stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/stat.h>
   5 #include <math.h>
   6
   7 #include "fio.h"
   8 #include "diskutil.h"
   9 #include "lib/ieee754.h"
  10 #include "json.h"
  11 #include "lib/getrusage.h"
  12 #include "idletime.h"
  13 #include "lib/pow2.h"
  14 #include "lib/output_buffer.h"
  15 #include "helper_thread.h"
  16 #include "smalloc.h"
  17 #include "zbd.h"
  18
  19 #define LOG_MSEC_SLACK  1
  20
  21 struct fio_sem *stat_sem;
  22
  23 void clear_rusage_stat(struct thread_data *td)
  24 {
  25         struct thread_stat *ts = &td->ts;
  26
  27         fio_getrusage(&td->ru_start);
  28         ts->usr_time = ts->sys_time = 0;
  29         ts->ctx = 0;
  30         ts->minf = ts->majf = 0;
  31 }
  32
  33 void update_rusage_stat(struct thread_data *td)
  34 {
  35         struct thread_stat *ts = &td->ts;
  36
  37         fio_getrusage(&td->ru_end);
  38         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
  39                                         &td->ru_end.ru_utime);
  40         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
  41                                         &td->ru_end.ru_stime);
  42         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  43                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  44         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  45         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  46
  47         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  48 }
  49
  50 /*
  51  * Given a latency, return the index of the corresponding bucket in
  52  * the structure tracking percentiles.
  53  *
  54  * (1) find the group (and error bits) that the value (latency)
  55  * belongs to by looking at its MSB. (2) find the bucket number in the
  56  * group by looking at the index bits.
  57  *
  58  */
  59 static unsigned int plat_val_to_idx(unsigned long long val)
  60 {
  61         unsigned int msb, error_bits, base, offset, idx;
  62
  63         /* Find MSB starting from bit 0 */
  64         if (val == 0)
  65                 msb = 0;
  66         else
  67                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
  68
  69         /*
  70          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  71          * all bits of the sample as index
  72          */
  73         if (msb <= FIO_IO_U_PLAT_BITS)
  74                 return val;
  75
  76         /* Compute the number of error bits to discard*/
  77         error_bits = msb - FIO_IO_U_PLAT_BITS;
  78
  79         /* Compute the number of buckets before the group */
  80         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  81
  82         /*
  83          * Discard the error bits and apply the mask to find the
  84          * index for the buckets in the group
  85          */
  86         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  87
  88         /* Make sure the index does not exceed (array size - 1) */
  89         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  90                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  91
  92         return idx;
  93 }
  94
  95 /*
  96  * Convert the given index of the bucket array to the value
  97  * represented by the bucket
  98  */
  99 static unsigned long long plat_idx_to_val(unsigned int idx)
 100 {
 101         unsigned int error_bits;
 102         unsigned long long k, base;
 103
 104         assert(idx < FIO_IO_U_PLAT_NR);
 105
 106         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 107          * all bits of the sample as index */
 108         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 109                 return idx;
 110
 111         /* Find the group and compute the minimum value of that group */
 112         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 113         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
 114
 115         /* Find its bucket number of the group */
 116         k = idx % FIO_IO_U_PLAT_VAL;
 117
 118         /* Return the mean of the range of the bucket */
 119         return base + ((k + 0.5) * (1 << error_bits));
 120 }
 121
 122 static int double_cmp(const void *a, const void *b)
 123 {
 124         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 125         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 126         int cmp = 0;
 127
 128         if (fa.u.f > fb.u.f)
 129                 cmp = 1;
 130         else if (fa.u.f < fb.u.f)
 131                 cmp = -1;
 132
 133         return cmp;
 134 }
 135
 136 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 137                                    fio_fp64_t *plist, unsigned long long **output,
 138                                    unsigned long long *maxv, unsigned long long *minv)
 139 {
 140         unsigned long long sum = 0;
 141         unsigned int len, i, j = 0;
 142         unsigned long long *ovals = NULL;
 143         bool is_last;
 144
 145         *minv = -1ULL;
 146         *maxv = 0;
 147
 148         len = 0;
 149         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 150                 len++;
 151
 152         if (!len)
 153                 return 0;
 154
 155         /*
 156          * Sort the percentile list. Note that it may already be sorted if
 157          * we are using the default values, but since it's a short list this
 158          * isn't a worry. Also note that this does not work for NaN values.
 159          */
 160         if (len > 1)
 161                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 162
 163         ovals = malloc(len * sizeof(*ovals));
 164         if (!ovals)
 165                 return 0;
 166
 167         /*
 168          * Calculate bucket values, note down max and min values
 169          */
 170         is_last = false;
 171         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 172                 sum += io_u_plat[i];
 173                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 174                         assert(plist[j].u.f <= 100.0);
 175
 176                         ovals[j] = plat_idx_to_val(i);
 177                         if (ovals[j] < *minv)
 178                                 *minv = ovals[j];
 179                         if (ovals[j] > *maxv)
 180                                 *maxv = ovals[j];
 181
 182                         is_last = (j == len - 1) != 0;
 183                         if (is_last)
 184                                 break;
 185
 186                         j++;
 187                 }
 188         }
 189
 190         *output = ovals;
 191         return len;
 192 }
 193
 194 /*
 195  * Find and display the p-th percentile of clat
 196  */
 197 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 198                                   fio_fp64_t *plist, unsigned int precision,
 199                                   const char *pre, struct buf_output *out)
 200 {
 201         unsigned int divisor, len, i, j = 0;
 202         unsigned long long minv, maxv;
 203         unsigned long long *ovals;
 204         int per_line, scale_down, time_width;
 205         bool is_last;
 206         char fmt[32];
 207
 208         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 209         if (!len || !ovals)
 210                 goto out;
 211
 212         /*
 213          * We default to nsecs, but if the value range is such that we
 214          * should scale down to usecs or msecs, do that.
 215          */
 216         if (minv > 2000000 && maxv > 99999999ULL) {
 217                 scale_down = 2;
 218                 divisor = 1000000;
 219                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 220         } else if (minv > 2000 && maxv > 99999) {
 221                 scale_down = 1;
 222                 divisor = 1000;
 223                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 224         } else {
 225                 scale_down = 0;
 226                 divisor = 1;
 227                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 228         }
 229
 230
 231         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 232         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 233                         precision, time_width);
 234         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 235         per_line = (80 - 7) / (precision + 10 + time_width);
 236
 237         for (j = 0; j < len; j++) {
 238                 /* for formatting */
 239                 if (j != 0 && (j % per_line) == 0)
 240                         log_buf(out, "     |");
 241
 242                 /* end of the list */
 243                 is_last = (j == len - 1) != 0;
 244
 245                 for (i = 0; i < scale_down; i++)
 246                         ovals[j] = (ovals[j] + 999) / 1000;
 247
 248                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 249
 250                 if (is_last)
 251                         break;
 252
 253                 if ((j % per_line) == per_line - 1)     /* for formatting */
 254                         log_buf(out, "\n");
 255         }
 256
 257 out:
 258         if (ovals)
 259                 free(ovals);
 260 }
 261
 262 bool calc_lat(struct io_stat *is, unsigned long long *min,
 263               unsigned long long *max, double *mean, double *dev)
 264 {
 265         double n = (double) is->samples;
 266
 267         if (n == 0)
 268                 return false;
 269
 270         *min = is->min_val;
 271         *max = is->max_val;
 272         *mean = is->mean.u.f;
 273
 274         if (n > 1.0)
 275                 *dev = sqrt(is->S.u.f / (n - 1.0));
 276         else
 277                 *dev = 0;
 278
 279         return true;
 280 }
 281
 282 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 283 {
 284         char *io, *agg, *min, *max;
 285         char *ioalt, *aggalt, *minalt, *maxalt;
 286         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 287         int i;
 288
 289         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 290
 291         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 292                 const int i2p = is_power_of_2(rs->kb_base);
 293
 294                 if (!rs->max_run[i])
 295                         continue;
 296
 297                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
 298                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
 299                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
 300                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 301                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 302                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 303                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 304                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 305                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 306                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 307                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 308                                 (unsigned long long) rs->min_run[i],
 309                                 (unsigned long long) rs->max_run[i]);
 310
 311                 free(io);
 312                 free(agg);
 313                 free(min);
 314                 free(max);
 315                 free(ioalt);
 316                 free(aggalt);
 317                 free(minalt);
 318                 free(maxalt);
 319         }
 320 }
 321
 322 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
 323 {
 324         int i;
 325
 326         /*
 327          * Do depth distribution calculations
 328          */
 329         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 330                 if (total) {
 331                         io_u_dist[i] = (double) map[i] / (double) total;
 332                         io_u_dist[i] *= 100.0;
 333                         if (io_u_dist[i] < 0.1 && map[i])
 334                                 io_u_dist[i] = 0.1;
 335                 } else
 336                         io_u_dist[i] = 0.0;
 337         }
 338 }
 339
 340 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 341                           uint64_t *src, int nr)
 342 {
 343         unsigned long total = ddir_rw_sum(ts->total_io_u);
 344         int i;
 345
 346         /*
 347          * Do latency distribution calculations
 348          */
 349         for (i = 0; i < nr; i++) {
 350                 if (total) {
 351                         dst[i] = (double) src[i] / (double) total;
 352                         dst[i] *= 100.0;
 353                         if (dst[i] < 0.01 && src[i])
 354                                 dst[i] = 0.01;
 355                 } else
 356                         dst[i] = 0.0;
 357         }
 358 }
 359
 360 /*
 361  * To keep the terse format unaltered, add all of the ns latency
 362  * buckets to the first us latency bucket
 363  */
 364 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 365 {
 366         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 367         int i;
 368
 369         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 370
 371         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 372                 ntotal += ts->io_u_lat_n[i];
 373
 374         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 375 }
 376
 377 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 378 {
 379         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 380 }
 381
 382 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 383 {
 384         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 385 }
 386
 387 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 388 {
 389         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 390 }
 391
 392 static void display_lat(const char *name, unsigned long long min,
 393                         unsigned long long max, double mean, double dev,
 394                         struct buf_output *out)
 395 {
 396         const char *base = "(nsec)";
 397         char *minp, *maxp;
 398
 399         if (nsec_to_msec(&min, &max, &mean, &dev))
 400                 base = "(msec)";
 401         else if (nsec_to_usec(&min, &max, &mean, &dev))
 402                 base = "(usec)";
 403
 404         minp = num2str(min, 6, 1, 0, N2S_NONE);
 405         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 406
 407         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 408                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 409
 410         free(minp);
 411         free(maxp);
 412 }
 413
 414 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 415                              int ddir, struct buf_output *out)
 416 {
 417         unsigned long runt;
 418         unsigned long long min, max, bw, iops;
 419         double mean, dev;
 420         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
 421         int i2p;
 422
 423         if (ddir_sync(ddir)) {
 424                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
 425                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
 426                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
 427                         show_clat_percentiles(ts->io_u_sync_plat,
 428                                                 ts->sync_stat.samples,
 429                                                 ts->percentile_list,
 430                                                 ts->percentile_precision,
 431                                                 io_ddir_name(ddir), out);
 432                 }
 433                 return;
 434         }
 435
 436         assert(ddir_rw(ddir));
 437
 438         if (!ts->runtime[ddir])
 439                 return;
 440
 441         i2p = is_power_of_2(rs->kb_base);
 442         runt = ts->runtime[ddir];
 443
 444         bw = (1000 * ts->io_bytes[ddir]) / runt;
 445         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
 446         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
 447         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
 448
 449         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 450         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
 451         if (ddir == DDIR_WRITE)
 452                 post_st = zbd_write_status(ts);
 453         else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
 454                 uint64_t total;
 455                 double hit;
 456
 457                 total = ts->cachehit + ts->cachemiss;
 458                 hit = (double) ts->cachehit / (double) total;
 459                 hit *= 100.0;
 460                 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
 461                         post_st = NULL;
 462         }
 463
 464         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
 465                         rs->unified_rw_rep ? "mixed" : io_ddir_name(ddir),
 466                         iops_p, bw_p, bw_p_alt, io_p,
 467                         (unsigned long long) ts->runtime[ddir],
 468                         post_st ? : "");
 469
 470         free(post_st);
 471         free(io_p);
 472         free(bw_p);
 473         free(bw_p_alt);
 474         free(iops_p);
 475
 476         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 477                 display_lat("slat", min, max, mean, dev, out);
 478         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 479                 display_lat("clat", min, max, mean, dev, out);
 480         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 481                 display_lat(" lat", min, max, mean, dev, out);
 482
 483         if (ts->clat_percentiles || ts->lat_percentiles) {
 484                 const char *name = ts->clat_percentiles ? "clat" : " lat";
 485                 uint64_t samples;
 486
 487                 if (ts->clat_percentiles)
 488                         samples = ts->clat_stat[ddir].samples;
 489                 else
 490                         samples = ts->lat_stat[ddir].samples;
 491
 492                 show_clat_percentiles(ts->io_u_plat[ddir],
 493                                         samples,
 494                                         ts->percentile_list,
 495                                         ts->percentile_precision, name, out);
 496         }
 497         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 498                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 499                 const char *bw_str;
 500
 501                 if ((rs->unit_base == 1) && i2p)
 502                         bw_str = "Kibit";
 503                 else if (rs->unit_base == 1)
 504                         bw_str = "kbit";
 505                 else if (i2p)
 506                         bw_str = "KiB";
 507                 else
 508                         bw_str = "kB";
 509
 510                 if (rs->agg[ddir]) {
 511                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 512                         if (p_of_agg > 100.0)
 513                                 p_of_agg = 100.0;
 514                 }
 515
 516                 if (rs->unit_base == 1) {
 517                         min *= 8.0;
 518                         max *= 8.0;
 519                         mean *= 8.0;
 520                         dev *= 8.0;
 521                 }
 522
 523                 if (mean > fkb_base * fkb_base) {
 524                         min /= fkb_base;
 525                         max /= fkb_base;
 526                         mean /= fkb_base;
 527                         dev /= fkb_base;
 528                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 529                 }
 530
 531                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 532                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 533                         bw_str, min, max, p_of_agg, mean, dev,
 534                         (&ts->bw_stat[ddir])->samples);
 535         }
 536         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 537                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 538                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 539                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 540         }
 541 }
 542
 543 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
 544                      const char *msg, struct buf_output *out)
 545 {
 546         bool new_line = true, shown = false;
 547         int i, line = 0;
 548
 549         for (i = 0; i < nr; i++) {
 550                 if (io_u_lat[i] <= 0.0)
 551                         continue;
 552                 shown = true;
 553                 if (new_line) {
 554                         if (line)
 555                                 log_buf(out, "\n");
 556                         log_buf(out, "  lat (%s)   : ", msg);
 557                         new_line = false;
 558                         line = 0;
 559                 }
 560                 if (line)
 561                         log_buf(out, ", ");
 562                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 563                 line++;
 564                 if (line == 5)
 565                         new_line = true;
 566         }
 567
 568         if (shown)
 569                 log_buf(out, "\n");
 570
 571         return true;
 572 }
 573
 574 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 575 {
 576         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 577                                  "250=", "500=", "750=", "1000=", };
 578
 579         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 580 }
 581
 582 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 583 {
 584         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 585                                  "250=", "500=", "750=", "1000=", };
 586
 587         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 588 }
 589
 590 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 591 {
 592         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 593                                  "250=", "500=", "750=", "1000=", "2000=",
 594                                  ">=2000=", };
 595
 596         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 597 }
 598
 599 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 600 {
 601         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 602         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 603         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 604
 605         stat_calc_lat_n(ts, io_u_lat_n);
 606         stat_calc_lat_u(ts, io_u_lat_u);
 607         stat_calc_lat_m(ts, io_u_lat_m);
 608
 609         show_lat_n(io_u_lat_n, out);
 610         show_lat_u(io_u_lat_u, out);
 611         show_lat_m(io_u_lat_m, out);
 612 }
 613
 614 static int block_state_category(int block_state)
 615 {
 616         switch (block_state) {
 617         case BLOCK_STATE_UNINIT:
 618                 return 0;
 619         case BLOCK_STATE_TRIMMED:
 620         case BLOCK_STATE_WRITTEN:
 621                 return 1;
 622         case BLOCK_STATE_WRITE_FAILURE:
 623         case BLOCK_STATE_TRIM_FAILURE:
 624                 return 2;
 625         default:
 626                 /* Silence compile warning on some BSDs and have a return */
 627                 assert(0);
 628                 return -1;
 629         }
 630 }
 631
 632 static int compare_block_infos(const void *bs1, const void *bs2)
 633 {
 634         uint64_t block1 = *(uint64_t *)bs1;
 635         uint64_t block2 = *(uint64_t *)bs2;
 636         int state1 = BLOCK_INFO_STATE(block1);
 637         int state2 = BLOCK_INFO_STATE(block2);
 638         int bscat1 = block_state_category(state1);
 639         int bscat2 = block_state_category(state2);
 640         int cycles1 = BLOCK_INFO_TRIMS(block1);
 641         int cycles2 = BLOCK_INFO_TRIMS(block2);
 642
 643         if (bscat1 < bscat2)
 644                 return -1;
 645         if (bscat1 > bscat2)
 646                 return 1;
 647
 648         if (cycles1 < cycles2)
 649                 return -1;
 650         if (cycles1 > cycles2)
 651                 return 1;
 652
 653         if (state1 < state2)
 654                 return -1;
 655         if (state1 > state2)
 656                 return 1;
 657
 658         assert(block1 == block2);
 659         return 0;
 660 }
 661
 662 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 663                                   fio_fp64_t *plist, unsigned int **percentiles,
 664                                   unsigned int *types)
 665 {
 666         int len = 0;
 667         int i, nr_uninit;
 668
 669         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 670
 671         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 672                 len++;
 673
 674         if (!len)
 675                 return 0;
 676
 677         /*
 678          * Sort the percentile list. Note that it may already be sorted if
 679          * we are using the default values, but since it's a short list this
 680          * isn't a worry. Also note that this does not work for NaN values.
 681          */
 682         if (len > 1)
 683                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 684
 685         /* Start only after the uninit entries end */
 686         for (nr_uninit = 0;
 687              nr_uninit < nr_block_infos
 688                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 689              nr_uninit ++)
 690                 ;
 691
 692         if (nr_uninit == nr_block_infos)
 693                 return 0;
 694
 695         *percentiles = calloc(len, sizeof(**percentiles));
 696
 697         for (i = 0; i < len; i++) {
 698                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 699                                 + nr_uninit;
 700                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 701         }
 702
 703         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 704         for (i = 0; i < nr_block_infos; i++)
 705                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 706
 707         return len;
 708 }
 709
 710 static const char *block_state_names[] = {
 711         [BLOCK_STATE_UNINIT] = "unwritten",
 712         [BLOCK_STATE_TRIMMED] = "trimmed",
 713         [BLOCK_STATE_WRITTEN] = "written",
 714         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 715         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 716 };
 717
 718 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 719                              fio_fp64_t *plist, struct buf_output *out)
 720 {
 721         int len, pos, i;
 722         unsigned int *percentiles = NULL;
 723         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 724
 725         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 726                                      &percentiles, block_state_counts);
 727
 728         log_buf(out, "  block lifetime percentiles :\n   |");
 729         pos = 0;
 730         for (i = 0; i < len; i++) {
 731                 uint32_t block_info = percentiles[i];
 732 #define LINE_LENGTH     75
 733                 char str[LINE_LENGTH];
 734                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 735                                      plist[i].u.f, block_info,
 736                                      i == len - 1 ? '\n' : ',');
 737                 assert(strln < LINE_LENGTH);
 738                 if (pos + strln > LINE_LENGTH) {
 739                         pos = 0;
 740                         log_buf(out, "\n   |");
 741                 }
 742                 log_buf(out, "%s", str);
 743                 pos += strln;
 744 #undef LINE_LENGTH
 745         }
 746         if (percentiles)
 747                 free(percentiles);
 748
 749         log_buf(out, "        states               :");
 750         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 751                 log_buf(out, " %s=%u%c",
 752                          block_state_names[i], block_state_counts[i],
 753                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 754 }
 755
 756 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 757 {
 758         char *p1, *p1alt, *p2;
 759         unsigned long long bw_mean, iops_mean;
 760         const int i2p = is_power_of_2(ts->kb_base);
 761
 762         if (!ts->ss_dur)
 763                 return;
 764
 765         bw_mean = steadystate_bw_mean(ts);
 766         iops_mean = steadystate_iops_mean(ts);
 767
 768         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
 769         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
 770         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
 771
 772         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 773                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
 774                 p1, p1alt, p2,
 775                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
 776                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
 777                 ts->ss_criterion.u.f,
 778                 ts->ss_state & FIO_SS_PCT ? "%" : "");
 779
 780         free(p1);
 781         free(p1alt);
 782         free(p2);
 783 }
 784
 785 static void show_thread_status_normal(struct thread_stat *ts,
 786                                       struct group_run_stats *rs,
 787                                       struct buf_output *out)
 788 {
 789         double usr_cpu, sys_cpu;
 790         unsigned long runtime;
 791         double io_u_dist[FIO_IO_U_MAP_NR];
 792         time_t time_p;
 793         char time_buf[32];
 794
 795         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 796                 return;
 797
 798         memset(time_buf, 0, sizeof(time_buf));
 799
 800         time(&time_p);
 801         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 802
 803         if (!ts->error) {
 804                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 805                                         ts->name, ts->groupid, ts->members,
 806                                         ts->error, (int) ts->pid, time_buf);
 807         } else {
 808                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 809                                         ts->name, ts->groupid, ts->members,
 810                                         ts->error, ts->verror, (int) ts->pid,
 811                                         time_buf);
 812         }
 813
 814         if (strlen(ts->description))
 815                 log_buf(out, "  Description  : [%s]\n", ts->description);
 816
 817         if (ts->io_bytes[DDIR_READ])
 818                 show_ddir_status(rs, ts, DDIR_READ, out);
 819         if (ts->io_bytes[DDIR_WRITE])
 820                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 821         if (ts->io_bytes[DDIR_TRIM])
 822                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 823
 824         show_latencies(ts, out);
 825
 826         if (ts->sync_stat.samples)
 827                 show_ddir_status(rs, ts, DDIR_SYNC, out);
 828
 829         runtime = ts->total_run_time;
 830         if (runtime) {
 831                 double runt = (double) runtime;
 832
 833                 usr_cpu = (double) ts->usr_time * 100 / runt;
 834                 sys_cpu = (double) ts->sys_time * 100 / runt;
 835         } else {
 836                 usr_cpu = 0;
 837                 sys_cpu = 0;
 838         }
 839
 840         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 841                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 842                         (unsigned long long) ts->ctx,
 843                         (unsigned long long) ts->majf,
 844                         (unsigned long long) ts->minf);
 845
 846         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 847         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 848                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 849                                         io_u_dist[1], io_u_dist[2],
 850                                         io_u_dist[3], io_u_dist[4],
 851                                         io_u_dist[5], io_u_dist[6]);
 852
 853         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 854         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 855                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 856                                         io_u_dist[1], io_u_dist[2],
 857                                         io_u_dist[3], io_u_dist[4],
 858                                         io_u_dist[5], io_u_dist[6]);
 859         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 860         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 861                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 862                                         io_u_dist[1], io_u_dist[2],
 863                                         io_u_dist[3], io_u_dist[4],
 864                                         io_u_dist[5], io_u_dist[6]);
 865         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
 866                                  " short=%llu,%llu,%llu,0"
 867                                  " dropped=%llu,%llu,%llu,0\n",
 868                                         (unsigned long long) ts->total_io_u[0],
 869                                         (unsigned long long) ts->total_io_u[1],
 870                                         (unsigned long long) ts->total_io_u[2],
 871                                         (unsigned long long) ts->total_io_u[3],
 872                                         (unsigned long long) ts->short_io_u[0],
 873                                         (unsigned long long) ts->short_io_u[1],
 874                                         (unsigned long long) ts->short_io_u[2],
 875                                         (unsigned long long) ts->drop_io_u[0],
 876                                         (unsigned long long) ts->drop_io_u[1],
 877                                         (unsigned long long) ts->drop_io_u[2]);
 878         if (ts->continue_on_error) {
 879                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 880                                         (unsigned long long)ts->total_err_count,
 881                                         ts->first_error,
 882                                         strerror(ts->first_error));
 883         }
 884         if (ts->latency_depth) {
 885                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 886                                         (unsigned long long)ts->latency_target,
 887                                         (unsigned long long)ts->latency_window,
 888                                         ts->latency_percentile.u.f,
 889                                         ts->latency_depth);
 890         }
 891
 892         if (ts->nr_block_infos)
 893                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 894                                   ts->percentile_list, out);
 895
 896         if (ts->ss_dur)
 897                 show_ss_normal(ts, out);
 898 }
 899
 900 static void show_ddir_status_terse(struct thread_stat *ts,
 901                                    struct group_run_stats *rs, int ddir,
 902                                    int ver, struct buf_output *out)
 903 {
 904         unsigned long long min, max, minv, maxv, bw, iops;
 905         unsigned long long *ovals = NULL;
 906         double mean, dev;
 907         unsigned int len;
 908         int i, bw_stat;
 909
 910         assert(ddir_rw(ddir));
 911
 912         iops = bw = 0;
 913         if (ts->runtime[ddir]) {
 914                 uint64_t runt = ts->runtime[ddir];
 915
 916                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 917                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 918         }
 919
 920         log_buf(out, ";%llu;%llu;%llu;%llu",
 921                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 922                                         (unsigned long long) ts->runtime[ddir]);
 923
 924         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 925                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 926         else
 927                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 928
 929         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 930                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 931         else
 932                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 933
 934         if (ts->clat_percentiles || ts->lat_percentiles) {
 935                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 936                                         ts->clat_stat[ddir].samples,
 937                                         ts->percentile_list, &ovals, &maxv,
 938                                         &minv);
 939         } else
 940                 len = 0;
 941
 942         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 943                 if (i >= len) {
 944                         log_buf(out, ";0%%=0");
 945                         continue;
 946                 }
 947                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
 948         }
 949
 950         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 951                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 952         else
 953                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 954
 955         if (ovals)
 956                 free(ovals);
 957
 958         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
 959         if (bw_stat) {
 960                 double p_of_agg = 100.0;
 961
 962                 if (rs->agg[ddir]) {
 963                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 964                         if (p_of_agg > 100.0)
 965                                 p_of_agg = 100.0;
 966                 }
 967
 968                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 969         } else
 970                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
 971
 972         if (ver == 5) {
 973                 if (bw_stat)
 974                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
 975                 else
 976                         log_buf(out, ";%lu", 0UL);
 977
 978                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
 979                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
 980                                 mean, dev, (&ts->iops_stat[ddir])->samples);
 981                 else
 982                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
 983         }
 984 }
 985
 986 static void add_ddir_status_json(struct thread_stat *ts,
 987                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 988 {
 989         unsigned long long min, max, minv, maxv;
 990         unsigned long long bw_bytes, bw;
 991         unsigned long long *ovals = NULL;
 992         double mean, dev, iops;
 993         unsigned int len;
 994         int i;
 995         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object = NULL;
 996         char buf[120];
 997         double p_of_agg = 100.0;
 998
 999         assert(ddir_rw(ddir) || ddir_sync(ddir));
1000
1001         if (ts->unified_rw_rep && ddir != DDIR_READ)
1002                 return;
1003
1004         dir_object = json_create_object();
1005         json_object_add_value_object(parent,
1006                 ts->unified_rw_rep ? "mixed" : io_ddir_name(ddir), dir_object);
1007
1008         if (ddir_rw(ddir)) {
1009                 bw_bytes = 0;
1010                 bw = 0;
1011                 iops = 0.0;
1012                 if (ts->runtime[ddir]) {
1013                         uint64_t runt = ts->runtime[ddir];
1014
1015                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1016                         bw = bw_bytes / 1024; /* KiB/s */
1017                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1018                 }
1019
1020                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1021                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1022                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1023                 json_object_add_value_int(dir_object, "bw", bw);
1024                 json_object_add_value_float(dir_object, "iops", iops);
1025                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1026                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1027                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1028                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1029
1030                 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
1031                         min = max = 0;
1032                         mean = dev = 0.0;
1033                 }
1034                 tmp_object = json_create_object();
1035                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1036                 json_object_add_value_int(tmp_object, "min", min);
1037                 json_object_add_value_int(tmp_object, "max", max);
1038                 json_object_add_value_float(tmp_object, "mean", mean);
1039                 json_object_add_value_float(tmp_object, "stddev", dev);
1040
1041                 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1042                         min = max = 0;
1043                         mean = dev = 0.0;
1044                 }
1045                 tmp_object = json_create_object();
1046                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1047                 json_object_add_value_int(tmp_object, "min", min);
1048                 json_object_add_value_int(tmp_object, "max", max);
1049                 json_object_add_value_float(tmp_object, "mean", mean);
1050                 json_object_add_value_float(tmp_object, "stddev", dev);
1051         } else {
1052                 if (!calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
1053                         min = max = 0;
1054                         mean = dev = 0.0;
1055                 }
1056
1057                 tmp_object = json_create_object();
1058                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1059                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1060                 json_object_add_value_int(tmp_object, "min", min);
1061                 json_object_add_value_int(tmp_object, "max", max);
1062                 json_object_add_value_float(tmp_object, "mean", mean);
1063                 json_object_add_value_float(tmp_object, "stddev", dev);
1064         }
1065
1066         if (ts->clat_percentiles || ts->lat_percentiles) {
1067                 if (ddir_rw(ddir)) {
1068                         uint64_t samples;
1069
1070                         if (ts->clat_percentiles)
1071                                 samples = ts->clat_stat[ddir].samples;
1072                         else
1073                                 samples = ts->lat_stat[ddir].samples;
1074
1075                         len = calc_clat_percentiles(ts->io_u_plat[ddir],
1076                                         samples, ts->percentile_list, &ovals,
1077                                         &maxv, &minv);
1078                 } else {
1079                         len = calc_clat_percentiles(ts->io_u_sync_plat,
1080                                         ts->sync_stat.samples,
1081                                         ts->percentile_list, &ovals, &maxv,
1082                                         &minv);
1083                 }
1084
1085                 if (len > FIO_IO_U_LIST_MAX_LEN)
1086                         len = FIO_IO_U_LIST_MAX_LEN;
1087         } else
1088                 len = 0;
1089
1090         percentile_object = json_create_object();
1091         json_object_add_value_object(tmp_object, "percentile", percentile_object);
1092         for (i = 0; i < len; i++) {
1093                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1094                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1095         }
1096
1097         if (output_format & FIO_OUTPUT_JSON_PLUS) {
1098                 clat_bins_object = json_create_object();
1099                 if (ts->clat_percentiles)
1100                         json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1101
1102                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1103                         if (ddir_rw(ddir)) {
1104                                 if (ts->io_u_plat[ddir][i]) {
1105                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1106                                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1107                                 }
1108                         } else {
1109                                 if (ts->io_u_sync_plat[i]) {
1110                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1111                                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_sync_plat[i]);
1112                                 }
1113                         }
1114                 }
1115         }
1116
1117         if (!ddir_rw(ddir))
1118                 return;
1119
1120         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1121                 min = max = 0;
1122                 mean = dev = 0.0;
1123         }
1124         tmp_object = json_create_object();
1125         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1126         json_object_add_value_int(tmp_object, "min", min);
1127         json_object_add_value_int(tmp_object, "max", max);
1128         json_object_add_value_float(tmp_object, "mean", mean);
1129         json_object_add_value_float(tmp_object, "stddev", dev);
1130         if (output_format & FIO_OUTPUT_JSON_PLUS && ts->lat_percentiles)
1131                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1132
1133         if (ovals)
1134                 free(ovals);
1135
1136         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1137                 if (rs->agg[ddir]) {
1138                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1139                         if (p_of_agg > 100.0)
1140                                 p_of_agg = 100.0;
1141                 }
1142         } else {
1143                 min = max = 0;
1144                 p_of_agg = mean = dev = 0.0;
1145         }
1146         json_object_add_value_int(dir_object, "bw_min", min);
1147         json_object_add_value_int(dir_object, "bw_max", max);
1148         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1149         json_object_add_value_float(dir_object, "bw_mean", mean);
1150         json_object_add_value_float(dir_object, "bw_dev", dev);
1151         json_object_add_value_int(dir_object, "bw_samples",
1152                                 (&ts->bw_stat[ddir])->samples);
1153
1154         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1155                 min = max = 0;
1156                 mean = dev = 0.0;
1157         }
1158         json_object_add_value_int(dir_object, "iops_min", min);
1159         json_object_add_value_int(dir_object, "iops_max", max);
1160         json_object_add_value_float(dir_object, "iops_mean", mean);
1161         json_object_add_value_float(dir_object, "iops_stddev", dev);
1162         json_object_add_value_int(dir_object, "iops_samples",
1163                                 (&ts->iops_stat[ddir])->samples);
1164
1165         if (ts->cachehit + ts->cachemiss) {
1166                 uint64_t total;
1167                 double hit;
1168
1169                 total = ts->cachehit + ts->cachemiss;
1170                 hit = (double) ts->cachehit / (double) total;
1171                 hit *= 100.0;
1172                 json_object_add_value_float(dir_object, "cachehit", hit);
1173         }
1174 }
1175
1176 static void show_thread_status_terse_all(struct thread_stat *ts,
1177                                          struct group_run_stats *rs, int ver,
1178                                          struct buf_output *out)
1179 {
1180         double io_u_dist[FIO_IO_U_MAP_NR];
1181         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1182         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1183         double usr_cpu, sys_cpu;
1184         int i;
1185
1186         /* General Info */
1187         if (ver == 2)
1188                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1189         else
1190                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1191                         ts->name, ts->groupid, ts->error);
1192
1193         /* Log Read Status */
1194         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1195         /* Log Write Status */
1196         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1197         /* Log Trim Status */
1198         if (ver == 2 || ver == 4 || ver == 5)
1199                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1200
1201         /* CPU Usage */
1202         if (ts->total_run_time) {
1203                 double runt = (double) ts->total_run_time;
1204
1205                 usr_cpu = (double) ts->usr_time * 100 / runt;
1206                 sys_cpu = (double) ts->sys_time * 100 / runt;
1207         } else {
1208                 usr_cpu = 0;
1209                 sys_cpu = 0;
1210         }
1211
1212         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1213                                                 (unsigned long long) ts->ctx,
1214                                                 (unsigned long long) ts->majf,
1215                                                 (unsigned long long) ts->minf);
1216
1217         /* Calc % distribution of IO depths, usecond, msecond latency */
1218         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1219         stat_calc_lat_nu(ts, io_u_lat_u);
1220         stat_calc_lat_m(ts, io_u_lat_m);
1221
1222         /* Only show fixed 7 I/O depth levels*/
1223         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1224                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1225                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1226
1227         /* Microsecond latency */
1228         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1229                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1230         /* Millisecond latency */
1231         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1232                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1233
1234         /* disk util stats, if any */
1235         if (ver >= 3 && is_running_backend())
1236                 show_disk_util(1, NULL, out);
1237
1238         /* Additional output if continue_on_error set - default off*/
1239         if (ts->continue_on_error)
1240                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1241         if (ver == 2)
1242                 log_buf(out, "\n");
1243
1244         /* Additional output if description is set */
1245         if (strlen(ts->description))
1246                 log_buf(out, ";%s", ts->description);
1247
1248         log_buf(out, "\n");
1249 }
1250
1251 static void json_add_job_opts(struct json_object *root, const char *name,
1252                               struct flist_head *opt_list)
1253 {
1254         struct json_object *dir_object;
1255         struct flist_head *entry;
1256         struct print_option *p;
1257
1258         if (flist_empty(opt_list))
1259                 return;
1260
1261         dir_object = json_create_object();
1262         json_object_add_value_object(root, name, dir_object);
1263
1264         flist_for_each(entry, opt_list) {
1265                 const char *pos = "";
1266
1267                 p = flist_entry(entry, struct print_option, list);
1268                 if (p->value)
1269                         pos = p->value;
1270                 json_object_add_value_string(dir_object, p->name, pos);
1271         }
1272 }
1273
1274 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1275                                                    struct group_run_stats *rs,
1276                                                    struct flist_head *opt_list)
1277 {
1278         struct json_object *root, *tmp;
1279         struct jobs_eta *je;
1280         double io_u_dist[FIO_IO_U_MAP_NR];
1281         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1282         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1283         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1284         double usr_cpu, sys_cpu;
1285         int i;
1286         size_t size;
1287
1288         root = json_create_object();
1289         json_object_add_value_string(root, "jobname", ts->name);
1290         json_object_add_value_int(root, "groupid", ts->groupid);
1291         json_object_add_value_int(root, "error", ts->error);
1292
1293         /* ETA Info */
1294         je = get_jobs_eta(true, &size);
1295         if (je) {
1296                 json_object_add_value_int(root, "eta", je->eta_sec);
1297                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1298         }
1299
1300         if (opt_list)
1301                 json_add_job_opts(root, "job options", opt_list);
1302
1303         add_ddir_status_json(ts, rs, DDIR_READ, root);
1304         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1305         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1306         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1307
1308         /* CPU Usage */
1309         if (ts->total_run_time) {
1310                 double runt = (double) ts->total_run_time;
1311
1312                 usr_cpu = (double) ts->usr_time * 100 / runt;
1313                 sys_cpu = (double) ts->sys_time * 100 / runt;
1314         } else {
1315                 usr_cpu = 0;
1316                 sys_cpu = 0;
1317         }
1318         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1319         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1320         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1321         json_object_add_value_int(root, "ctx", ts->ctx);
1322         json_object_add_value_int(root, "majf", ts->majf);
1323         json_object_add_value_int(root, "minf", ts->minf);
1324
1325         /* Calc % distribution of IO depths */
1326         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1327         tmp = json_create_object();
1328         json_object_add_value_object(root, "iodepth_level", tmp);
1329         /* Only show fixed 7 I/O depth levels*/
1330         for (i = 0; i < 7; i++) {
1331                 char name[20];
1332                 if (i < 6)
1333                         snprintf(name, 20, "%d", 1 << i);
1334                 else
1335                         snprintf(name, 20, ">=%d", 1 << i);
1336                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1337         }
1338
1339         /* Calc % distribution of submit IO depths */
1340         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1341         tmp = json_create_object();
1342         json_object_add_value_object(root, "iodepth_submit", tmp);
1343         /* Only show fixed 7 I/O depth levels*/
1344         for (i = 0; i < 7; i++) {
1345                 char name[20];
1346                 if (i == 0)
1347                         snprintf(name, 20, "0");
1348                 else if (i < 6)
1349                         snprintf(name, 20, "%d", 1 << (i+1));
1350                 else
1351                         snprintf(name, 20, ">=%d", 1 << i);
1352                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1353         }
1354
1355         /* Calc % distribution of completion IO depths */
1356         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1357         tmp = json_create_object();
1358         json_object_add_value_object(root, "iodepth_complete", tmp);
1359         /* Only show fixed 7 I/O depth levels*/
1360         for (i = 0; i < 7; i++) {
1361                 char name[20];
1362                 if (i == 0)
1363                         snprintf(name, 20, "0");
1364                 else if (i < 6)
1365                         snprintf(name, 20, "%d", 1 << (i+1));
1366                 else
1367                         snprintf(name, 20, ">=%d", 1 << i);
1368                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1369         }
1370
1371         /* Calc % distribution of nsecond, usecond, msecond latency */
1372         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1373         stat_calc_lat_n(ts, io_u_lat_n);
1374         stat_calc_lat_u(ts, io_u_lat_u);
1375         stat_calc_lat_m(ts, io_u_lat_m);
1376
1377         /* Nanosecond latency */
1378         tmp = json_create_object();
1379         json_object_add_value_object(root, "latency_ns", tmp);
1380         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1381                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1382                                  "250", "500", "750", "1000", };
1383                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1384         }
1385         /* Microsecond latency */
1386         tmp = json_create_object();
1387         json_object_add_value_object(root, "latency_us", tmp);
1388         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1389                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1390                                  "250", "500", "750", "1000", };
1391                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1392         }
1393         /* Millisecond latency */
1394         tmp = json_create_object();
1395         json_object_add_value_object(root, "latency_ms", tmp);
1396         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1397                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1398                                  "250", "500", "750", "1000", "2000",
1399                                  ">=2000", };
1400                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1401         }
1402
1403         /* Additional output if continue_on_error set - default off*/
1404         if (ts->continue_on_error) {
1405                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1406                 json_object_add_value_int(root, "first_error", ts->first_error);
1407         }
1408
1409         if (ts->latency_depth) {
1410                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1411                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1412                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1413                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1414         }
1415
1416         /* Additional output if description is set */
1417         if (strlen(ts->description))
1418                 json_object_add_value_string(root, "desc", ts->description);
1419
1420         if (ts->nr_block_infos) {
1421                 /* Block error histogram and types */
1422                 int len;
1423                 unsigned int *percentiles = NULL;
1424                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1425
1426                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1427                                              ts->percentile_list,
1428                                              &percentiles, block_state_counts);
1429
1430                 if (len) {
1431                         struct json_object *block, *percentile_object, *states;
1432                         int state;
1433                         block = json_create_object();
1434                         json_object_add_value_object(root, "block", block);
1435
1436                         percentile_object = json_create_object();
1437                         json_object_add_value_object(block, "percentiles",
1438                                                      percentile_object);
1439                         for (i = 0; i < len; i++) {
1440                                 char buf[20];
1441                                 snprintf(buf, sizeof(buf), "%f",
1442                                          ts->percentile_list[i].u.f);
1443                                 json_object_add_value_int(percentile_object,
1444                                                           (const char *)buf,
1445                                                           percentiles[i]);
1446                         }
1447
1448                         states = json_create_object();
1449                         json_object_add_value_object(block, "states", states);
1450                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1451                                 json_object_add_value_int(states,
1452                                         block_state_names[state],
1453                                         block_state_counts[state]);
1454                         }
1455                         free(percentiles);
1456                 }
1457         }
1458
1459         if (ts->ss_dur) {
1460                 struct json_object *data;
1461                 struct json_array *iops, *bw;
1462                 int j, k, l;
1463                 char ss_buf[64];
1464
1465                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1466                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1467                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1468                         (float) ts->ss_limit.u.f,
1469                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1470
1471                 tmp = json_create_object();
1472                 json_object_add_value_object(root, "steadystate", tmp);
1473                 json_object_add_value_string(tmp, "ss", ss_buf);
1474                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1475                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1476
1477                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1478                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1479                 json_object_add_value_string(tmp, "criterion", ss_buf);
1480                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1481                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1482
1483                 data = json_create_object();
1484                 json_object_add_value_object(tmp, "data", data);
1485                 bw = json_create_array();
1486                 iops = json_create_array();
1487
1488                 /*
1489                 ** if ss was attained or the buffer is not full,
1490                 ** ss->head points to the first element in the list.
1491                 ** otherwise it actually points to the second element
1492                 ** in the list
1493                 */
1494                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1495                         j = ts->ss_head;
1496                 else
1497                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1498                 for (l = 0; l < ts->ss_dur; l++) {
1499                         k = (j + l) % ts->ss_dur;
1500                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1501                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1502                 }
1503                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1504                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1505                 json_object_add_value_array(data, "iops", iops);
1506                 json_object_add_value_array(data, "bw", bw);
1507         }
1508
1509         return root;
1510 }
1511
1512 static void show_thread_status_terse(struct thread_stat *ts,
1513                                      struct group_run_stats *rs,
1514                                      struct buf_output *out)
1515 {
1516         if (terse_version >= 2 && terse_version <= 5)
1517                 show_thread_status_terse_all(ts, rs, terse_version, out);
1518         else
1519                 log_err("fio: bad terse version!? %d\n", terse_version);
1520 }
1521
1522 struct json_object *show_thread_status(struct thread_stat *ts,
1523                                        struct group_run_stats *rs,
1524                                        struct flist_head *opt_list,
1525                                        struct buf_output *out)
1526 {
1527         struct json_object *ret = NULL;
1528
1529         if (output_format & FIO_OUTPUT_TERSE)
1530                 show_thread_status_terse(ts, rs,  out);
1531         if (output_format & FIO_OUTPUT_JSON)
1532                 ret = show_thread_status_json(ts, rs, opt_list);
1533         if (output_format & FIO_OUTPUT_NORMAL)
1534                 show_thread_status_normal(ts, rs,  out);
1535
1536         return ret;
1537 }
1538
1539 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1540 {
1541         double mean, S;
1542
1543         dst->min_val = min(dst->min_val, src->min_val);
1544         dst->max_val = max(dst->max_val, src->max_val);
1545
1546         /*
1547          * Compute new mean and S after the merge
1548          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1549          *  #Parallel_algorithm>
1550          */
1551         if (first) {
1552                 mean = src->mean.u.f;
1553                 S = src->S.u.f;
1554         } else {
1555                 double delta = src->mean.u.f - dst->mean.u.f;
1556
1557                 mean = ((src->mean.u.f * src->samples) +
1558                         (dst->mean.u.f * dst->samples)) /
1559                         (dst->samples + src->samples);
1560
1561                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1562                         (dst->samples * src->samples) /
1563                         (dst->samples + src->samples);
1564         }
1565
1566         dst->samples += src->samples;
1567         dst->mean.u.f = mean;
1568         dst->S.u.f = S;
1569
1570 }
1571
1572 /*
1573  * We sum two kinds of stats - one that is time based, in which case we
1574  * apply the proper summing technique, and then one that is iops/bw
1575  * numbers. For group_reporting, we should just add those up, not make
1576  * them the mean of everything.
1577  */
1578 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first,
1579                      bool pure_sum)
1580 {
1581         if (src->samples == 0)
1582                 return;
1583
1584         if (!pure_sum) {
1585                 __sum_stat(dst, src, first);
1586                 return;
1587         }
1588
1589         if (first) {
1590                 dst->min_val = src->min_val;
1591                 dst->max_val = src->max_val;
1592                 dst->samples = src->samples;
1593                 dst->mean.u.f = src->mean.u.f;
1594                 dst->S.u.f = src->S.u.f;
1595         } else {
1596                 dst->min_val += src->min_val;
1597                 dst->max_val += src->max_val;
1598                 dst->samples += src->samples;
1599                 dst->mean.u.f += src->mean.u.f;
1600                 dst->S.u.f += src->S.u.f;
1601         }
1602 }
1603
1604 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1605 {
1606         int i;
1607
1608         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1609                 if (dst->max_run[i] < src->max_run[i])
1610                         dst->max_run[i] = src->max_run[i];
1611                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1612                         dst->min_run[i] = src->min_run[i];
1613                 if (dst->max_bw[i] < src->max_bw[i])
1614                         dst->max_bw[i] = src->max_bw[i];
1615                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1616                         dst->min_bw[i] = src->min_bw[i];
1617
1618                 dst->iobytes[i] += src->iobytes[i];
1619                 dst->agg[i] += src->agg[i];
1620         }
1621
1622         if (!dst->kb_base)
1623                 dst->kb_base = src->kb_base;
1624         if (!dst->unit_base)
1625                 dst->unit_base = src->unit_base;
1626         if (!dst->sig_figs)
1627                 dst->sig_figs = src->sig_figs;
1628 }
1629
1630 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1631                       bool first)
1632 {
1633         int l, k;
1634
1635         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1636                 if (!dst->unified_rw_rep) {
1637                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first, false);
1638                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first, false);
1639                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first, false);
1640                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first, true);
1641                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first, true);
1642
1643                         dst->io_bytes[l] += src->io_bytes[l];
1644
1645                         if (dst->runtime[l] < src->runtime[l])
1646                                 dst->runtime[l] = src->runtime[l];
1647                 } else {
1648                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first, false);
1649                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first, false);
1650                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first, false);
1651                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first, true);
1652                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first, true);
1653
1654                         dst->io_bytes[0] += src->io_bytes[l];
1655
1656                         if (dst->runtime[0] < src->runtime[l])
1657                                 dst->runtime[0] = src->runtime[l];
1658
1659                         /*
1660                          * We're summing to the same destination, so override
1661                          * 'first' after the first iteration of the loop
1662                          */
1663                         first = false;
1664                 }
1665         }
1666
1667         sum_stat(&dst->sync_stat, &src->sync_stat, first, false);
1668         dst->usr_time += src->usr_time;
1669         dst->sys_time += src->sys_time;
1670         dst->ctx += src->ctx;
1671         dst->majf += src->majf;
1672         dst->minf += src->minf;
1673
1674         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
1675                 dst->io_u_map[k] += src->io_u_map[k];
1676                 dst->io_u_submit[k] += src->io_u_submit[k];
1677                 dst->io_u_complete[k] += src->io_u_complete[k];
1678         }
1679         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++) {
1680                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1681                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1682                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1683         }
1684         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1685                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
1686
1687         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1688                 if (!dst->unified_rw_rep) {
1689                         dst->total_io_u[k] += src->total_io_u[k];
1690                         dst->short_io_u[k] += src->short_io_u[k];
1691                         dst->drop_io_u[k] += src->drop_io_u[k];
1692                 } else {
1693                         dst->total_io_u[0] += src->total_io_u[k];
1694                         dst->short_io_u[0] += src->short_io_u[k];
1695                         dst->drop_io_u[0] += src->drop_io_u[k];
1696                 }
1697         }
1698
1699         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
1700
1701         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1702                 int m;
1703
1704                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1705                         if (!dst->unified_rw_rep)
1706                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1707                         else
1708                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1709                 }
1710         }
1711
1712         dst->total_run_time += src->total_run_time;
1713         dst->total_submit += src->total_submit;
1714         dst->total_complete += src->total_complete;
1715         dst->nr_zone_resets += src->nr_zone_resets;
1716         dst->cachehit += src->cachehit;
1717         dst->cachemiss += src->cachemiss;
1718 }
1719
1720 void init_group_run_stat(struct group_run_stats *gs)
1721 {
1722         int i;
1723         memset(gs, 0, sizeof(*gs));
1724
1725         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1726                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1727 }
1728
1729 void init_thread_stat(struct thread_stat *ts)
1730 {
1731         int j;
1732
1733         memset(ts, 0, sizeof(*ts));
1734
1735         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1736                 ts->lat_stat[j].min_val = -1UL;
1737                 ts->clat_stat[j].min_val = -1UL;
1738                 ts->slat_stat[j].min_val = -1UL;
1739                 ts->bw_stat[j].min_val = -1UL;
1740                 ts->iops_stat[j].min_val = -1UL;
1741         }
1742         ts->sync_stat.min_val = -1UL;
1743         ts->groupid = -1;
1744 }
1745
1746 void __show_run_stats(void)
1747 {
1748         struct group_run_stats *runstats, *rs;
1749         struct thread_data *td;
1750         struct thread_stat *threadstats, *ts;
1751         int i, j, k, nr_ts, last_ts, idx;
1752         bool kb_base_warned = false;
1753         bool unit_base_warned = false;
1754         struct json_object *root = NULL;
1755         struct json_array *array = NULL;
1756         struct buf_output output[FIO_OUTPUT_NR];
1757         struct flist_head **opt_lists;
1758
1759         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1760
1761         for (i = 0; i < groupid + 1; i++)
1762                 init_group_run_stat(&runstats[i]);
1763
1764         /*
1765          * find out how many threads stats we need. if group reporting isn't
1766          * enabled, it's one-per-td.
1767          */
1768         nr_ts = 0;
1769         last_ts = -1;
1770         for_each_td(td, i) {
1771                 if (!td->o.group_reporting) {
1772                         nr_ts++;
1773                         continue;
1774                 }
1775                 if (last_ts == td->groupid)
1776                         continue;
1777                 if (!td->o.stats)
1778                         continue;
1779
1780                 last_ts = td->groupid;
1781                 nr_ts++;
1782         }
1783
1784         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1785         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1786
1787         for (i = 0; i < nr_ts; i++) {
1788                 init_thread_stat(&threadstats[i]);
1789                 opt_lists[i] = NULL;
1790         }
1791
1792         j = 0;
1793         last_ts = -1;
1794         idx = 0;
1795         for_each_td(td, i) {
1796                 if (!td->o.stats)
1797                         continue;
1798                 if (idx && (!td->o.group_reporting ||
1799                     (td->o.group_reporting && last_ts != td->groupid))) {
1800                         idx = 0;
1801                         j++;
1802                 }
1803
1804                 last_ts = td->groupid;
1805
1806                 ts = &threadstats[j];
1807
1808                 ts->clat_percentiles = td->o.clat_percentiles;
1809                 ts->lat_percentiles = td->o.lat_percentiles;
1810                 ts->percentile_precision = td->o.percentile_precision;
1811                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1812                 opt_lists[j] = &td->opt_list;
1813
1814                 idx++;
1815                 ts->members++;
1816
1817                 if (ts->groupid == -1) {
1818                         /*
1819                          * These are per-group shared already
1820                          */
1821                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1822                         if (td->o.description)
1823                                 strncpy(ts->description, td->o.description,
1824                                                 FIO_JOBDESC_SIZE - 1);
1825                         else
1826                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1827
1828                         /*
1829                          * If multiple entries in this group, this is
1830                          * the first member.
1831                          */
1832                         ts->thread_number = td->thread_number;
1833                         ts->groupid = td->groupid;
1834
1835                         /*
1836                          * first pid in group, not very useful...
1837                          */
1838                         ts->pid = td->pid;
1839
1840                         ts->kb_base = td->o.kb_base;
1841                         ts->unit_base = td->o.unit_base;
1842                         ts->sig_figs = td->o.sig_figs;
1843                         ts->unified_rw_rep = td->o.unified_rw_rep;
1844                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1845                         log_info("fio: kb_base differs for jobs in group, using"
1846                                  " %u as the base\n", ts->kb_base);
1847                         kb_base_warned = true;
1848                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1849                         log_info("fio: unit_base differs for jobs in group, using"
1850                                  " %u as the base\n", ts->unit_base);
1851                         unit_base_warned = true;
1852                 }
1853
1854                 ts->continue_on_error = td->o.continue_on_error;
1855                 ts->total_err_count += td->total_err_count;
1856                 ts->first_error = td->first_error;
1857                 if (!ts->error) {
1858                         if (!td->error && td->o.continue_on_error &&
1859                             td->first_error) {
1860                                 ts->error = td->first_error;
1861                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1862                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1863                         } else  if (td->error) {
1864                                 ts->error = td->error;
1865                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1866                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1867                         }
1868                 }
1869
1870                 ts->latency_depth = td->latency_qd;
1871                 ts->latency_target = td->o.latency_target;
1872                 ts->latency_percentile = td->o.latency_percentile;
1873                 ts->latency_window = td->o.latency_window;
1874
1875                 ts->nr_block_infos = td->ts.nr_block_infos;
1876                 for (k = 0; k < ts->nr_block_infos; k++)
1877                         ts->block_infos[k] = td->ts.block_infos[k];
1878
1879                 sum_thread_stats(ts, &td->ts, idx == 1);
1880
1881                 if (td->o.ss_dur) {
1882                         ts->ss_state = td->ss.state;
1883                         ts->ss_dur = td->ss.dur;
1884                         ts->ss_head = td->ss.head;
1885                         ts->ss_bw_data = td->ss.bw_data;
1886                         ts->ss_iops_data = td->ss.iops_data;
1887                         ts->ss_limit.u.f = td->ss.limit;
1888                         ts->ss_slope.u.f = td->ss.slope;
1889                         ts->ss_deviation.u.f = td->ss.deviation;
1890                         ts->ss_criterion.u.f = td->ss.criterion;
1891                 }
1892                 else
1893                         ts->ss_dur = ts->ss_state = 0;
1894         }
1895
1896         for (i = 0; i < nr_ts; i++) {
1897                 unsigned long long bw;
1898
1899                 ts = &threadstats[i];
1900                 if (ts->groupid == -1)
1901                         continue;
1902                 rs = &runstats[ts->groupid];
1903                 rs->kb_base = ts->kb_base;
1904                 rs->unit_base = ts->unit_base;
1905                 rs->sig_figs = ts->sig_figs;
1906                 rs->unified_rw_rep += ts->unified_rw_rep;
1907
1908                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1909                         if (!ts->runtime[j])
1910                                 continue;
1911                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1912                                 rs->min_run[j] = ts->runtime[j];
1913                         if (ts->runtime[j] > rs->max_run[j])
1914                                 rs->max_run[j] = ts->runtime[j];
1915
1916                         bw = 0;
1917                         if (ts->runtime[j])
1918                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1919                         if (bw < rs->min_bw[j])
1920                                 rs->min_bw[j] = bw;
1921                         if (bw > rs->max_bw[j])
1922                                 rs->max_bw[j] = bw;
1923
1924                         rs->iobytes[j] += ts->io_bytes[j];
1925                 }
1926         }
1927
1928         for (i = 0; i < groupid + 1; i++) {
1929                 int ddir;
1930
1931                 rs = &runstats[i];
1932
1933                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1934                         if (rs->max_run[ddir])
1935                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1936                                                 rs->max_run[ddir];
1937                 }
1938         }
1939
1940         for (i = 0; i < FIO_OUTPUT_NR; i++)
1941                 buf_output_init(&output[i]);
1942
1943         /*
1944          * don't overwrite last signal output
1945          */
1946         if (output_format & FIO_OUTPUT_NORMAL)
1947                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1948         if (output_format & FIO_OUTPUT_JSON) {
1949                 struct thread_data *global;
1950                 char time_buf[32];
1951                 struct timeval now;
1952                 unsigned long long ms_since_epoch;
1953                 time_t tv_sec;
1954
1955                 gettimeofday(&now, NULL);
1956                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1957                                  (unsigned long long)(now.tv_usec) / 1000;
1958
1959                 tv_sec = now.tv_sec;
1960                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
1961                 if (time_buf[strlen(time_buf) - 1] == '\n')
1962                         time_buf[strlen(time_buf) - 1] = '\0';
1963
1964                 root = json_create_object();
1965                 json_object_add_value_string(root, "fio version", fio_version_string);
1966                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1967                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1968                 json_object_add_value_string(root, "time", time_buf);
1969                 global = get_global_options();
1970                 json_add_job_opts(root, "global options", &global->opt_list);
1971                 array = json_create_array();
1972                 json_object_add_value_array(root, "jobs", array);
1973         }
1974
1975         if (is_backend)
1976                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1977
1978         for (i = 0; i < nr_ts; i++) {
1979                 ts = &threadstats[i];
1980                 rs = &runstats[ts->groupid];
1981
1982                 if (is_backend) {
1983                         fio_server_send_job_options(opt_lists[i], i);
1984                         fio_server_send_ts(ts, rs);
1985                 } else {
1986                         if (output_format & FIO_OUTPUT_TERSE)
1987                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1988                         if (output_format & FIO_OUTPUT_JSON) {
1989                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1990                                 json_array_add_value_object(array, tmp);
1991                         }
1992                         if (output_format & FIO_OUTPUT_NORMAL)
1993                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1994                 }
1995         }
1996         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1997                 /* disk util stats, if any */
1998                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1999
2000                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2001
2002                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2003                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2004                 json_free_object(root);
2005         }
2006
2007         for (i = 0; i < groupid + 1; i++) {
2008                 rs = &runstats[i];
2009
2010                 rs->groupid = i;
2011                 if (is_backend)
2012                         fio_server_send_gs(rs);
2013                 else if (output_format & FIO_OUTPUT_NORMAL)
2014                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2015         }
2016
2017         if (is_backend)
2018                 fio_server_send_du();
2019         else if (output_format & FIO_OUTPUT_NORMAL) {
2020                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2021                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2022         }
2023
2024         for (i = 0; i < FIO_OUTPUT_NR; i++) {
2025                 struct buf_output *out = &output[i];
2026
2027                 log_info_buf(out->buf, out->buflen);
2028                 buf_output_free(out);
2029         }
2030
2031         fio_idle_prof_cleanup();
2032
2033         log_info_flush();
2034         free(runstats);
2035         free(threadstats);
2036         free(opt_lists);
2037 }
2038
2039 void __show_running_run_stats(void)
2040 {
2041         struct thread_data *td;
2042         unsigned long long *rt;
2043         struct timespec ts;
2044         int i;
2045
2046         fio_sem_down(stat_sem);
2047
2048         rt = malloc(thread_number * sizeof(unsigned long long));
2049         fio_gettime(&ts, NULL);
2050
2051         for_each_td(td, i) {
2052                 td->update_rusage = 1;
2053                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
2054                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
2055                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2056                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2057
2058                 rt[i] = mtime_since(&td->start, &ts);
2059                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2060                         td->ts.runtime[DDIR_READ] += rt[i];
2061                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2062                         td->ts.runtime[DDIR_WRITE] += rt[i];
2063                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2064                         td->ts.runtime[DDIR_TRIM] += rt[i];
2065         }
2066
2067         for_each_td(td, i) {
2068                 if (td->runstate >= TD_EXITED)
2069                         continue;
2070                 if (td->rusage_sem) {
2071                         td->update_rusage = 1;
2072                         fio_sem_down(td->rusage_sem);
2073                 }
2074                 td->update_rusage = 0;
2075         }
2076
2077         __show_run_stats();
2078
2079         for_each_td(td, i) {
2080                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2081                         td->ts.runtime[DDIR_READ] -= rt[i];
2082                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2083                         td->ts.runtime[DDIR_WRITE] -= rt[i];
2084                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2085                         td->ts.runtime[DDIR_TRIM] -= rt[i];
2086         }
2087
2088         free(rt);
2089         fio_sem_up(stat_sem);
2090 }
2091
2092 static bool status_interval_init;
2093 static struct timespec status_time;
2094 static bool status_file_disabled;
2095
2096 #define FIO_STATUS_FILE         "fio-dump-status"
2097
2098 static int check_status_file(void)
2099 {
2100         struct stat sb;
2101         const char *temp_dir;
2102         char fio_status_file_path[PATH_MAX];
2103
2104         if (status_file_disabled)
2105                 return 0;
2106
2107         temp_dir = getenv("TMPDIR");
2108         if (temp_dir == NULL) {
2109                 temp_dir = getenv("TEMP");
2110                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2111                         temp_dir = NULL;
2112         }
2113         if (temp_dir == NULL)
2114                 temp_dir = "/tmp";
2115
2116         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2117
2118         if (stat(fio_status_file_path, &sb))
2119                 return 0;
2120
2121         if (unlink(fio_status_file_path) < 0) {
2122                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2123                                                         strerror(errno));
2124                 log_err("fio: disabling status file updates\n");
2125                 status_file_disabled = true;
2126         }
2127
2128         return 1;
2129 }
2130
2131 void check_for_running_stats(void)
2132 {
2133         if (status_interval) {
2134                 if (!status_interval_init) {
2135                         fio_gettime(&status_time, NULL);
2136                         status_interval_init = true;
2137                 } else if (mtime_since_now(&status_time) >= status_interval) {
2138                         show_running_run_stats();
2139                         fio_gettime(&status_time, NULL);
2140                         return;
2141                 }
2142         }
2143         if (check_status_file()) {
2144                 show_running_run_stats();
2145                 return;
2146         }
2147 }
2148
2149 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2150 {
2151         double val = data;
2152         double delta;
2153
2154         if (data > is->max_val)
2155                 is->max_val = data;
2156         if (data < is->min_val)
2157                 is->min_val = data;
2158
2159         delta = val - is->mean.u.f;
2160         if (delta) {
2161                 is->mean.u.f += delta / (is->samples + 1.0);
2162                 is->S.u.f += delta * (val - is->mean.u.f);
2163         }
2164
2165         is->samples++;
2166 }
2167
2168 /*
2169  * Return a struct io_logs, which is added to the tail of the log
2170  * list for 'iolog'.
2171  */
2172 static struct io_logs *get_new_log(struct io_log *iolog)
2173 {
2174         size_t new_size, new_samples;
2175         struct io_logs *cur_log;
2176
2177         /*
2178          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2179          * forever
2180          */
2181         if (!iolog->cur_log_max)
2182                 new_samples = DEF_LOG_ENTRIES;
2183         else {
2184                 new_samples = iolog->cur_log_max * 2;
2185                 if (new_samples > MAX_LOG_ENTRIES)
2186                         new_samples = MAX_LOG_ENTRIES;
2187         }
2188
2189         new_size = new_samples * log_entry_sz(iolog);
2190
2191         cur_log = smalloc(sizeof(*cur_log));
2192         if (cur_log) {
2193                 INIT_FLIST_HEAD(&cur_log->list);
2194                 cur_log->log = malloc(new_size);
2195                 if (cur_log->log) {
2196                         cur_log->nr_samples = 0;
2197                         cur_log->max_samples = new_samples;
2198                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2199                         iolog->cur_log_max = new_samples;
2200                         return cur_log;
2201                 }
2202                 sfree(cur_log);
2203         }
2204
2205         return NULL;
2206 }
2207
2208 /*
2209  * Add and return a new log chunk, or return current log if big enough
2210  */
2211 static struct io_logs *regrow_log(struct io_log *iolog)
2212 {
2213         struct io_logs *cur_log;
2214         int i;
2215
2216         if (!iolog || iolog->disabled)
2217                 goto disable;
2218
2219         cur_log = iolog_cur_log(iolog);
2220         if (!cur_log) {
2221                 cur_log = get_new_log(iolog);
2222                 if (!cur_log)
2223                         return NULL;
2224         }
2225
2226         if (cur_log->nr_samples < cur_log->max_samples)
2227                 return cur_log;
2228
2229         /*
2230          * No room for a new sample. If we're compressing on the fly, flush
2231          * out the current chunk
2232          */
2233         if (iolog->log_gz) {
2234                 if (iolog_cur_flush(iolog, cur_log)) {
2235                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2236                         return NULL;
2237                 }
2238         }
2239
2240         /*
2241          * Get a new log array, and add to our list
2242          */
2243         cur_log = get_new_log(iolog);
2244         if (!cur_log) {
2245                 log_err("fio: failed extending iolog! Will stop logging.\n");
2246                 return NULL;
2247         }
2248
2249         if (!iolog->pending || !iolog->pending->nr_samples)
2250                 return cur_log;
2251
2252         /*
2253          * Flush pending items to new log
2254          */
2255         for (i = 0; i < iolog->pending->nr_samples; i++) {
2256                 struct io_sample *src, *dst;
2257
2258                 src = get_sample(iolog, iolog->pending, i);
2259                 dst = get_sample(iolog, cur_log, i);
2260                 memcpy(dst, src, log_entry_sz(iolog));
2261         }
2262         cur_log->nr_samples = iolog->pending->nr_samples;
2263
2264         iolog->pending->nr_samples = 0;
2265         return cur_log;
2266 disable:
2267         if (iolog)
2268                 iolog->disabled = true;
2269         return NULL;
2270 }
2271
2272 void regrow_logs(struct thread_data *td)
2273 {
2274         regrow_log(td->slat_log);
2275         regrow_log(td->clat_log);
2276         regrow_log(td->clat_hist_log);
2277         regrow_log(td->lat_log);
2278         regrow_log(td->bw_log);
2279         regrow_log(td->iops_log);
2280         td->flags &= ~TD_F_REGROW_LOGS;
2281 }
2282
2283 static struct io_logs *get_cur_log(struct io_log *iolog)
2284 {
2285         struct io_logs *cur_log;
2286
2287         cur_log = iolog_cur_log(iolog);
2288         if (!cur_log) {
2289                 cur_log = get_new_log(iolog);
2290                 if (!cur_log)
2291                         return NULL;
2292         }
2293
2294         if (cur_log->nr_samples < cur_log->max_samples)
2295                 return cur_log;
2296
2297         /*
2298          * Out of space. If we're in IO offload mode, or we're not doing
2299          * per unit logging (hence logging happens outside of the IO thread
2300          * as well), add a new log chunk inline. If we're doing inline
2301          * submissions, flag 'td' as needing a log regrow and we'll take
2302          * care of it on the submission side.
2303          */
2304         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2305             !per_unit_log(iolog))
2306                 return regrow_log(iolog);
2307
2308         if (iolog->td)
2309                 iolog->td->flags |= TD_F_REGROW_LOGS;
2310         if (iolog->pending)
2311                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2312         return iolog->pending;
2313 }
2314
2315 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2316                              enum fio_ddir ddir, unsigned long long bs,
2317                              unsigned long t, uint64_t offset)
2318 {
2319         struct io_logs *cur_log;
2320
2321         if (iolog->disabled)
2322                 return;
2323         if (flist_empty(&iolog->io_logs))
2324                 iolog->avg_last[ddir] = t;
2325
2326         cur_log = get_cur_log(iolog);
2327         if (cur_log) {
2328                 struct io_sample *s;
2329
2330                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2331
2332                 s->data = data;
2333                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2334                 io_sample_set_ddir(iolog, s, ddir);
2335                 s->bs = bs;
2336
2337                 if (iolog->log_offset) {
2338                         struct io_sample_offset *so = (void *) s;
2339
2340                         so->offset = offset;
2341                 }
2342
2343                 cur_log->nr_samples++;
2344                 return;
2345         }
2346
2347         iolog->disabled = true;
2348 }
2349
2350 static inline void reset_io_stat(struct io_stat *ios)
2351 {
2352         ios->min_val = -1ULL;
2353         ios->max_val = ios->samples = 0;
2354         ios->mean.u.f = ios->S.u.f = 0;
2355 }
2356
2357 void reset_io_stats(struct thread_data *td)
2358 {
2359         struct thread_stat *ts = &td->ts;
2360         int i, j;
2361
2362         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2363                 reset_io_stat(&ts->clat_stat[i]);
2364                 reset_io_stat(&ts->slat_stat[i]);
2365                 reset_io_stat(&ts->lat_stat[i]);
2366                 reset_io_stat(&ts->bw_stat[i]);
2367                 reset_io_stat(&ts->iops_stat[i]);
2368
2369                 ts->io_bytes[i] = 0;
2370                 ts->runtime[i] = 0;
2371                 ts->total_io_u[i] = 0;
2372                 ts->short_io_u[i] = 0;
2373                 ts->drop_io_u[i] = 0;
2374
2375                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2376                         ts->io_u_plat[i][j] = 0;
2377                         if (!i)
2378                                 ts->io_u_sync_plat[j] = 0;
2379                 }
2380         }
2381
2382         ts->total_io_u[DDIR_SYNC] = 0;
2383
2384         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2385                 ts->io_u_map[i] = 0;
2386                 ts->io_u_submit[i] = 0;
2387                 ts->io_u_complete[i] = 0;
2388         }
2389
2390         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2391                 ts->io_u_lat_n[i] = 0;
2392         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2393                 ts->io_u_lat_u[i] = 0;
2394         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2395                 ts->io_u_lat_m[i] = 0;
2396
2397         ts->total_submit = 0;
2398         ts->total_complete = 0;
2399         ts->nr_zone_resets = 0;
2400         ts->cachehit = ts->cachemiss = 0;
2401 }
2402
2403 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2404                               unsigned long elapsed, bool log_max)
2405 {
2406         /*
2407          * Note an entry in the log. Use the mean from the logged samples,
2408          * making sure to properly round up. Only write a log entry if we
2409          * had actual samples done.
2410          */
2411         if (iolog->avg_window[ddir].samples) {
2412                 union io_sample_data data;
2413
2414                 if (log_max)
2415                         data.val = iolog->avg_window[ddir].max_val;
2416                 else
2417                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2418
2419                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2420         }
2421
2422         reset_io_stat(&iolog->avg_window[ddir]);
2423 }
2424
2425 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2426                              bool log_max)
2427 {
2428         int ddir;
2429
2430         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2431                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2432 }
2433
2434 static unsigned long add_log_sample(struct thread_data *td,
2435                                     struct io_log *iolog,
2436                                     union io_sample_data data,
2437                                     enum fio_ddir ddir, unsigned long long bs,
2438                                     uint64_t offset)
2439 {
2440         unsigned long elapsed, this_window;
2441
2442         if (!ddir_rw(ddir))
2443                 return 0;
2444
2445         elapsed = mtime_since_now(&td->epoch);
2446
2447         /*
2448          * If no time averaging, just add the log sample.
2449          */
2450         if (!iolog->avg_msec) {
2451                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2452                 return 0;
2453         }
2454
2455         /*
2456          * Add the sample. If the time period has passed, then
2457          * add that entry to the log and clear.
2458          */
2459         add_stat_sample(&iolog->avg_window[ddir], data.val);
2460
2461         /*
2462          * If period hasn't passed, adding the above sample is all we
2463          * need to do.
2464          */
2465         this_window = elapsed - iolog->avg_last[ddir];
2466         if (elapsed < iolog->avg_last[ddir])
2467                 return iolog->avg_last[ddir] - elapsed;
2468         else if (this_window < iolog->avg_msec) {
2469                 unsigned long diff = iolog->avg_msec - this_window;
2470
2471                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2472                         return diff;
2473         }
2474
2475         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2476
2477         iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2478         return iolog->avg_msec;
2479 }
2480
2481 void finalize_logs(struct thread_data *td, bool unit_logs)
2482 {
2483         unsigned long elapsed;
2484
2485         elapsed = mtime_since_now(&td->epoch);
2486
2487         if (td->clat_log && unit_logs)
2488                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2489         if (td->slat_log && unit_logs)
2490                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2491         if (td->lat_log && unit_logs)
2492                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2493         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2494                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2495         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2496                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2497 }
2498
2499 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned long long bs)
2500 {
2501         struct io_log *iolog;
2502
2503         if (!ddir_rw(ddir))
2504                 return;
2505
2506         iolog = agg_io_log[ddir];
2507         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2508 }
2509
2510 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2511 {
2512         unsigned int idx = plat_val_to_idx(nsec);
2513         assert(idx < FIO_IO_U_PLAT_NR);
2514
2515         ts->io_u_sync_plat[idx]++;
2516         add_stat_sample(&ts->sync_stat, nsec);
2517 }
2518
2519 static void add_clat_percentile_sample(struct thread_stat *ts,
2520                                 unsigned long long nsec, enum fio_ddir ddir)
2521 {
2522         unsigned int idx = plat_val_to_idx(nsec);
2523         assert(idx < FIO_IO_U_PLAT_NR);
2524
2525         ts->io_u_plat[ddir][idx]++;
2526 }
2527
2528 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2529                      unsigned long long nsec, unsigned long long bs,
2530                      uint64_t offset)
2531 {
2532         const bool needs_lock = td_async_processing(td);
2533         unsigned long elapsed, this_window;
2534         struct thread_stat *ts = &td->ts;
2535         struct io_log *iolog = td->clat_hist_log;
2536
2537         if (needs_lock)
2538                 __td_io_u_lock(td);
2539
2540         add_stat_sample(&ts->clat_stat[ddir], nsec);
2541
2542         if (td->clat_log)
2543                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2544                                offset);
2545
2546         if (ts->clat_percentiles)
2547                 add_clat_percentile_sample(ts, nsec, ddir);
2548
2549         if (iolog && iolog->hist_msec) {
2550                 struct io_hist *hw = &iolog->hist_window[ddir];
2551
2552                 hw->samples++;
2553                 elapsed = mtime_since_now(&td->epoch);
2554                 if (!hw->hist_last)
2555                         hw->hist_last = elapsed;
2556                 this_window = elapsed - hw->hist_last;
2557
2558                 if (this_window >= iolog->hist_msec) {
2559                         uint64_t *io_u_plat;
2560                         struct io_u_plat_entry *dst;
2561
2562                         /*
2563                          * Make a byte-for-byte copy of the latency histogram
2564                          * stored in td->ts.io_u_plat[ddir], recording it in a
2565                          * log sample. Note that the matching call to free() is
2566                          * located in iolog.c after printing this sample to the
2567                          * log file.
2568                          */
2569                         io_u_plat = (uint64_t *) td->ts.io_u_plat[ddir];
2570                         dst = malloc(sizeof(struct io_u_plat_entry));
2571                         memcpy(&(dst->io_u_plat), io_u_plat,
2572                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2573                         flist_add(&dst->list, &hw->list);
2574                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2575                                                 elapsed, offset);
2576
2577                         /*
2578                          * Update the last time we recorded as being now, minus
2579                          * any drift in time we encountered before actually
2580                          * making the record.
2581                          */
2582                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2583                         hw->samples = 0;
2584                 }
2585         }
2586
2587         if (needs_lock)
2588                 __td_io_u_unlock(td);
2589 }
2590
2591 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2592                      unsigned long usec, unsigned long long bs, uint64_t offset)
2593 {
2594         const bool needs_lock = td_async_processing(td);
2595         struct thread_stat *ts = &td->ts;
2596
2597         if (!ddir_rw(ddir))
2598                 return;
2599
2600         if (needs_lock)
2601                 __td_io_u_lock(td);
2602
2603         add_stat_sample(&ts->slat_stat[ddir], usec);
2604
2605         if (td->slat_log)
2606                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2607
2608         if (needs_lock)
2609                 __td_io_u_unlock(td);
2610 }
2611
2612 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2613                     unsigned long long nsec, unsigned long long bs,
2614                     uint64_t offset)
2615 {
2616         const bool needs_lock = td_async_processing(td);
2617         struct thread_stat *ts = &td->ts;
2618
2619         if (!ddir_rw(ddir))
2620                 return;
2621
2622         if (needs_lock)
2623                 __td_io_u_lock(td);
2624
2625         add_stat_sample(&ts->lat_stat[ddir], nsec);
2626
2627         if (td->lat_log)
2628                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2629                                offset);
2630
2631         if (ts->lat_percentiles)
2632                 add_clat_percentile_sample(ts, nsec, ddir);
2633
2634         if (needs_lock)
2635                 __td_io_u_unlock(td);
2636 }
2637
2638 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2639                    unsigned int bytes, unsigned long long spent)
2640 {
2641         const bool needs_lock = td_async_processing(td);
2642         struct thread_stat *ts = &td->ts;
2643         unsigned long rate;
2644
2645         if (spent)
2646                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2647         else
2648                 rate = 0;
2649
2650         if (needs_lock)
2651                 __td_io_u_lock(td);
2652
2653         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2654
2655         if (td->bw_log)
2656                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2657                                bytes, io_u->offset);
2658
2659         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2660
2661         if (needs_lock)
2662                 __td_io_u_unlock(td);
2663 }
2664
2665 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2666                          struct timespec *t, unsigned int avg_time,
2667                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2668                          struct io_stat *stat, struct io_log *log,
2669                          bool is_kb)
2670 {
2671         const bool needs_lock = td_async_processing(td);
2672         unsigned long spent, rate;
2673         enum fio_ddir ddir;
2674         unsigned long next, next_log;
2675
2676         next_log = avg_time;
2677
2678         spent = mtime_since(parent_tv, t);
2679         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2680                 return avg_time - spent;
2681
2682         if (needs_lock)
2683                 __td_io_u_lock(td);
2684
2685         /*
2686          * Compute both read and write rates for the interval.
2687          */
2688         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2689                 uint64_t delta;
2690
2691                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2692                 if (!delta)
2693                         continue; /* No entries for interval */
2694
2695                 if (spent) {
2696                         if (is_kb)
2697                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2698                         else
2699                                 rate = (delta * 1000) / spent;
2700                 } else
2701                         rate = 0;
2702
2703                 add_stat_sample(&stat[ddir], rate);
2704
2705                 if (log) {
2706                         unsigned long long bs = 0;
2707
2708                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2709                                 bs = td->o.min_bs[ddir];
2710
2711                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2712                         next_log = min(next_log, next);
2713                 }
2714
2715                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2716         }
2717
2718         timespec_add_msec(parent_tv, avg_time);
2719
2720         if (needs_lock)
2721                 __td_io_u_unlock(td);
2722
2723         if (spent <= avg_time)
2724                 next = avg_time;
2725         else
2726                 next = avg_time - (1 + spent - avg_time);
2727
2728         return min(next, next_log);
2729 }
2730
2731 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2732 {
2733         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2734                                 td->this_io_bytes, td->stat_io_bytes,
2735                                 td->ts.bw_stat, td->bw_log, true);
2736 }
2737
2738 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2739                      unsigned int bytes)
2740 {
2741         const bool needs_lock = td_async_processing(td);
2742         struct thread_stat *ts = &td->ts;
2743
2744         if (needs_lock)
2745                 __td_io_u_lock(td);
2746
2747         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2748
2749         if (td->iops_log)
2750                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2751                                bytes, io_u->offset);
2752
2753         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2754
2755         if (needs_lock)
2756                 __td_io_u_unlock(td);
2757 }
2758
2759 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2760 {
2761         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2762                                 td->this_io_blocks, td->stat_io_blocks,
2763                                 td->ts.iops_stat, td->iops_log, false);
2764 }
2765
2766 /*
2767  * Returns msecs to next event
2768  */
2769 int calc_log_samples(void)
2770 {
2771         struct thread_data *td;
2772         unsigned int next = ~0U, tmp;
2773         struct timespec now;
2774         int i;
2775
2776         fio_gettime(&now, NULL);
2777
2778         for_each_td(td, i) {
2779                 if (!td->o.stats)
2780                         continue;
2781                 if (in_ramp_time(td) ||
2782                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2783                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2784                         continue;
2785                 }
2786                 if (!td->bw_log ||
2787                         (td->bw_log && !per_unit_log(td->bw_log))) {
2788                         tmp = add_bw_samples(td, &now);
2789                         if (tmp < next)
2790                                 next = tmp;
2791                 }
2792                 if (!td->iops_log ||
2793                         (td->iops_log && !per_unit_log(td->iops_log))) {
2794                         tmp = add_iops_samples(td, &now);
2795                         if (tmp < next)
2796                                 next = tmp;
2797                 }
2798         }
2799
2800         return next == ~0U ? 0 : next;
2801 }
2802
2803 void stat_init(void)
2804 {
2805         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
2806 }
2807
2808 void stat_exit(void)
2809 {
2810         /*
2811          * When we have the mutex, we know out-of-band access to it
2812          * have ended.
2813          */
2814         fio_sem_down(stat_sem);
2815         fio_sem_remove(stat_sem);
2816 }
2817
2818 /*
2819  * Called from signal handler. Wake up status thread.
2820  */
2821 void show_running_run_stats(void)
2822 {
2823         helper_do_stat();
2824 }
2825
2826 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2827 {
2828         /* Ignore io_u's which span multiple blocks--they will just get
2829          * inaccurate counts. */
2830         int idx = (io_u->offset - io_u->file->file_offset)
2831                         / td->o.bs[DDIR_TRIM];
2832         uint32_t *info = &td->ts.block_infos[idx];
2833         assert(idx < td->ts.nr_block_infos);
2834         return info;
2835 }