stat.c

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