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