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