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