stat.c

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