stat.c

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