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