stat,
2156                                       uint64_t *io_u_plat)
2157 {
2158         int i, dst_index;
2159
2160         if (!io_stat->samples)
2161                 return 0;
2162
2163         if (alloc_or_get_clat_prio_index(dst, dst_ddir, ioprio, &dst_index))
2164                 return 1;
2165
2166         sum_stat(&dst->clat_prio[dst_ddir][dst_index].clat_stat, io_stat,
2167                  false);
2168
2169         for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
2170                 dst->clat_prio[dst_ddir][dst_index].io_u_plat[i] += io_u_plat[i];
2171
2172         return 0;
2173 }
2174
2175 static int sum_clat_prio_stats_src_single_prio(struct thread_stat *dst,
2176                                                struct thread_stat *src,
2177                                                enum fio_ddir dst_ddir,
2178                                                enum fio_ddir src_ddir)
2179 {
2180         struct io_stat *io_stat;
2181         uint64_t *io_u_plat;
2182
2183         /*
2184          * If src ts has no clat_prio_stat array, then all I/Os were submitted
2185          * using src->ioprio. Thus, the global samples in src->clat_stat (or
2186          * src->lat_stat) can be used as the 'per prio' samples for src->ioprio.
2187          */
2188         assert(!src->clat_prio[src_ddir]);
2189         assert(src->nr_clat_prio[src_ddir] == 0);
2190
2191         if (src->lat_percentiles) {
2192                 io_u_plat = src->io_u_plat[FIO_LAT][src_ddir];
2193                 io_stat = &src->lat_stat[src_ddir];
2194         } else {
2195                 io_u_plat = src->io_u_plat[FIO_CLAT][src_ddir];
2196                 io_stat = &src->clat_stat[src_ddir];
2197         }
2198
2199         return clat_prio_stat_add_samples(dst, dst_ddir, src->ioprio, io_stat,
2200                                           io_u_plat);
2201 }
2202
2203 static int sum_clat_prio_stats_src_multi_prio(struct thread_stat *dst,
2204                                               struct thread_stat *src,
2205                                               enum fio_ddir dst_ddir,
2206                                               enum fio_ddir src_ddir)
2207 {
2208         int i;
2209
2210         /*
2211          * If src ts has a clat_prio_stat array, then there are multiple prios
2212          * in use (i.e. src ts had cmdprio_percentage or cmdprio_bssplit set).
2213          * The samples for the default prio will exist in the src->clat_prio
2214          * array, just like the samples for any other prio.
2215          */
2216         assert(src->clat_prio[src_ddir]);
2217         assert(src->nr_clat_prio[src_ddir]);
2218
2219         /* If the dst ts doesn't yet have a clat_prio array, simply memcpy. */
2220         if (!dst->clat_prio[dst_ddir])
2221                 return clat_prio_stats_copy(dst, src, dst_ddir, src_ddir);
2222
2223         /* The dst ts already has a clat_prio_array, add src stats into it. */
2224         for (i = 0; i < src->nr_clat_prio[src_ddir]; i++) {
2225                 struct io_stat *io_stat = &src->clat_prio[src_ddir][i].clat_stat;
2226                 uint64_t *io_u_plat = src->clat_prio[src_ddir][i].io_u_plat;
2227                 uint32_t ioprio = src->clat_prio[src_ddir][i].ioprio;
2228
2229                 if (clat_prio_stat_add_samples(dst, dst_ddir, ioprio, io_stat, io_u_plat))
2230                         return 1;
2231         }
2232
2233         return 0;
2234 }
2235
2236 static int sum_clat_prio_stats(struct thread_stat *dst, struct thread_stat *src,
2237                                enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2238 {
2239         if (dst->disable_prio_stat)
2240                 return 0;
2241
2242         if (!src->clat_prio[src_ddir])
2243                 return sum_clat_prio_stats_src_single_prio(dst, src, dst_ddir,
2244                                                            src_ddir);
2245
2246         return sum_clat_prio_stats_src_multi_prio(dst, src, dst_ddir, src_ddir);
2247 }
2248
2249 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2250 {
2251         int k, l, m;
2252
2253         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2254                 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2255                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2256                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2257                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2258                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2259                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2260                         sum_clat_prio_stats(dst, src, l, l);
2261
2262                         dst->io_bytes[l] += src->io_bytes[l];
2263
2264                         if (dst->runtime[l] < src->runtime[l])
2265                                 dst->runtime[l] = src->runtime[l];
2266                 } else {
2267                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2268                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2269                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2270                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2271                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2272                         sum_clat_prio_stats(dst, src, 0, l);
2273
2274                         dst->io_bytes[0] += src->io_bytes[l];
2275
2276                         if (dst->runtime[0] < src->runtime[l])
2277                                 dst->runtime[0] = src->runtime[l];
2278                 }
2279         }
2280
2281         sum_stat(&dst->sync_stat, &src->sync_stat, false);
2282         dst->usr_time += src->usr_time;
2283         dst->sys_time += src->sys_time;
2284         dst->ctx += src->ctx;
2285         dst->majf += src->majf;
2286         dst->minf += src->minf;
2287
2288         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2289                 dst->io_u_map[k] += src->io_u_map[k];
2290                 dst->io_u_submit[k] += src->io_u_submit[k];
2291                 dst->io_u_complete[k] += src->io_u_complete[k];
2292         }
2293
2294         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2295                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2296         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2297                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2298         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2299                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2300
2301         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2302                 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2303                         dst->total_io_u[k] += src->total_io_u[k];
2304                         dst->short_io_u[k] += src->short_io_u[k];
2305                         dst->drop_io_u[k] += src->drop_io_u[k];
2306                 } else {
2307                         dst->total_io_u[0] += src->total_io_u[k];
2308                         dst->short_io_u[0] += src->short_io_u[k];
2309                         dst->drop_io_u[0] += src->drop_io_u[k];
2310                 }
2311         }
2312
2313         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2314
2315         for (k = 0; k < FIO_LAT_CNT; k++)
2316                 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2317                         for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2318                                 if (dst->unified_rw_rep != UNIFIED_MIXED)
2319                                         dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2320                                 else
2321                                         dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2322
2323         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2324                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2325
2326         dst->total_run_time += src->total_run_time;
2327         dst->total_submit += src->total_submit;
2328         dst->total_complete += src->total_complete;
2329         dst->nr_zone_resets += src->nr_zone_resets;
2330         dst->cachehit += src->cachehit;
2331         dst->cachemiss += src->cachemiss;
2332 }
2333
2334 void init_group_run_stat(struct group_run_stats *gs)
2335 {
2336         int i;
2337         memset(gs, 0, sizeof(*gs));
2338
2339         for (i = 0; i < DDIR_RWDIR_CNT; i++)
2340                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2341 }
2342
2343 void init_thread_stat_min_vals(struct thread_stat *ts)
2344 {
2345         int i;
2346
2347         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2348                 ts->clat_stat[i].min_val = ULONG_MAX;
2349                 ts->slat_stat[i].min_val = ULONG_MAX;
2350                 ts->lat_stat[i].min_val = ULONG_MAX;
2351                 ts->bw_stat[i].min_val = ULONG_MAX;
2352                 ts->iops_stat[i].min_val = ULONG_MAX;
2353         }
2354         ts->sync_stat.min_val = ULONG_MAX;
2355 }
2356
2357 void init_thread_stat(struct thread_stat *ts)
2358 {
2359         memset(ts, 0, sizeof(*ts));
2360
2361         init_thread_stat_min_vals(ts);
2362         ts->groupid = -1;
2363 }
2364
2365 static void init_per_prio_stats(struct thread_stat *threadstats, int nr_ts)
2366 {
2367         struct thread_data *td;
2368         struct thread_stat *ts;
2369         int i, j, last_ts, idx;
2370         enum fio_ddir ddir;
2371
2372         j = 0;
2373         last_ts = -1;
2374         idx = 0;
2375
2376         /*
2377          * Loop through all tds, if a td requires per prio stats, temporarily
2378          * store a 1 in ts->disable_prio_stat, and then do an additional
2379          * loop at the end where we invert the ts->disable_prio_stat values.
2380          */
2381         for_each_td(td, i) {
2382                 if (!td->o.stats)
2383                         continue;
2384                 if (idx &&
2385                     (!td->o.group_reporting ||
2386                      (td->o.group_reporting && last_ts != td->groupid))) {
2387                         idx = 0;
2388                         j++;
2389                 }
2390
2391                 last_ts = td->groupid;
2392                 ts = &threadstats[j];
2393
2394                 /* idx == 0 means first td in group, or td is not in a group. */
2395                 if (idx == 0)
2396                         ts->ioprio = td->ioprio;
2397                 else if (td->ioprio != ts->ioprio)
2398                         ts->disable_prio_stat = 1;
2399
2400                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2401                         if (td->ts.clat_prio[ddir]) {
2402                                 ts->disable_prio_stat = 1;
2403                                 break;
2404                         }
2405                 }
2406
2407                 idx++;
2408         }
2409
2410         /* Loop through all dst threadstats and fixup the values. */
2411         for (i = 0; i < nr_ts; i++) {
2412                 ts = &threadstats[i];
2413                 ts->disable_prio_stat = !ts->disable_prio_stat;
2414         }
2415 }
2416
2417 void __show_run_stats(void)
2418 {
2419         struct group_run_stats *runstats, *rs;
2420         struct thread_data *td;
2421         struct thread_stat *threadstats, *ts;
2422         int i, j, k, nr_ts, last_ts, idx;
2423         bool kb_base_warned = false;
2424         bool unit_base_warned = false;
2425         struct json_object *root = NULL;
2426         struct json_array *array = NULL;
2427         struct buf_output output[FIO_OUTPUT_NR];
2428         struct flist_head **opt_lists;
2429
2430         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2431
2432         for (i = 0; i < groupid + 1; i++)
2433                 init_group_run_stat(&runstats[i]);
2434
2435         /*
2436          * find out how many threads stats we need. if group reporting isn't
2437          * enabled, it's one-per-td.
2438          */
2439         nr_ts = 0;
2440         last_ts = -1;
2441         for_each_td(td, i) {
2442                 if (!td->o.group_reporting) {
2443                         nr_ts++;
2444                         continue;
2445                 }
2446                 if (last_ts == td->groupid)
2447                         continue;
2448                 if (!td->o.stats)
2449                         continue;
2450
2451                 last_ts = td->groupid;
2452                 nr_ts++;
2453         }
2454
2455         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2456         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2457
2458         for (i = 0; i < nr_ts; i++) {
2459                 init_thread_stat(&threadstats[i]);
2460                 opt_lists[i] = NULL;
2461         }
2462
2463         init_per_prio_stats(threadstats, nr_ts);
2464
2465         j = 0;
2466         last_ts = -1;
2467         idx = 0;
2468         for_each_td(td, i) {
2469                 if (!td->o.stats)
2470                         continue;
2471                 if (idx && (!td->o.group_reporting ||
2472                     (td->o.group_reporting && last_ts != td->groupid))) {
2473                         idx = 0;
2474                         j++;
2475                 }
2476
2477                 last_ts = td->groupid;
2478
2479                 ts = &threadstats[j];
2480
2481                 ts->clat_percentiles = td->o.clat_percentiles;
2482                 ts->lat_percentiles = td->o.lat_percentiles;
2483                 ts->slat_percentiles = td->o.slat_percentiles;
2484                 ts->percentile_precision = td->o.percentile_precision;
2485                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2486                 opt_lists[j] = &td->opt_list;
2487
2488                 idx++;
2489
2490                 if (ts->groupid == -1) {
2491                         /*
2492                          * These are per-group shared already
2493                          */
2494                         snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2495                         if (td->o.description)
2496                                 snprintf(ts->description,
2497                                          sizeof(ts->description), "%s",
2498                                          td->o.description);
2499                         else
2500                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2501
2502                         /*
2503                          * If multiple entries in this group, this is
2504                          * the first member.
2505                          */
2506                         ts->thread_number = td->thread_number;
2507                         ts->groupid = td->groupid;
2508
2509                         /*
2510                          * first pid in group, not very useful...
2511                          */
2512                         ts->pid = td->pid;
2513
2514                         ts->kb_base = td->o.kb_base;
2515                         ts->unit_base = td->o.unit_base;
2516                         ts->sig_figs = td->o.sig_figs;
2517                         ts->unified_rw_rep = td->o.unified_rw_rep;
2518                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2519                         log_info("fio: kb_base differs for jobs in group, using"
2520                                  " %u as the base\n", ts->kb_base);
2521                         kb_base_warned = true;
2522                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2523                         log_info("fio: unit_base differs for jobs in group, using"
2524                                  " %u as the base\n", ts->unit_base);
2525                         unit_base_warned = true;
2526                 }
2527
2528                 ts->continue_on_error = td->o.continue_on_error;
2529                 ts->total_err_count += td->total_err_count;
2530                 ts->first_error = td->first_error;
2531                 if (!ts->error) {
2532                         if (!td->error && td->o.continue_on_error &&
2533                             td->first_error) {
2534                                 ts->error = td->first_error;
2535                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2536                                          td->verror);
2537                         } else  if (td->error) {
2538                                 ts->error = td->error;
2539                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2540                                          td->verror);
2541                         }
2542                 }
2543
2544                 ts->latency_depth = td->latency_qd;
2545                 ts->latency_target = td->o.latency_target;
2546                 ts->latency_percentile = td->o.latency_percentile;
2547                 ts->latency_window = td->o.latency_window;
2548
2549                 ts->nr_block_infos = td->ts.nr_block_infos;
2550                 for (k = 0; k < ts->nr_block_infos; k++)
2551                         ts->block_infos[k] = td->ts.block_infos[k];
2552
2553                 sum_thread_stats(ts, &td->ts);
2554
2555                 ts->members++;
2556
2557                 if (td->o.ss_dur) {
2558                         ts->ss_state = td->ss.state;
2559                         ts->ss_dur = td->ss.dur;
2560                         ts->ss_head = td->ss.head;
2561                         ts->ss_bw_data = td->ss.bw_data;
2562                         ts->ss_iops_data = td->ss.iops_data;
2563                         ts->ss_limit.u.f = td->ss.limit;
2564                         ts->ss_slope.u.f = td->ss.slope;
2565                         ts->ss_deviation.u.f = td->ss.deviation;
2566                         ts->ss_criterion.u.f = td->ss.criterion;
2567                 }
2568                 else
2569                         ts->ss_dur = ts->ss_state = 0;
2570         }
2571
2572         for (i = 0; i < nr_ts; i++) {
2573                 unsigned long long bw;
2574
2575                 ts = &threadstats[i];
2576                 if (ts->groupid == -1)
2577                         continue;
2578                 rs = &runstats[ts->groupid];
2579                 rs->kb_base = ts->kb_base;
2580                 rs->unit_base = ts->unit_base;
2581                 rs->sig_figs = ts->sig_figs;
2582                 rs->unified_rw_rep |= ts->unified_rw_rep;
2583
2584                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2585                         if (!ts->runtime[j])
2586                                 continue;
2587                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2588                                 rs->min_run[j] = ts->runtime[j];
2589                         if (ts->runtime[j] > rs->max_run[j])
2590                                 rs->max_run[j] = ts->runtime[j];
2591
2592                         bw = 0;
2593                         if (ts->runtime[j])
2594                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2595                         if (bw < rs->min_bw[j])
2596                                 rs->min_bw[j] = bw;
2597                         if (bw > rs->max_bw[j])
2598                                 rs->max_bw[j] = bw;
2599
2600                         rs->iobytes[j] += ts->io_bytes[j];
2601                 }
2602         }
2603
2604         for (i = 0; i < groupid + 1; i++) {
2605                 enum fio_ddir ddir;
2606
2607                 rs = &runstats[i];
2608
2609                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2610                         if (rs->max_run[ddir])
2611                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2612                                                 rs->max_run[ddir];
2613                 }
2614         }
2615
2616         for (i = 0; i < FIO_OUTPUT_NR; i++)
2617                 buf_output_init(&output[i]);
2618
2619         /*
2620          * don't overwrite last signal output
2621          */
2622         if (output_format & FIO_OUTPUT_NORMAL)
2623                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2624         if (output_format & FIO_OUTPUT_JSON) {
2625                 struct thread_data *global;
2626                 char time_buf[32];
2627                 struct timeval now;
2628                 unsigned long long ms_since_epoch;
2629                 time_t tv_sec;
2630
2631                 gettimeofday(&now, NULL);
2632                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2633                                  (unsigned long long)(now.tv_usec) / 1000;
2634
2635                 tv_sec = now.tv_sec;
2636                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2637                 if (time_buf[strlen(time_buf) - 1] == '\n')
2638                         time_buf[strlen(time_buf) - 1] = '\0';
2639
2640                 root = json_create_object();
2641                 json_object_add_value_string(root, "fio version", fio_version_string);
2642                 json_object_add_value_int(root, "timestamp", now.tv_sec);
2643                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2644                 json_object_add_value_string(root, "time", time_buf);
2645                 global = get_global_options();
2646                 json_add_job_opts(root, "global options", &global->opt_list);
2647                 array = json_create_array();
2648                 json_object_add_value_array(root, "jobs", array);
2649         }
2650
2651         if (is_backend)
2652                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2653
2654         for (i = 0; i < nr_ts; i++) {
2655                 ts = &threadstats[i];
2656                 rs = &runstats[ts->groupid];
2657
2658                 if (is_backend) {
2659                         fio_server_send_job_options(opt_lists[i], i);
2660                         fio_server_send_ts(ts, rs);
2661                 } else {
2662                         if (output_format & FIO_OUTPUT_TERSE)
2663                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2664                         if (output_format & FIO_OUTPUT_JSON) {
2665                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2666                                 json_array_add_value_object(array, tmp);
2667                         }
2668                         if (output_format & FIO_OUTPUT_NORMAL)
2669                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2670                 }
2671         }
2672         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2673                 /* disk util stats, if any */
2674                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2675
2676                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2677
2678                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2679                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2680                 json_free_object(root);
2681         }
2682
2683         for (i = 0; i < groupid + 1; i++) {
2684                 rs = &runstats[i];
2685
2686                 rs->groupid = i;
2687                 if (is_backend)
2688                         fio_server_send_gs(rs);
2689                 else if (output_format & FIO_OUTPUT_NORMAL)
2690                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2691         }
2692
2693         if (is_backend)
2694                 fio_server_send_du();
2695         else if (output_format & FIO_OUTPUT_NORMAL) {
2696                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2697                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2698         }
2699
2700         for (i = 0; i < FIO_OUTPUT_NR; i++) {
2701                 struct buf_output *out = &output[i];
2702
2703                 log_info_buf(out->buf, out->buflen);
2704                 buf_output_free(out);
2705         }
2706
2707         fio_idle_prof_cleanup();
2708
2709         log_info_flush();
2710         free(runstats);
2711
2712         /* free arrays allocated by sum_thread_stats(), if any */
2713         for (i = 0; i < nr_ts; i++) {
2714                 ts = &threadstats[i];
2715                 free_clat_prio_stats(ts);
2716         }
2717         free(threadstats);
2718         free(opt_lists);
2719 }
2720
2721 int __show_running_run_stats(void)
2722 {
2723         struct thread_data *td;
2724         unsigned long long *rt;
2725         struct timespec ts;
2726         int i;
2727
2728         fio_sem_down(stat_sem);
2729
2730         rt = malloc(thread_number * sizeof(unsigned long long));
2731         fio_gettime(&ts, NULL);
2732
2733         for_each_td(td, i) {
2734                 if (td->runstate >= TD_EXITED)
2735                         continue;
2736
2737                 td->update_rusage = 1;
2738                 for_each_rw_ddir(ddir) {
2739                         td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2740                 }
2741                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2742
2743                 rt[i] = mtime_since(&td->start, &ts);
2744                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2745                         td->ts.runtime[DDIR_READ] += rt[i];
2746                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2747                         td->ts.runtime[DDIR_WRITE] += rt[i];
2748                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2749                         td->ts.runtime[DDIR_TRIM] += rt[i];
2750         }
2751
2752         for_each_td(td, i) {
2753                 if (td->runstate >= TD_EXITED)
2754                         continue;
2755                 if (td->rusage_sem) {
2756                         td->update_rusage = 1;
2757                         fio_sem_down(td->rusage_sem);
2758                 }
2759                 td->update_rusage = 0;
2760         }
2761
2762         __show_run_stats();
2763
2764         for_each_td(td, i) {
2765                 if (td->runstate >= TD_EXITED)
2766                         continue;
2767
2768                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2769                         td->ts.runtime[DDIR_READ] -= rt[i];
2770                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2771                         td->ts.runtime[DDIR_WRITE] -= rt[i];
2772                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2773                         td->ts.runtime[DDIR_TRIM] -= rt[i];
2774         }
2775
2776         free(rt);
2777         fio_sem_up(stat_sem);
2778
2779         return 0;
2780 }
2781
2782 static bool status_file_disabled;
2783
2784 #define FIO_STATUS_FILE         "fio-dump-status"
2785
2786 static int check_status_file(void)
2787 {
2788         struct stat sb;
2789         const char *temp_dir;
2790         char fio_status_file_path[PATH_MAX];
2791
2792         if (status_file_disabled)
2793                 return 0;
2794
2795         temp_dir = getenv("TMPDIR");
2796         if (temp_dir == NULL) {
2797                 temp_dir = getenv("TEMP");
2798                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2799                         temp_dir = NULL;
2800         }
2801         if (temp_dir == NULL)
2802                 temp_dir = "/tmp";
2803 #ifdef __COVERITY__
2804         __coverity_tainted_data_sanitize__(temp_dir);
2805 #endif
2806
2807         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2808
2809         if (stat(fio_status_file_path, &sb))
2810                 return 0;
2811
2812         if (unlink(fio_status_file_path) < 0) {
2813                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2814                                                         strerror(errno));
2815                 log_err("fio: disabling status file updates\n");
2816                 status_file_disabled = true;
2817         }
2818
2819         return 1;
2820 }
2821
2822 void check_for_running_stats(void)
2823 {
2824         if (check_status_file()) {
2825                 show_running_run_stats();
2826                 return;
2827         }
2828 }
2829
2830 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2831 {
2832         double val = data;
2833         double delta;
2834
2835         if (data > is->max_val)
2836                 is->max_val = data;
2837         if (data < is->min_val)
2838                 is->min_val = data;
2839
2840         delta = val - is->mean.u.f;
2841         if (delta) {
2842                 is->mean.u.f += delta / (is->samples + 1.0);
2843                 is->S.u.f += delta * (val - is->mean.u.f);
2844         }
2845
2846         is->samples++;
2847 }
2848
2849 static inline void add_stat_prio_sample(struct clat_prio_stat *clat_prio,
2850                                         unsigned short clat_prio_index,
2851                                         unsigned long long nsec)
2852 {
2853         if (clat_prio)
2854                 add_stat_sample(&clat_prio[clat_prio_index].clat_stat, nsec);
2855 }
2856
2857 /*
2858  * Return a struct io_logs, which is added to the tail of the log
2859  * list for 'iolog'.
2860  */
2861 static struct io_logs *get_new_log(struct io_log *iolog)
2862 {
2863         size_t new_samples;
2864         struct io_logs *cur_log;
2865
2866         /*
2867          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2868          * forever
2869          */
2870         if (!iolog->cur_log_max) {
2871                 new_samples = iolog->td->o.log_entries;
2872         } else {
2873                 new_samples = iolog->cur_log_max * 2;
2874                 if (new_samples > MAX_LOG_ENTRIES)
2875                         new_samples = MAX_LOG_ENTRIES;
2876         }
2877
2878         cur_log = smalloc(sizeof(*cur_log));
2879         if (cur_log) {
2880                 INIT_FLIST_HEAD(&cur_log->list);
2881                 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2882                 if (cur_log->log) {
2883                         cur_log->nr_samples = 0;
2884                         cur_log->max_samples = new_samples;
2885                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2886                         iolog->cur_log_max = new_samples;
2887                         return cur_log;
2888                 }
2889                 sfree(cur_log);
2890         }
2891
2892         return NULL;
2893 }
2894
2895 /*
2896  * Add and return a new log chunk, or return current log if big enough
2897  */
2898 static struct io_logs *regrow_log(struct io_log *iolog)
2899 {
2900         struct io_logs *cur_log;
2901         int i;
2902
2903         if (!iolog || iolog->disabled)
2904                 goto disable;
2905
2906         cur_log = iolog_cur_log(iolog);
2907         if (!cur_log) {
2908                 cur_log = get_new_log(iolog);
2909                 if (!cur_log)
2910                         return NULL;
2911         }
2912
2913         if (cur_log->nr_samples < cur_log->max_samples)
2914                 return cur_log;
2915
2916         /*
2917          * No room for a new sample. If we're compressing on the fly, flush
2918          * out the current chunk
2919          */
2920         if (iolog->log_gz) {
2921                 if (iolog_cur_flush(iolog, cur_log)) {
2922                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2923                         return NULL;
2924                 }
2925         }
2926
2927         /*
2928          * Get a new log array, and add to our list
2929          */
2930         cur_log = get_new_log(iolog);
2931         if (!cur_log) {
2932                 log_err("fio: failed extending iolog! Will stop logging.\n");
2933                 return NULL;
2934         }
2935
2936         if (!iolog->pending || !iolog->pending->nr_samples)
2937                 return cur_log;
2938
2939         /*
2940          * Flush pending items to new log
2941          */
2942         for (i = 0; i < iolog->pending->nr_samples; i++) {
2943                 struct io_sample *src, *dst;
2944
2945                 src = get_sample(iolog, iolog->pending, i);
2946                 dst = get_sample(iolog, cur_log, i);
2947                 memcpy(dst, src, log_entry_sz(iolog));
2948         }
2949         cur_log->nr_samples = iolog->pending->nr_samples;
2950
2951         iolog->pending->nr_samples = 0;
2952         return cur_log;
2953 disable:
2954         if (iolog)
2955                 iolog->disabled = true;
2956         return NULL;
2957 }
2958
2959 void regrow_logs(struct thread_data *td)
2960 {
2961         regrow_log(td->slat_log);
2962         regrow_log(td->clat_log);
2963         regrow_log(td->clat_hist_log);
2964         regrow_log(td->lat_log);
2965         regrow_log(td->bw_log);
2966         regrow_log(td->iops_log);
2967         td->flags &= ~TD_F_REGROW_LOGS;
2968 }
2969
2970 void regrow_agg_logs(void)
2971 {
2972         enum fio_ddir ddir;
2973
2974         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2975                 regrow_log(agg_io_log[ddir]);
2976 }
2977
2978 static struct io_logs *get_cur_log(struct io_log *iolog)
2979 {
2980         struct io_logs *cur_log;
2981
2982         cur_log = iolog_cur_log(iolog);
2983         if (!cur_log) {
2984                 cur_log = get_new_log(iolog);
2985                 if (!cur_log)
2986                         return NULL;
2987         }
2988
2989         if (cur_log->nr_samples < cur_log->max_samples)
2990                 return cur_log;
2991
2992         /*
2993          * Out of space. If we're in IO offload mode, or we're not doing
2994          * per unit logging (hence logging happens outside of the IO thread
2995          * as well), add a new log chunk inline. If we're doing inline
2996          * submissions, flag 'td' as needing a log regrow and we'll take
2997          * care of it on the submission side.
2998          */
2999         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
3000             !per_unit_log(iolog))
3001                 return regrow_log(iolog);
3002
3003         if (iolog->td)
3004                 iolog->td->flags |= TD_F_REGROW_LOGS;
3005         if (iolog->pending)
3006                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
3007         return iolog->pending;
3008 }
3009
3010 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
3011                              enum fio_ddir ddir, unsigned long long bs,
3012                              unsigned long t, uint64_t offset,
3013                              unsigned int priority)
3014 {
3015         struct io_logs *cur_log;
3016
3017         if (iolog->disabled)
3018                 return;
3019         if (flist_empty(&iolog->io_logs))
3020                 iolog->avg_last[ddir] = t;
3021
3022         cur_log = get_cur_log(iolog);
3023         if (cur_log) {
3024                 struct io_sample *s;
3025
3026                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
3027
3028                 s->data = data;
3029                 s->time = t + (iolog->td ? iolog->td->alternate_epoch : 0);
3030                 io_sample_set_ddir(iolog, s, ddir);
3031                 s->bs = bs;
3032                 s->priority = priority;
3033
3034                 if (iolog->log_offset) {
3035                         struct io_sample_offset *so = (void *) s;
3036
3037                         so->offset = offset;
3038                 }
3039
3040                 cur_log->nr_samples++;
3041                 return;
3042         }
3043
3044         iolog->disabled = true;
3045 }
3046
3047 static inline void reset_io_stat(struct io_stat *ios)
3048 {
3049         ios->min_val = -1ULL;
3050         ios->max_val = ios->samples = 0;
3051         ios->mean.u.f = ios->S.u.f = 0;
3052 }
3053
3054 static inline void reset_io_u_plat(uint64_t *io_u_plat)
3055 {
3056         int i;
3057
3058         for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
3059                 io_u_plat[i] = 0;
3060 }
3061
3062 static inline void reset_clat_prio_stats(struct thread_stat *ts)
3063 {
3064         enum fio_ddir ddir;
3065         int i;
3066
3067         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3068                 if (!ts->clat_prio[ddir])
3069                         continue;
3070
3071                 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
3072                         reset_io_stat(&ts->clat_prio[ddir][i].clat_stat);
3073                         reset_io_u_plat(ts->clat_prio[ddir][i].io_u_plat);
3074                 }
3075         }
3076 }
3077
3078 void reset_io_stats(struct thread_data *td)
3079 {
3080         struct thread_stat *ts = &td->ts;
3081         int i, j;
3082
3083         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
3084                 reset_io_stat(&ts->clat_stat[i]);
3085                 reset_io_stat(&ts->slat_stat[i]);
3086                 reset_io_stat(&ts->lat_stat[i]);
3087                 reset_io_stat(&ts->bw_stat[i]);
3088                 reset_io_stat(&ts->iops_stat[i]);
3089
3090                 ts->io_bytes[i] = 0;
3091                 ts->runtime[i] = 0;
3092                 ts->total_io_u[i] = 0;
3093                 ts->short_io_u[i] = 0;
3094                 ts->drop_io_u[i] = 0;
3095         }
3096
3097         for (i = 0; i < FIO_LAT_CNT; i++)
3098                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
3099                         reset_io_u_plat(ts->io_u_plat[i][j]);
3100
3101         reset_clat_prio_stats(ts);
3102
3103         ts->total_io_u[DDIR_SYNC] = 0;
3104         reset_io_u_plat(ts->io_u_sync_plat);
3105
3106         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
3107                 ts->io_u_map[i] = 0;
3108                 ts->io_u_submit[i] = 0;
3109                 ts->io_u_complete[i] = 0;
3110         }
3111
3112         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
3113                 ts->io_u_lat_n[i] = 0;
3114         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
3115                 ts->io_u_lat_u[i] = 0;
3116         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
3117                 ts->io_u_lat_m[i] = 0;
3118
3119         ts->total_submit = 0;
3120         ts->total_complete = 0;
3121         ts->nr_zone_resets = 0;
3122         ts->cachehit = ts->cachemiss = 0;
3123 }
3124
3125 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
3126                               unsigned long elapsed, bool log_max)
3127 {
3128         /*
3129          * Note an entry in the log. Use the mean from the logged samples,
3130          * making sure to properly round up. Only write a log entry if we
3131          * had actual samples done.
3132          */
3133         if (iolog->avg_window[ddir].samples) {
3134                 union io_sample_data data;
3135
3136                 if (log_max)
3137                         data.val = iolog->avg_window[ddir].max_val;
3138                 else
3139                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
3140
3141                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
3142         }
3143
3144         reset_io_stat(&iolog->avg_window[ddir]);
3145 }
3146
3147 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
3148                              bool log_max)
3149 {
3150         enum fio_ddir ddir;
3151
3152         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
3153                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
3154 }
3155
3156 static unsigned long add_log_sample(struct thread_data *td,
3157                                     struct io_log *iolog,
3158                                     union io_sample_data data,
3159                                     enum fio_ddir ddir, unsigned long long bs,
3160                                     uint64_t offset, unsigned int ioprio)
3161 {
3162         unsigned long elapsed, this_window;
3163
3164         if (!ddir_rw(ddir))
3165                 return 0;
3166
3167         elapsed = mtime_since_now(&td->epoch);
3168
3169         /*
3170          * If no time averaging, just add the log sample.
3171          */
3172         if (!iolog->avg_msec) {
3173                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
3174                                  ioprio);
3175                 return 0;
3176         }
3177
3178         /*
3179          * Add the sample. If the time period has passed, then
3180          * add that entry to the log and clear.
3181          */
3182         add_stat_sample(&iolog->avg_window[ddir], data.val);
3183
3184         /*
3185          * If period hasn't passed, adding the above sample is all we
3186          * need to do.
3187          */
3188         this_window = elapsed - iolog->avg_last[ddir];
3189         if (elapsed < iolog->avg_last[ddir])
3190                 return iolog->avg_last[ddir] - elapsed;
3191         else if (this_window < iolog->avg_msec) {
3192                 unsigned long diff = iolog->avg_msec - this_window;
3193
3194                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
3195                         return diff;
3196         }
3197
3198         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
3199
3200         iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
3201
3202         return iolog->avg_msec;
3203 }
3204
3205 void finalize_logs(struct thread_data *td, bool unit_logs)
3206 {
3207         unsigned long elapsed;
3208
3209         elapsed = mtime_since_now(&td->epoch);
3210
3211         if (td->clat_log && unit_logs)
3212                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
3213         if (td->slat_log && unit_logs)
3214                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
3215         if (td->lat_log && unit_logs)
3216                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
3217         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
3218                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
3219         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
3220                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
3221 }
3222
3223 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
3224                     unsigned long long bs)
3225 {
3226         struct io_log *iolog;
3227
3228         if (!ddir_rw(ddir))
3229                 return;
3230
3231         iolog = agg_io_log[ddir];
3232         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
3233 }
3234
3235 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
3236 {
3237         unsigned int idx = plat_val_to_idx(nsec);
3238         assert(idx < FIO_IO_U_PLAT_NR);
3239
3240         ts->io_u_sync_plat[idx]++;
3241         add_stat_sample(&ts->sync_stat, nsec);
3242 }
3243
3244 static inline void add_lat_percentile_sample(struct thread_stat *ts,
3245                                              unsigned long long nsec,
3246                                              enum fio_ddir ddir,
3247                                              enum fio_lat lat)
3248 {
3249         unsigned int idx = plat_val_to_idx(nsec);
3250         assert(idx < FIO_IO_U_PLAT_NR);
3251
3252         ts->io_u_plat[lat][ddir][idx]++;
3253 }
3254
3255 static inline void
3256 add_lat_percentile_prio_sample(struct thread_stat *ts, unsigned long long nsec,
3257                                enum fio_ddir ddir,
3258                                unsigned short clat_prio_index)
3259 {
3260         unsigned int idx = plat_val_to_idx(nsec);
3261
3262         if (ts->clat_prio[ddir])
3263                 ts->clat_prio[ddir][clat_prio_index].io_u_plat[idx]++;
3264 }
3265
3266 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
3267                      unsigned long long nsec, unsigned long long bs,
3268                      uint64_t offset, unsigned int ioprio,
3269                      unsigned short clat_prio_index)
3270 {
3271         const bool needs_lock = td_async_processing(td);
3272         unsigned long elapsed, this_window;
3273         struct thread_stat *ts = &td->ts;
3274         struct io_log *iolog = td->clat_hist_log;
3275
3276         if (needs_lock)
3277                 __td_io_u_lock(td);
3278
3279         add_stat_sample(&ts->clat_stat[ddir], nsec);
3280
3281         /*
3282          * When lat_percentiles=1 (default 0), the reported per priority
3283          * percentiles and stats are used for describing total latency values,
3284          * even though the variable names themselves start with clat_.
3285          *
3286          * Because of the above definition, add a prio stat sample only when
3287          * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3288          * when lat_percentiles=1.
3289          */
3290         if (!ts->lat_percentiles)
3291                 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3292                                      nsec);
3293
3294         if (td->clat_log)
3295                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3296                                offset, ioprio);
3297
3298         if (ts->clat_percentiles) {
3299                 /*
3300                  * Because of the above definition, add a prio lat percentile
3301                  * sample only when lat_percentiles=0. add_lat_sample() will add
3302                  * the prio lat percentile sample when lat_percentiles=1.
3303                  */
3304                 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3305                 if (!ts->lat_percentiles)
3306                         add_lat_percentile_prio_sample(ts, nsec, ddir,
3307                                                        clat_prio_index);
3308         }
3309
3310         if (iolog && iolog->hist_msec) {
3311                 struct io_hist *hw = &iolog->hist_window[ddir];
3312
3313                 hw->samples++;
3314                 elapsed = mtime_since_now(&td->epoch);
3315                 if (!hw->hist_last)
3316                         hw->hist_last = elapsed;
3317                 this_window = elapsed - hw->hist_last;
3318
3319                 if (this_window >= iolog->hist_msec) {
3320                         uint64_t *io_u_plat;
3321                         struct io_u_plat_entry *dst;
3322
3323                         /*
3324                          * Make a byte-for-byte copy of the latency histogram
3325                          * stored in td->ts.io_u_plat[ddir], recording it in a
3326                          * log sample. Note that the matching call to free() is
3327                          * located in iolog.c after printing this sample to the
3328                          * log file.
3329                          */
3330                         io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3331                         dst = malloc(sizeof(struct io_u_plat_entry));
3332                         memcpy(&(dst->io_u_plat), io_u_plat,
3333                                 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3334                         flist_add(&dst->list, &hw->list);
3335                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3336                                          elapsed, offset, ioprio);
3337
3338                         /*
3339                          * Update the last time we recorded as being now, minus
3340                          * any drift in time we encountered before actually
3341                          * making the record.
3342                          */
3343                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3344                         hw->samples = 0;
3345                 }
3346         }
3347
3348         if (needs_lock)
3349                 __td_io_u_unlock(td);
3350 }
3351
3352 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3353                      unsigned long long nsec, unsigned long long bs,
3354                      uint64_t offset, unsigned int ioprio)
3355 {
3356         const bool needs_lock = td_async_processing(td);
3357         struct thread_stat *ts = &td->ts;
3358
3359         if (!ddir_rw(ddir))
3360                 return;
3361
3362         if (needs_lock)
3363                 __td_io_u_lock(td);
3364
3365         add_stat_sample(&ts->slat_stat[ddir], nsec);
3366
3367         if (td->slat_log)
3368                 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3369                                offset, ioprio);
3370
3371         if (ts->slat_percentiles)
3372                 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3373
3374         if (needs_lock)
3375                 __td_io_u_unlock(td);
3376 }
3377
3378 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3379                     unsigned long long nsec, unsigned long long bs,
3380                     uint64_t offset, unsigned int ioprio,
3381                     unsigned short clat_prio_index)
3382 {
3383         const bool needs_lock = td_async_processing(td);
3384         struct thread_stat *ts = &td->ts;
3385
3386         if (!ddir_rw(ddir))
3387                 return;
3388
3389         if (needs_lock)
3390                 __td_io_u_lock(td);
3391
3392         add_stat_sample(&ts->lat_stat[ddir], nsec);
3393
3394         if (td->lat_log)
3395                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3396                                offset, ioprio);
3397
3398         /*
3399          * When lat_percentiles=1 (default 0), the reported per priority
3400          * percentiles and stats are used for describing total latency values,
3401          * even though the variable names themselves start with clat_.
3402          *
3403          * Because of the above definition, add a prio stat and prio lat
3404          * percentile sample only when lat_percentiles=1. add_clat_sample() will
3405          * add the prio stat and prio lat percentile sample when
3406          * lat_percentiles=0.
3407          */
3408         if (ts->lat_percentiles) {
3409                 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3410                 add_lat_percentile_prio_sample(ts, nsec, ddir, clat_prio_index);
3411                 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3412                                      nsec);
3413         }
3414         if (needs_lock)
3415                 __td_io_u_unlock(td);
3416 }
3417
3418 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3419                    unsigned int bytes, unsigned long long spent)
3420 {
3421         const bool needs_lock = td_async_processing(td);
3422         struct thread_stat *ts = &td->ts;
3423         unsigned long rate;
3424
3425         if (spent)
3426                 rate = (unsigned long) (bytes * 1000000ULL / spent);
3427         else
3428                 rate = 0;
3429
3430         if (needs_lock)
3431                 __td_io_u_lock(td);
3432
3433         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3434
3435         if (td->bw_log)
3436                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3437                                bytes, io_u->offset, io_u->ioprio);
3438
3439         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3440
3441         if (needs_lock)
3442                 __td_io_u_unlock(td);
3443 }
3444
3445 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3446                          struct timespec *t, unsigned int avg_time,
3447                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3448                          struct io_stat *stat, struct io_log *log,
3449                          bool is_kb)
3450 {
3451         const bool needs_lock = td_async_processing(td);
3452         unsigned long spent, rate;
3453         enum fio_ddir ddir;
3454         unsigned long next, next_log;
3455
3456         next_log = avg_time;
3457
3458         spent = mtime_since(parent_tv, t);
3459         if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3460                 return avg_time - spent;
3461
3462         if (needs_lock)
3463                 __td_io_u_lock(td);
3464
3465         /*
3466          * Compute both read and write rates for the interval.
3467          */
3468         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3469                 uint64_t delta;
3470
3471                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3472                 if (!delta)
3473                         continue; /* No entries for interval */
3474
3475                 if (spent) {
3476                         if (is_kb)
3477                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
3478                         else
3479                                 rate = (delta * 1000) / spent;
3480                 } else
3481                         rate = 0;
3482
3483                 add_stat_sample(&stat[ddir], rate);
3484
3485                 if (log) {
3486                         unsigned long long bs = 0;
3487
3488                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3489                                 bs = td->o.min_bs[ddir];
3490
3491                         next = add_log_sample(td, log, sample_val(rate), ddir,
3492                                               bs, 0, 0);
3493                         next_log = min(next_log, next);
3494                 }
3495
3496                 stat_io_bytes[ddir] = this_io_bytes[ddir];
3497         }
3498
3499         *parent_tv = *t;
3500
3501         if (needs_lock)
3502                 __td_io_u_unlock(td);
3503
3504         if (spent <= avg_time)
3505                 next = avg_time;
3506         else
3507                 next = avg_time - (1 + spent - avg_time);
3508
3509         return min(next, next_log);
3510 }
3511
3512 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3513 {
3514         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3515                                 td->this_io_bytes, td->stat_io_bytes,
3516                                 td->ts.bw_stat, td->bw_log, true);
3517 }
3518
3519 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3520                      unsigned int bytes)
3521 {
3522         const bool needs_lock = td_async_processing(td);
3523         struct thread_stat *ts = &td->ts;
3524
3525         if (needs_lock)
3526                 __td_io_u_lock(td);
3527
3528         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3529
3530         if (td->iops_log)
3531                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3532                                bytes, io_u->offset, io_u->ioprio);
3533
3534         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3535
3536         if (needs_lock)
3537                 __td_io_u_unlock(td);
3538 }
3539
3540 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3541 {
3542         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3543                                 td->this_io_blocks, td->stat_io_blocks,
3544                                 td->ts.iops_stat, td->iops_log, false);
3545 }
3546
3547 /*
3548  * Returns msecs to next event
3549  */
3550 int calc_log_samples(void)
3551 {
3552         struct thread_data *td;
3553         unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3554         struct timespec now;
3555         int i;
3556         long elapsed_time = 0;
3557
3558         fio_gettime(&now, NULL);
3559
3560         for_each_td(td, i) {
3561                 elapsed_time = mtime_since_now(&td->epoch);
3562
3563                 if (!td->o.stats)
3564                         continue;
3565                 if (in_ramp_time(td) ||
3566                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3567                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3568                         continue;
3569                 }
3570                 if (!td->bw_log ||
3571                         (td->bw_log && !per_unit_log(td->bw_log))) {
3572                         tmp = add_bw_samples(td, &now);
3573
3574                         if (td->bw_log)
3575                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3576                 }
3577                 if (!td->iops_log ||
3578                         (td->iops_log && !per_unit_log(td->iops_log))) {
3579                         tmp = add_iops_samples(td, &now);
3580
3581                         if (td->iops_log)
3582                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3583                 }
3584
3585                 if (tmp < next)
3586                         next = tmp;
3587         }
3588
3589         /* if log_avg_msec_min has not been changed, set it to 0 */
3590         if (log_avg_msec_min == -1U)
3591                 log_avg_msec_min = 0;
3592
3593         if (log_avg_msec_min == 0)
3594                 next_mod = elapsed_time;
3595         else
3596                 next_mod = elapsed_time % log_avg_msec_min;
3597
3598         /* correction to keep the time on the log avg msec boundary */
3599         next = min(next, (log_avg_msec_min - next_mod));
3600
3601         return next == ~0U ? 0 : next;
3602 }
3603
3604 void stat_init(void)
3605 {
3606         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3607 }
3608
3609 void stat_exit(void)
3610 {
3611         /*
3612          * When we have the mutex, we know out-of-band access to it
3613          * have ended.
3614          */
3615         fio_sem_down(stat_sem);
3616         fio_sem_remove(stat_sem);
3617 }
3618
3619 /*
3620  * Called from signal handler. Wake up status thread.
3621  */
3622 void show_running_run_stats(void)
3623 {
3624         helper_do_stat();
3625 }
3626
3627 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3628 {
3629         /* Ignore io_u's which span multiple blocks--they will just get
3630          * inaccurate counts. */
3631         int idx = (io_u->offset - io_u->file->file_offset)
3632                         / td->o.bs[DDIR_TRIM];
3633         uint32_t *info = &td->ts.block_infos[idx];
3634         assert(idx < td->ts.nr_block_infos);
3635         return info;
3636 }
3637