stat.c

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