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