stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/stat.h>
   5 #include <math.h>
   6
   7 #include "fio.h"
   8 #include "diskutil.h"
   9 #include "lib/ieee754.h"
  10 #include "json.h"
  11 #include "lib/getrusage.h"
  12 #include "idletime.h"
  13 #include "lib/pow2.h"
  14 #include "lib/output_buffer.h"
  15 #include "helper_thread.h"
  16 #include "smalloc.h"
  17 #include "zbd.h"
  18
  19 #define LOG_MSEC_SLACK  1
  20
  21 struct fio_sem *stat_sem;
  22
  23 void clear_rusage_stat(struct thread_data *td)
  24 {
  25         struct thread_stat *ts = &td->ts;
  26
  27         fio_getrusage(&td->ru_start);
  28         ts->usr_time = ts->sys_time = 0;
  29         ts->ctx = 0;
  30         ts->minf = ts->majf = 0;
  31 }
  32
  33 void update_rusage_stat(struct thread_data *td)
  34 {
  35         struct thread_stat *ts = &td->ts;
  36
  37         fio_getrusage(&td->ru_end);
  38         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
  39                                         &td->ru_end.ru_utime);
  40         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
  41                                         &td->ru_end.ru_stime);
  42         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  43                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  44         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  45         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  46
  47         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  48 }
  49
  50 /*
  51  * Given a latency, return the index of the corresponding bucket in
  52  * the structure tracking percentiles.
  53  *
  54  * (1) find the group (and error bits) that the value (latency)
  55  * belongs to by looking at its MSB. (2) find the bucket number in the
  56  * group by looking at the index bits.
  57  *
  58  */
  59 static unsigned int plat_val_to_idx(unsigned long long val)
  60 {
  61         unsigned int msb, error_bits, base, offset, idx;
  62
  63         /* Find MSB starting from bit 0 */
  64         if (val == 0)
  65                 msb = 0;
  66         else
  67                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
  68
  69         /*
  70          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  71          * all bits of the sample as index
  72          */
  73         if (msb <= FIO_IO_U_PLAT_BITS)
  74                 return val;
  75
  76         /* Compute the number of error bits to discard*/
  77         error_bits = msb - FIO_IO_U_PLAT_BITS;
  78
  79         /* Compute the number of buckets before the group */
  80         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  81
  82         /*
  83          * Discard the error bits and apply the mask to find the
  84          * index for the buckets in the group
  85          */
  86         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  87
  88         /* Make sure the index does not exceed (array size - 1) */
  89         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  90                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  91
  92         return idx;
  93 }
  94
  95 /*
  96  * Convert the given index of the bucket array to the value
  97  * represented by the bucket
  98  */
  99 static unsigned long long plat_idx_to_val(unsigned int idx)
 100 {
 101         unsigned int error_bits;
 102         unsigned long long k, base;
 103
 104         assert(idx < FIO_IO_U_PLAT_NR);
 105
 106         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 107          * all bits of the sample as index */
 108         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 109                 return idx;
 110
 111         /* Find the group and compute the minimum value of that group */
 112         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 113         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
 114
 115         /* Find its bucket number of the group */
 116         k = idx % FIO_IO_U_PLAT_VAL;
 117
 118         /* Return the mean of the range of the bucket */
 119         return base + ((k + 0.5) * (1 << error_bits));
 120 }
 121
 122 static int double_cmp(const void *a, const void *b)
 123 {
 124         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 125         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 126         int cmp = 0;
 127
 128         if (fa.u.f > fb.u.f)
 129                 cmp = 1;
 130         else if (fa.u.f < fb.u.f)
 131                 cmp = -1;
 132
 133         return cmp;
 134 }
 135
 136 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 137                                    fio_fp64_t *plist, unsigned long long **output,
 138                                    unsigned long long *maxv, unsigned long long *minv)
 139 {
 140         unsigned long long sum = 0;
 141         unsigned int len, i, j = 0;
 142         unsigned long long *ovals = NULL;
 143         bool is_last;
 144
 145         *minv = -1ULL;
 146         *maxv = 0;
 147
 148         len = 0;
 149         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 150                 len++;
 151
 152         if (!len)
 153                 return 0;
 154
 155         /*
 156          * Sort the percentile list. Note that it may already be sorted if
 157          * we are using the default values, but since it's a short list this
 158          * isn't a worry. Also note that this does not work for NaN values.
 159          */
 160         if (len > 1)
 161                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 162
 163         ovals = malloc(len * sizeof(*ovals));
 164         if (!ovals)
 165                 return 0;
 166
 167         /*
 168          * Calculate bucket values, note down max and min values
 169          */
 170         is_last = false;
 171         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 172                 sum += io_u_plat[i];
 173                 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
 174                         assert(plist[j].u.f <= 100.0);
 175
 176                         ovals[j] = plat_idx_to_val(i);
 177                         if (ovals[j] < *minv)
 178                                 *minv = ovals[j];
 179                         if (ovals[j] > *maxv)
 180                                 *maxv = ovals[j];
 181
 182                         is_last = (j == len - 1) != 0;
 183                         if (is_last)
 184                                 break;
 185
 186                         j++;
 187                 }
 188         }
 189
 190         if (!is_last)
 191                 log_err("fio: error calculating latency percentiles\n");
 192
 193         *output = ovals;
 194         return len;
 195 }
 196
 197 /*
 198  * Find and display the p-th percentile of clat
 199  */
 200 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 201                                   fio_fp64_t *plist, unsigned int precision,
 202                                   const char *pre, struct buf_output *out)
 203 {
 204         unsigned int divisor, len, i, j = 0;
 205         unsigned long long minv, maxv;
 206         unsigned long long *ovals;
 207         int per_line, scale_down, time_width;
 208         bool is_last;
 209         char fmt[32];
 210
 211         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 212         if (!len || !ovals)
 213                 goto out;
 214
 215         /*
 216          * We default to nsecs, but if the value range is such that we
 217          * should scale down to usecs or msecs, do that.
 218          */
 219         if (minv > 2000000 && maxv > 99999999ULL) {
 220                 scale_down = 2;
 221                 divisor = 1000000;
 222                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 223         } else if (minv > 2000 && maxv > 99999) {
 224                 scale_down = 1;
 225                 divisor = 1000;
 226                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 227         } else {
 228                 scale_down = 0;
 229                 divisor = 1;
 230                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 231         }
 232
 233
 234         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 235         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 236                         precision, time_width);
 237         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 238         per_line = (80 - 7) / (precision + 10 + time_width);
 239
 240         for (j = 0; j < len; j++) {
 241                 /* for formatting */
 242                 if (j != 0 && (j % per_line) == 0)
 243                         log_buf(out, "     |");
 244
 245                 /* end of the list */
 246                 is_last = (j == len - 1) != 0;
 247
 248                 for (i = 0; i < scale_down; i++)
 249                         ovals[j] = (ovals[j] + 999) / 1000;
 250
 251                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 252
 253                 if (is_last)
 254                         break;
 255
 256                 if ((j % per_line) == per_line - 1)     /* for formatting */
 257                         log_buf(out, "\n");
 258         }
 259
 260 out:
 261         if (ovals)
 262                 free(ovals);
 263 }
 264
 265 bool calc_lat(struct io_stat *is, unsigned long long *min,
 266               unsigned long long *max, double *mean, double *dev)
 267 {
 268         double n = (double) is->samples;
 269
 270         if (n == 0)
 271                 return false;
 272
 273         *min = is->min_val;
 274         *max = is->max_val;
 275         *mean = is->mean.u.f;
 276
 277         if (n > 1.0)
 278                 *dev = sqrt(is->S.u.f / (n - 1.0));
 279         else
 280                 *dev = 0;
 281
 282         return true;
 283 }
 284
 285 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 286 {
 287         char *io, *agg, *min, *max;
 288         char *ioalt, *aggalt, *minalt, *maxalt;
 289         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 290         int i;
 291
 292         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 293
 294         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 295                 const int i2p = is_power_of_2(rs->kb_base);
 296
 297                 if (!rs->max_run[i])
 298                         continue;
 299
 300                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
 301                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
 302                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
 303                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 304                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 305                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 306                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 307                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 308                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 309                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 310                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 311                                 (unsigned long long) rs->min_run[i],
 312                                 (unsigned long long) rs->max_run[i]);
 313
 314                 free(io);
 315                 free(agg);
 316                 free(min);
 317                 free(max);
 318                 free(ioalt);
 319                 free(aggalt);
 320                 free(minalt);
 321                 free(maxalt);
 322         }
 323 }
 324
 325 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
 326 {
 327         int i;
 328
 329         /*
 330          * Do depth distribution calculations
 331          */
 332         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 333                 if (total) {
 334                         io_u_dist[i] = (double) map[i] / (double) total;
 335                         io_u_dist[i] *= 100.0;
 336                         if (io_u_dist[i] < 0.1 && map[i])
 337                                 io_u_dist[i] = 0.1;
 338                 } else
 339                         io_u_dist[i] = 0.0;
 340         }
 341 }
 342
 343 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 344                           uint64_t *src, int nr)
 345 {
 346         unsigned long total = ddir_rw_sum(ts->total_io_u);
 347         int i;
 348
 349         /*
 350          * Do latency distribution calculations
 351          */
 352         for (i = 0; i < nr; i++) {
 353                 if (total) {
 354                         dst[i] = (double) src[i] / (double) total;
 355                         dst[i] *= 100.0;
 356                         if (dst[i] < 0.01 && src[i])
 357                                 dst[i] = 0.01;
 358                 } else
 359                         dst[i] = 0.0;
 360         }
 361 }
 362
 363 /*
 364  * To keep the terse format unaltered, add all of the ns latency
 365  * buckets to the first us latency bucket
 366  */
 367 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 368 {
 369         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 370         int i;
 371
 372         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 373
 374         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 375                 ntotal += ts->io_u_lat_n[i];
 376
 377         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 378 }
 379
 380 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 381 {
 382         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 383 }
 384
 385 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 386 {
 387         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 388 }
 389
 390 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 391 {
 392         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 393 }
 394
 395 static void display_lat(const char *name, unsigned long long min,
 396                         unsigned long long max, double mean, double dev,
 397                         struct buf_output *out)
 398 {
 399         const char *base = "(nsec)";
 400         char *minp, *maxp;
 401
 402         if (nsec_to_msec(&min, &max, &mean, &dev))
 403                 base = "(msec)";
 404         else if (nsec_to_usec(&min, &max, &mean, &dev))
 405                 base = "(usec)";
 406
 407         minp = num2str(min, 6, 1, 0, N2S_NONE);
 408         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 409
 410         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 411                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 412
 413         free(minp);
 414         free(maxp);
 415 }
 416
 417 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 418                              int ddir, struct buf_output *out)
 419 {
 420         unsigned long runt;
 421         unsigned long long min, max, bw, iops;
 422         double mean, dev;
 423         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
 424         int i2p;
 425
 426         if (ddir_sync(ddir)) {
 427                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
 428                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
 429                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
 430                         show_clat_percentiles(ts->io_u_sync_plat,
 431                                                 ts->sync_stat.samples,
 432                                                 ts->percentile_list,
 433                                                 ts->percentile_precision,
 434                                                 io_ddir_name(ddir), out);
 435                 }
 436                 return;
 437         }
 438
 439         assert(ddir_rw(ddir));
 440
 441         if (!ts->runtime[ddir])
 442                 return;
 443
 444         i2p = is_power_of_2(rs->kb_base);
 445         runt = ts->runtime[ddir];
 446
 447         bw = (1000 * ts->io_bytes[ddir]) / runt;
 448         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
 449         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
 450         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
 451
 452         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 453         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
 454         if (ddir == DDIR_WRITE)
 455                 post_st = zbd_write_status(ts);
 456         else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
 457                 uint64_t total;
 458                 double hit;
 459
 460                 total = ts->cachehit + ts->cachemiss;
 461                 hit = (double) ts->cachehit / (double) total;
 462                 hit *= 100.0;
 463                 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
 464                         post_st = NULL;
 465         }
 466
 467         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
 468                         rs->unified_rw_rep ? "mixed" : io_ddir_name(ddir),
 469                         iops_p, bw_p, bw_p_alt, io_p,
 470                         (unsigned long long) ts->runtime[ddir],
 471                         post_st ? : "");
 472
 473         free(post_st);
 474         free(io_p);
 475         free(bw_p);
 476         free(bw_p_alt);
 477         free(iops_p);
 478
 479         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 480                 display_lat("slat", min, max, mean, dev, out);
 481         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 482                 display_lat("clat", min, max, mean, dev, out);
 483         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 484                 display_lat(" lat", min, max, mean, dev, out);
 485
 486         if (ts->clat_percentiles || ts->lat_percentiles) {
 487                 const char *name = ts->clat_percentiles ? "clat" : " lat";
 488                 uint64_t samples;
 489
 490                 if (ts->clat_percentiles)
 491                         samples = ts->clat_stat[ddir].samples;
 492                 else
 493                         samples = ts->lat_stat[ddir].samples;
 494
 495                 show_clat_percentiles(ts->io_u_plat[ddir],
 496                                         samples,
 497                                         ts->percentile_list,
 498                                         ts->percentile_precision, name, out);
 499         }
 500         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 501                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 502                 const char *bw_str;
 503
 504                 if ((rs->unit_base == 1) && i2p)
 505                         bw_str = "Kibit";
 506                 else if (rs->unit_base == 1)
 507                         bw_str = "kbit";
 508                 else if (i2p)
 509                         bw_str = "KiB";
 510                 else
 511                         bw_str = "kB";
 512
 513                 if (rs->agg[ddir]) {
 514                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 515                         if (p_of_agg > 100.0)
 516                                 p_of_agg = 100.0;
 517                 }
 518
 519                 if (rs->unit_base == 1) {
 520                         min *= 8.0;
 521                         max *= 8.0;
 522                         mean *= 8.0;
 523                         dev *= 8.0;
 524                 }
 525
 526                 if (mean > fkb_base * fkb_base) {
 527                         min /= fkb_base;
 528                         max /= fkb_base;
 529                         mean /= fkb_base;
 530                         dev /= fkb_base;
 531                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 532                 }
 533
 534                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 535                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 536                         bw_str, min, max, p_of_agg, mean, dev,
 537                         (&ts->bw_stat[ddir])->samples);
 538         }
 539         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 540                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 541                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 542                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 543         }
 544 }
 545
 546 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
 547                      const char *msg, struct buf_output *out)
 548 {
 549         bool new_line = true, shown = false;
 550         int i, line = 0;
 551
 552         for (i = 0; i < nr; i++) {
 553                 if (io_u_lat[i] <= 0.0)
 554                         continue;
 555                 shown = true;
 556                 if (new_line) {
 557                         if (line)
 558                                 log_buf(out, "\n");
 559                         log_buf(out, "  lat (%s)   : ", msg);
 560                         new_line = false;
 561                         line = 0;
 562                 }
 563                 if (line)
 564                         log_buf(out, ", ");
 565                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 566                 line++;
 567                 if (line == 5)
 568                         new_line = true;
 569         }
 570
 571         if (shown)
 572                 log_buf(out, "\n");
 573
 574         return true;
 575 }
 576
 577 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 578 {
 579         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 580                                  "250=", "500=", "750=", "1000=", };
 581
 582         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 583 }
 584
 585 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 586 {
 587         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 588                                  "250=", "500=", "750=", "1000=", };
 589
 590         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 591 }
 592
 593 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 594 {
 595         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 596                                  "250=", "500=", "750=", "1000=", "2000=",
 597                                  ">=2000=", };
 598
 599         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 600 }
 601
 602 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 603 {
 604         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 605         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 606         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 607
 608         stat_calc_lat_n(ts, io_u_lat_n);
 609         stat_calc_lat_u(ts, io_u_lat_u);
 610         stat_calc_lat_m(ts, io_u_lat_m);
 611
 612         show_lat_n(io_u_lat_n, out);
 613         show_lat_u(io_u_lat_u, out);
 614         show_lat_m(io_u_lat_m, out);
 615 }
 616
 617 static int block_state_category(int block_state)
 618 {
 619         switch (block_state) {
 620         case BLOCK_STATE_UNINIT:
 621                 return 0;
 622         case BLOCK_STATE_TRIMMED:
 623         case BLOCK_STATE_WRITTEN:
 624                 return 1;
 625         case BLOCK_STATE_WRITE_FAILURE:
 626         case BLOCK_STATE_TRIM_FAILURE:
 627                 return 2;
 628         default:
 629                 /* Silence compile warning on some BSDs and have a return */
 630                 assert(0);
 631                 return -1;
 632         }
 633 }
 634
 635 static int compare_block_infos(const void *bs1, const void *bs2)
 636 {
 637         uint64_t block1 = *(uint64_t *)bs1;
 638         uint64_t block2 = *(uint64_t *)bs2;
 639         int state1 = BLOCK_INFO_STATE(block1);
 640         int state2 = BLOCK_INFO_STATE(block2);
 641         int bscat1 = block_state_category(state1);
 642         int bscat2 = block_state_category(state2);
 643         int cycles1 = BLOCK_INFO_TRIMS(block1);
 644         int cycles2 = BLOCK_INFO_TRIMS(block2);
 645
 646         if (bscat1 < bscat2)
 647                 return -1;
 648         if (bscat1 > bscat2)
 649                 return 1;
 650
 651         if (cycles1 < cycles2)
 652                 return -1;
 653         if (cycles1 > cycles2)
 654                 return 1;
 655
 656         if (state1 < state2)
 657                 return -1;
 658         if (state1 > state2)
 659                 return 1;
 660
 661         assert(block1 == block2);
 662         return 0;
 663 }
 664
 665 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 666                                   fio_fp64_t *plist, unsigned int **percentiles,
 667                                   unsigned int *types)
 668 {
 669         int len = 0;
 670         int i, nr_uninit;
 671
 672         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 673
 674         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 675                 len++;
 676
 677         if (!len)
 678                 return 0;
 679
 680         /*
 681          * Sort the percentile list. Note that it may already be sorted if
 682          * we are using the default values, but since it's a short list this
 683          * isn't a worry. Also note that this does not work for NaN values.
 684          */
 685         if (len > 1)
 686                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 687
 688         /* Start only after the uninit entries end */
 689         for (nr_uninit = 0;
 690              nr_uninit < nr_block_infos
 691                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 692              nr_uninit ++)
 693                 ;
 694
 695         if (nr_uninit == nr_block_infos)
 696                 return 0;
 697
 698         *percentiles = calloc(len, sizeof(**percentiles));
 699
 700         for (i = 0; i < len; i++) {
 701                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 702                                 + nr_uninit;
 703                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 704         }
 705
 706         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 707         for (i = 0; i < nr_block_infos; i++)
 708                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 709
 710         return len;
 711 }
 712
 713 static const char *block_state_names[] = {
 714         [BLOCK_STATE_UNINIT] = "unwritten",
 715         [BLOCK_STATE_TRIMMED] = "trimmed",
 716         [BLOCK_STATE_WRITTEN] = "written",
 717         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 718         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 719 };
 720
 721 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 722                              fio_fp64_t *plist, struct buf_output *out)
 723 {
 724         int len, pos, i;
 725         unsigned int *percentiles = NULL;
 726         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 727
 728         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 729                                      &percentiles, block_state_counts);
 730
 731         log_buf(out, "  block lifetime percentiles :\n   |");
 732         pos = 0;
 733         for (i = 0; i < len; i++) {
 734                 uint32_t block_info = percentiles[i];
 735 #define LINE_LENGTH     75
 736                 char str[LINE_LENGTH];
 737                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 738                                      plist[i].u.f, block_info,
 739                                      i == len - 1 ? '\n' : ',');
 740                 assert(strln < LINE_LENGTH);
 741                 if (pos + strln > LINE_LENGTH) {
 742                         pos = 0;
 743                         log_buf(out, "\n   |");
 744                 }
 745                 log_buf(out, "%s", str);
 746                 pos += strln;
 747 #undef LINE_LENGTH
 748         }
 749         if (percentiles)
 750                 free(percentiles);
 751
 752         log_buf(out, "        states               :");
 753         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 754                 log_buf(out, " %s=%u%c",
 755                          block_state_names[i], block_state_counts[i],
 756                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 757 }
 758
 759 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 760 {
 761         char *p1, *p1alt, *p2;
 762         unsigned long long bw_mean, iops_mean;
 763         const int i2p = is_power_of_2(ts->kb_base);
 764
 765         if (!ts->ss_dur)
 766                 return;
 767
 768         bw_mean = steadystate_bw_mean(ts);
 769         iops_mean = steadystate_iops_mean(ts);
 770
 771         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
 772         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
 773         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
 774
 775         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 776                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
 777                 p1, p1alt, p2,
 778                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
 779                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
 780                 ts->ss_criterion.u.f,
 781                 ts->ss_state & FIO_SS_PCT ? "%" : "");
 782
 783         free(p1);
 784         free(p1alt);
 785         free(p2);
 786 }
 787
 788 static void show_thread_status_normal(struct thread_stat *ts,
 789                                       struct group_run_stats *rs,
 790                                       struct buf_output *out)
 791 {
 792         double usr_cpu, sys_cpu;
 793         unsigned long runtime;
 794         double io_u_dist[FIO_IO_U_MAP_NR];
 795         time_t time_p;
 796         char time_buf[32];
 797
 798         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 799                 return;
 800
 801         memset(time_buf, 0, sizeof(time_buf));
 802
 803         time(&time_p);
 804         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 805
 806         if (!ts->error) {
 807                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 808                                         ts->name, ts->groupid, ts->members,
 809                                         ts->error, (int) ts->pid, time_buf);
 810         } else {
 811                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 812                                         ts->name, ts->groupid, ts->members,
 813                                         ts->error, ts->verror, (int) ts->pid,
 814                                         time_buf);
 815         }
 816
 817         if (strlen(ts->description))
 818                 log_buf(out, "  Description  : [%s]\n", ts->description);
 819
 820         if (ts->io_bytes[DDIR_READ])
 821                 show_ddir_status(rs, ts, DDIR_READ, out);
 822         if (ts->io_bytes[DDIR_WRITE])
 823                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 824         if (ts->io_bytes[DDIR_TRIM])
 825                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 826
 827         show_latencies(ts, out);
 828
 829         if (ts->sync_stat.samples)
 830                 show_ddir_status(rs, ts, DDIR_SYNC, out);
 831
 832         runtime = ts->total_run_time;
 833         if (runtime) {
 834                 double runt = (double) runtime;
 835
 836                 usr_cpu = (double) ts->usr_time * 100 / runt;
 837                 sys_cpu = (double) ts->sys_time * 100 / runt;
 838         } else {
 839                 usr_cpu = 0;
 840                 sys_cpu = 0;
 841         }
 842
 843         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 844                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 845                         (unsigned long long) ts->ctx,
 846                         (unsigned long long) ts->majf,
 847                         (unsigned long long) ts->minf);
 848
 849         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 850         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 851                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 852                                         io_u_dist[1], io_u_dist[2],
 853                                         io_u_dist[3], io_u_dist[4],
 854                                         io_u_dist[5], io_u_dist[6]);
 855
 856         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 857         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 858                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 859                                         io_u_dist[1], io_u_dist[2],
 860                                         io_u_dist[3], io_u_dist[4],
 861                                         io_u_dist[5], io_u_dist[6]);
 862         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 863         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 864                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 865                                         io_u_dist[1], io_u_dist[2],
 866                                         io_u_dist[3], io_u_dist[4],
 867                                         io_u_dist[5], io_u_dist[6]);
 868         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
 869                                  " short=%llu,%llu,%llu,0"
 870                                  " dropped=%llu,%llu,%llu,0\n",
 871                                         (unsigned long long) ts->total_io_u[0],
 872                                         (unsigned long long) ts->total_io_u[1],
 873                                         (unsigned long long) ts->total_io_u[2],
 874                                         (unsigned long long) ts->total_io_u[3],
 875                                         (unsigned long long) ts->short_io_u[0],
 876                                         (unsigned long long) ts->short_io_u[1],
 877                                         (unsigned long long) ts->short_io_u[2],
 878                                         (unsigned long long) ts->drop_io_u[0],
 879                                         (unsigned long long) ts->drop_io_u[1],
 880                                         (unsigned long long) ts->drop_io_u[2]);
 881         if (ts->continue_on_error) {
 882                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 883                                         (unsigned long long)ts->total_err_count,
 884                                         ts->first_error,
 885                                         strerror(ts->first_error));
 886         }
 887         if (ts->latency_depth) {
 888                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 889                                         (unsigned long long)ts->latency_target,
 890                                         (unsigned long long)ts->latency_window,
 891                                         ts->latency_percentile.u.f,
 892                                         ts->latency_depth);
 893         }
 894
 895         if (ts->nr_block_infos)
 896                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 897                                   ts->percentile_list, out);
 898
 899         if (ts->ss_dur)
 900                 show_ss_normal(ts, out);
 901 }
 902
 903 static void show_ddir_status_terse(struct thread_stat *ts,
 904                                    struct group_run_stats *rs, int ddir,
 905                                    int ver, struct buf_output *out)
 906 {
 907         unsigned long long min, max, minv, maxv, bw, iops;
 908         unsigned long long *ovals = NULL;
 909         double mean, dev;
 910         unsigned int len;
 911         int i, bw_stat;
 912
 913         assert(ddir_rw(ddir));
 914
 915         iops = bw = 0;
 916         if (ts->runtime[ddir]) {
 917                 uint64_t runt = ts->runtime[ddir];
 918
 919                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 920                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 921         }
 922
 923         log_buf(out, ";%llu;%llu;%llu;%llu",
 924                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 925                                         (unsigned long long) ts->runtime[ddir]);
 926
 927         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 928                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 929         else
 930                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 931
 932         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 933                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 934         else
 935                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 936
 937         if (ts->clat_percentiles || ts->lat_percentiles) {
 938                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 939                                         ts->clat_stat[ddir].samples,
 940                                         ts->percentile_list, &ovals, &maxv,
 941                                         &minv);
 942         } else
 943                 len = 0;
 944
 945         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 946                 if (i >= len) {
 947                         log_buf(out, ";0%%=0");
 948                         continue;
 949                 }
 950                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
 951         }
 952
 953         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 954                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 955         else
 956                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 957
 958         if (ovals)
 959                 free(ovals);
 960
 961         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
 962         if (bw_stat) {
 963                 double p_of_agg = 100.0;
 964
 965                 if (rs->agg[ddir]) {
 966                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 967                         if (p_of_agg > 100.0)
 968                                 p_of_agg = 100.0;
 969                 }
 970
 971                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 972         } else
 973                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
 974
 975         if (ver == 5) {
 976                 if (bw_stat)
 977                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
 978                 else
 979                         log_buf(out, ";%lu", 0UL);
 980
 981                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
 982                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
 983                                 mean, dev, (&ts->iops_stat[ddir])->samples);
 984                 else
 985                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
 986         }
 987 }
 988
 989 static void add_ddir_status_json(struct thread_stat *ts,
 990                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 991 {
 992         unsigned long long min, max, minv, maxv;
 993         unsigned long long bw_bytes, bw;
 994         unsigned long long *ovals = NULL;
 995         double mean, dev, iops;
 996         unsigned int len;
 997         int i;
 998         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object = NULL;
 999         char buf[120];
1000         double p_of_agg = 100.0;
1001
1002         assert(ddir_rw(ddir) || ddir_sync(ddir));
1003
1004         if (ts->unified_rw_rep && ddir != DDIR_READ)
1005                 return;
1006
1007         dir_object = json_create_object();
1008         json_object_add_value_object(parent,
1009                 ts->unified_rw_rep ? "mixed" : io_ddir_name(ddir), dir_object);
1010
1011         if (ddir_rw(ddir)) {
1012                 bw_bytes = 0;
1013                 bw = 0;
1014                 iops = 0.0;
1015                 if (ts->runtime[ddir]) {
1016                         uint64_t runt = ts->runtime[ddir];
1017
1018                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1019                         bw = bw_bytes / 1024; /* KiB/s */
1020                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1021                 }
1022
1023                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1024                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1025                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1026                 json_object_add_value_int(dir_object, "bw", bw);
1027                 json_object_add_value_float(dir_object, "iops", iops);
1028                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1029                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1030                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1031                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1032
1033                 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
1034                         min = max = 0;
1035                         mean = dev = 0.0;
1036                 }
1037                 tmp_object = json_create_object();
1038                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1039                 json_object_add_value_int(tmp_object, "min", min);
1040                 json_object_add_value_int(tmp_object, "max", max);
1041                 json_object_add_value_float(tmp_object, "mean", mean);
1042                 json_object_add_value_float(tmp_object, "stddev", dev);
1043
1044                 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1045                         min = max = 0;
1046                         mean = dev = 0.0;
1047                 }
1048                 tmp_object = json_create_object();
1049                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1050                 json_object_add_value_int(tmp_object, "min", min);
1051                 json_object_add_value_int(tmp_object, "max", max);
1052                 json_object_add_value_float(tmp_object, "mean", mean);
1053                 json_object_add_value_float(tmp_object, "stddev", dev);
1054         } else {
1055                 if (!calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
1056                         min = max = 0;
1057                         mean = dev = 0.0;
1058                 }
1059
1060                 tmp_object = json_create_object();
1061                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1062                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1063                 json_object_add_value_int(tmp_object, "min", min);
1064                 json_object_add_value_int(tmp_object, "max", max);
1065                 json_object_add_value_float(tmp_object, "mean", mean);
1066                 json_object_add_value_float(tmp_object, "stddev", dev);
1067         }
1068
1069         if (ts->clat_percentiles || ts->lat_percentiles) {
1070                 if (ddir_rw(ddir)) {
1071                         uint64_t samples;
1072
1073                         if (ts->clat_percentiles)
1074                                 samples = ts->clat_stat[ddir].samples;
1075                         else
1076                                 samples = ts->lat_stat[ddir].samples;
1077
1078                         len = calc_clat_percentiles(ts->io_u_plat[ddir],
1079                                         samples, ts->percentile_list, &ovals,
1080                                         &maxv, &minv);
1081                 } else {
1082                         len = calc_clat_percentiles(ts->io_u_sync_plat,
1083                                         ts->sync_stat.samples,
1084                                         ts->percentile_list, &ovals, &maxv,
1085                                         &minv);
1086                 }
1087
1088                 if (len > FIO_IO_U_LIST_MAX_LEN)
1089                         len = FIO_IO_U_LIST_MAX_LEN;
1090         } else
1091                 len = 0;
1092
1093         percentile_object = json_create_object();
1094         if (ts->clat_percentiles)
1095                 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1096         for (i = 0; i < len; i++) {
1097                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1098                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1099         }
1100
1101         if (output_format & FIO_OUTPUT_JSON_PLUS) {
1102                 clat_bins_object = json_create_object();
1103                 if (ts->clat_percentiles)
1104                         json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1105
1106                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1107                         if (ddir_rw(ddir)) {
1108                                 if (ts->io_u_plat[ddir][i]) {
1109                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1110                                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1111                                 }
1112                         } else {
1113                                 if (ts->io_u_sync_plat[i]) {
1114                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1115                                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_sync_plat[i]);
1116                                 }
1117                         }
1118                 }
1119         }
1120
1121         if (!ddir_rw(ddir))
1122                 return;
1123
1124         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1125                 min = max = 0;
1126                 mean = dev = 0.0;
1127         }
1128         tmp_object = json_create_object();
1129         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1130         json_object_add_value_int(tmp_object, "min", min);
1131         json_object_add_value_int(tmp_object, "max", max);
1132         json_object_add_value_float(tmp_object, "mean", mean);
1133         json_object_add_value_float(tmp_object, "stddev", dev);
1134         if (ts->lat_percentiles)
1135                 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1136         if (output_format & FIO_OUTPUT_JSON_PLUS && ts->lat_percentiles)
1137                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1138
1139         if (ovals)
1140                 free(ovals);
1141
1142         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1143                 if (rs->agg[ddir]) {
1144                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1145                         if (p_of_agg > 100.0)
1146                                 p_of_agg = 100.0;
1147                 }
1148         } else {
1149                 min = max = 0;
1150                 p_of_agg = mean = dev = 0.0;
1151         }
1152         json_object_add_value_int(dir_object, "bw_min", min);
1153         json_object_add_value_int(dir_object, "bw_max", max);
1154         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1155         json_object_add_value_float(dir_object, "bw_mean", mean);
1156         json_object_add_value_float(dir_object, "bw_dev", dev);
1157         json_object_add_value_int(dir_object, "bw_samples",
1158                                 (&ts->bw_stat[ddir])->samples);
1159
1160         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1161                 min = max = 0;
1162                 mean = dev = 0.0;
1163         }
1164         json_object_add_value_int(dir_object, "iops_min", min);
1165         json_object_add_value_int(dir_object, "iops_max", max);
1166         json_object_add_value_float(dir_object, "iops_mean", mean);
1167         json_object_add_value_float(dir_object, "iops_stddev", dev);
1168         json_object_add_value_int(dir_object, "iops_samples",
1169                                 (&ts->iops_stat[ddir])->samples);
1170
1171         if (ts->cachehit + ts->cachemiss) {
1172                 uint64_t total;
1173                 double hit;
1174
1175                 total = ts->cachehit + ts->cachemiss;
1176                 hit = (double) ts->cachehit / (double) total;
1177                 hit *= 100.0;
1178                 json_object_add_value_float(dir_object, "cachehit", hit);
1179         }
1180 }
1181
1182 static void show_thread_status_terse_all(struct thread_stat *ts,
1183                                          struct group_run_stats *rs, int ver,
1184                                          struct buf_output *out)
1185 {
1186         double io_u_dist[FIO_IO_U_MAP_NR];
1187         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1188         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1189         double usr_cpu, sys_cpu;
1190         int i;
1191
1192         /* General Info */
1193         if (ver == 2)
1194                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1195         else
1196                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1197                         ts->name, ts->groupid, ts->error);
1198
1199         /* Log Read Status */
1200         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1201         /* Log Write Status */
1202         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1203         /* Log Trim Status */
1204         if (ver == 2 || ver == 4 || ver == 5)
1205                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1206
1207         /* CPU Usage */
1208         if (ts->total_run_time) {
1209                 double runt = (double) ts->total_run_time;
1210
1211                 usr_cpu = (double) ts->usr_time * 100 / runt;
1212                 sys_cpu = (double) ts->sys_time * 100 / runt;
1213         } else {
1214                 usr_cpu = 0;
1215                 sys_cpu = 0;
1216         }
1217
1218         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1219                                                 (unsigned long long) ts->ctx,
1220                                                 (unsigned long long) ts->majf,
1221                                                 (unsigned long long) ts->minf);
1222
1223         /* Calc % distribution of IO depths, usecond, msecond latency */
1224         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1225         stat_calc_lat_nu(ts, io_u_lat_u);
1226         stat_calc_lat_m(ts, io_u_lat_m);
1227
1228         /* Only show fixed 7 I/O depth levels*/
1229         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1230                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1231                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1232
1233         /* Microsecond latency */
1234         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1235                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1236         /* Millisecond latency */
1237         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1238                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1239
1240         /* disk util stats, if any */
1241         if (ver >= 3 && is_running_backend())
1242                 show_disk_util(1, NULL, out);
1243
1244         /* Additional output if continue_on_error set - default off*/
1245         if (ts->continue_on_error)
1246                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1247         if (ver == 2)
1248                 log_buf(out, "\n");
1249
1250         /* Additional output if description is set */
1251         if (strlen(ts->description))
1252                 log_buf(out, ";%s", ts->description);
1253
1254         log_buf(out, "\n");
1255 }
1256
1257 static void json_add_job_opts(struct json_object *root, const char *name,
1258                               struct flist_head *opt_list)
1259 {
1260         struct json_object *dir_object;
1261         struct flist_head *entry;
1262         struct print_option *p;
1263
1264         if (flist_empty(opt_list))
1265                 return;
1266
1267         dir_object = json_create_object();
1268         json_object_add_value_object(root, name, dir_object);
1269
1270         flist_for_each(entry, opt_list) {
1271                 const char *pos = "";
1272
1273                 p = flist_entry(entry, struct print_option, list);
1274                 if (p->value)
1275                         pos = p->value;
1276                 json_object_add_value_string(dir_object, p->name, pos);
1277         }
1278 }
1279
1280 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1281                                                    struct group_run_stats *rs,
1282                                                    struct flist_head *opt_list)
1283 {
1284         struct json_object *root, *tmp;
1285         struct jobs_eta *je;
1286         double io_u_dist[FIO_IO_U_MAP_NR];
1287         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1288         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1289         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1290         double usr_cpu, sys_cpu;
1291         int i;
1292         size_t size;
1293
1294         root = json_create_object();
1295         json_object_add_value_string(root, "jobname", ts->name);
1296         json_object_add_value_int(root, "groupid", ts->groupid);
1297         json_object_add_value_int(root, "error", ts->error);
1298
1299         /* ETA Info */
1300         je = get_jobs_eta(true, &size);
1301         if (je) {
1302                 json_object_add_value_int(root, "eta", je->eta_sec);
1303                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1304         }
1305
1306         if (opt_list)
1307                 json_add_job_opts(root, "job options", opt_list);
1308
1309         add_ddir_status_json(ts, rs, DDIR_READ, root);
1310         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1311         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1312         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1313
1314         /* CPU Usage */
1315         if (ts->total_run_time) {
1316                 double runt = (double) ts->total_run_time;
1317
1318                 usr_cpu = (double) ts->usr_time * 100 / runt;
1319                 sys_cpu = (double) ts->sys_time * 100 / runt;
1320         } else {
1321                 usr_cpu = 0;
1322                 sys_cpu = 0;
1323         }
1324         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1325         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1326         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1327         json_object_add_value_int(root, "ctx", ts->ctx);
1328         json_object_add_value_int(root, "majf", ts->majf);
1329         json_object_add_value_int(root, "minf", ts->minf);
1330
1331         /* Calc % distribution of IO depths */
1332         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1333         tmp = json_create_object();
1334         json_object_add_value_object(root, "iodepth_level", tmp);
1335         /* Only show fixed 7 I/O depth levels*/
1336         for (i = 0; i < 7; i++) {
1337                 char name[20];
1338                 if (i < 6)
1339                         snprintf(name, 20, "%d", 1 << i);
1340                 else
1341                         snprintf(name, 20, ">=%d", 1 << i);
1342                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1343         }
1344
1345         /* Calc % distribution of submit IO depths */
1346         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1347         tmp = json_create_object();
1348         json_object_add_value_object(root, "iodepth_submit", tmp);
1349         /* Only show fixed 7 I/O depth levels*/
1350         for (i = 0; i < 7; i++) {
1351                 char name[20];
1352                 if (i == 0)
1353                         snprintf(name, 20, "0");
1354                 else if (i < 6)
1355                         snprintf(name, 20, "%d", 1 << (i+1));
1356                 else
1357                         snprintf(name, 20, ">=%d", 1 << i);
1358                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1359         }
1360
1361         /* Calc % distribution of completion IO depths */
1362         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1363         tmp = json_create_object();
1364         json_object_add_value_object(root, "iodepth_complete", tmp);
1365         /* Only show fixed 7 I/O depth levels*/
1366         for (i = 0; i < 7; i++) {
1367                 char name[20];
1368                 if (i == 0)
1369                         snprintf(name, 20, "0");
1370                 else if (i < 6)
1371                         snprintf(name, 20, "%d", 1 << (i+1));
1372                 else
1373                         snprintf(name, 20, ">=%d", 1 << i);
1374                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1375         }
1376
1377         /* Calc % distribution of nsecond, usecond, msecond latency */
1378         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1379         stat_calc_lat_n(ts, io_u_lat_n);
1380         stat_calc_lat_u(ts, io_u_lat_u);
1381         stat_calc_lat_m(ts, io_u_lat_m);
1382
1383         /* Nanosecond latency */
1384         tmp = json_create_object();
1385         json_object_add_value_object(root, "latency_ns", tmp);
1386         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1387                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1388                                  "250", "500", "750", "1000", };
1389                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1390         }
1391         /* Microsecond latency */
1392         tmp = json_create_object();
1393         json_object_add_value_object(root, "latency_us", tmp);
1394         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1395                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1396                                  "250", "500", "750", "1000", };
1397                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1398         }
1399         /* Millisecond latency */
1400         tmp = json_create_object();
1401         json_object_add_value_object(root, "latency_ms", tmp);
1402         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1403                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1404                                  "250", "500", "750", "1000", "2000",
1405                                  ">=2000", };
1406                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1407         }
1408
1409         /* Additional output if continue_on_error set - default off*/
1410         if (ts->continue_on_error) {
1411                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1412                 json_object_add_value_int(root, "first_error", ts->first_error);
1413         }
1414
1415         if (ts->latency_depth) {
1416                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1417                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1418                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1419                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1420         }
1421
1422         /* Additional output if description is set */
1423         if (strlen(ts->description))
1424                 json_object_add_value_string(root, "desc", ts->description);
1425
1426         if (ts->nr_block_infos) {
1427                 /* Block error histogram and types */
1428                 int len;
1429                 unsigned int *percentiles = NULL;
1430                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1431
1432                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1433                                              ts->percentile_list,
1434                                              &percentiles, block_state_counts);
1435
1436                 if (len) {
1437                         struct json_object *block, *percentile_object, *states;
1438                         int state;
1439                         block = json_create_object();
1440                         json_object_add_value_object(root, "block", block);
1441
1442                         percentile_object = json_create_object();
1443                         json_object_add_value_object(block, "percentiles",
1444                                                      percentile_object);
1445                         for (i = 0; i < len; i++) {
1446                                 char buf[20];
1447                                 snprintf(buf, sizeof(buf), "%f",
1448                                          ts->percentile_list[i].u.f);
1449                                 json_object_add_value_int(percentile_object,
1450                                                           (const char *)buf,
1451                                                           percentiles[i]);
1452                         }
1453
1454                         states = json_create_object();
1455                         json_object_add_value_object(block, "states", states);
1456                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1457                                 json_object_add_value_int(states,
1458                                         block_state_names[state],
1459                                         block_state_counts[state]);
1460                         }
1461                         free(percentiles);
1462                 }
1463         }
1464
1465         if (ts->ss_dur) {
1466                 struct json_object *data;
1467                 struct json_array *iops, *bw;
1468                 int j, k, l;
1469                 char ss_buf[64];
1470
1471                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1472                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1473                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1474                         (float) ts->ss_limit.u.f,
1475                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1476
1477                 tmp = json_create_object();
1478                 json_object_add_value_object(root, "steadystate", tmp);
1479                 json_object_add_value_string(tmp, "ss", ss_buf);
1480                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1481                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1482
1483                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1484                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1485                 json_object_add_value_string(tmp, "criterion", ss_buf);
1486                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1487                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1488
1489                 data = json_create_object();
1490                 json_object_add_value_object(tmp, "data", data);
1491                 bw = json_create_array();
1492                 iops = json_create_array();
1493
1494                 /*
1495                 ** if ss was attained or the buffer is not full,
1496                 ** ss->head points to the first element in the list.
1497                 ** otherwise it actually points to the second element
1498                 ** in the list
1499                 */
1500                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1501                         j = ts->ss_head;
1502                 else
1503                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1504                 for (l = 0; l < ts->ss_dur; l++) {
1505                         k = (j + l) % ts->ss_dur;
1506                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1507                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1508                 }
1509                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1510                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1511                 json_object_add_value_array(data, "iops", iops);
1512                 json_object_add_value_array(data, "bw", bw);
1513         }
1514
1515         return root;
1516 }
1517
1518 static void show_thread_status_terse(struct thread_stat *ts,
1519                                      struct group_run_stats *rs,
1520                                      struct buf_output *out)
1521 {
1522         if (terse_version >= 2 && terse_version <= 5)
1523                 show_thread_status_terse_all(ts, rs, terse_version, out);
1524         else
1525                 log_err("fio: bad terse version!? %d\n", terse_version);
1526 }
1527
1528 struct json_object *show_thread_status(struct thread_stat *ts,
1529                                        struct group_run_stats *rs,
1530                                        struct flist_head *opt_list,
1531                                        struct buf_output *out)
1532 {
1533         struct json_object *ret = NULL;
1534
1535         if (output_format & FIO_OUTPUT_TERSE)
1536                 show_thread_status_terse(ts, rs,  out);
1537         if (output_format & FIO_OUTPUT_JSON)
1538                 ret = show_thread_status_json(ts, rs, opt_list);
1539         if (output_format & FIO_OUTPUT_NORMAL)
1540                 show_thread_status_normal(ts, rs,  out);
1541
1542         return ret;
1543 }
1544
1545 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1546 {
1547         double mean, S;
1548
1549         dst->min_val = min(dst->min_val, src->min_val);
1550         dst->max_val = max(dst->max_val, src->max_val);
1551
1552         /*
1553          * Compute new mean and S after the merge
1554          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1555          *  #Parallel_algorithm>
1556          */
1557         if (first) {
1558                 mean = src->mean.u.f;
1559                 S = src->S.u.f;
1560         } else {
1561                 double delta = src->mean.u.f - dst->mean.u.f;
1562
1563                 mean = ((src->mean.u.f * src->samples) +
1564                         (dst->mean.u.f * dst->samples)) /
1565                         (dst->samples + src->samples);
1566
1567                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1568                         (dst->samples * src->samples) /
1569                         (dst->samples + src->samples);
1570         }
1571
1572         dst->samples += src->samples;
1573         dst->mean.u.f = mean;
1574         dst->S.u.f = S;
1575
1576 }
1577
1578 /*
1579  * We sum two kinds of stats - one that is time based, in which case we
1580  * apply the proper summing technique, and then one that is iops/bw
1581  * numbers. For group_reporting, we should just add those up, not make
1582  * them the mean of everything.
1583  */
1584 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first,
1585                      bool pure_sum)
1586 {
1587         if (src->samples == 0)
1588                 return;
1589
1590         if (!pure_sum) {
1591                 __sum_stat(dst, src, first);
1592                 return;
1593         }
1594
1595         if (first) {
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         } else {
1602                 dst->min_val += src->min_val;
1603                 dst->max_val += src->max_val;
1604                 dst->samples += src->samples;
1605                 dst->mean.u.f += src->mean.u.f;
1606                 dst->S.u.f += src->S.u.f;
1607         }
1608 }
1609
1610 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1611 {
1612         int i;
1613
1614         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1615                 if (dst->max_run[i] < src->max_run[i])
1616                         dst->max_run[i] = src->max_run[i];
1617                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1618                         dst->min_run[i] = src->min_run[i];
1619                 if (dst->max_bw[i] < src->max_bw[i])
1620                         dst->max_bw[i] = src->max_bw[i];
1621                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1622                         dst->min_bw[i] = src->min_bw[i];
1623
1624                 dst->iobytes[i] += src->iobytes[i];
1625                 dst->agg[i] += src->agg[i];
1626         }
1627
1628         if (!dst->kb_base)
1629                 dst->kb_base = src->kb_base;
1630         if (!dst->unit_base)
1631                 dst->unit_base = src->unit_base;
1632         if (!dst->sig_figs)
1633                 dst->sig_figs = src->sig_figs;
1634 }
1635
1636 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1637                       bool first)
1638 {
1639         int l, k;
1640
1641         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1642                 if (!dst->unified_rw_rep) {
1643                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first, false);
1644                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first, false);
1645                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first, false);
1646                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first, true);
1647                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first, true);
1648
1649                         dst->io_bytes[l] += src->io_bytes[l];
1650
1651                         if (dst->runtime[l] < src->runtime[l])
1652                                 dst->runtime[l] = src->runtime[l];
1653                 } else {
1654                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first, false);
1655                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first, false);
1656                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first, false);
1657                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first, true);
1658                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first, true);
1659
1660                         dst->io_bytes[0] += src->io_bytes[l];
1661
1662                         if (dst->runtime[0] < src->runtime[l])
1663                                 dst->runtime[0] = src->runtime[l];
1664
1665                         /*
1666                          * We're summing to the same destination, so override
1667                          * 'first' after the first iteration of the loop
1668                          */
1669                         first = false;
1670                 }
1671         }
1672
1673         sum_stat(&dst->sync_stat, &src->sync_stat, first, false);
1674         dst->usr_time += src->usr_time;
1675         dst->sys_time += src->sys_time;
1676         dst->ctx += src->ctx;
1677         dst->majf += src->majf;
1678         dst->minf += src->minf;
1679
1680         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
1681                 dst->io_u_map[k] += src->io_u_map[k];
1682                 dst->io_u_submit[k] += src->io_u_submit[k];
1683                 dst->io_u_complete[k] += src->io_u_complete[k];
1684         }
1685
1686         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1687                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1688         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1689                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1690         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1691                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1692
1693         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1694                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
1695
1696         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1697                 if (!dst->unified_rw_rep) {
1698                         dst->total_io_u[k] += src->total_io_u[k];
1699                         dst->short_io_u[k] += src->short_io_u[k];
1700                         dst->drop_io_u[k] += src->drop_io_u[k];
1701                 } else {
1702                         dst->total_io_u[0] += src->total_io_u[k];
1703                         dst->short_io_u[0] += src->short_io_u[k];
1704                         dst->drop_io_u[0] += src->drop_io_u[k];
1705                 }
1706         }
1707
1708         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
1709
1710         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1711                 int m;
1712
1713                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1714                         if (!dst->unified_rw_rep)
1715                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1716                         else
1717                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1718                 }
1719         }
1720
1721         dst->total_run_time += src->total_run_time;
1722         dst->total_submit += src->total_submit;
1723         dst->total_complete += src->total_complete;
1724         dst->nr_zone_resets += src->nr_zone_resets;
1725         dst->cachehit += src->cachehit;
1726         dst->cachemiss += src->cachemiss;
1727 }
1728
1729 void init_group_run_stat(struct group_run_stats *gs)
1730 {
1731         int i;
1732         memset(gs, 0, sizeof(*gs));
1733
1734         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1735                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1736 }
1737
1738 void init_thread_stat(struct thread_stat *ts)
1739 {
1740         int j;
1741
1742         memset(ts, 0, sizeof(*ts));
1743
1744         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1745                 ts->lat_stat[j].min_val = -1UL;
1746                 ts->clat_stat[j].min_val = -1UL;
1747                 ts->slat_stat[j].min_val = -1UL;
1748                 ts->bw_stat[j].min_val = -1UL;
1749                 ts->iops_stat[j].min_val = -1UL;
1750         }
1751         ts->sync_stat.min_val = -1UL;
1752         ts->groupid = -1;
1753 }
1754
1755 void __show_run_stats(void)
1756 {
1757         struct group_run_stats *runstats, *rs;
1758         struct thread_data *td;
1759         struct thread_stat *threadstats, *ts;
1760         int i, j, k, nr_ts, last_ts, idx;
1761         bool kb_base_warned = false;
1762         bool unit_base_warned = false;
1763         struct json_object *root = NULL;
1764         struct json_array *array = NULL;
1765         struct buf_output output[FIO_OUTPUT_NR];
1766         struct flist_head **opt_lists;
1767
1768         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1769
1770         for (i = 0; i < groupid + 1; i++)
1771                 init_group_run_stat(&runstats[i]);
1772
1773         /*
1774          * find out how many threads stats we need. if group reporting isn't
1775          * enabled, it's one-per-td.
1776          */
1777         nr_ts = 0;
1778         last_ts = -1;
1779         for_each_td(td, i) {
1780                 if (!td->o.group_reporting) {
1781                         nr_ts++;
1782                         continue;
1783                 }
1784                 if (last_ts == td->groupid)
1785                         continue;
1786                 if (!td->o.stats)
1787                         continue;
1788
1789                 last_ts = td->groupid;
1790                 nr_ts++;
1791         }
1792
1793         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1794         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1795
1796         for (i = 0; i < nr_ts; i++) {
1797                 init_thread_stat(&threadstats[i]);
1798                 opt_lists[i] = NULL;
1799         }
1800
1801         j = 0;
1802         last_ts = -1;
1803         idx = 0;
1804         for_each_td(td, i) {
1805                 if (!td->o.stats)
1806                         continue;
1807                 if (idx && (!td->o.group_reporting ||
1808                     (td->o.group_reporting && last_ts != td->groupid))) {
1809                         idx = 0;
1810                         j++;
1811                 }
1812
1813                 last_ts = td->groupid;
1814
1815                 ts = &threadstats[j];
1816
1817                 ts->clat_percentiles = td->o.clat_percentiles;
1818                 ts->lat_percentiles = td->o.lat_percentiles;
1819                 ts->percentile_precision = td->o.percentile_precision;
1820                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1821                 opt_lists[j] = &td->opt_list;
1822
1823                 idx++;
1824                 ts->members++;
1825
1826                 if (ts->groupid == -1) {
1827                         /*
1828                          * These are per-group shared already
1829                          */
1830                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1831                         if (td->o.description)
1832                                 strncpy(ts->description, td->o.description,
1833                                                 FIO_JOBDESC_SIZE - 1);
1834                         else
1835                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1836
1837                         /*
1838                          * If multiple entries in this group, this is
1839                          * the first member.
1840                          */
1841                         ts->thread_number = td->thread_number;
1842                         ts->groupid = td->groupid;
1843
1844                         /*
1845                          * first pid in group, not very useful...
1846                          */
1847                         ts->pid = td->pid;
1848
1849                         ts->kb_base = td->o.kb_base;
1850                         ts->unit_base = td->o.unit_base;
1851                         ts->sig_figs = td->o.sig_figs;
1852                         ts->unified_rw_rep = td->o.unified_rw_rep;
1853                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1854                         log_info("fio: kb_base differs for jobs in group, using"
1855                                  " %u as the base\n", ts->kb_base);
1856                         kb_base_warned = true;
1857                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1858                         log_info("fio: unit_base differs for jobs in group, using"
1859                                  " %u as the base\n", ts->unit_base);
1860                         unit_base_warned = true;
1861                 }
1862
1863                 ts->continue_on_error = td->o.continue_on_error;
1864                 ts->total_err_count += td->total_err_count;
1865                 ts->first_error = td->first_error;
1866                 if (!ts->error) {
1867                         if (!td->error && td->o.continue_on_error &&
1868                             td->first_error) {
1869                                 ts->error = td->first_error;
1870                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1871                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1872                         } else  if (td->error) {
1873                                 ts->error = td->error;
1874                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1875                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1876                         }
1877                 }
1878
1879                 ts->latency_depth = td->latency_qd;
1880                 ts->latency_target = td->o.latency_target;
1881                 ts->latency_percentile = td->o.latency_percentile;
1882                 ts->latency_window = td->o.latency_window;
1883
1884                 ts->nr_block_infos = td->ts.nr_block_infos;
1885                 for (k = 0; k < ts->nr_block_infos; k++)
1886                         ts->block_infos[k] = td->ts.block_infos[k];
1887
1888                 sum_thread_stats(ts, &td->ts, idx == 1);
1889
1890                 if (td->o.ss_dur) {
1891                         ts->ss_state = td->ss.state;
1892                         ts->ss_dur = td->ss.dur;
1893                         ts->ss_head = td->ss.head;
1894                         ts->ss_bw_data = td->ss.bw_data;
1895                         ts->ss_iops_data = td->ss.iops_data;
1896                         ts->ss_limit.u.f = td->ss.limit;
1897                         ts->ss_slope.u.f = td->ss.slope;
1898                         ts->ss_deviation.u.f = td->ss.deviation;
1899                         ts->ss_criterion.u.f = td->ss.criterion;
1900                 }
1901                 else
1902                         ts->ss_dur = ts->ss_state = 0;
1903         }
1904
1905         for (i = 0; i < nr_ts; i++) {
1906                 unsigned long long bw;
1907
1908                 ts = &threadstats[i];
1909                 if (ts->groupid == -1)
1910                         continue;
1911                 rs = &runstats[ts->groupid];
1912                 rs->kb_base = ts->kb_base;
1913                 rs->unit_base = ts->unit_base;
1914                 rs->sig_figs = ts->sig_figs;
1915                 rs->unified_rw_rep += ts->unified_rw_rep;
1916
1917                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1918                         if (!ts->runtime[j])
1919                                 continue;
1920                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1921                                 rs->min_run[j] = ts->runtime[j];
1922                         if (ts->runtime[j] > rs->max_run[j])
1923                                 rs->max_run[j] = ts->runtime[j];
1924
1925                         bw = 0;
1926                         if (ts->runtime[j])
1927                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1928                         if (bw < rs->min_bw[j])
1929                                 rs->min_bw[j] = bw;
1930                         if (bw > rs->max_bw[j])
1931                                 rs->max_bw[j] = bw;
1932
1933                         rs->iobytes[j] += ts->io_bytes[j];
1934                 }
1935         }
1936
1937         for (i = 0; i < groupid + 1; i++) {
1938                 int ddir;
1939
1940                 rs = &runstats[i];
1941
1942                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1943                         if (rs->max_run[ddir])
1944                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1945                                                 rs->max_run[ddir];
1946                 }
1947         }
1948
1949         for (i = 0; i < FIO_OUTPUT_NR; i++)
1950                 buf_output_init(&output[i]);
1951
1952         /*
1953          * don't overwrite last signal output
1954          */
1955         if (output_format & FIO_OUTPUT_NORMAL)
1956                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1957         if (output_format & FIO_OUTPUT_JSON) {
1958                 struct thread_data *global;
1959                 char time_buf[32];
1960                 struct timeval now;
1961                 unsigned long long ms_since_epoch;
1962                 time_t tv_sec;
1963
1964                 gettimeofday(&now, NULL);
1965                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1966                                  (unsigned long long)(now.tv_usec) / 1000;
1967
1968                 tv_sec = now.tv_sec;
1969                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
1970                 if (time_buf[strlen(time_buf) - 1] == '\n')
1971                         time_buf[strlen(time_buf) - 1] = '\0';
1972
1973                 root = json_create_object();
1974                 json_object_add_value_string(root, "fio version", fio_version_string);
1975                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1976                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1977                 json_object_add_value_string(root, "time", time_buf);
1978                 global = get_global_options();
1979                 json_add_job_opts(root, "global options", &global->opt_list);
1980                 array = json_create_array();
1981                 json_object_add_value_array(root, "jobs", array);
1982         }
1983
1984         if (is_backend)
1985                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1986
1987         for (i = 0; i < nr_ts; i++) {
1988                 ts = &threadstats[i];
1989                 rs = &runstats[ts->groupid];
1990
1991                 if (is_backend) {
1992                         fio_server_send_job_options(opt_lists[i], i);
1993                         fio_server_send_ts(ts, rs);
1994                 } else {
1995                         if (output_format & FIO_OUTPUT_TERSE)
1996                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1997                         if (output_format & FIO_OUTPUT_JSON) {
1998                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1999                                 json_array_add_value_object(array, tmp);
2000                         }
2001                         if (output_format & FIO_OUTPUT_NORMAL)
2002                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2003                 }
2004         }
2005         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2006                 /* disk util stats, if any */
2007                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2008
2009                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2010
2011                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2012                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2013                 json_free_object(root);
2014         }
2015
2016         for (i = 0; i < groupid + 1; i++) {
2017                 rs = &runstats[i];
2018
2019                 rs->groupid = i;
2020                 if (is_backend)
2021                         fio_server_send_gs(rs);
2022                 else if (output_format & FIO_OUTPUT_NORMAL)
2023                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2024         }
2025
2026         if (is_backend)
2027                 fio_server_send_du();
2028         else if (output_format & FIO_OUTPUT_NORMAL) {
2029                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2030                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2031         }
2032
2033         for (i = 0; i < FIO_OUTPUT_NR; i++) {
2034                 struct buf_output *out = &output[i];
2035
2036                 log_info_buf(out->buf, out->buflen);
2037                 buf_output_free(out);
2038         }
2039
2040         fio_idle_prof_cleanup();
2041
2042         log_info_flush();
2043         free(runstats);
2044         free(threadstats);
2045         free(opt_lists);
2046 }
2047
2048 void __show_running_run_stats(void)
2049 {
2050         struct thread_data *td;
2051         unsigned long long *rt;
2052         struct timespec ts;
2053         int i;
2054
2055         fio_sem_down(stat_sem);
2056
2057         rt = malloc(thread_number * sizeof(unsigned long long));
2058         fio_gettime(&ts, NULL);
2059
2060         for_each_td(td, i) {
2061                 td->update_rusage = 1;
2062                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
2063                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
2064                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2065                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2066
2067                 rt[i] = mtime_since(&td->start, &ts);
2068                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2069                         td->ts.runtime[DDIR_READ] += rt[i];
2070                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2071                         td->ts.runtime[DDIR_WRITE] += rt[i];
2072                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2073                         td->ts.runtime[DDIR_TRIM] += rt[i];
2074         }
2075
2076         for_each_td(td, i) {
2077                 if (td->runstate >= TD_EXITED)
2078                         continue;
2079                 if (td->rusage_sem) {
2080                         td->update_rusage = 1;
2081                         fio_sem_down(td->rusage_sem);
2082                 }
2083                 td->update_rusage = 0;
2084         }
2085
2086         __show_run_stats();
2087
2088         for_each_td(td, i) {
2089                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2090                         td->ts.runtime[DDIR_READ] -= rt[i];
2091                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2092                         td->ts.runtime[DDIR_WRITE] -= rt[i];
2093                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2094                         td->ts.runtime[DDIR_TRIM] -= rt[i];
2095         }
2096
2097         free(rt);
2098         fio_sem_up(stat_sem);
2099 }
2100
2101 static bool status_interval_init;
2102 static struct timespec status_time;
2103 static bool status_file_disabled;
2104
2105 #define FIO_STATUS_FILE         "fio-dump-status"
2106
2107 static int check_status_file(void)
2108 {
2109         struct stat sb;
2110         const char *temp_dir;
2111         char fio_status_file_path[PATH_MAX];
2112
2113         if (status_file_disabled)
2114                 return 0;
2115
2116         temp_dir = getenv("TMPDIR");
2117         if (temp_dir == NULL) {
2118                 temp_dir = getenv("TEMP");
2119                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2120                         temp_dir = NULL;
2121         }
2122         if (temp_dir == NULL)
2123                 temp_dir = "/tmp";
2124
2125         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2126
2127         if (stat(fio_status_file_path, &sb))
2128                 return 0;
2129
2130         if (unlink(fio_status_file_path) < 0) {
2131                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2132                                                         strerror(errno));
2133                 log_err("fio: disabling status file updates\n");
2134                 status_file_disabled = true;
2135         }
2136
2137         return 1;
2138 }
2139
2140 void check_for_running_stats(void)
2141 {
2142         if (status_interval) {
2143                 if (!status_interval_init) {
2144                         fio_gettime(&status_time, NULL);
2145                         status_interval_init = true;
2146                 } else if (mtime_since_now(&status_time) >= status_interval) {
2147                         show_running_run_stats();
2148                         fio_gettime(&status_time, NULL);
2149                         return;
2150                 }
2151         }
2152         if (check_status_file()) {
2153                 show_running_run_stats();
2154                 return;
2155         }
2156 }
2157
2158 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2159 {
2160         double val = data;
2161         double delta;
2162
2163         if (data > is->max_val)
2164                 is->max_val = data;
2165         if (data < is->min_val)
2166                 is->min_val = data;
2167
2168         delta = val - is->mean.u.f;
2169         if (delta) {
2170                 is->mean.u.f += delta / (is->samples + 1.0);
2171                 is->S.u.f += delta * (val - is->mean.u.f);
2172         }
2173
2174         is->samples++;
2175 }
2176
2177 /*
2178  * Return a struct io_logs, which is added to the tail of the log
2179  * list for 'iolog'.
2180  */
2181 static struct io_logs *get_new_log(struct io_log *iolog)
2182 {
2183         size_t new_size, new_samples;
2184         struct io_logs *cur_log;
2185
2186         /*
2187          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2188          * forever
2189          */
2190         if (!iolog->cur_log_max)
2191                 new_samples = DEF_LOG_ENTRIES;
2192         else {
2193                 new_samples = iolog->cur_log_max * 2;
2194                 if (new_samples > MAX_LOG_ENTRIES)
2195                         new_samples = MAX_LOG_ENTRIES;
2196         }
2197
2198         new_size = new_samples * log_entry_sz(iolog);
2199
2200         cur_log = smalloc(sizeof(*cur_log));
2201         if (cur_log) {
2202                 INIT_FLIST_HEAD(&cur_log->list);
2203                 cur_log->log = malloc(new_size);
2204                 if (cur_log->log) {
2205                         cur_log->nr_samples = 0;
2206                         cur_log->max_samples = new_samples;
2207                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2208                         iolog->cur_log_max = new_samples;
2209                         return cur_log;
2210                 }
2211                 sfree(cur_log);
2212         }
2213
2214         return NULL;
2215 }
2216
2217 /*
2218  * Add and return a new log chunk, or return current log if big enough
2219  */
2220 static struct io_logs *regrow_log(struct io_log *iolog)
2221 {
2222         struct io_logs *cur_log;
2223         int i;
2224
2225         if (!iolog || iolog->disabled)
2226                 goto disable;
2227
2228         cur_log = iolog_cur_log(iolog);
2229         if (!cur_log) {
2230                 cur_log = get_new_log(iolog);
2231                 if (!cur_log)
2232                         return NULL;
2233         }
2234
2235         if (cur_log->nr_samples < cur_log->max_samples)
2236                 return cur_log;
2237
2238         /*
2239          * No room for a new sample. If we're compressing on the fly, flush
2240          * out the current chunk
2241          */
2242         if (iolog->log_gz) {
2243                 if (iolog_cur_flush(iolog, cur_log)) {
2244                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2245                         return NULL;
2246                 }
2247         }
2248
2249         /*
2250          * Get a new log array, and add to our list
2251          */
2252         cur_log = get_new_log(iolog);
2253         if (!cur_log) {
2254                 log_err("fio: failed extending iolog! Will stop logging.\n");
2255                 return NULL;
2256         }
2257
2258         if (!iolog->pending || !iolog->pending->nr_samples)
2259                 return cur_log;
2260
2261         /*
2262          * Flush pending items to new log
2263          */
2264         for (i = 0; i < iolog->pending->nr_samples; i++) {
2265                 struct io_sample *src, *dst;
2266
2267                 src = get_sample(iolog, iolog->pending, i);
2268                 dst = get_sample(iolog, cur_log, i);
2269                 memcpy(dst, src, log_entry_sz(iolog));
2270         }
2271         cur_log->nr_samples = iolog->pending->nr_samples;
2272
2273         iolog->pending->nr_samples = 0;
2274         return cur_log;
2275 disable:
2276         if (iolog)
2277                 iolog->disabled = true;
2278         return NULL;
2279 }
2280
2281 void regrow_logs(struct thread_data *td)
2282 {
2283         regrow_log(td->slat_log);
2284         regrow_log(td->clat_log);
2285         regrow_log(td->clat_hist_log);
2286         regrow_log(td->lat_log);
2287         regrow_log(td->bw_log);
2288         regrow_log(td->iops_log);
2289         td->flags &= ~TD_F_REGROW_LOGS;
2290 }
2291
2292 static struct io_logs *get_cur_log(struct io_log *iolog)
2293 {
2294         struct io_logs *cur_log;
2295
2296         cur_log = iolog_cur_log(iolog);
2297         if (!cur_log) {
2298                 cur_log = get_new_log(iolog);
2299                 if (!cur_log)
2300                         return NULL;
2301         }
2302
2303         if (cur_log->nr_samples < cur_log->max_samples)
2304                 return cur_log;
2305
2306         /*
2307          * Out of space. If we're in IO offload mode, or we're not doing
2308          * per unit logging (hence logging happens outside of the IO thread
2309          * as well), add a new log chunk inline. If we're doing inline
2310          * submissions, flag 'td' as needing a log regrow and we'll take
2311          * care of it on the submission side.
2312          */
2313         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2314             !per_unit_log(iolog))
2315                 return regrow_log(iolog);
2316
2317         if (iolog->td)
2318                 iolog->td->flags |= TD_F_REGROW_LOGS;
2319         if (iolog->pending)
2320                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2321         return iolog->pending;
2322 }
2323
2324 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2325                              enum fio_ddir ddir, unsigned long long bs,
2326                              unsigned long t, uint64_t offset)
2327 {
2328         struct io_logs *cur_log;
2329
2330         if (iolog->disabled)
2331                 return;
2332         if (flist_empty(&iolog->io_logs))
2333                 iolog->avg_last[ddir] = t;
2334
2335         cur_log = get_cur_log(iolog);
2336         if (cur_log) {
2337                 struct io_sample *s;
2338
2339                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2340
2341                 s->data = data;
2342                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2343                 io_sample_set_ddir(iolog, s, ddir);
2344                 s->bs = bs;
2345
2346                 if (iolog->log_offset) {
2347                         struct io_sample_offset *so = (void *) s;
2348
2349                         so->offset = offset;
2350                 }
2351
2352                 cur_log->nr_samples++;
2353                 return;
2354         }
2355
2356         iolog->disabled = true;
2357 }
2358
2359 static inline void reset_io_stat(struct io_stat *ios)
2360 {
2361         ios->min_val = -1ULL;
2362         ios->max_val = ios->samples = 0;
2363         ios->mean.u.f = ios->S.u.f = 0;
2364 }
2365
2366 void reset_io_stats(struct thread_data *td)
2367 {
2368         struct thread_stat *ts = &td->ts;
2369         int i, j;
2370
2371         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2372                 reset_io_stat(&ts->clat_stat[i]);
2373                 reset_io_stat(&ts->slat_stat[i]);
2374                 reset_io_stat(&ts->lat_stat[i]);
2375                 reset_io_stat(&ts->bw_stat[i]);
2376                 reset_io_stat(&ts->iops_stat[i]);
2377
2378                 ts->io_bytes[i] = 0;
2379                 ts->runtime[i] = 0;
2380                 ts->total_io_u[i] = 0;
2381                 ts->short_io_u[i] = 0;
2382                 ts->drop_io_u[i] = 0;
2383
2384                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2385                         ts->io_u_plat[i][j] = 0;
2386                         if (!i)
2387                                 ts->io_u_sync_plat[j] = 0;
2388                 }
2389         }
2390
2391         ts->total_io_u[DDIR_SYNC] = 0;
2392
2393         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2394                 ts->io_u_map[i] = 0;
2395                 ts->io_u_submit[i] = 0;
2396                 ts->io_u_complete[i] = 0;
2397         }
2398
2399         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2400                 ts->io_u_lat_n[i] = 0;
2401         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2402                 ts->io_u_lat_u[i] = 0;
2403         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2404                 ts->io_u_lat_m[i] = 0;
2405
2406         ts->total_submit = 0;
2407         ts->total_complete = 0;
2408         ts->nr_zone_resets = 0;
2409         ts->cachehit = ts->cachemiss = 0;
2410 }
2411
2412 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2413                               unsigned long elapsed, bool log_max)
2414 {
2415         /*
2416          * Note an entry in the log. Use the mean from the logged samples,
2417          * making sure to properly round up. Only write a log entry if we
2418          * had actual samples done.
2419          */
2420         if (iolog->avg_window[ddir].samples) {
2421                 union io_sample_data data;
2422
2423                 if (log_max)
2424                         data.val = iolog->avg_window[ddir].max_val;
2425                 else
2426                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2427
2428                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2429         }
2430
2431         reset_io_stat(&iolog->avg_window[ddir]);
2432 }
2433
2434 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2435                              bool log_max)
2436 {
2437         int ddir;
2438
2439         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2440                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2441 }
2442
2443 static unsigned long add_log_sample(struct thread_data *td,
2444                                     struct io_log *iolog,
2445                                     union io_sample_data data,
2446                                     enum fio_ddir ddir, unsigned long long bs,
2447                                     uint64_t offset)
2448 {
2449         unsigned long elapsed, this_window;
2450
2451         if (!ddir_rw(ddir))
2452                 return 0;
2453
2454         elapsed = mtime_since_now(&td->epoch);
2455
2456         /*
2457          * If no time averaging, just add the log sample.
2458          */
2459         if (!iolog->avg_msec) {
2460                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2461                 return 0;
2462         }
2463
2464         /*
2465          * Add the sample. If the time period has passed, then
2466          * add that entry to the log and clear.
2467          */
2468         add_stat_sample(&iolog->avg_window[ddir], data.val);
2469
2470         /*
2471          * If period hasn't passed, adding the above sample is all we
2472          * need to do.
2473          */
2474         this_window = elapsed - iolog->avg_last[ddir];
2475         if (elapsed < iolog->avg_last[ddir])
2476                 return iolog->avg_last[ddir] - elapsed;
2477         else if (this_window < iolog->avg_msec) {
2478                 unsigned long diff = iolog->avg_msec - this_window;
2479
2480                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2481                         return diff;
2482         }
2483
2484         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2485
2486         iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2487         return iolog->avg_msec;
2488 }
2489
2490 void finalize_logs(struct thread_data *td, bool unit_logs)
2491 {
2492         unsigned long elapsed;
2493
2494         elapsed = mtime_since_now(&td->epoch);
2495
2496         if (td->clat_log && unit_logs)
2497                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2498         if (td->slat_log && unit_logs)
2499                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2500         if (td->lat_log && unit_logs)
2501                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2502         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2503                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2504         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2505                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2506 }
2507
2508 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned long long bs)
2509 {
2510         struct io_log *iolog;
2511
2512         if (!ddir_rw(ddir))
2513                 return;
2514
2515         iolog = agg_io_log[ddir];
2516         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2517 }
2518
2519 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2520 {
2521         unsigned int idx = plat_val_to_idx(nsec);
2522         assert(idx < FIO_IO_U_PLAT_NR);
2523
2524         ts->io_u_sync_plat[idx]++;
2525         add_stat_sample(&ts->sync_stat, nsec);
2526 }
2527
2528 static void add_clat_percentile_sample(struct thread_stat *ts,
2529                                 unsigned long long nsec, enum fio_ddir ddir)
2530 {
2531         unsigned int idx = plat_val_to_idx(nsec);
2532         assert(idx < FIO_IO_U_PLAT_NR);
2533
2534         ts->io_u_plat[ddir][idx]++;
2535 }
2536
2537 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2538                      unsigned long long nsec, unsigned long long bs,
2539                      uint64_t offset)
2540 {
2541         const bool needs_lock = td_async_processing(td);
2542         unsigned long elapsed, this_window;
2543         struct thread_stat *ts = &td->ts;
2544         struct io_log *iolog = td->clat_hist_log;
2545
2546         if (needs_lock)
2547                 __td_io_u_lock(td);
2548
2549         add_stat_sample(&ts->clat_stat[ddir], nsec);
2550
2551         if (td->clat_log)
2552                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2553                                offset);
2554
2555         if (ts->clat_percentiles)
2556                 add_clat_percentile_sample(ts, nsec, ddir);
2557
2558         if (iolog && iolog->hist_msec) {
2559                 struct io_hist *hw = &iolog->hist_window[ddir];
2560
2561                 hw->samples++;
2562                 elapsed = mtime_since_now(&td->epoch);
2563                 if (!hw->hist_last)
2564                         hw->hist_last = elapsed;
2565                 this_window = elapsed - hw->hist_last;
2566
2567                 if (this_window >= iolog->hist_msec) {
2568                         uint64_t *io_u_plat;
2569                         struct io_u_plat_entry *dst;
2570
2571                         /*
2572                          * Make a byte-for-byte copy of the latency histogram
2573                          * stored in td->ts.io_u_plat[ddir], recording it in a
2574                          * log sample. Note that the matching call to free() is
2575                          * located in iolog.c after printing this sample to the
2576                          * log file.
2577                          */
2578                         io_u_plat = (uint64_t *) td->ts.io_u_plat[ddir];
2579                         dst = malloc(sizeof(struct io_u_plat_entry));
2580                         memcpy(&(dst->io_u_plat), io_u_plat,
2581                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2582                         flist_add(&dst->list, &hw->list);
2583                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2584                                                 elapsed, offset);
2585
2586                         /*
2587                          * Update the last time we recorded as being now, minus
2588                          * any drift in time we encountered before actually
2589                          * making the record.
2590                          */
2591                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2592                         hw->samples = 0;
2593                 }
2594         }
2595
2596         if (needs_lock)
2597                 __td_io_u_unlock(td);
2598 }
2599
2600 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2601                      unsigned long usec, unsigned long long bs, uint64_t offset)
2602 {
2603         const bool needs_lock = td_async_processing(td);
2604         struct thread_stat *ts = &td->ts;
2605
2606         if (!ddir_rw(ddir))
2607                 return;
2608
2609         if (needs_lock)
2610                 __td_io_u_lock(td);
2611
2612         add_stat_sample(&ts->slat_stat[ddir], usec);
2613
2614         if (td->slat_log)
2615                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2616
2617         if (needs_lock)
2618                 __td_io_u_unlock(td);
2619 }
2620
2621 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2622                     unsigned long long nsec, unsigned long long bs,
2623                     uint64_t offset)
2624 {
2625         const bool needs_lock = td_async_processing(td);
2626         struct thread_stat *ts = &td->ts;
2627
2628         if (!ddir_rw(ddir))
2629                 return;
2630
2631         if (needs_lock)
2632                 __td_io_u_lock(td);
2633
2634         add_stat_sample(&ts->lat_stat[ddir], nsec);
2635
2636         if (td->lat_log)
2637                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2638                                offset);
2639
2640         if (ts->lat_percentiles)
2641                 add_clat_percentile_sample(ts, nsec, ddir);
2642
2643         if (needs_lock)
2644                 __td_io_u_unlock(td);
2645 }
2646
2647 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2648                    unsigned int bytes, unsigned long long spent)
2649 {
2650         const bool needs_lock = td_async_processing(td);
2651         struct thread_stat *ts = &td->ts;
2652         unsigned long rate;
2653
2654         if (spent)
2655                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2656         else
2657                 rate = 0;
2658
2659         if (needs_lock)
2660                 __td_io_u_lock(td);
2661
2662         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2663
2664         if (td->bw_log)
2665                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2666                                bytes, io_u->offset);
2667
2668         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2669
2670         if (needs_lock)
2671                 __td_io_u_unlock(td);
2672 }
2673
2674 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2675                          struct timespec *t, unsigned int avg_time,
2676                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2677                          struct io_stat *stat, struct io_log *log,
2678                          bool is_kb)
2679 {
2680         const bool needs_lock = td_async_processing(td);
2681         unsigned long spent, rate;
2682         enum fio_ddir ddir;
2683         unsigned long next, next_log;
2684
2685         next_log = avg_time;
2686
2687         spent = mtime_since(parent_tv, t);
2688         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2689                 return avg_time - spent;
2690
2691         if (needs_lock)
2692                 __td_io_u_lock(td);
2693
2694         /*
2695          * Compute both read and write rates for the interval.
2696          */
2697         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2698                 uint64_t delta;
2699
2700                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2701                 if (!delta)
2702                         continue; /* No entries for interval */
2703
2704                 if (spent) {
2705                         if (is_kb)
2706                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2707                         else
2708                                 rate = (delta * 1000) / spent;
2709                 } else
2710                         rate = 0;
2711
2712                 add_stat_sample(&stat[ddir], rate);
2713
2714                 if (log) {
2715                         unsigned long long bs = 0;
2716
2717                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2718                                 bs = td->o.min_bs[ddir];
2719
2720                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2721                         next_log = min(next_log, next);
2722                 }
2723
2724                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2725         }
2726
2727         timespec_add_msec(parent_tv, avg_time);
2728
2729         if (needs_lock)
2730                 __td_io_u_unlock(td);
2731
2732         if (spent <= avg_time)
2733                 next = avg_time;
2734         else
2735                 next = avg_time - (1 + spent - avg_time);
2736
2737         return min(next, next_log);
2738 }
2739
2740 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2741 {
2742         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2743                                 td->this_io_bytes, td->stat_io_bytes,
2744                                 td->ts.bw_stat, td->bw_log, true);
2745 }
2746
2747 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2748                      unsigned int bytes)
2749 {
2750         const bool needs_lock = td_async_processing(td);
2751         struct thread_stat *ts = &td->ts;
2752
2753         if (needs_lock)
2754                 __td_io_u_lock(td);
2755
2756         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2757
2758         if (td->iops_log)
2759                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2760                                bytes, io_u->offset);
2761
2762         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2763
2764         if (needs_lock)
2765                 __td_io_u_unlock(td);
2766 }
2767
2768 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2769 {
2770         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2771                                 td->this_io_blocks, td->stat_io_blocks,
2772                                 td->ts.iops_stat, td->iops_log, false);
2773 }
2774
2775 /*
2776  * Returns msecs to next event
2777  */
2778 int calc_log_samples(void)
2779 {
2780         struct thread_data *td;
2781         unsigned int next = ~0U, tmp;
2782         struct timespec now;
2783         int i;
2784
2785         fio_gettime(&now, NULL);
2786
2787         for_each_td(td, i) {
2788                 if (!td->o.stats)
2789                         continue;
2790                 if (in_ramp_time(td) ||
2791                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2792                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2793                         continue;
2794                 }
2795                 if (!td->bw_log ||
2796                         (td->bw_log && !per_unit_log(td->bw_log))) {
2797                         tmp = add_bw_samples(td, &now);
2798                         if (tmp < next)
2799                                 next = tmp;
2800                 }
2801                 if (!td->iops_log ||
2802                         (td->iops_log && !per_unit_log(td->iops_log))) {
2803                         tmp = add_iops_samples(td, &now);
2804                         if (tmp < next)
2805                                 next = tmp;
2806                 }
2807         }
2808
2809         return next == ~0U ? 0 : next;
2810 }
2811
2812 void stat_init(void)
2813 {
2814         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
2815 }
2816
2817 void stat_exit(void)
2818 {
2819         /*
2820          * When we have the mutex, we know out-of-band access to it
2821          * have ended.
2822          */
2823         fio_sem_down(stat_sem);
2824         fio_sem_remove(stat_sem);
2825 }
2826
2827 /*
2828  * Called from signal handler. Wake up status thread.
2829  */
2830 void show_running_run_stats(void)
2831 {
2832         helper_do_stat();
2833 }
2834
2835 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2836 {
2837         /* Ignore io_u's which span multiple blocks--they will just get
2838          * inaccurate counts. */
2839         int idx = (io_u->offset - io_u->file->file_offset)
2840                         / td->o.bs[DDIR_TRIM];
2841         uint32_t *info = &td->ts.block_infos[idx];
2842         assert(idx < td->ts.nr_block_infos);
2843         return info;
2844 }