stat.c

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