stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <stdlib.h>
   4 #include <sys/time.h>
   5 #include <sys/stat.h>
   6 #include <math.h>
   7
   8 #include "fio.h"
   9 #include "diskutil.h"
  10 #include "lib/ieee754.h"
  11 #include "json.h"
  12 #include "lib/getrusage.h"
  13 #include "idletime.h"
  14 #include "lib/pow2.h"
  15 #include "lib/output_buffer.h"
  16 #include "helper_thread.h"
  17 #include "smalloc.h"
  18 #include "zbd.h"
  19 #include "oslib/asprintf.h"
  20
  21 #ifdef WIN32
  22 #define LOG_MSEC_SLACK  2
  23 #else
  24 #define LOG_MSEC_SLACK  1
  25 #endif
  26
  27 struct fio_sem *stat_sem;
  28
  29 void clear_rusage_stat(struct thread_data *td)
  30 {
  31         struct thread_stat *ts = &td->ts;
  32
  33         fio_getrusage(&td->ru_start);
  34         ts->usr_time = ts->sys_time = 0;
  35         ts->ctx = 0;
  36         ts->minf = ts->majf = 0;
  37 }
  38
  39 void update_rusage_stat(struct thread_data *td)
  40 {
  41         struct thread_stat *ts = &td->ts;
  42
  43         fio_getrusage(&td->ru_end);
  44         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
  45                                         &td->ru_end.ru_utime);
  46         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
  47                                         &td->ru_end.ru_stime);
  48         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  49                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  50         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  51         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  52
  53         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  54 }
  55
  56 /*
  57  * Given a latency, return the index of the corresponding bucket in
  58  * the structure tracking percentiles.
  59  *
  60  * (1) find the group (and error bits) that the value (latency)
  61  * belongs to by looking at its MSB. (2) find the bucket number in the
  62  * group by looking at the index bits.
  63  *
  64  */
  65 static unsigned int plat_val_to_idx(unsigned long long val)
  66 {
  67         unsigned int msb, error_bits, base, offset, idx;
  68
  69         /* Find MSB starting from bit 0 */
  70         if (val == 0)
  71                 msb = 0;
  72         else
  73                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
  74
  75         /*
  76          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  77          * all bits of the sample as index
  78          */
  79         if (msb <= FIO_IO_U_PLAT_BITS)
  80                 return val;
  81
  82         /* Compute the number of error bits to discard*/
  83         error_bits = msb - FIO_IO_U_PLAT_BITS;
  84
  85         /* Compute the number of buckets before the group */
  86         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  87
  88         /*
  89          * Discard the error bits and apply the mask to find the
  90          * index for the buckets in the group
  91          */
  92         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  93
  94         /* Make sure the index does not exceed (array size - 1) */
  95         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  96                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  97
  98         return idx;
  99 }
 100
 101 /*
 102  * Convert the given index of the bucket array to the value
 103  * represented by the bucket
 104  */
 105 static unsigned long long plat_idx_to_val(unsigned int idx)
 106 {
 107         unsigned int error_bits;
 108         unsigned long long k, base;
 109
 110         assert(idx < FIO_IO_U_PLAT_NR);
 111
 112         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 113          * all bits of the sample as index */
 114         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 115                 return idx;
 116
 117         /* Find the group and compute the minimum value of that group */
 118         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 119         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
 120
 121         /* Find its bucket number of the group */
 122         k = idx % FIO_IO_U_PLAT_VAL;
 123
 124         /* Return the mean of the range of the bucket */
 125         return base + ((k + 0.5) * (1 << error_bits));
 126 }
 127
 128 static int double_cmp(const void *a, const void *b)
 129 {
 130         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 131         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 132         int cmp = 0;
 133
 134         if (fa.u.f > fb.u.f)
 135                 cmp = 1;
 136         else if (fa.u.f < fb.u.f)
 137                 cmp = -1;
 138
 139         return cmp;
 140 }
 141
 142 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 143                                    fio_fp64_t *plist, unsigned long long **output,
 144                                    unsigned long long *maxv, unsigned long long *minv)
 145 {
 146         unsigned long long sum = 0;
 147         unsigned int len, i, j = 0;
 148         unsigned long long *ovals = NULL;
 149         bool is_last;
 150
 151         *minv = -1ULL;
 152         *maxv = 0;
 153
 154         len = 0;
 155         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 156                 len++;
 157
 158         if (!len)
 159                 return 0;
 160
 161         /*
 162          * Sort the percentile list. Note that it may already be sorted if
 163          * we are using the default values, but since it's a short list this
 164          * isn't a worry. Also note that this does not work for NaN values.
 165          */
 166         if (len > 1)
 167                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 168
 169         ovals = malloc(len * sizeof(*ovals));
 170         if (!ovals)
 171                 return 0;
 172
 173         /*
 174          * Calculate bucket values, note down max and min values
 175          */
 176         is_last = false;
 177         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 178                 sum += io_u_plat[i];
 179                 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
 180                         assert(plist[j].u.f <= 100.0);
 181
 182                         ovals[j] = plat_idx_to_val(i);
 183                         if (ovals[j] < *minv)
 184                                 *minv = ovals[j];
 185                         if (ovals[j] > *maxv)
 186                                 *maxv = ovals[j];
 187
 188                         is_last = (j == len - 1) != 0;
 189                         if (is_last)
 190                                 break;
 191
 192                         j++;
 193                 }
 194         }
 195
 196         if (!is_last)
 197                 log_err("fio: error calculating latency percentiles\n");
 198
 199         *output = ovals;
 200         return len;
 201 }
 202
 203 /*
 204  * Find and display the p-th percentile of clat
 205  */
 206 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 207                                   fio_fp64_t *plist, unsigned int precision,
 208                                   const char *pre, struct buf_output *out)
 209 {
 210         unsigned int divisor, len, i, j = 0;
 211         unsigned long long minv, maxv;
 212         unsigned long long *ovals;
 213         int per_line, scale_down, time_width;
 214         bool is_last;
 215         char fmt[32];
 216
 217         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 218         if (!len || !ovals)
 219                 return;
 220
 221         /*
 222          * We default to nsecs, but if the value range is such that we
 223          * should scale down to usecs or msecs, do that.
 224          */
 225         if (minv > 2000000 && maxv > 99999999ULL) {
 226                 scale_down = 2;
 227                 divisor = 1000000;
 228                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 229         } else if (minv > 2000 && maxv > 99999) {
 230                 scale_down = 1;
 231                 divisor = 1000;
 232                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 233         } else {
 234                 scale_down = 0;
 235                 divisor = 1;
 236                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 237         }
 238
 239
 240         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 241         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 242                         precision, time_width);
 243         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 244         per_line = (80 - 7) / (precision + 10 + time_width);
 245
 246         for (j = 0; j < len; j++) {
 247                 /* for formatting */
 248                 if (j != 0 && (j % per_line) == 0)
 249                         log_buf(out, "     |");
 250
 251                 /* end of the list */
 252                 is_last = (j == len - 1) != 0;
 253
 254                 for (i = 0; i < scale_down; i++)
 255                         ovals[j] = (ovals[j] + 999) / 1000;
 256
 257                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 258
 259                 if (is_last)
 260                         break;
 261
 262                 if ((j % per_line) == per_line - 1)     /* for formatting */
 263                         log_buf(out, "\n");
 264         }
 265
 266         free(ovals);
 267 }
 268
 269 static int get_nr_prios_with_samples(struct thread_stat *ts, enum fio_ddir ddir)
 270 {
 271         int i, nr_prios_with_samples = 0;
 272
 273         for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
 274                 if (ts->clat_prio[ddir][i].clat_stat.samples)
 275                         nr_prios_with_samples++;
 276         }
 277
 278         return nr_prios_with_samples;
 279 }
 280
 281 bool calc_lat(struct io_stat *is, unsigned long long *min,
 282               unsigned long long *max, double *mean, double *dev)
 283 {
 284         double n = (double) is->samples;
 285
 286         if (n == 0)
 287                 return false;
 288
 289         *min = is->min_val;
 290         *max = is->max_val;
 291         *mean = is->mean.u.f;
 292
 293         if (n > 1.0)
 294                 *dev = sqrt(is->S.u.f / (n - 1.0));
 295         else
 296                 *dev = 0;
 297
 298         return true;
 299 }
 300
 301 void show_mixed_group_stats(struct group_run_stats *rs, struct buf_output *out)
 302 {
 303         char *io, *agg, *min, *max;
 304         char *ioalt, *aggalt, *minalt, *maxalt;
 305         uint64_t io_mix = 0, agg_mix = 0, min_mix = -1, max_mix = 0;
 306         uint64_t min_run = -1, max_run = 0;
 307         const int i2p = is_power_of_2(rs->kb_base);
 308         int i;
 309
 310         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 311                 if (!rs->max_run[i])
 312                         continue;
 313                 io_mix += rs->iobytes[i];
 314                 agg_mix += rs->agg[i];
 315                 min_mix = min_mix < rs->min_bw[i] ? min_mix : rs->min_bw[i];
 316                 max_mix = max_mix > rs->max_bw[i] ? max_mix : rs->max_bw[i];
 317                 min_run = min_run < rs->min_run[i] ? min_run : rs->min_run[i];
 318                 max_run = max_run > rs->max_run[i] ? max_run : rs->max_run[i];
 319         }
 320         io = num2str(io_mix, rs->sig_figs, 1, i2p, N2S_BYTE);
 321         ioalt = num2str(io_mix, rs->sig_figs, 1, !i2p, N2S_BYTE);
 322         agg = num2str(agg_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 323         aggalt = num2str(agg_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 324         min = num2str(min_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 325         minalt = num2str(min_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 326         max = num2str(max_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 327         maxalt = num2str(max_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 328         log_buf(out, "  MIXED: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 329                         agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 330                         (unsigned long long) min_run,
 331                         (unsigned long long) max_run);
 332         free(io);
 333         free(agg);
 334         free(min);
 335         free(max);
 336         free(ioalt);
 337         free(aggalt);
 338         free(minalt);
 339         free(maxalt);
 340 }
 341
 342 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 343 {
 344         char *io, *agg, *min, *max;
 345         char *ioalt, *aggalt, *minalt, *maxalt;
 346         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 347         int i;
 348
 349         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 350
 351         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 352                 const int i2p = is_power_of_2(rs->kb_base);
 353
 354                 if (!rs->max_run[i])
 355                         continue;
 356
 357                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
 358                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
 359                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
 360                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 361                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 362                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 363                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 364                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 365                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 366                                 (rs->unified_rw_rep == UNIFIED_MIXED) ? "  MIXED" : str[i],
 367                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 368                                 (unsigned long long) rs->min_run[i],
 369                                 (unsigned long long) rs->max_run[i]);
 370
 371                 free(io);
 372                 free(agg);
 373                 free(min);
 374                 free(max);
 375                 free(ioalt);
 376                 free(aggalt);
 377                 free(minalt);
 378                 free(maxalt);
 379         }
 380
 381         /* Need to aggregate statistics to show mixed values */
 382         if (rs->unified_rw_rep == UNIFIED_BOTH)
 383                 show_mixed_group_stats(rs, out);
 384 }
 385
 386 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
 387 {
 388         int i;
 389
 390         /*
 391          * Do depth distribution calculations
 392          */
 393         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 394                 if (total) {
 395                         io_u_dist[i] = (double) map[i] / (double) total;
 396                         io_u_dist[i] *= 100.0;
 397                         if (io_u_dist[i] < 0.1 && map[i])
 398                                 io_u_dist[i] = 0.1;
 399                 } else
 400                         io_u_dist[i] = 0.0;
 401         }
 402 }
 403
 404 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 405                           uint64_t *src, int nr)
 406 {
 407         unsigned long total = ddir_rw_sum(ts->total_io_u);
 408         int i;
 409
 410         /*
 411          * Do latency distribution calculations
 412          */
 413         for (i = 0; i < nr; i++) {
 414                 if (total) {
 415                         dst[i] = (double) src[i] / (double) total;
 416                         dst[i] *= 100.0;
 417                         if (dst[i] < 0.01 && src[i])
 418                                 dst[i] = 0.01;
 419                 } else
 420                         dst[i] = 0.0;
 421         }
 422 }
 423
 424 /*
 425  * To keep the terse format unaltered, add all of the ns latency
 426  * buckets to the first us latency bucket
 427  */
 428 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 429 {
 430         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 431         int i;
 432
 433         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 434
 435         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 436                 ntotal += ts->io_u_lat_n[i];
 437
 438         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 439 }
 440
 441 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 442 {
 443         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 444 }
 445
 446 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 447 {
 448         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 449 }
 450
 451 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 452 {
 453         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 454 }
 455
 456 static void display_lat(const char *name, unsigned long long min,
 457                         unsigned long long max, double mean, double dev,
 458                         struct buf_output *out)
 459 {
 460         const char *base = "(nsec)";
 461         char *minp, *maxp;
 462
 463         if (nsec_to_msec(&min, &max, &mean, &dev))
 464                 base = "(msec)";
 465         else if (nsec_to_usec(&min, &max, &mean, &dev))
 466                 base = "(usec)";
 467
 468         minp = num2str(min, 6, 1, 0, N2S_NONE);
 469         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 470
 471         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 472                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 473
 474         free(minp);
 475         free(maxp);
 476 }
 477
 478 static struct thread_stat *gen_mixed_ddir_stats_from_ts(struct thread_stat *ts)
 479 {
 480         struct thread_stat *ts_lcl;
 481
 482         /*
 483          * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
 484          * Trims (ddir = 2)
 485          */
 486         ts_lcl = malloc(sizeof(struct thread_stat));
 487         if (!ts_lcl) {
 488                 log_err("fio: failed to allocate local thread stat\n");
 489                 return NULL;
 490         }
 491
 492         init_thread_stat(ts_lcl);
 493
 494         /* calculate mixed stats  */
 495         ts_lcl->unified_rw_rep = UNIFIED_MIXED;
 496         ts_lcl->lat_percentiles = ts->lat_percentiles;
 497         ts_lcl->clat_percentiles = ts->clat_percentiles;
 498         ts_lcl->slat_percentiles = ts->slat_percentiles;
 499         ts_lcl->percentile_precision = ts->percentile_precision;
 500         memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
 501
 502         sum_thread_stats(ts_lcl, ts);
 503
 504         return ts_lcl;
 505 }
 506
 507 static double convert_agg_kbytes_percent(struct group_run_stats *rs,
 508                                          enum fio_ddir ddir, int mean)
 509 {
 510         double p_of_agg = 100.0;
 511         if (rs && rs->agg[ddir] > 1024) {
 512                 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
 513
 514                 if (p_of_agg > 100.0)
 515                         p_of_agg = 100.0;
 516         }
 517         return p_of_agg;
 518 }
 519
 520 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 521                              enum fio_ddir ddir, struct buf_output *out)
 522 {
 523         unsigned long runt;
 524         unsigned long long min, max, bw, iops;
 525         double mean, dev;
 526         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
 527         int i2p, i;
 528         const char *clat_type = ts->lat_percentiles ? "lat" : "clat";
 529
 530         if (ddir_sync(ddir)) {
 531                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
 532                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
 533                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
 534                         show_clat_percentiles(ts->io_u_sync_plat,
 535                                                 ts->sync_stat.samples,
 536                                                 ts->percentile_list,
 537                                                 ts->percentile_precision,
 538                                                 io_ddir_name(ddir), out);
 539                 }
 540                 return;
 541         }
 542
 543         assert(ddir_rw(ddir));
 544
 545         if (!ts->runtime[ddir])
 546                 return;
 547
 548         i2p = is_power_of_2(rs->kb_base);
 549         runt = ts->runtime[ddir];
 550
 551         bw = (1000 * ts->io_bytes[ddir]) / runt;
 552         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
 553         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
 554         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
 555
 556         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 557         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
 558         if (ddir == DDIR_WRITE)
 559                 post_st = zbd_write_status(ts);
 560         else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
 561                 uint64_t total;
 562                 double hit;
 563
 564                 total = ts->cachehit + ts->cachemiss;
 565                 hit = (double) ts->cachehit / (double) total;
 566                 hit *= 100.0;
 567                 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
 568                         post_st = NULL;
 569         }
 570
 571         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
 572                         (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
 573                         iops_p, bw_p, bw_p_alt, io_p,
 574                         (unsigned long long) ts->runtime[ddir],
 575                         post_st ? : "");
 576
 577         free(post_st);
 578         free(io_p);
 579         free(bw_p);
 580         free(bw_p_alt);
 581         free(iops_p);
 582
 583         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 584                 display_lat("slat", min, max, mean, dev, out);
 585         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 586                 display_lat("clat", min, max, mean, dev, out);
 587         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 588                 display_lat(" lat", min, max, mean, dev, out);
 589
 590         /* Only print per prio stats if there are >= 2 prios with samples */
 591         if (get_nr_prios_with_samples(ts, ddir) >= 2) {
 592                 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
 593                         if (calc_lat(&ts->clat_prio[ddir][i].clat_stat, &min,
 594                                      &max, &mean, &dev)) {
 595                                 char buf[64];
 596
 597                                 snprintf(buf, sizeof(buf),
 598                                          "%s prio %u/%u",
 599                                          clat_type,
 600                                          ts->clat_prio[ddir][i].ioprio >> 13,
 601                                          ts->clat_prio[ddir][i].ioprio & 7);
 602                                 display_lat(buf, min, max, mean, dev, out);
 603                         }
 604                 }
 605         }
 606
 607         if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
 608                 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
 609                                         ts->slat_stat[ddir].samples,
 610                                         ts->percentile_list,
 611                                         ts->percentile_precision, "slat", out);
 612         if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
 613                 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
 614                                         ts->clat_stat[ddir].samples,
 615                                         ts->percentile_list,
 616                                         ts->percentile_precision, "clat", out);
 617         if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
 618                 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
 619                                         ts->lat_stat[ddir].samples,
 620                                         ts->percentile_list,
 621                                         ts->percentile_precision, "lat", out);
 622
 623         if (ts->clat_percentiles || ts->lat_percentiles) {
 624                 char prio_name[64];
 625                 uint64_t samples;
 626
 627                 if (ts->lat_percentiles)
 628                         samples = ts->lat_stat[ddir].samples;
 629                 else
 630                         samples = ts->clat_stat[ddir].samples;
 631
 632                 /* Only print per prio stats if there are >= 2 prios with samples */
 633                 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
 634                         for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
 635                                 uint64_t prio_samples = ts->clat_prio[ddir][i].clat_stat.samples;
 636
 637                                 if (prio_samples > 0) {
 638                                         snprintf(prio_name, sizeof(prio_name),
 639                                                  "%s prio %u/%u (%.2f%% of IOs)",
 640                                                  clat_type,
 641                                                  ts->clat_prio[ddir][i].ioprio >> 13,
 642                                                  ts->clat_prio[ddir][i].ioprio & 7,
 643                                                  100. * (double) prio_samples / (double) samples);
 644                                         show_clat_percentiles(ts->clat_prio[ddir][i].io_u_plat,
 645                                                               prio_samples, ts->percentile_list,
 646                                                               ts->percentile_precision,
 647                                                               prio_name, out);
 648                                 }
 649                         }
 650                 }
 651         }
 652
 653         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 654                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 655                 const char *bw_str;
 656
 657                 if ((rs->unit_base == 1) && i2p)
 658                         bw_str = "Kibit";
 659                 else if (rs->unit_base == 1)
 660                         bw_str = "kbit";
 661                 else if (i2p)
 662                         bw_str = "KiB";
 663                 else
 664                         bw_str = "kB";
 665
 666                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
 667
 668                 if (rs->unit_base == 1) {
 669                         min *= 8.0;
 670                         max *= 8.0;
 671                         mean *= 8.0;
 672                         dev *= 8.0;
 673                 }
 674
 675                 if (mean > fkb_base * fkb_base) {
 676                         min /= fkb_base;
 677                         max /= fkb_base;
 678                         mean /= fkb_base;
 679                         dev /= fkb_base;
 680                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 681                 }
 682
 683                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 684                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 685                         bw_str, min, max, p_of_agg, mean, dev,
 686                         (&ts->bw_stat[ddir])->samples);
 687         }
 688         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 689                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 690                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 691                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 692         }
 693 }
 694
 695 static void show_mixed_ddir_status(struct group_run_stats *rs,
 696                                    struct thread_stat *ts,
 697                                    struct buf_output *out)
 698 {
 699         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
 700
 701         if (ts_lcl)
 702                 show_ddir_status(rs, ts_lcl, DDIR_READ, out);
 703
 704         free_clat_prio_stats(ts_lcl);
 705         free(ts_lcl);
 706 }
 707
 708 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
 709                      const char *msg, struct buf_output *out)
 710 {
 711         bool new_line = true, shown = false;
 712         int i, line = 0;
 713
 714         for (i = 0; i < nr; i++) {
 715                 if (io_u_lat[i] <= 0.0)
 716                         continue;
 717                 shown = true;
 718                 if (new_line) {
 719                         if (line)
 720                                 log_buf(out, "\n");
 721                         log_buf(out, "  lat (%s)   : ", msg);
 722                         new_line = false;
 723                         line = 0;
 724                 }
 725                 if (line)
 726                         log_buf(out, ", ");
 727                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 728                 line++;
 729                 if (line == 5)
 730                         new_line = true;
 731         }
 732
 733         if (shown)
 734                 log_buf(out, "\n");
 735
 736         return true;
 737 }
 738
 739 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 740 {
 741         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 742                                  "250=", "500=", "750=", "1000=", };
 743
 744         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 745 }
 746
 747 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 748 {
 749         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 750                                  "250=", "500=", "750=", "1000=", };
 751
 752         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 753 }
 754
 755 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 756 {
 757         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 758                                  "250=", "500=", "750=", "1000=", "2000=",
 759                                  ">=2000=", };
 760
 761         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 762 }
 763
 764 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 765 {
 766         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 767         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 768         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 769
 770         stat_calc_lat_n(ts, io_u_lat_n);
 771         stat_calc_lat_u(ts, io_u_lat_u);
 772         stat_calc_lat_m(ts, io_u_lat_m);
 773
 774         show_lat_n(io_u_lat_n, out);
 775         show_lat_u(io_u_lat_u, out);
 776         show_lat_m(io_u_lat_m, out);
 777 }
 778
 779 static int block_state_category(int block_state)
 780 {
 781         switch (block_state) {
 782         case BLOCK_STATE_UNINIT:
 783                 return 0;
 784         case BLOCK_STATE_TRIMMED:
 785         case BLOCK_STATE_WRITTEN:
 786                 return 1;
 787         case BLOCK_STATE_WRITE_FAILURE:
 788         case BLOCK_STATE_TRIM_FAILURE:
 789                 return 2;
 790         default:
 791                 /* Silence compile warning on some BSDs and have a return */
 792                 assert(0);
 793                 return -1;
 794         }
 795 }
 796
 797 static int compare_block_infos(const void *bs1, const void *bs2)
 798 {
 799         uint64_t block1 = *(uint64_t *)bs1;
 800         uint64_t block2 = *(uint64_t *)bs2;
 801         int state1 = BLOCK_INFO_STATE(block1);
 802         int state2 = BLOCK_INFO_STATE(block2);
 803         int bscat1 = block_state_category(state1);
 804         int bscat2 = block_state_category(state2);
 805         int cycles1 = BLOCK_INFO_TRIMS(block1);
 806         int cycles2 = BLOCK_INFO_TRIMS(block2);
 807
 808         if (bscat1 < bscat2)
 809                 return -1;
 810         if (bscat1 > bscat2)
 811                 return 1;
 812
 813         if (cycles1 < cycles2)
 814                 return -1;
 815         if (cycles1 > cycles2)
 816                 return 1;
 817
 818         if (state1 < state2)
 819                 return -1;
 820         if (state1 > state2)
 821                 return 1;
 822
 823         assert(block1 == block2);
 824         return 0;
 825 }
 826
 827 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 828                                   fio_fp64_t *plist, unsigned int **percentiles,
 829                                   unsigned int *types)
 830 {
 831         int len = 0;
 832         int i, nr_uninit;
 833
 834         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 835
 836         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 837                 len++;
 838
 839         if (!len)
 840                 return 0;
 841
 842         /*
 843          * Sort the percentile list. Note that it may already be sorted if
 844          * we are using the default values, but since it's a short list this
 845          * isn't a worry. Also note that this does not work for NaN values.
 846          */
 847         if (len > 1)
 848                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 849
 850         /* Start only after the uninit entries end */
 851         for (nr_uninit = 0;
 852              nr_uninit < nr_block_infos
 853                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 854              nr_uninit ++)
 855                 ;
 856
 857         if (nr_uninit == nr_block_infos)
 858                 return 0;
 859
 860         *percentiles = calloc(len, sizeof(**percentiles));
 861
 862         for (i = 0; i < len; i++) {
 863                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 864                                 + nr_uninit;
 865                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 866         }
 867
 868         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 869         for (i = 0; i < nr_block_infos; i++)
 870                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 871
 872         return len;
 873 }
 874
 875 static const char *block_state_names[] = {
 876         [BLOCK_STATE_UNINIT] = "unwritten",
 877         [BLOCK_STATE_TRIMMED] = "trimmed",
 878         [BLOCK_STATE_WRITTEN] = "written",
 879         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 880         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 881 };
 882
 883 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 884                              fio_fp64_t *plist, struct buf_output *out)
 885 {
 886         int len, pos, i;
 887         unsigned int *percentiles = NULL;
 888         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 889
 890         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 891                                      &percentiles, block_state_counts);
 892
 893         log_buf(out, "  block lifetime percentiles :\n   |");
 894         pos = 0;
 895         for (i = 0; i < len; i++) {
 896                 uint32_t block_info = percentiles[i];
 897 #define LINE_LENGTH     75
 898                 char str[LINE_LENGTH];
 899                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 900                                      plist[i].u.f, block_info,
 901                                      i == len - 1 ? '\n' : ',');
 902                 assert(strln < LINE_LENGTH);
 903                 if (pos + strln > LINE_LENGTH) {
 904                         pos = 0;
 905                         log_buf(out, "\n   |");
 906                 }
 907                 log_buf(out, "%s", str);
 908                 pos += strln;
 909 #undef LINE_LENGTH
 910         }
 911         if (percentiles)
 912                 free(percentiles);
 913
 914         log_buf(out, "        states               :");
 915         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 916                 log_buf(out, " %s=%u%c",
 917                          block_state_names[i], block_state_counts[i],
 918                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 919 }
 920
 921 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 922 {
 923         char *p1, *p1alt, *p2;
 924         unsigned long long bw_mean, iops_mean;
 925         const int i2p = is_power_of_2(ts->kb_base);
 926
 927         if (!ts->ss_dur)
 928                 return;
 929
 930         bw_mean = steadystate_bw_mean(ts);
 931         iops_mean = steadystate_iops_mean(ts);
 932
 933         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
 934         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
 935         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
 936
 937         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 938                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
 939                 p1, p1alt, p2,
 940                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
 941                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
 942                 ts->ss_criterion.u.f,
 943                 ts->ss_state & FIO_SS_PCT ? "%" : "");
 944
 945         free(p1);
 946         free(p1alt);
 947         free(p2);
 948 }
 949
 950 static void show_agg_stats(struct disk_util_agg *agg, int terse,
 951                            struct buf_output *out)
 952 {
 953         if (!agg->slavecount)
 954                 return;
 955
 956         if (!terse) {
 957                 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
 958                          "aggrticks=%llu/%llu, aggrin_queue=%llu, "
 959                          "aggrutil=%3.2f%%",
 960                         (unsigned long long) agg->ios[0] / agg->slavecount,
 961                         (unsigned long long) agg->ios[1] / agg->slavecount,
 962                         (unsigned long long) agg->merges[0] / agg->slavecount,
 963                         (unsigned long long) agg->merges[1] / agg->slavecount,
 964                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 965                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 966                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 967                         agg->max_util.u.f);
 968         } else {
 969                 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
 970                         (unsigned long long) agg->ios[0] / agg->slavecount,
 971                         (unsigned long long) agg->ios[1] / agg->slavecount,
 972                         (unsigned long long) agg->merges[0] / agg->slavecount,
 973                         (unsigned long long) agg->merges[1] / agg->slavecount,
 974                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 975                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 976                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 977                         agg->max_util.u.f);
 978         }
 979 }
 980
 981 static void aggregate_slaves_stats(struct disk_util *masterdu)
 982 {
 983         struct disk_util_agg *agg = &masterdu->agg;
 984         struct disk_util_stat *dus;
 985         struct flist_head *entry;
 986         struct disk_util *slavedu;
 987         double util;
 988
 989         flist_for_each(entry, &masterdu->slaves) {
 990                 slavedu = flist_entry(entry, struct disk_util, slavelist);
 991                 dus = &slavedu->dus;
 992                 agg->ios[0] += dus->s.ios[0];
 993                 agg->ios[1] += dus->s.ios[1];
 994                 agg->merges[0] += dus->s.merges[0];
 995                 agg->merges[1] += dus->s.merges[1];
 996                 agg->sectors[0] += dus->s.sectors[0];
 997                 agg->sectors[1] += dus->s.sectors[1];
 998                 agg->ticks[0] += dus->s.ticks[0];
 999                 agg->ticks[1] += dus->s.ticks[1];
1000                 agg->time_in_queue += dus->s.time_in_queue;
1001                 agg->slavecount++;
1002
1003                 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
1004                 /* System utilization is the utilization of the
1005                  * component with the highest utilization.
1006                  */
1007                 if (util > agg->max_util.u.f)
1008                         agg->max_util.u.f = util;
1009
1010         }
1011
1012         if (agg->max_util.u.f > 100.0)
1013                 agg->max_util.u.f = 100.0;
1014 }
1015
1016 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
1017                      int terse, struct buf_output *out)
1018 {
1019         double util = 0;
1020
1021         if (dus->s.msec)
1022                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1023         if (util > 100.0)
1024                 util = 100.0;
1025
1026         if (!terse) {
1027                 if (agg->slavecount)
1028                         log_buf(out, "  ");
1029
1030                 log_buf(out, "  %s: ios=%llu/%llu, merge=%llu/%llu, "
1031                          "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
1032                                 dus->name,
1033                                 (unsigned long long) dus->s.ios[0],
1034                                 (unsigned long long) dus->s.ios[1],
1035                                 (unsigned long long) dus->s.merges[0],
1036                                 (unsigned long long) dus->s.merges[1],
1037                                 (unsigned long long) dus->s.ticks[0],
1038                                 (unsigned long long) dus->s.ticks[1],
1039                                 (unsigned long long) dus->s.time_in_queue,
1040                                 util);
1041         } else {
1042                 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1043                                 dus->name,
1044                                 (unsigned long long) dus->s.ios[0],
1045                                 (unsigned long long) dus->s.ios[1],
1046                                 (unsigned long long) dus->s.merges[0],
1047                                 (unsigned long long) dus->s.merges[1],
1048                                 (unsigned long long) dus->s.ticks[0],
1049                                 (unsigned long long) dus->s.ticks[1],
1050                                 (unsigned long long) dus->s.time_in_queue,
1051                                 util);
1052         }
1053
1054         /*
1055          * If the device has slaves, aggregate the stats for
1056          * those slave devices also.
1057          */
1058         show_agg_stats(agg, terse, out);
1059
1060         if (!terse)
1061                 log_buf(out, "\n");
1062 }
1063
1064 void json_array_add_disk_util(struct disk_util_stat *dus,
1065                 struct disk_util_agg *agg, struct json_array *array)
1066 {
1067         struct json_object *obj;
1068         double util = 0;
1069
1070         if (dus->s.msec)
1071                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1072         if (util > 100.0)
1073                 util = 100.0;
1074
1075         obj = json_create_object();
1076         json_array_add_value_object(array, obj);
1077
1078         json_object_add_value_string(obj, "name", (const char *)dus->name);
1079         json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1080         json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1081         json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1082         json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1083         json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1084         json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1085         json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1086         json_object_add_value_float(obj, "util", util);
1087
1088         /*
1089          * If the device has slaves, aggregate the stats for
1090          * those slave devices also.
1091          */
1092         if (!agg->slavecount)
1093                 return;
1094         json_object_add_value_int(obj, "aggr_read_ios",
1095                                 agg->ios[0] / agg->slavecount);
1096         json_object_add_value_int(obj, "aggr_write_ios",
1097                                 agg->ios[1] / agg->slavecount);
1098         json_object_add_value_int(obj, "aggr_read_merges",
1099                                 agg->merges[0] / agg->slavecount);
1100         json_object_add_value_int(obj, "aggr_write_merge",
1101                                 agg->merges[1] / agg->slavecount);
1102         json_object_add_value_int(obj, "aggr_read_ticks",
1103                                 agg->ticks[0] / agg->slavecount);
1104         json_object_add_value_int(obj, "aggr_write_ticks",
1105                                 agg->ticks[1] / agg->slavecount);
1106         json_object_add_value_int(obj, "aggr_in_queue",
1107                                 agg->time_in_queue / agg->slavecount);
1108         json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1109 }
1110
1111 static void json_object_add_disk_utils(struct json_object *obj,
1112                                        struct flist_head *head)
1113 {
1114         struct json_array *array = json_create_array();
1115         struct flist_head *entry;
1116         struct disk_util *du;
1117
1118         json_object_add_value_array(obj, "disk_util", array);
1119
1120         flist_for_each(entry, head) {
1121                 du = flist_entry(entry, struct disk_util, list);
1122
1123                 aggregate_slaves_stats(du);
1124                 json_array_add_disk_util(&du->dus, &du->agg, array);
1125         }
1126 }
1127
1128 void show_disk_util(int terse, struct json_object *parent,
1129                     struct buf_output *out)
1130 {
1131         struct flist_head *entry;
1132         struct disk_util *du;
1133         bool do_json;
1134
1135         if (!is_running_backend())
1136                 return;
1137
1138         if (flist_empty(&disk_list))
1139                 return;
1140
1141         if ((output_format & FIO_OUTPUT_JSON) && parent)
1142                 do_json = true;
1143         else
1144                 do_json = false;
1145
1146         if (!terse && !do_json)
1147                 log_buf(out, "\nDisk stats (read/write):\n");
1148
1149         if (do_json) {
1150                 json_object_add_disk_utils(parent, &disk_list);
1151         } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1152                 flist_for_each(entry, &disk_list) {
1153                         du = flist_entry(entry, struct disk_util, list);
1154
1155                         aggregate_slaves_stats(du);
1156                         print_disk_util(&du->dus, &du->agg, terse, out);
1157                 }
1158         }
1159 }
1160
1161 static void show_thread_status_normal(struct thread_stat *ts,
1162                                       struct group_run_stats *rs,
1163                                       struct buf_output *out)
1164 {
1165         double usr_cpu, sys_cpu;
1166         unsigned long runtime;
1167         double io_u_dist[FIO_IO_U_MAP_NR];
1168         time_t time_p;
1169         char time_buf[32];
1170
1171         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1172                 return;
1173
1174         memset(time_buf, 0, sizeof(time_buf));
1175
1176         time(&time_p);
1177         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1178
1179         if (!ts->error) {
1180                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1181                                         ts->name, ts->groupid, ts->members,
1182                                         ts->error, (int) ts->pid, time_buf);
1183         } else {
1184                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1185                                         ts->name, ts->groupid, ts->members,
1186                                         ts->error, ts->verror, (int) ts->pid,
1187                                         time_buf);
1188         }
1189
1190         if (strlen(ts->description))
1191                 log_buf(out, "  Description  : [%s]\n", ts->description);
1192
1193         for_each_rw_ddir(ddir) {
1194                 if (ts->io_bytes[ddir])
1195                         show_ddir_status(rs, ts, ddir, out);
1196         }
1197
1198         if (ts->unified_rw_rep == UNIFIED_BOTH)
1199                 show_mixed_ddir_status(rs, ts, out);
1200
1201         show_latencies(ts, out);
1202
1203         if (ts->sync_stat.samples)
1204                 show_ddir_status(rs, ts, DDIR_SYNC, out);
1205
1206         runtime = ts->total_run_time;
1207         if (runtime) {
1208                 double runt = (double) runtime;
1209
1210                 usr_cpu = (double) ts->usr_time * 100 / runt;
1211                 sys_cpu = (double) ts->sys_time * 100 / runt;
1212         } else {
1213                 usr_cpu = 0;
1214                 sys_cpu = 0;
1215         }
1216
1217         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1218                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1219                         (unsigned long long) ts->ctx,
1220                         (unsigned long long) ts->majf,
1221                         (unsigned long long) ts->minf);
1222
1223         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1224         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1225                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1226                                         io_u_dist[1], io_u_dist[2],
1227                                         io_u_dist[3], io_u_dist[4],
1228                                         io_u_dist[5], io_u_dist[6]);
1229
1230         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1231         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1232                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1233                                         io_u_dist[1], io_u_dist[2],
1234                                         io_u_dist[3], io_u_dist[4],
1235                                         io_u_dist[5], io_u_dist[6]);
1236         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1237         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1238                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1239                                         io_u_dist[1], io_u_dist[2],
1240                                         io_u_dist[3], io_u_dist[4],
1241                                         io_u_dist[5], io_u_dist[6]);
1242         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
1243                                  " short=%llu,%llu,%llu,0"
1244                                  " dropped=%llu,%llu,%llu,0\n",
1245                                         (unsigned long long) ts->total_io_u[0],
1246                                         (unsigned long long) ts->total_io_u[1],
1247                                         (unsigned long long) ts->total_io_u[2],
1248                                         (unsigned long long) ts->total_io_u[3],
1249                                         (unsigned long long) ts->short_io_u[0],
1250                                         (unsigned long long) ts->short_io_u[1],
1251                                         (unsigned long long) ts->short_io_u[2],
1252                                         (unsigned long long) ts->drop_io_u[0],
1253                                         (unsigned long long) ts->drop_io_u[1],
1254                                         (unsigned long long) ts->drop_io_u[2]);
1255         if (ts->continue_on_error) {
1256                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
1257                                         (unsigned long long)ts->total_err_count,
1258                                         ts->first_error,
1259                                         strerror(ts->first_error));
1260         }
1261         if (ts->latency_depth) {
1262                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1263                                         (unsigned long long)ts->latency_target,
1264                                         (unsigned long long)ts->latency_window,
1265                                         ts->latency_percentile.u.f,
1266                                         ts->latency_depth);
1267         }
1268
1269         if (ts->nr_block_infos)
1270                 show_block_infos(ts->nr_block_infos, ts->block_infos,
1271                                   ts->percentile_list, out);
1272
1273         if (ts->ss_dur)
1274                 show_ss_normal(ts, out);
1275 }
1276
1277 static void show_ddir_status_terse(struct thread_stat *ts,
1278                                    struct group_run_stats *rs,
1279                                    enum fio_ddir ddir, int ver,
1280                                    struct buf_output *out)
1281 {
1282         unsigned long long min, max, minv, maxv, bw, iops;
1283         unsigned long long *ovals = NULL;
1284         double mean, dev;
1285         unsigned int len;
1286         int i, bw_stat;
1287
1288         assert(ddir_rw(ddir));
1289
1290         iops = bw = 0;
1291         if (ts->runtime[ddir]) {
1292                 uint64_t runt = ts->runtime[ddir];
1293
1294                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1295                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1296         }
1297
1298         log_buf(out, ";%llu;%llu;%llu;%llu",
1299                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1300                                         (unsigned long long) ts->runtime[ddir]);
1301
1302         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1303                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1304         else
1305                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1306
1307         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1308                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1309         else
1310                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1311
1312         if (ts->lat_percentiles) {
1313                 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1314                                         ts->lat_stat[ddir].samples,
1315                                         ts->percentile_list, &ovals, &maxv,
1316                                         &minv);
1317         } else if (ts->clat_percentiles) {
1318                 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1319                                         ts->clat_stat[ddir].samples,
1320                                         ts->percentile_list, &ovals, &maxv,
1321                                         &minv);
1322         } else {
1323                 len = 0;
1324         }
1325
1326         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1327                 if (i >= len) {
1328                         log_buf(out, ";0%%=0");
1329                         continue;
1330                 }
1331                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1332         }
1333
1334         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1335                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1336         else
1337                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1338
1339         free(ovals);
1340
1341         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1342         if (bw_stat) {
1343                 double p_of_agg = 100.0;
1344
1345                 if (rs->agg[ddir]) {
1346                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1347                         if (p_of_agg > 100.0)
1348                                 p_of_agg = 100.0;
1349                 }
1350
1351                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1352         } else {
1353                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1354         }
1355
1356         if (ver == 5) {
1357                 if (bw_stat)
1358                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1359                 else
1360                         log_buf(out, ";%lu", 0UL);
1361
1362                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1363                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1364                                 mean, dev, (&ts->iops_stat[ddir])->samples);
1365                 else
1366                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1367         }
1368 }
1369
1370 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1371                                    struct group_run_stats *rs,
1372                                    int ver, struct buf_output *out)
1373 {
1374         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1375
1376         if (ts_lcl)
1377                 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1378
1379         free_clat_prio_stats(ts_lcl);
1380         free(ts_lcl);
1381 }
1382
1383 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1384                                              uint32_t percentiles,
1385                                              struct io_stat *lat_stat,
1386                                              uint64_t *io_u_plat)
1387 {
1388         char buf[120];
1389         double mean, dev;
1390         unsigned int i, len;
1391         struct json_object *lat_object, *percentile_object, *clat_bins_object;
1392         unsigned long long min, max, maxv, minv, *ovals = NULL;
1393
1394         if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1395                 min = max = 0;
1396                 mean = dev = 0.0;
1397         }
1398         lat_object = json_create_object();
1399         json_object_add_value_int(lat_object, "min", min);
1400         json_object_add_value_int(lat_object, "max", max);
1401         json_object_add_value_float(lat_object, "mean", mean);
1402         json_object_add_value_float(lat_object, "stddev", dev);
1403         json_object_add_value_int(lat_object, "N", lat_stat->samples);
1404
1405         if (percentiles && lat_stat->samples) {
1406                 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1407                                 ts->percentile_list, &ovals, &maxv, &minv);
1408
1409                 if (len > FIO_IO_U_LIST_MAX_LEN)
1410                         len = FIO_IO_U_LIST_MAX_LEN;
1411
1412                 percentile_object = json_create_object();
1413                 json_object_add_value_object(lat_object, "percentile", percentile_object);
1414                 for (i = 0; i < len; i++) {
1415                         snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1416                         json_object_add_value_int(percentile_object, buf, ovals[i]);
1417                 }
1418                 free(ovals);
1419
1420                 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1421                         clat_bins_object = json_create_object();
1422                         json_object_add_value_object(lat_object, "bins", clat_bins_object);
1423
1424                         for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1425                                 if (io_u_plat[i]) {
1426                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1427                                         json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1428                                 }
1429                 }
1430         }
1431
1432         return lat_object;
1433 }
1434
1435 static void add_ddir_status_json(struct thread_stat *ts,
1436                                  struct group_run_stats *rs, enum fio_ddir ddir,
1437                                  struct json_object *parent)
1438 {
1439         unsigned long long min, max;
1440         unsigned long long bw_bytes, bw;
1441         double mean, dev, iops;
1442         struct json_object *dir_object, *tmp_object;
1443         double p_of_agg = 100.0;
1444
1445         assert(ddir_rw(ddir) || ddir_sync(ddir));
1446
1447         if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1448                 return;
1449
1450         dir_object = json_create_object();
1451         json_object_add_value_object(parent,
1452                 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1453
1454         if (ddir_rw(ddir)) {
1455                 bw_bytes = 0;
1456                 bw = 0;
1457                 iops = 0.0;
1458                 if (ts->runtime[ddir]) {
1459                         uint64_t runt = ts->runtime[ddir];
1460
1461                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1462                         bw = bw_bytes / 1024; /* KiB/s */
1463                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1464                 }
1465
1466                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1467                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1468                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1469                 json_object_add_value_int(dir_object, "bw", bw);
1470                 json_object_add_value_float(dir_object, "iops", iops);
1471                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1472                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1473                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1474                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1475
1476                 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1477                                 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1478                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1479
1480                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1481                                 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1482                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1483
1484                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1485                                 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1486                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1487         } else {
1488                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1489                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1490                                 &ts->sync_stat, ts->io_u_sync_plat);
1491                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1492         }
1493
1494         if (!ddir_rw(ddir))
1495                 return;
1496
1497         /* Only include per prio stats if there are >= 2 prios with samples */
1498         if (get_nr_prios_with_samples(ts, ddir) >= 2) {
1499                 struct json_array *array = json_create_array();
1500                 const char *obj_name;
1501                 int i;
1502
1503                 if (ts->lat_percentiles)
1504                         obj_name = "lat_ns";
1505                 else
1506                         obj_name = "clat_ns";
1507
1508                 json_object_add_value_array(dir_object, "prios", array);
1509
1510                 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
1511                         if (ts->clat_prio[ddir][i].clat_stat.samples > 0) {
1512                                 struct json_object *obj = json_create_object();
1513                                 unsigned long long class, level;
1514
1515                                 class = ts->clat_prio[ddir][i].ioprio >> 13;
1516                                 json_object_add_value_int(obj, "prioclass", class);
1517                                 level = ts->clat_prio[ddir][i].ioprio & 7;
1518                                 json_object_add_value_int(obj, "prio", level);
1519
1520                                 tmp_object = add_ddir_lat_json(ts,
1521                                                                ts->clat_percentiles | ts->lat_percentiles,
1522                                                                &ts->clat_prio[ddir][i].clat_stat,
1523                                                                ts->clat_prio[ddir][i].io_u_plat);
1524                                 json_object_add_value_object(obj, obj_name, tmp_object);
1525                                 json_array_add_value_object(array, obj);
1526                         }
1527                 }
1528         }
1529
1530         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1531                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1532         } else {
1533                 min = max = 0;
1534                 p_of_agg = mean = dev = 0.0;
1535         }
1536
1537         json_object_add_value_int(dir_object, "bw_min", min);
1538         json_object_add_value_int(dir_object, "bw_max", max);
1539         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1540         json_object_add_value_float(dir_object, "bw_mean", mean);
1541         json_object_add_value_float(dir_object, "bw_dev", dev);
1542         json_object_add_value_int(dir_object, "bw_samples",
1543                                 (&ts->bw_stat[ddir])->samples);
1544
1545         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1546                 min = max = 0;
1547                 mean = dev = 0.0;
1548         }
1549         json_object_add_value_int(dir_object, "iops_min", min);
1550         json_object_add_value_int(dir_object, "iops_max", max);
1551         json_object_add_value_float(dir_object, "iops_mean", mean);
1552         json_object_add_value_float(dir_object, "iops_stddev", dev);
1553         json_object_add_value_int(dir_object, "iops_samples",
1554                                 (&ts->iops_stat[ddir])->samples);
1555
1556         if (ts->cachehit + ts->cachemiss) {
1557                 uint64_t total;
1558                 double hit;
1559
1560                 total = ts->cachehit + ts->cachemiss;
1561                 hit = (double) ts->cachehit / (double) total;
1562                 hit *= 100.0;
1563                 json_object_add_value_float(dir_object, "cachehit", hit);
1564         }
1565 }
1566
1567 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1568                 struct group_run_stats *rs, struct json_object *parent)
1569 {
1570         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1571
1572         /* add the aggregated stats to json parent */
1573         if (ts_lcl)
1574                 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1575
1576         free_clat_prio_stats(ts_lcl);
1577         free(ts_lcl);
1578 }
1579
1580 static void show_thread_status_terse_all(struct thread_stat *ts,
1581                                          struct group_run_stats *rs, int ver,
1582                                          struct buf_output *out)
1583 {
1584         double io_u_dist[FIO_IO_U_MAP_NR];
1585         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1586         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1587         double usr_cpu, sys_cpu;
1588         int i;
1589
1590         /* General Info */
1591         if (ver == 2)
1592                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1593         else
1594                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1595                         ts->name, ts->groupid, ts->error);
1596
1597         /* Log Read Status, or mixed if unified_rw_rep = 1 */
1598         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1599         if (ts->unified_rw_rep != UNIFIED_MIXED) {
1600                 /* Log Write Status */
1601                 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1602                 /* Log Trim Status */
1603                 if (ver == 2 || ver == 4 || ver == 5)
1604                         show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1605         }
1606         if (ts->unified_rw_rep == UNIFIED_BOTH)
1607                 show_mixed_ddir_status_terse(ts, rs, ver, out);
1608         /* CPU Usage */
1609         if (ts->total_run_time) {
1610                 double runt = (double) ts->total_run_time;
1611
1612                 usr_cpu = (double) ts->usr_time * 100 / runt;
1613                 sys_cpu = (double) ts->sys_time * 100 / runt;
1614         } else {
1615                 usr_cpu = 0;
1616                 sys_cpu = 0;
1617         }
1618
1619         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1620                                                 (unsigned long long) ts->ctx,
1621                                                 (unsigned long long) ts->majf,
1622                                                 (unsigned long long) ts->minf);
1623
1624         /* Calc % distribution of IO depths, usecond, msecond latency */
1625         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1626         stat_calc_lat_nu(ts, io_u_lat_u);
1627         stat_calc_lat_m(ts, io_u_lat_m);
1628
1629         /* Only show fixed 7 I/O depth levels*/
1630         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1631                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1632                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1633
1634         /* Microsecond latency */
1635         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1636                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1637         /* Millisecond latency */
1638         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1639                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1640
1641         /* disk util stats, if any */
1642         if (ver >= 3 && is_running_backend())
1643                 show_disk_util(1, NULL, out);
1644
1645         /* Additional output if continue_on_error set - default off*/
1646         if (ts->continue_on_error)
1647                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1648
1649         /* Additional output if description is set */
1650         if (strlen(ts->description)) {
1651                 if (ver == 2)
1652                         log_buf(out, "\n");
1653                 log_buf(out, ";%s", ts->description);
1654         }
1655
1656         log_buf(out, "\n");
1657 }
1658
1659 static void json_add_job_opts(struct json_object *root, const char *name,
1660                               struct flist_head *opt_list)
1661 {
1662         struct json_object *dir_object;
1663         struct flist_head *entry;
1664         struct print_option *p;
1665
1666         if (flist_empty(opt_list))
1667                 return;
1668
1669         dir_object = json_create_object();
1670         json_object_add_value_object(root, name, dir_object);
1671
1672         flist_for_each(entry, opt_list) {
1673                 p = flist_entry(entry, struct print_option, list);
1674                 json_object_add_value_string(dir_object, p->name, p->value);
1675         }
1676 }
1677
1678 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1679                                                    struct group_run_stats *rs,
1680                                                    struct flist_head *opt_list)
1681 {
1682         struct json_object *root, *tmp;
1683         struct jobs_eta *je;
1684         double io_u_dist[FIO_IO_U_MAP_NR];
1685         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1686         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1687         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1688         double usr_cpu, sys_cpu;
1689         int i;
1690         size_t size;
1691
1692         root = json_create_object();
1693         json_object_add_value_string(root, "jobname", ts->name);
1694         json_object_add_value_int(root, "groupid", ts->groupid);
1695         json_object_add_value_int(root, "error", ts->error);
1696
1697         /* ETA Info */
1698         je = get_jobs_eta(true, &size);
1699         if (je) {
1700                 json_object_add_value_int(root, "eta", je->eta_sec);
1701                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1702                 free(je);
1703         }
1704
1705         if (opt_list)
1706                 json_add_job_opts(root, "job options", opt_list);
1707
1708         add_ddir_status_json(ts, rs, DDIR_READ, root);
1709         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1710         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1711         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1712
1713         if (ts->unified_rw_rep == UNIFIED_BOTH)
1714                 add_mixed_ddir_status_json(ts, rs, root);
1715
1716         /* CPU Usage */
1717         if (ts->total_run_time) {
1718                 double runt = (double) ts->total_run_time;
1719
1720                 usr_cpu = (double) ts->usr_time * 100 / runt;
1721                 sys_cpu = (double) ts->sys_time * 100 / runt;
1722         } else {
1723                 usr_cpu = 0;
1724                 sys_cpu = 0;
1725         }
1726         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1727         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1728         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1729         json_object_add_value_int(root, "ctx", ts->ctx);
1730         json_object_add_value_int(root, "majf", ts->majf);
1731         json_object_add_value_int(root, "minf", ts->minf);
1732
1733         /* Calc % distribution of IO depths */
1734         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1735         tmp = json_create_object();
1736         json_object_add_value_object(root, "iodepth_level", tmp);
1737         /* Only show fixed 7 I/O depth levels*/
1738         for (i = 0; i < 7; i++) {
1739                 char name[20];
1740                 if (i < 6)
1741                         snprintf(name, 20, "%d", 1 << i);
1742                 else
1743                         snprintf(name, 20, ">=%d", 1 << i);
1744                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1745         }
1746
1747         /* Calc % distribution of submit IO depths */
1748         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1749         tmp = json_create_object();
1750         json_object_add_value_object(root, "iodepth_submit", tmp);
1751         /* Only show fixed 7 I/O depth levels*/
1752         for (i = 0; i < 7; i++) {
1753                 char name[20];
1754                 if (i == 0)
1755                         snprintf(name, 20, "0");
1756                 else if (i < 6)
1757                         snprintf(name, 20, "%d", 1 << (i+1));
1758                 else
1759                         snprintf(name, 20, ">=%d", 1 << i);
1760                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1761         }
1762
1763         /* Calc % distribution of completion IO depths */
1764         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1765         tmp = json_create_object();
1766         json_object_add_value_object(root, "iodepth_complete", tmp);
1767         /* Only show fixed 7 I/O depth levels*/
1768         for (i = 0; i < 7; i++) {
1769                 char name[20];
1770                 if (i == 0)
1771                         snprintf(name, 20, "0");
1772                 else if (i < 6)
1773                         snprintf(name, 20, "%d", 1 << (i+1));
1774                 else
1775                         snprintf(name, 20, ">=%d", 1 << i);
1776                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1777         }
1778
1779         /* Calc % distribution of nsecond, usecond, msecond latency */
1780         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1781         stat_calc_lat_n(ts, io_u_lat_n);
1782         stat_calc_lat_u(ts, io_u_lat_u);
1783         stat_calc_lat_m(ts, io_u_lat_m);
1784
1785         /* Nanosecond latency */
1786         tmp = json_create_object();
1787         json_object_add_value_object(root, "latency_ns", tmp);
1788         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1789                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1790                                  "250", "500", "750", "1000", };
1791                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1792         }
1793         /* Microsecond latency */
1794         tmp = json_create_object();
1795         json_object_add_value_object(root, "latency_us", tmp);
1796         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1797                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1798                                  "250", "500", "750", "1000", };
1799                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1800         }
1801         /* Millisecond latency */
1802         tmp = json_create_object();
1803         json_object_add_value_object(root, "latency_ms", tmp);
1804         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1805                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1806                                  "250", "500", "750", "1000", "2000",
1807                                  ">=2000", };
1808                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1809         }
1810
1811         /* Additional output if continue_on_error set - default off*/
1812         if (ts->continue_on_error) {
1813                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1814                 json_object_add_value_int(root, "first_error", ts->first_error);
1815         }
1816
1817         if (ts->latency_depth) {
1818                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1819                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1820                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1821                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1822         }
1823
1824         /* Additional output if description is set */
1825         if (strlen(ts->description))
1826                 json_object_add_value_string(root, "desc", ts->description);
1827
1828         if (ts->nr_block_infos) {
1829                 /* Block error histogram and types */
1830                 int len;
1831                 unsigned int *percentiles = NULL;
1832                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1833
1834                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1835                                              ts->percentile_list,
1836                                              &percentiles, block_state_counts);
1837
1838                 if (len) {
1839                         struct json_object *block, *percentile_object, *states;
1840                         int state;
1841                         block = json_create_object();
1842                         json_object_add_value_object(root, "block", block);
1843
1844                         percentile_object = json_create_object();
1845                         json_object_add_value_object(block, "percentiles",
1846                                                      percentile_object);
1847                         for (i = 0; i < len; i++) {
1848                                 char buf[20];
1849                                 snprintf(buf, sizeof(buf), "%f",
1850                                          ts->percentile_list[i].u.f);
1851                                 json_object_add_value_int(percentile_object,
1852                                                           buf,
1853                                                           percentiles[i]);
1854                         }
1855
1856                         states = json_create_object();
1857                         json_object_add_value_object(block, "states", states);
1858                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1859                                 json_object_add_value_int(states,
1860                                         block_state_names[state],
1861                                         block_state_counts[state]);
1862                         }
1863                         free(percentiles);
1864                 }
1865         }
1866
1867         if (ts->ss_dur) {
1868                 struct json_object *data;
1869                 struct json_array *iops, *bw;
1870                 int j, k, l;
1871                 char ss_buf[64];
1872
1873                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1874                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1875                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1876                         (float) ts->ss_limit.u.f,
1877                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1878
1879                 tmp = json_create_object();
1880                 json_object_add_value_object(root, "steadystate", tmp);
1881                 json_object_add_value_string(tmp, "ss", ss_buf);
1882                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1883                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1884
1885                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1886                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1887                 json_object_add_value_string(tmp, "criterion", ss_buf);
1888                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1889                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1890
1891                 data = json_create_object();
1892                 json_object_add_value_object(tmp, "data", data);
1893                 bw = json_create_array();
1894                 iops = json_create_array();
1895
1896                 /*
1897                 ** if ss was attained or the buffer is not full,
1898                 ** ss->head points to the first element in the list.
1899                 ** otherwise it actually points to the second element
1900                 ** in the list
1901                 */
1902                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1903                         j = ts->ss_head;
1904                 else
1905                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1906                 for (l = 0; l < ts->ss_dur; l++) {
1907                         k = (j + l) % ts->ss_dur;
1908                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1909                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1910                 }
1911                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1912                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1913                 json_object_add_value_array(data, "iops", iops);
1914                 json_object_add_value_array(data, "bw", bw);
1915         }
1916
1917         return root;
1918 }
1919
1920 static void show_thread_status_terse(struct thread_stat *ts,
1921                                      struct group_run_stats *rs,
1922                                      struct buf_output *out)
1923 {
1924         if (terse_version >= 2 && terse_version <= 5)
1925                 show_thread_status_terse_all(ts, rs, terse_version, out);
1926         else
1927                 log_err("fio: bad terse version!? %d\n", terse_version);
1928 }
1929
1930 struct json_object *show_thread_status(struct thread_stat *ts,
1931                                        struct group_run_stats *rs,
1932                                        struct flist_head *opt_list,
1933                                        struct buf_output *out)
1934 {
1935         struct json_object *ret = NULL;
1936
1937         if (output_format & FIO_OUTPUT_TERSE)
1938                 show_thread_status_terse(ts, rs,  out);
1939         if (output_format & FIO_OUTPUT_JSON)
1940                 ret = show_thread_status_json(ts, rs, opt_list);
1941         if (output_format & FIO_OUTPUT_NORMAL)
1942                 show_thread_status_normal(ts, rs,  out);
1943
1944         return ret;
1945 }
1946
1947 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1948 {
1949         double mean, S;
1950
1951         dst->min_val = min(dst->min_val, src->min_val);
1952         dst->max_val = max(dst->max_val, src->max_val);
1953
1954         /*
1955          * Compute new mean and S after the merge
1956          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1957          *  #Parallel_algorithm>
1958          */
1959         if (first) {
1960                 mean = src->mean.u.f;
1961                 S = src->S.u.f;
1962         } else {
1963                 double delta = src->mean.u.f - dst->mean.u.f;
1964
1965                 mean = ((src->mean.u.f * src->samples) +
1966                         (dst->mean.u.f * dst->samples)) /
1967                         (dst->samples + src->samples);
1968
1969                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1970                         (dst->samples * src->samples) /
1971                         (dst->samples + src->samples);
1972         }
1973
1974         dst->samples += src->samples;
1975         dst->mean.u.f = mean;
1976         dst->S.u.f = S;
1977
1978 }
1979
1980 /*
1981  * We sum two kinds of stats - one that is time based, in which case we
1982  * apply the proper summing technique, and then one that is iops/bw
1983  * numbers. For group_reporting, we should just add those up, not make
1984  * them the mean of everything.
1985  */
1986 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
1987 {
1988         bool first = dst->samples == 0;
1989
1990         if (src->samples == 0)
1991                 return;
1992
1993         if (!pure_sum) {
1994                 __sum_stat(dst, src, first);
1995                 return;
1996         }
1997
1998         if (first) {
1999                 dst->min_val = src->min_val;
2000                 dst->max_val = src->max_val;
2001                 dst->samples = src->samples;
2002                 dst->mean.u.f = src->mean.u.f;
2003                 dst->S.u.f = src->S.u.f;
2004         } else {
2005                 dst->min_val += src->min_val;
2006                 dst->max_val += src->max_val;
2007                 dst->samples += src->samples;
2008                 dst->mean.u.f += src->mean.u.f;
2009                 dst->S.u.f += src->S.u.f;
2010         }
2011 }
2012
2013 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
2014 {
2015         int i;
2016
2017         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2018                 if (dst->max_run[i] < src->max_run[i])
2019                         dst->max_run[i] = src->max_run[i];
2020                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
2021                         dst->min_run[i] = src->min_run[i];
2022                 if (dst->max_bw[i] < src->max_bw[i])
2023                         dst->max_bw[i] = src->max_bw[i];
2024                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
2025                         dst->min_bw[i] = src->min_bw[i];
2026
2027                 dst->iobytes[i] += src->iobytes[i];
2028                 dst->agg[i] += src->agg[i];
2029         }
2030
2031         if (!dst->kb_base)
2032                 dst->kb_base = src->kb_base;
2033         if (!dst->unit_base)
2034                 dst->unit_base = src->unit_base;
2035         if (!dst->sig_figs)
2036                 dst->sig_figs = src->sig_figs;
2037 }
2038
2039 /*
2040  * Free the clat_prio_stat arrays allocated by alloc_clat_prio_stat_ddir().
2041  */
2042 void free_clat_prio_stats(struct thread_stat *ts)
2043 {
2044         enum fio_ddir ddir;
2045
2046         if (!ts)
2047                 return;
2048
2049         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2050                 sfree(ts->clat_prio[ddir]);
2051                 ts->clat_prio[ddir] = NULL;
2052                 ts->nr_clat_prio[ddir] = 0;
2053         }
2054 }
2055
2056 /*
2057  * Allocate a clat_prio_stat array. The array has to be allocated/freed using
2058  * smalloc/sfree, so that it is accessible by the process/thread summing the
2059  * thread_stats.
2060  */
2061 int alloc_clat_prio_stat_ddir(struct thread_stat *ts, enum fio_ddir ddir,
2062                               int nr_prios)
2063 {
2064         struct clat_prio_stat *clat_prio;
2065         int i;
2066
2067         clat_prio = scalloc(nr_prios, sizeof(*ts->clat_prio[ddir]));
2068         if (!clat_prio) {
2069                 log_err("fio: failed to allocate ts clat data\n");
2070                 return 1;
2071         }
2072
2073         for (i = 0; i < nr_prios; i++)
2074                 clat_prio[i].clat_stat.min_val = ULONG_MAX;
2075
2076         ts->clat_prio[ddir] = clat_prio;
2077         ts->nr_clat_prio[ddir] = nr_prios;
2078
2079         return 0;
2080 }
2081
2082 static int grow_clat_prio_stat(struct thread_stat *dst, enum fio_ddir ddir)
2083 {
2084         int curr_len = dst->nr_clat_prio[ddir];
2085         void *new_arr;
2086
2087         new_arr = scalloc(curr_len + 1, sizeof(*dst->clat_prio[ddir]));
2088         if (!new_arr) {
2089                 log_err("fio: failed to grow clat prio array\n");
2090                 return 1;
2091         }
2092
2093         memcpy(new_arr, dst->clat_prio[ddir],
2094                curr_len * sizeof(*dst->clat_prio[ddir]));
2095         sfree(dst->clat_prio[ddir]);
2096
2097         dst->clat_prio[ddir] = new_arr;
2098         dst->clat_prio[ddir][curr_len].clat_stat.min_val = ULONG_MAX;
2099         dst->nr_clat_prio[ddir]++;
2100
2101         return 0;
2102 }
2103
2104 static int find_clat_prio_index(struct thread_stat *dst, enum fio_ddir ddir,
2105                                 uint32_t ioprio)
2106 {
2107         int i, nr_prios = dst->nr_clat_prio[ddir];
2108
2109         for (i = 0; i < nr_prios; i++) {
2110                 if (dst->clat_prio[ddir][i].ioprio == ioprio)
2111                         return i;
2112         }
2113
2114         return -1;
2115 }
2116
2117 static int alloc_or_get_clat_prio_index(struct thread_stat *dst,
2118                                         enum fio_ddir ddir, uint32_t ioprio,
2119                                         int *idx)
2120 {
2121         int index = find_clat_prio_index(dst, ddir, ioprio);
2122
2123         if (index == -1) {
2124                 index = dst->nr_clat_prio[ddir];
2125
2126                 if (grow_clat_prio_stat(dst, ddir))
2127                         return 1;
2128
2129                 dst->clat_prio[ddir][index].ioprio = ioprio;
2130         }
2131
2132         *idx = index;
2133
2134         return 0;
2135 }
2136
2137 static int clat_prio_stats_copy(struct thread_stat *dst, struct thread_stat *src,
2138                                 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2139 {
2140         size_t sz = sizeof(*src->clat_prio[src_ddir]) *
2141                 src->nr_clat_prio[src_ddir];
2142
2143         dst->clat_prio[dst_ddir] = smalloc(sz);
2144         if (!dst->clat_prio[dst_ddir]) {
2145                 log_err("fio: failed to alloc clat prio array\n");
2146                 return 1;
2147         }
2148
2149         memcpy(dst->clat_prio[dst_ddir], src->clat_prio[src_ddir], sz);
2150         dst->nr_clat_prio[dst_ddir] = src->nr_clat_prio[src_ddir];
2151
2152         return 0;
2153 }
2154
2155 static int clat_prio_stat_add_samples(struct thread_stat *dst,
2156                                       enum fio_ddir dst_ddir, uint32_t ioprio,
2157                                       struct io_stat *io_stat,
2158                                       uint64_t *io_u_plat)
2159 {
2160         int i, dst_index;
2161
2162         if (!io_stat->samples)
2163                 return 0;
2164
2165         if (alloc_or_get_clat_prio_index(dst, dst_ddir, ioprio, &dst_index))
2166                 return 1;
2167
2168         sum_stat(&dst->clat_prio[dst_ddir][dst_index].clat_stat, io_stat,
2169                  false);
2170
2171         for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
2172                 dst->clat_prio[dst_ddir][dst_index].io_u_plat[i] += io_u_plat[i];
2173
2174         return 0;
2175 }
2176
2177 static int sum_clat_prio_stats_src_single_prio(struct thread_stat *dst,
2178                                                struct thread_stat *src,
2179                                                enum fio_ddir dst_ddir,
2180                                                enum fio_ddir src_ddir)
2181 {
2182         struct io_stat *io_stat;
2183         uint64_t *io_u_plat;
2184
2185         /*
2186          * If src ts has no clat_prio_stat array, then all I/Os were submitted
2187          * using src->ioprio. Thus, the global samples in src->clat_stat (or
2188          * src->lat_stat) can be used as the 'per prio' samples for src->ioprio.
2189          */
2190         assert(!src->clat_prio[src_ddir]);
2191         assert(src->nr_clat_prio[src_ddir] == 0);
2192
2193         if (src->lat_percentiles) {
2194                 io_u_plat = src->io_u_plat[FIO_LAT][src_ddir];
2195                 io_stat = &src->lat_stat[src_ddir];
2196         } else {
2197                 io_u_plat = src->io_u_plat[FIO_CLAT][src_ddir];
2198                 io_stat = &src->clat_stat[src_ddir];
2199         }
2200
2201         return clat_prio_stat_add_samples(dst, dst_ddir, src->ioprio, io_stat,
2202                                           io_u_plat);
2203 }
2204
2205 static int sum_clat_prio_stats_src_multi_prio(struct thread_stat *dst,
2206                                               struct thread_stat *src,
2207                                               enum fio_ddir dst_ddir,
2208                                               enum fio_ddir src_ddir)
2209 {
2210         int i;
2211
2212         /*
2213          * If src ts has a clat_prio_stat array, then there are multiple prios
2214          * in use (i.e. src ts had cmdprio_percentage or cmdprio_bssplit set).
2215          * The samples for the default prio will exist in the src->clat_prio
2216          * array, just like the samples for any other prio.
2217          */
2218         assert(src->clat_prio[src_ddir]);
2219         assert(src->nr_clat_prio[src_ddir]);
2220
2221         /* If the dst ts doesn't yet have a clat_prio array, simply memcpy. */
2222         if (!dst->clat_prio[dst_ddir])
2223                 return clat_prio_stats_copy(dst, src, dst_ddir, src_ddir);
2224
2225         /* The dst ts already has a clat_prio_array, add src stats into it. */
2226         for (i = 0; i < src->nr_clat_prio[src_ddir]; i++) {
2227                 struct io_stat *io_stat = &src->clat_prio[src_ddir][i].clat_stat;
2228                 uint64_t *io_u_plat = src->clat_prio[src_ddir][i].io_u_plat;
2229                 uint32_t ioprio = src->clat_prio[src_ddir][i].ioprio;
2230
2231                 if (clat_prio_stat_add_samples(dst, dst_ddir, ioprio, io_stat, io_u_plat))
2232                         return 1;
2233         }
2234
2235         return 0;
2236 }
2237
2238 static int sum_clat_prio_stats(struct thread_stat *dst, struct thread_stat *src,
2239                                enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2240 {
2241         if (dst->disable_prio_stat)
2242                 return 0;
2243
2244         if (!src->clat_prio[src_ddir])
2245                 return sum_clat_prio_stats_src_single_prio(dst, src, dst_ddir,
2246                                                            src_ddir);
2247
2248         return sum_clat_prio_stats_src_multi_prio(dst, src, dst_ddir, src_ddir);
2249 }
2250
2251 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2252 {
2253         int k, l, m;
2254
2255         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2256                 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2257                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2258                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2259                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2260                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2261                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2262                         sum_clat_prio_stats(dst, src, l, l);
2263
2264                         dst->io_bytes[l] += src->io_bytes[l];
2265
2266                         if (dst->runtime[l] < src->runtime[l])
2267                                 dst->runtime[l] = src->runtime[l];
2268                 } else {
2269                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2270                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2271                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2272                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2273                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2274                         sum_clat_prio_stats(dst, src, 0, l);
2275
2276                         dst->io_bytes[0] += src->io_bytes[l];
2277
2278                         if (dst->runtime[0] < src->runtime[l])
2279                                 dst->runtime[0] = src->runtime[l];
2280                 }
2281         }
2282
2283         sum_stat(&dst->sync_stat, &src->sync_stat, false);
2284         dst->usr_time += src->usr_time;
2285         dst->sys_time += src->sys_time;
2286         dst->ctx += src->ctx;
2287         dst->majf += src->majf;
2288         dst->minf += src->minf;
2289
2290         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2291                 dst->io_u_map[k] += src->io_u_map[k];
2292                 dst->io_u_submit[k] += src->io_u_submit[k];
2293                 dst->io_u_complete[k] += src->io_u_complete[k];
2294         }
2295
2296         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2297                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2298         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2299                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2300         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2301                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2302
2303         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2304                 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2305                         dst->total_io_u[k] += src->total_io_u[k];
2306                         dst->short_io_u[k] += src->short_io_u[k];
2307                         dst->drop_io_u[k] += src->drop_io_u[k];
2308                 } else {
2309                         dst->total_io_u[0] += src->total_io_u[k];
2310                         dst->short_io_u[0] += src->short_io_u[k];
2311                         dst->drop_io_u[0] += src->drop_io_u[k];
2312                 }
2313         }
2314
2315         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2316
2317         for (k = 0; k < FIO_LAT_CNT; k++)
2318                 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2319                         for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2320                                 if (dst->unified_rw_rep != UNIFIED_MIXED)
2321                                         dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2322                                 else
2323                                         dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2324
2325         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2326                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2327
2328         dst->total_run_time += src->total_run_time;
2329         dst->total_submit += src->total_submit;
2330         dst->total_complete += src->total_complete;
2331         dst->nr_zone_resets += src->nr_zone_resets;
2332         dst->cachehit += src->cachehit;
2333         dst->cachemiss += src->cachemiss;
2334 }
2335
2336 void init_group_run_stat(struct group_run_stats *gs)
2337 {
2338         int i;
2339         memset(gs, 0, sizeof(*gs));
2340
2341         for (i = 0; i < DDIR_RWDIR_CNT; i++)
2342                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2343 }
2344
2345 void init_thread_stat_min_vals(struct thread_stat *ts)
2346 {
2347         int i;
2348
2349         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2350                 ts->clat_stat[i].min_val = ULONG_MAX;
2351                 ts->slat_stat[i].min_val = ULONG_MAX;
2352                 ts->lat_stat[i].min_val = ULONG_MAX;
2353                 ts->bw_stat[i].min_val = ULONG_MAX;
2354                 ts->iops_stat[i].min_val = ULONG_MAX;
2355         }
2356         ts->sync_stat.min_val = ULONG_MAX;
2357 }
2358
2359 void init_thread_stat(struct thread_stat *ts)
2360 {
2361         memset(ts, 0, sizeof(*ts));
2362
2363         init_thread_stat_min_vals(ts);
2364         ts->groupid = -1;
2365 }
2366
2367 static void init_per_prio_stats(struct thread_stat *threadstats, int nr_ts)
2368 {
2369         struct thread_data *td;
2370         struct thread_stat *ts;
2371         int i, j, last_ts, idx;
2372         enum fio_ddir ddir;
2373
2374         j = 0;
2375         last_ts = -1;
2376         idx = 0;
2377
2378         /*
2379          * Loop through all tds, if a td requires per prio stats, temporarily
2380          * store a 1 in ts->disable_prio_stat, and then do an additional
2381          * loop at the end where we invert the ts->disable_prio_stat values.
2382          */
2383         for_each_td(td, i) {
2384                 if (!td->o.stats)
2385                         continue;
2386                 if (idx &&
2387                     (!td->o.group_reporting ||
2388                      (td->o.group_reporting && last_ts != td->groupid))) {
2389                         idx = 0;
2390                         j++;
2391                 }
2392
2393                 last_ts = td->groupid;
2394                 ts = &threadstats[j];
2395
2396                 /* idx == 0 means first td in group, or td is not in a group. */
2397                 if (idx == 0)
2398                         ts->ioprio = td->ioprio;
2399                 else if (td->ioprio != ts->ioprio)
2400                         ts->disable_prio_stat = 1;
2401
2402                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2403                         if (td->ts.clat_prio[ddir]) {
2404                                 ts->disable_prio_stat = 1;
2405                                 break;
2406                         }
2407                 }
2408
2409                 idx++;
2410         }
2411
2412         /* Loop through all dst threadstats and fixup the values. */
2413         for (i = 0; i < nr_ts; i++) {
2414                 ts = &threadstats[i];
2415                 ts->disable_prio_stat = !ts->disable_prio_stat;
2416         }
2417 }
2418
2419 void __show_run_stats(void)
2420 {
2421         struct group_run_stats *runstats, *rs;
2422         struct thread_data *td;
2423         struct thread_stat *threadstats, *ts;
2424         int i, j, k, nr_ts, last_ts, idx;
2425         bool kb_base_warned = false;
2426         bool unit_base_warned = false;
2427         struct json_object *root = NULL;
2428         struct json_array *array = NULL;
2429         struct buf_output output[FIO_OUTPUT_NR];
2430         struct flist_head **opt_lists;
2431
2432         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2433
2434         for (i = 0; i < groupid + 1; i++)
2435                 init_group_run_stat(&runstats[i]);
2436
2437         /*
2438          * find out how many threads stats we need. if group reporting isn't
2439          * enabled, it's one-per-td.
2440          */
2441         nr_ts = 0;
2442         last_ts = -1;
2443         for_each_td(td, i) {
2444                 if (!td->o.group_reporting) {
2445                         nr_ts++;
2446                         continue;
2447                 }
2448                 if (last_ts == td->groupid)
2449                         continue;
2450                 if (!td->o.stats)
2451                         continue;
2452
2453                 last_ts = td->groupid;
2454                 nr_ts++;
2455         }
2456
2457         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2458         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2459
2460         for (i = 0; i < nr_ts; i++) {
2461                 init_thread_stat(&threadstats[i]);
2462                 opt_lists[i] = NULL;
2463         }
2464
2465         init_per_prio_stats(threadstats, nr_ts);
2466
2467         j = 0;
2468         last_ts = -1;
2469         idx = 0;
2470         for_each_td(td, i) {
2471                 if (!td->o.stats)
2472                         continue;
2473                 if (idx && (!td->o.group_reporting ||
2474                     (td->o.group_reporting && last_ts != td->groupid))) {
2475                         idx = 0;
2476                         j++;
2477                 }
2478
2479                 last_ts = td->groupid;
2480
2481                 ts = &threadstats[j];
2482
2483                 ts->clat_percentiles = td->o.clat_percentiles;
2484                 ts->lat_percentiles = td->o.lat_percentiles;
2485                 ts->slat_percentiles = td->o.slat_percentiles;
2486                 ts->percentile_precision = td->o.percentile_precision;
2487                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2488                 opt_lists[j] = &td->opt_list;
2489
2490                 idx++;
2491
2492                 if (ts->groupid == -1) {
2493                         /*
2494                          * These are per-group shared already
2495                          */
2496                         snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2497                         if (td->o.description)
2498                                 snprintf(ts->description,
2499                                          sizeof(ts->description), "%s",
2500                                          td->o.description);
2501                         else
2502                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2503
2504                         /*
2505                          * If multiple entries in this group, this is
2506                          * the first member.
2507                          */
2508                         ts->thread_number = td->thread_number;
2509                         ts->groupid = td->groupid;
2510
2511                         /*
2512                          * first pid in group, not very useful...
2513                          */
2514                         ts->pid = td->pid;
2515
2516                         ts->kb_base = td->o.kb_base;
2517                         ts->unit_base = td->o.unit_base;
2518                         ts->sig_figs = td->o.sig_figs;
2519                         ts->unified_rw_rep = td->o.unified_rw_rep;
2520                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2521                         log_info("fio: kb_base differs for jobs in group, using"
2522                                  " %u as the base\n", ts->kb_base);
2523                         kb_base_warned = true;
2524                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2525                         log_info("fio: unit_base differs for jobs in group, using"
2526                                  " %u as the base\n", ts->unit_base);
2527                         unit_base_warned = true;
2528                 }
2529
2530                 ts->continue_on_error = td->o.continue_on_error;
2531                 ts->total_err_count += td->total_err_count;
2532                 ts->first_error = td->first_error;
2533                 if (!ts->error) {
2534                         if (!td->error && td->o.continue_on_error &&
2535                             td->first_error) {
2536                                 ts->error = td->first_error;
2537                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2538                                          td->verror);
2539                         } else  if (td->error) {
2540                                 ts->error = td->error;
2541                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2542                                          td->verror);
2543                         }
2544                 }
2545
2546                 ts->latency_depth = td->latency_qd;
2547                 ts->latency_target = td->o.latency_target;
2548                 ts->latency_percentile = td->o.latency_percentile;
2549                 ts->latency_window = td->o.latency_window;
2550
2551                 ts->nr_block_infos = td->ts.nr_block_infos;
2552                 for (k = 0; k < ts->nr_block_infos; k++)
2553                         ts->block_infos[k] = td->ts.block_infos[k];
2554
2555                 sum_thread_stats(ts, &td->ts);
2556
2557                 ts->members++;
2558
2559                 if (td->o.ss_dur) {
2560                         ts->ss_state = td->ss.state;
2561                         ts->ss_dur = td->ss.dur;
2562                         ts->ss_head = td->ss.head;
2563                         ts->ss_bw_data = td->ss.bw_data;
2564                         ts->ss_iops_data = td->ss.iops_data;
2565                         ts->ss_limit.u.f = td->ss.limit;
2566                         ts->ss_slope.u.f = td->ss.slope;
2567                         ts->ss_deviation.u.f = td->ss.deviation;
2568                         ts->ss_criterion.u.f = td->ss.criterion;
2569                 }
2570                 else
2571                         ts->ss_dur = ts->ss_state = 0;
2572         }
2573
2574         for (i = 0; i < nr_ts; i++) {
2575                 unsigned long long bw;
2576
2577                 ts = &threadstats[i];
2578                 if (ts->groupid == -1)
2579                         continue;
2580                 rs = &runstats[ts->groupid];
2581                 rs->kb_base = ts->kb_base;
2582                 rs->unit_base = ts->unit_base;
2583                 rs->sig_figs = ts->sig_figs;
2584                 rs->unified_rw_rep |= ts->unified_rw_rep;
2585
2586                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2587                         if (!ts->runtime[j])
2588                                 continue;
2589                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2590                                 rs->min_run[j] = ts->runtime[j];
2591                         if (ts->runtime[j] > rs->max_run[j])
2592                                 rs->max_run[j] = ts->runtime[j];
2593
2594                         bw = 0;
2595                         if (ts->runtime[j])
2596                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2597                         if (bw < rs->min_bw[j])
2598                                 rs->min_bw[j] = bw;
2599                         if (bw > rs->max_bw[j])
2600                                 rs->max_bw[j] = bw;
2601
2602                         rs->iobytes[j] += ts->io_bytes[j];
2603                 }
2604         }
2605
2606         for (i = 0; i < groupid + 1; i++) {
2607                 enum fio_ddir ddir;
2608
2609                 rs = &runstats[i];
2610
2611                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2612                         if (rs->max_run[ddir])
2613                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2614                                                 rs->max_run[ddir];
2615                 }
2616         }
2617
2618         for (i = 0; i < FIO_OUTPUT_NR; i++)
2619                 buf_output_init(&output[i]);
2620
2621         /*
2622          * don't overwrite last signal output
2623          */
2624         if (output_format & FIO_OUTPUT_NORMAL)
2625                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2626         if (output_format & FIO_OUTPUT_JSON) {
2627                 struct thread_data *global;
2628                 char time_buf[32];
2629                 struct timeval now;
2630                 unsigned long long ms_since_epoch;
2631                 time_t tv_sec;
2632
2633                 gettimeofday(&now, NULL);
2634                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2635                                  (unsigned long long)(now.tv_usec) / 1000;
2636
2637                 tv_sec = now.tv_sec;
2638                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2639                 if (time_buf[strlen(time_buf) - 1] == '\n')
2640                         time_buf[strlen(time_buf) - 1] = '\0';
2641
2642                 root = json_create_object();
2643                 json_object_add_value_string(root, "fio version", fio_version_string);
2644                 json_object_add_value_int(root, "timestamp", now.tv_sec);
2645                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2646                 json_object_add_value_string(root, "time", time_buf);
2647                 global = get_global_options();
2648                 json_add_job_opts(root, "global options", &global->opt_list);
2649                 array = json_create_array();
2650                 json_object_add_value_array(root, "jobs", array);
2651         }
2652
2653         if (is_backend)
2654                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2655
2656         for (i = 0; i < nr_ts; i++) {
2657                 ts = &threadstats[i];
2658                 rs = &runstats[ts->groupid];
2659
2660                 if (is_backend) {
2661                         fio_server_send_job_options(opt_lists[i], i);
2662                         fio_server_send_ts(ts, rs);
2663                 } else {
2664                         if (output_format & FIO_OUTPUT_TERSE)
2665                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2666                         if (output_format & FIO_OUTPUT_JSON) {
2667                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2668                                 json_array_add_value_object(array, tmp);
2669                         }
2670                         if (output_format & FIO_OUTPUT_NORMAL)
2671                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2672                 }
2673         }
2674         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2675                 /* disk util stats, if any */
2676                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2677
2678                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2679
2680                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2681                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2682                 json_free_object(root);
2683         }
2684
2685         for (i = 0; i < groupid + 1; i++) {
2686                 rs = &runstats[i];
2687
2688                 rs->groupid = i;
2689                 if (is_backend)
2690                         fio_server_send_gs(rs);
2691                 else if (output_format & FIO_OUTPUT_NORMAL)
2692                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2693         }
2694
2695         if (is_backend)
2696                 fio_server_send_du();
2697         else if (output_format & FIO_OUTPUT_NORMAL) {
2698                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2699                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2700         }
2701
2702         for (i = 0; i < FIO_OUTPUT_NR; i++) {
2703                 struct buf_output *out = &output[i];
2704
2705                 log_info_buf(out->buf, out->buflen);
2706                 buf_output_free(out);
2707         }
2708
2709         fio_idle_prof_cleanup();
2710
2711         log_info_flush();
2712         free(runstats);
2713
2714         /* free arrays allocated by sum_thread_stats(), if any */
2715         for (i = 0; i < nr_ts; i++) {
2716                 ts = &threadstats[i];
2717                 free_clat_prio_stats(ts);
2718         }
2719         free(threadstats);
2720         free(opt_lists);
2721 }
2722
2723 int __show_running_run_stats(void)
2724 {
2725         struct thread_data *td;
2726         unsigned long long *rt;
2727         struct timespec ts;
2728         int i;
2729
2730         fio_sem_down(stat_sem);
2731
2732         rt = malloc(thread_number * sizeof(unsigned long long));
2733         fio_gettime(&ts, NULL);
2734
2735         for_each_td(td, i) {
2736                 if (td->runstate >= TD_EXITED)
2737                         continue;
2738
2739                 td->update_rusage = 1;
2740                 for_each_rw_ddir(ddir) {
2741                         td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2742                 }
2743                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2744
2745                 rt[i] = mtime_since(&td->start, &ts);
2746                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2747                         td->ts.runtime[DDIR_READ] += rt[i];
2748                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2749                         td->ts.runtime[DDIR_WRITE] += rt[i];
2750                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2751                         td->ts.runtime[DDIR_TRIM] += rt[i];
2752         }
2753
2754         for_each_td(td, i) {
2755                 if (td->runstate >= TD_EXITED)
2756                         continue;
2757                 if (td->rusage_sem) {
2758                         td->update_rusage = 1;
2759                         fio_sem_down(td->rusage_sem);
2760                 }
2761                 td->update_rusage = 0;
2762         }
2763
2764         __show_run_stats();
2765
2766         for_each_td(td, i) {
2767                 if (td->runstate >= TD_EXITED)
2768                         continue;
2769
2770                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2771                         td->ts.runtime[DDIR_READ] -= rt[i];
2772                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2773                         td->ts.runtime[DDIR_WRITE] -= rt[i];
2774                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2775                         td->ts.runtime[DDIR_TRIM] -= rt[i];
2776         }
2777
2778         free(rt);
2779         fio_sem_up(stat_sem);
2780
2781         return 0;
2782 }
2783
2784 static bool status_file_disabled;
2785
2786 #define FIO_STATUS_FILE         "fio-dump-status"
2787
2788 static int check_status_file(void)
2789 {
2790         struct stat sb;
2791         const char *temp_dir;
2792         char fio_status_file_path[PATH_MAX];
2793
2794         if (status_file_disabled)
2795                 return 0;
2796
2797         temp_dir = getenv("TMPDIR");
2798         if (temp_dir == NULL) {
2799                 temp_dir = getenv("TEMP");
2800                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2801                         temp_dir = NULL;
2802         }
2803         if (temp_dir == NULL)
2804                 temp_dir = "/tmp";
2805 #ifdef __COVERITY__
2806         __coverity_tainted_data_sanitize__(temp_dir);
2807 #endif
2808
2809         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2810
2811         if (stat(fio_status_file_path, &sb))
2812                 return 0;
2813
2814         if (unlink(fio_status_file_path) < 0) {
2815                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2816                                                         strerror(errno));
2817                 log_err("fio: disabling status file updates\n");
2818                 status_file_disabled = true;
2819         }
2820
2821         return 1;
2822 }
2823
2824 void check_for_running_stats(void)
2825 {
2826         if (check_status_file()) {
2827                 show_running_run_stats();
2828                 return;
2829         }
2830 }
2831
2832 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2833 {
2834         double val = data;
2835         double delta;
2836
2837         if (data > is->max_val)
2838                 is->max_val = data;
2839         if (data < is->min_val)
2840                 is->min_val = data;
2841
2842         delta = val - is->mean.u.f;
2843         if (delta) {
2844                 is->mean.u.f += delta / (is->samples + 1.0);
2845                 is->S.u.f += delta * (val - is->mean.u.f);
2846         }
2847
2848         is->samples++;
2849 }
2850
2851 static inline void add_stat_prio_sample(struct clat_prio_stat *clat_prio,
2852                                         unsigned short clat_prio_index,
2853                                         unsigned long long nsec)
2854 {
2855         if (clat_prio)
2856                 add_stat_sample(&clat_prio[clat_prio_index].clat_stat, nsec);
2857 }
2858
2859 /*
2860  * Return a struct io_logs, which is added to the tail of the log
2861  * list for 'iolog'.
2862  */
2863 static struct io_logs *get_new_log(struct io_log *iolog)
2864 {
2865         size_t new_samples;
2866         struct io_logs *cur_log;
2867
2868         /*
2869          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2870          * forever
2871          */
2872         if (!iolog->cur_log_max) {
2873                 if (iolog->td)
2874                         new_samples = iolog->td->o.log_entries;
2875                 else
2876                         new_samples = DEF_LOG_ENTRIES;
2877         } else {
2878                 new_samples = iolog->cur_log_max * 2;
2879                 if (new_samples > MAX_LOG_ENTRIES)
2880                         new_samples = MAX_LOG_ENTRIES;
2881         }
2882
2883         cur_log = smalloc(sizeof(*cur_log));
2884         if (cur_log) {
2885                 INIT_FLIST_HEAD(&cur_log->list);
2886                 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2887                 if (cur_log->log) {
2888                         cur_log->nr_samples = 0;
2889                         cur_log->max_samples = new_samples;
2890                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2891                         iolog->cur_log_max = new_samples;
2892                         return cur_log;
2893                 }
2894                 sfree(cur_log);
2895         }
2896
2897         return NULL;
2898 }
2899
2900 /*
2901  * Add and return a new log chunk, or return current log if big enough
2902  */
2903 static struct io_logs *regrow_log(struct io_log *iolog)
2904 {
2905         struct io_logs *cur_log;
2906         int i;
2907
2908         if (!iolog || iolog->disabled)
2909                 goto disable;
2910
2911         cur_log = iolog_cur_log(iolog);
2912         if (!cur_log) {
2913                 cur_log = get_new_log(iolog);
2914                 if (!cur_log)
2915                         return NULL;
2916         }
2917
2918         if (cur_log->nr_samples < cur_log->max_samples)
2919                 return cur_log;
2920
2921         /*
2922          * No room for a new sample. If we're compressing on the fly, flush
2923          * out the current chunk
2924          */
2925         if (iolog->log_gz) {
2926                 if (iolog_cur_flush(iolog, cur_log)) {
2927                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2928                         return NULL;
2929                 }
2930         }
2931
2932         /*
2933          * Get a new log array, and add to our list
2934          */
2935         cur_log = get_new_log(iolog);
2936         if (!cur_log) {
2937                 log_err("fio: failed extending iolog! Will stop logging.\n");
2938                 return NULL;
2939         }
2940
2941         if (!iolog->pending || !iolog->pending->nr_samples)
2942                 return cur_log;
2943
2944         /*
2945          * Flush pending items to new log
2946          */
2947         for (i = 0; i < iolog->pending->nr_samples; i++) {
2948                 struct io_sample *src, *dst;
2949
2950                 src = get_sample(iolog, iolog->pending, i);
2951                 dst = get_sample(iolog, cur_log, i);
2952                 memcpy(dst, src, log_entry_sz(iolog));
2953         }
2954         cur_log->nr_samples = iolog->pending->nr_samples;
2955
2956         iolog->pending->nr_samples = 0;
2957         return cur_log;
2958 disable:
2959         if (iolog)
2960                 iolog->disabled = true;
2961         return NULL;
2962 }
2963
2964 void regrow_logs(struct thread_data *td)
2965 {
2966         regrow_log(td->slat_log);
2967         regrow_log(td->clat_log);
2968         regrow_log(td->clat_hist_log);
2969         regrow_log(td->lat_log);
2970         regrow_log(td->bw_log);
2971         regrow_log(td->iops_log);
2972         td->flags &= ~TD_F_REGROW_LOGS;
2973 }
2974
2975 void regrow_agg_logs(void)
2976 {
2977         enum fio_ddir ddir;
2978
2979         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2980                 regrow_log(agg_io_log[ddir]);
2981 }
2982
2983 static struct io_logs *get_cur_log(struct io_log *iolog)
2984 {
2985         struct io_logs *cur_log;
2986
2987         cur_log = iolog_cur_log(iolog);
2988         if (!cur_log) {
2989                 cur_log = get_new_log(iolog);
2990                 if (!cur_log)
2991                         return NULL;
2992         }
2993
2994         if (cur_log->nr_samples < cur_log->max_samples)
2995                 return cur_log;
2996
2997         /*
2998          * Out of space. If we're in IO offload mode, or we're not doing
2999          * per unit logging (hence logging happens outside of the IO thread
3000          * as well), add a new log chunk inline. If we're doing inline
3001          * submissions, flag 'td' as needing a log regrow and we'll take
3002          * care of it on the submission side.
3003          */
3004         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
3005             !per_unit_log(iolog))
3006                 return regrow_log(iolog);
3007
3008         if (iolog->td)
3009                 iolog->td->flags |= TD_F_REGROW_LOGS;
3010         if (iolog->pending)
3011                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
3012         return iolog->pending;
3013 }
3014
3015 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
3016                              enum fio_ddir ddir, unsigned long long bs,
3017                              unsigned long t, uint64_t offset,
3018                              unsigned int priority)
3019 {
3020         struct io_logs *cur_log;
3021
3022         if (iolog->disabled)
3023                 return;
3024         if (flist_empty(&iolog->io_logs))
3025                 iolog->avg_last[ddir] = t;
3026
3027         cur_log = get_cur_log(iolog);
3028         if (cur_log) {
3029                 struct io_sample *s;
3030
3031                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
3032
3033                 s->data = data;
3034                 s->time = t + (iolog->td ? iolog->td->alternate_epoch : 0);
3035                 io_sample_set_ddir(iolog, s, ddir);
3036                 s->bs = bs;
3037                 s->priority = priority;
3038
3039                 if (iolog->log_offset) {
3040                         struct io_sample_offset *so = (void *) s;
3041
3042                         so->offset = offset;
3043                 }
3044
3045                 cur_log->nr_samples++;
3046                 return;
3047         }
3048
3049         iolog->disabled = true;
3050 }
3051
3052 static inline void reset_io_stat(struct io_stat *ios)
3053 {
3054         ios->min_val = -1ULL;
3055         ios->max_val = ios->samples = 0;
3056         ios->mean.u.f = ios->S.u.f = 0;
3057 }
3058
3059 static inline void reset_io_u_plat(uint64_t *io_u_plat)
3060 {
3061         int i;
3062
3063         for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
3064                 io_u_plat[i] = 0;
3065 }
3066
3067 static inline void reset_clat_prio_stats(struct thread_stat *ts)
3068 {
3069         enum fio_ddir ddir;
3070         int i;
3071
3072         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3073                 if (!ts->clat_prio[ddir])
3074                         continue;
3075
3076                 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
3077                         reset_io_stat(&ts->clat_prio[ddir][i].clat_stat);
3078                         reset_io_u_plat(ts->clat_prio[ddir][i].io_u_plat);
3079                 }
3080         }
3081 }
3082
3083 void reset_io_stats(struct thread_data *td)
3084 {
3085         struct thread_stat *ts = &td->ts;
3086         int i, j;
3087
3088         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
3089                 reset_io_stat(&ts->clat_stat[i]);
3090                 reset_io_stat(&ts->slat_stat[i]);
3091                 reset_io_stat(&ts->lat_stat[i]);
3092                 reset_io_stat(&ts->bw_stat[i]);
3093                 reset_io_stat(&ts->iops_stat[i]);
3094
3095                 ts->io_bytes[i] = 0;
3096                 ts->runtime[i] = 0;
3097                 ts->total_io_u[i] = 0;
3098                 ts->short_io_u[i] = 0;
3099                 ts->drop_io_u[i] = 0;
3100         }
3101
3102         for (i = 0; i < FIO_LAT_CNT; i++)
3103                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
3104                         reset_io_u_plat(ts->io_u_plat[i][j]);
3105
3106         reset_clat_prio_stats(ts);
3107
3108         ts->total_io_u[DDIR_SYNC] = 0;
3109         reset_io_u_plat(ts->io_u_sync_plat);
3110
3111         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
3112                 ts->io_u_map[i] = 0;
3113                 ts->io_u_submit[i] = 0;
3114                 ts->io_u_complete[i] = 0;
3115         }
3116
3117         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
3118                 ts->io_u_lat_n[i] = 0;
3119         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
3120                 ts->io_u_lat_u[i] = 0;
3121         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
3122                 ts->io_u_lat_m[i] = 0;
3123
3124         ts->total_submit = 0;
3125         ts->total_complete = 0;
3126         ts->nr_zone_resets = 0;
3127         ts->cachehit = ts->cachemiss = 0;
3128 }
3129
3130 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
3131                               unsigned long elapsed, bool log_max)
3132 {
3133         /*
3134          * Note an entry in the log. Use the mean from the logged samples,
3135          * making sure to properly round up. Only write a log entry if we
3136          * had actual samples done.
3137          */
3138         if (iolog->avg_window[ddir].samples) {
3139                 union io_sample_data data;
3140
3141                 if (log_max)
3142                         data.val = iolog->avg_window[ddir].max_val;
3143                 else
3144                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
3145
3146                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
3147         }
3148
3149         reset_io_stat(&iolog->avg_window[ddir]);
3150 }
3151
3152 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
3153                              bool log_max)
3154 {
3155         enum fio_ddir ddir;
3156
3157         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
3158                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
3159 }
3160
3161 static unsigned long add_log_sample(struct thread_data *td,
3162                                     struct io_log *iolog,
3163                                     union io_sample_data data,
3164                                     enum fio_ddir ddir, unsigned long long bs,
3165                                     uint64_t offset, unsigned int ioprio)
3166 {
3167         unsigned long elapsed, this_window;
3168
3169         if (!ddir_rw(ddir))
3170                 return 0;
3171
3172         elapsed = mtime_since_now(&td->epoch);
3173
3174         /*
3175          * If no time averaging, just add the log sample.
3176          */
3177         if (!iolog->avg_msec) {
3178                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
3179                                  ioprio);
3180                 return 0;
3181         }
3182
3183         /*
3184          * Add the sample. If the time period has passed, then
3185          * add that entry to the log and clear.
3186          */
3187         add_stat_sample(&iolog->avg_window[ddir], data.val);
3188
3189         /*
3190          * If period hasn't passed, adding the above sample is all we
3191          * need to do.
3192          */
3193         this_window = elapsed - iolog->avg_last[ddir];
3194         if (elapsed < iolog->avg_last[ddir])
3195                 return iolog->avg_last[ddir] - elapsed;
3196         else if (this_window < iolog->avg_msec) {
3197                 unsigned long diff = iolog->avg_msec - this_window;
3198
3199                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
3200                         return diff;
3201         }
3202
3203         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
3204
3205         iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
3206
3207         return iolog->avg_msec;
3208 }
3209
3210 void finalize_logs(struct thread_data *td, bool unit_logs)
3211 {
3212         unsigned long elapsed;
3213
3214         elapsed = mtime_since_now(&td->epoch);
3215
3216         if (td->clat_log && unit_logs)
3217                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
3218         if (td->slat_log && unit_logs)
3219                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
3220         if (td->lat_log && unit_logs)
3221                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
3222         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
3223                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
3224         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
3225                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
3226 }
3227
3228 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
3229                     unsigned long long bs)
3230 {
3231         struct io_log *iolog;
3232
3233         if (!ddir_rw(ddir))
3234                 return;
3235
3236         iolog = agg_io_log[ddir];
3237         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
3238 }
3239
3240 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
3241 {
3242         unsigned int idx = plat_val_to_idx(nsec);
3243         assert(idx < FIO_IO_U_PLAT_NR);
3244
3245         ts->io_u_sync_plat[idx]++;
3246         add_stat_sample(&ts->sync_stat, nsec);
3247 }
3248
3249 static inline void add_lat_percentile_sample(struct thread_stat *ts,
3250                                              unsigned long long nsec,
3251                                              enum fio_ddir ddir,
3252                                              enum fio_lat lat)
3253 {
3254         unsigned int idx = plat_val_to_idx(nsec);
3255         assert(idx < FIO_IO_U_PLAT_NR);
3256
3257         ts->io_u_plat[lat][ddir][idx]++;
3258 }
3259
3260 static inline void
3261 add_lat_percentile_prio_sample(struct thread_stat *ts, unsigned long long nsec,
3262                                enum fio_ddir ddir,
3263                                unsigned short clat_prio_index)
3264 {
3265         unsigned int idx = plat_val_to_idx(nsec);
3266
3267         if (ts->clat_prio[ddir])
3268                 ts->clat_prio[ddir][clat_prio_index].io_u_plat[idx]++;
3269 }
3270
3271 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
3272                      unsigned long long nsec, unsigned long long bs,
3273                      uint64_t offset, unsigned int ioprio,
3274                      unsigned short clat_prio_index)
3275 {
3276         const bool needs_lock = td_async_processing(td);
3277         unsigned long elapsed, this_window;
3278         struct thread_stat *ts = &td->ts;
3279         struct io_log *iolog = td->clat_hist_log;
3280
3281         if (needs_lock)
3282                 __td_io_u_lock(td);
3283
3284         add_stat_sample(&ts->clat_stat[ddir], nsec);
3285
3286         /*
3287          * When lat_percentiles=1 (default 0), the reported per priority
3288          * percentiles and stats are used for describing total latency values,
3289          * even though the variable names themselves start with clat_.
3290          *
3291          * Because of the above definition, add a prio stat sample only when
3292          * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3293          * when lat_percentiles=1.
3294          */
3295         if (!ts->lat_percentiles)
3296                 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3297                                      nsec);
3298
3299         if (td->clat_log)
3300                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3301                                offset, ioprio);
3302
3303         if (ts->clat_percentiles) {
3304                 /*
3305                  * Because of the above definition, add a prio lat percentile
3306                  * sample only when lat_percentiles=0. add_lat_sample() will add
3307                  * the prio lat percentile sample when lat_percentiles=1.
3308                  */
3309                 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3310                 if (!ts->lat_percentiles)
3311                         add_lat_percentile_prio_sample(ts, nsec, ddir,
3312                                                        clat_prio_index);
3313         }
3314
3315         if (iolog && iolog->hist_msec) {
3316                 struct io_hist *hw = &iolog->hist_window[ddir];
3317
3318                 hw->samples++;
3319                 elapsed = mtime_since_now(&td->epoch);
3320                 if (!hw->hist_last)
3321                         hw->hist_last = elapsed;
3322                 this_window = elapsed - hw->hist_last;
3323
3324                 if (this_window >= iolog->hist_msec) {
3325                         uint64_t *io_u_plat;
3326                         struct io_u_plat_entry *dst;
3327
3328                         /*
3329                          * Make a byte-for-byte copy of the latency histogram
3330                          * stored in td->ts.io_u_plat[ddir], recording it in a
3331                          * log sample. Note that the matching call to free() is
3332                          * located in iolog.c after printing this sample to the
3333                          * log file.
3334                          */
3335                         io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3336                         dst = malloc(sizeof(struct io_u_plat_entry));
3337                         memcpy(&(dst->io_u_plat), io_u_plat,
3338                                 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3339                         flist_add(&dst->list, &hw->list);
3340                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3341                                          elapsed, offset, ioprio);
3342
3343                         /*
3344                          * Update the last time we recorded as being now, minus
3345                          * any drift in time we encountered before actually
3346                          * making the record.
3347                          */
3348                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3349                         hw->samples = 0;
3350                 }
3351         }
3352
3353         if (needs_lock)
3354                 __td_io_u_unlock(td);
3355 }
3356
3357 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3358                      unsigned long long nsec, unsigned long long bs,
3359                      uint64_t offset, unsigned int ioprio)
3360 {
3361         const bool needs_lock = td_async_processing(td);
3362         struct thread_stat *ts = &td->ts;
3363
3364         if (!ddir_rw(ddir))
3365                 return;
3366
3367         if (needs_lock)
3368                 __td_io_u_lock(td);
3369
3370         add_stat_sample(&ts->slat_stat[ddir], nsec);
3371
3372         if (td->slat_log)
3373                 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3374                                offset, ioprio);
3375
3376         if (ts->slat_percentiles)
3377                 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3378
3379         if (needs_lock)
3380                 __td_io_u_unlock(td);
3381 }
3382
3383 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3384                     unsigned long long nsec, unsigned long long bs,
3385                     uint64_t offset, unsigned int ioprio,
3386                     unsigned short clat_prio_index)
3387 {
3388         const bool needs_lock = td_async_processing(td);
3389         struct thread_stat *ts = &td->ts;
3390
3391         if (!ddir_rw(ddir))
3392                 return;
3393
3394         if (needs_lock)
3395                 __td_io_u_lock(td);
3396
3397         add_stat_sample(&ts->lat_stat[ddir], nsec);
3398
3399         if (td->lat_log)
3400                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3401                                offset, ioprio);
3402
3403         /*
3404          * When lat_percentiles=1 (default 0), the reported per priority
3405          * percentiles and stats are used for describing total latency values,
3406          * even though the variable names themselves start with clat_.
3407          *
3408          * Because of the above definition, add a prio stat and prio lat
3409          * percentile sample only when lat_percentiles=1. add_clat_sample() will
3410          * add the prio stat and prio lat percentile sample when
3411          * lat_percentiles=0.
3412          */
3413         if (ts->lat_percentiles) {
3414                 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3415                 add_lat_percentile_prio_sample(ts, nsec, ddir, clat_prio_index);
3416                 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3417                                      nsec);
3418         }
3419         if (needs_lock)
3420                 __td_io_u_unlock(td);
3421 }
3422
3423 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3424                    unsigned int bytes, unsigned long long spent)
3425 {
3426         const bool needs_lock = td_async_processing(td);
3427         struct thread_stat *ts = &td->ts;
3428         unsigned long rate;
3429
3430         if (spent)
3431                 rate = (unsigned long) (bytes * 1000000ULL / spent);
3432         else
3433                 rate = 0;
3434
3435         if (needs_lock)
3436                 __td_io_u_lock(td);
3437
3438         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3439
3440         if (td->bw_log)
3441                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3442                                bytes, io_u->offset, io_u->ioprio);
3443
3444         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3445
3446         if (needs_lock)
3447                 __td_io_u_unlock(td);
3448 }
3449
3450 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3451                          struct timespec *t, unsigned int avg_time,
3452                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3453                          struct io_stat *stat, struct io_log *log,
3454                          bool is_kb)
3455 {
3456         const bool needs_lock = td_async_processing(td);
3457         unsigned long spent, rate;
3458         enum fio_ddir ddir;
3459         unsigned long next, next_log;
3460
3461         next_log = avg_time;
3462
3463         spent = mtime_since(parent_tv, t);
3464         if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3465                 return avg_time - spent;
3466
3467         if (needs_lock)
3468                 __td_io_u_lock(td);
3469
3470         /*
3471          * Compute both read and write rates for the interval.
3472          */
3473         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3474                 uint64_t delta;
3475
3476                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3477                 if (!delta)
3478                         continue; /* No entries for interval */
3479
3480                 if (spent) {
3481                         if (is_kb)
3482                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
3483                         else
3484                                 rate = (delta * 1000) / spent;
3485                 } else
3486                         rate = 0;
3487
3488                 add_stat_sample(&stat[ddir], rate);
3489
3490                 if (log) {
3491                         unsigned long long bs = 0;
3492
3493                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3494                                 bs = td->o.min_bs[ddir];
3495
3496                         next = add_log_sample(td, log, sample_val(rate), ddir,
3497                                               bs, 0, 0);
3498                         next_log = min(next_log, next);
3499                 }
3500
3501                 stat_io_bytes[ddir] = this_io_bytes[ddir];
3502         }
3503
3504         *parent_tv = *t;
3505
3506         if (needs_lock)
3507                 __td_io_u_unlock(td);
3508
3509         if (spent <= avg_time)
3510                 next = avg_time;
3511         else
3512                 next = avg_time - (1 + spent - avg_time);
3513
3514         return min(next, next_log);
3515 }
3516
3517 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3518 {
3519         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3520                                 td->this_io_bytes, td->stat_io_bytes,
3521                                 td->ts.bw_stat, td->bw_log, true);
3522 }
3523
3524 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3525                      unsigned int bytes)
3526 {
3527         const bool needs_lock = td_async_processing(td);
3528         struct thread_stat *ts = &td->ts;
3529
3530         if (needs_lock)
3531                 __td_io_u_lock(td);
3532
3533         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3534
3535         if (td->iops_log)
3536                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3537                                bytes, io_u->offset, io_u->ioprio);
3538
3539         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3540
3541         if (needs_lock)
3542                 __td_io_u_unlock(td);
3543 }
3544
3545 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3546 {
3547         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3548                                 td->this_io_blocks, td->stat_io_blocks,
3549                                 td->ts.iops_stat, td->iops_log, false);
3550 }
3551
3552 /*
3553  * Returns msecs to next event
3554  */
3555 int calc_log_samples(void)
3556 {
3557         struct thread_data *td;
3558         unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3559         struct timespec now;
3560         int i;
3561         long elapsed_time = 0;
3562
3563         fio_gettime(&now, NULL);
3564
3565         for_each_td(td, i) {
3566                 elapsed_time = mtime_since_now(&td->epoch);
3567
3568                 if (!td->o.stats)
3569                         continue;
3570                 if (in_ramp_time(td) ||
3571                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3572                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3573                         continue;
3574                 }
3575                 if (!td->bw_log ||
3576                         (td->bw_log && !per_unit_log(td->bw_log))) {
3577                         tmp = add_bw_samples(td, &now);
3578
3579                         if (td->bw_log)
3580                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3581                 }
3582                 if (!td->iops_log ||
3583                         (td->iops_log && !per_unit_log(td->iops_log))) {
3584                         tmp = add_iops_samples(td, &now);
3585
3586                         if (td->iops_log)
3587                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3588                 }
3589
3590                 if (tmp < next)
3591                         next = tmp;
3592         }
3593
3594         /* if log_avg_msec_min has not been changed, set it to 0 */
3595         if (log_avg_msec_min == -1U)
3596                 log_avg_msec_min = 0;
3597
3598         if (log_avg_msec_min == 0)
3599                 next_mod = elapsed_time;
3600         else
3601                 next_mod = elapsed_time % log_avg_msec_min;
3602
3603         /* correction to keep the time on the log avg msec boundary */
3604         next = min(next, (log_avg_msec_min - next_mod));
3605
3606         return next == ~0U ? 0 : next;
3607 }
3608
3609 void stat_init(void)
3610 {
3611         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3612 }
3613
3614 void stat_exit(void)
3615 {
3616         /*
3617          * When we have the mutex, we know out-of-band access to it
3618          * have ended.
3619          */
3620         fio_sem_down(stat_sem);
3621         fio_sem_remove(stat_sem);
3622 }
3623
3624 /*
3625  * Called from signal handler. Wake up status thread.
3626  */
3627 void show_running_run_stats(void)
3628 {
3629         helper_do_stat();
3630 }
3631
3632 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3633 {
3634         /* Ignore io_u's which span multiple blocks--they will just get
3635          * inaccurate counts. */
3636         int idx = (io_u->offset - io_u->file->file_offset)
3637                         / td->o.bs[DDIR_TRIM];
3638         uint32_t *info = &td->ts.block_infos[idx];
3639         assert(idx < td->ts.nr_block_infos);
3640         return info;
3641 }
3642