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 #include "oslib/asprintf.h"
  19
  20 #define LOG_MSEC_SLACK  1
  21
  22 struct fio_sem *stat_sem;
  23
  24 void clear_rusage_stat(struct thread_data *td)
  25 {
  26         struct thread_stat *ts = &td->ts;
  27
  28         fio_getrusage(&td->ru_start);
  29         ts->usr_time = ts->sys_time = 0;
  30         ts->ctx = 0;
  31         ts->minf = ts->majf = 0;
  32 }
  33
  34 void update_rusage_stat(struct thread_data *td)
  35 {
  36         struct thread_stat *ts = &td->ts;
  37
  38         fio_getrusage(&td->ru_end);
  39         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
  40                                         &td->ru_end.ru_utime);
  41         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
  42                                         &td->ru_end.ru_stime);
  43         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  44                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  45         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  46         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  47
  48         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  49 }
  50
  51 /*
  52  * Given a latency, return the index of the corresponding bucket in
  53  * the structure tracking percentiles.
  54  *
  55  * (1) find the group (and error bits) that the value (latency)
  56  * belongs to by looking at its MSB. (2) find the bucket number in the
  57  * group by looking at the index bits.
  58  *
  59  */
  60 static unsigned int plat_val_to_idx(unsigned long long val)
  61 {
  62         unsigned int msb, error_bits, base, offset, idx;
  63
  64         /* Find MSB starting from bit 0 */
  65         if (val == 0)
  66                 msb = 0;
  67         else
  68                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
  69
  70         /*
  71          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  72          * all bits of the sample as index
  73          */
  74         if (msb <= FIO_IO_U_PLAT_BITS)
  75                 return val;
  76
  77         /* Compute the number of error bits to discard*/
  78         error_bits = msb - FIO_IO_U_PLAT_BITS;
  79
  80         /* Compute the number of buckets before the group */
  81         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  82
  83         /*
  84          * Discard the error bits and apply the mask to find the
  85          * index for the buckets in the group
  86          */
  87         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  88
  89         /* Make sure the index does not exceed (array size - 1) */
  90         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  91                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  92
  93         return idx;
  94 }
  95
  96 /*
  97  * Convert the given index of the bucket array to the value
  98  * represented by the bucket
  99  */
 100 static unsigned long long plat_idx_to_val(unsigned int idx)
 101 {
 102         unsigned int error_bits;
 103         unsigned long long k, base;
 104
 105         assert(idx < FIO_IO_U_PLAT_NR);
 106
 107         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 108          * all bits of the sample as index */
 109         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 110                 return idx;
 111
 112         /* Find the group and compute the minimum value of that group */
 113         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 114         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
 115
 116         /* Find its bucket number of the group */
 117         k = idx % FIO_IO_U_PLAT_VAL;
 118
 119         /* Return the mean of the range of the bucket */
 120         return base + ((k + 0.5) * (1 << error_bits));
 121 }
 122
 123 static int double_cmp(const void *a, const void *b)
 124 {
 125         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 126         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 127         int cmp = 0;
 128
 129         if (fa.u.f > fb.u.f)
 130                 cmp = 1;
 131         else if (fa.u.f < fb.u.f)
 132                 cmp = -1;
 133
 134         return cmp;
 135 }
 136
 137 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 138                                    fio_fp64_t *plist, unsigned long long **output,
 139                                    unsigned long long *maxv, unsigned long long *minv)
 140 {
 141         unsigned long long sum = 0;
 142         unsigned int len, i, j = 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(plist, len, sizeof(plist[0]), double_cmp);
 163
 164         ovals = malloc(len * sizeof(*ovals));
 165         if (!ovals)
 166                 return 0;
 167
 168         /*
 169          * Calculate bucket values, note down max and min values
 170          */
 171         is_last = false;
 172         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 173                 sum += io_u_plat[i];
 174                 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
 175                         assert(plist[j].u.f <= 100.0);
 176
 177                         ovals[j] = plat_idx_to_val(i);
 178                         if (ovals[j] < *minv)
 179                                 *minv = ovals[j];
 180                         if (ovals[j] > *maxv)
 181                                 *maxv = ovals[j];
 182
 183                         is_last = (j == len - 1) != 0;
 184                         if (is_last)
 185                                 break;
 186
 187                         j++;
 188                 }
 189         }
 190
 191         if (!is_last)
 192                 log_err("fio: error calculating latency percentiles\n");
 193
 194         *output = ovals;
 195         return len;
 196 }
 197
 198 /*
 199  * Find and display the p-th percentile of clat
 200  */
 201 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 202                                   fio_fp64_t *plist, unsigned int precision,
 203                                   const char *pre, struct buf_output *out)
 204 {
 205         unsigned int divisor, len, i, j = 0;
 206         unsigned long long minv, maxv;
 207         unsigned long long *ovals;
 208         int per_line, scale_down, time_width;
 209         bool is_last;
 210         char fmt[32];
 211
 212         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 213         if (!len || !ovals)
 214                 goto out;
 215
 216         /*
 217          * We default to nsecs, but if the value range is such that we
 218          * should scale down to usecs or msecs, do that.
 219          */
 220         if (minv > 2000000 && maxv > 99999999ULL) {
 221                 scale_down = 2;
 222                 divisor = 1000000;
 223                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 224         } else if (minv > 2000 && maxv > 99999) {
 225                 scale_down = 1;
 226                 divisor = 1000;
 227                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 228         } else {
 229                 scale_down = 0;
 230                 divisor = 1;
 231                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 232         }
 233
 234
 235         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 236         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 237                         precision, time_width);
 238         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 239         per_line = (80 - 7) / (precision + 10 + time_width);
 240
 241         for (j = 0; j < len; j++) {
 242                 /* for formatting */
 243                 if (j != 0 && (j % per_line) == 0)
 244                         log_buf(out, "     |");
 245
 246                 /* end of the list */
 247                 is_last = (j == len - 1) != 0;
 248
 249                 for (i = 0; i < scale_down; i++)
 250                         ovals[j] = (ovals[j] + 999) / 1000;
 251
 252                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 253
 254                 if (is_last)
 255                         break;
 256
 257                 if ((j % per_line) == per_line - 1)     /* for formatting */
 258                         log_buf(out, "\n");
 259         }
 260
 261 out:
 262         free(ovals);
 263 }
 264
 265 bool calc_lat(struct io_stat *is, unsigned long long *min,
 266               unsigned long long *max, double *mean, double *dev)
 267 {
 268         double n = (double) is->samples;
 269
 270         if (n == 0)
 271                 return false;
 272
 273         *min = is->min_val;
 274         *max = is->max_val;
 275         *mean = is->mean.u.f;
 276
 277         if (n > 1.0)
 278                 *dev = sqrt(is->S.u.f / (n - 1.0));
 279         else
 280                 *dev = 0;
 281
 282         return true;
 283 }
 284
 285 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 286 {
 287         char *io, *agg, *min, *max;
 288         char *ioalt, *aggalt, *minalt, *maxalt;
 289         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 290         int i;
 291
 292         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 293
 294         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 295                 const int i2p = is_power_of_2(rs->kb_base);
 296
 297                 if (!rs->max_run[i])
 298                         continue;
 299
 300                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
 301                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
 302                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
 303                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 304                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 305                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 306                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 307                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 308                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 309                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 310                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 311                                 (unsigned long long) rs->min_run[i],
 312                                 (unsigned long long) rs->max_run[i]);
 313
 314                 free(io);
 315                 free(agg);
 316                 free(min);
 317                 free(max);
 318                 free(ioalt);
 319                 free(aggalt);
 320                 free(minalt);
 321                 free(maxalt);
 322         }
 323 }
 324
 325 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
 326 {
 327         int i;
 328
 329         /*
 330          * Do depth distribution calculations
 331          */
 332         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 333                 if (total) {
 334                         io_u_dist[i] = (double) map[i] / (double) total;
 335                         io_u_dist[i] *= 100.0;
 336                         if (io_u_dist[i] < 0.1 && map[i])
 337                                 io_u_dist[i] = 0.1;
 338                 } else
 339                         io_u_dist[i] = 0.0;
 340         }
 341 }
 342
 343 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 344                           uint64_t *src, int nr)
 345 {
 346         unsigned long total = ddir_rw_sum(ts->total_io_u);
 347         int i;
 348
 349         /*
 350          * Do latency distribution calculations
 351          */
 352         for (i = 0; i < nr; i++) {
 353                 if (total) {
 354                         dst[i] = (double) src[i] / (double) total;
 355                         dst[i] *= 100.0;
 356                         if (dst[i] < 0.01 && src[i])
 357                                 dst[i] = 0.01;
 358                 } else
 359                         dst[i] = 0.0;
 360         }
 361 }
 362
 363 /*
 364  * To keep the terse format unaltered, add all of the ns latency
 365  * buckets to the first us latency bucket
 366  */
 367 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 368 {
 369         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 370         int i;
 371
 372         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 373
 374         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 375                 ntotal += ts->io_u_lat_n[i];
 376
 377         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 378 }
 379
 380 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 381 {
 382         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 383 }
 384
 385 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 386 {
 387         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 388 }
 389
 390 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 391 {
 392         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 393 }
 394
 395 static void display_lat(const char *name, unsigned long long min,
 396                         unsigned long long max, double mean, double dev,
 397                         struct buf_output *out)
 398 {
 399         const char *base = "(nsec)";
 400         char *minp, *maxp;
 401
 402         if (nsec_to_msec(&min, &max, &mean, &dev))
 403                 base = "(msec)";
 404         else if (nsec_to_usec(&min, &max, &mean, &dev))
 405                 base = "(usec)";
 406
 407         minp = num2str(min, 6, 1, 0, N2S_NONE);
 408         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 409
 410         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 411                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 412
 413         free(minp);
 414         free(maxp);
 415 }
 416
 417 static double convert_agg_kbytes_percent(struct group_run_stats *rs, int ddir, int mean)
 418 {
 419         double p_of_agg = 100.0;
 420         if (rs && rs->agg[ddir] > 1024) {
 421                 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024.0);
 422
 423                 if (p_of_agg > 100.0)
 424                         p_of_agg = 100.0;
 425         }
 426         return p_of_agg;
 427 }
 428
 429 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 430                              int ddir, struct buf_output *out)
 431 {
 432         unsigned long runt;
 433         unsigned long long min, max, bw, iops;
 434         double mean, dev;
 435         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
 436         int i2p;
 437
 438         if (ddir_sync(ddir)) {
 439                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
 440                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
 441                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
 442                         show_clat_percentiles(ts->io_u_sync_plat,
 443                                                 ts->sync_stat.samples,
 444                                                 ts->percentile_list,
 445                                                 ts->percentile_precision,
 446                                                 io_ddir_name(ddir), out);
 447                 }
 448                 return;
 449         }
 450
 451         assert(ddir_rw(ddir));
 452
 453         if (!ts->runtime[ddir])
 454                 return;
 455
 456         i2p = is_power_of_2(rs->kb_base);
 457         runt = ts->runtime[ddir];
 458
 459         bw = (1000 * ts->io_bytes[ddir]) / runt;
 460         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
 461         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
 462         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
 463
 464         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 465         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
 466         if (ddir == DDIR_WRITE)
 467                 post_st = zbd_write_status(ts);
 468         else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
 469                 uint64_t total;
 470                 double hit;
 471
 472                 total = ts->cachehit + ts->cachemiss;
 473                 hit = (double) ts->cachehit / (double) total;
 474                 hit *= 100.0;
 475                 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
 476                         post_st = NULL;
 477         }
 478
 479         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
 480                         rs->unified_rw_rep ? "mixed" : io_ddir_name(ddir),
 481                         iops_p, bw_p, bw_p_alt, io_p,
 482                         (unsigned long long) ts->runtime[ddir],
 483                         post_st ? : "");
 484
 485         free(post_st);
 486         free(io_p);
 487         free(bw_p);
 488         free(bw_p_alt);
 489         free(iops_p);
 490
 491         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 492                 display_lat("slat", min, max, mean, dev, out);
 493         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 494                 display_lat("clat", min, max, mean, dev, out);
 495         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 496                 display_lat(" lat", min, max, mean, dev, out);
 497         if (calc_lat(&ts->clat_high_prio_stat[ddir], &min, &max, &mean, &dev)) {
 498                 display_lat(ts->lat_percentiles ? "high prio_lat" : "high prio_clat",
 499                                 min, max, mean, dev, out);
 500                 if (calc_lat(&ts->clat_low_prio_stat[ddir], &min, &max, &mean, &dev))
 501                         display_lat(ts->lat_percentiles ? "low prio_lat" : "low prio_clat",
 502                                         min, max, mean, dev, out);
 503         }
 504
 505         if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
 506                 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
 507                                         ts->slat_stat[ddir].samples,
 508                                         ts->percentile_list,
 509                                         ts->percentile_precision, "slat", out);
 510         if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
 511                 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
 512                                         ts->clat_stat[ddir].samples,
 513                                         ts->percentile_list,
 514                                         ts->percentile_precision, "clat", out);
 515         if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
 516                 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
 517                                         ts->lat_stat[ddir].samples,
 518                                         ts->percentile_list,
 519                                         ts->percentile_precision, "lat", out);
 520
 521         if (ts->clat_percentiles || ts->lat_percentiles) {
 522                 const char *name = ts->lat_percentiles ? "lat" : "clat";
 523                 char prio_name[32];
 524                 uint64_t samples;
 525
 526                 if (ts->lat_percentiles)
 527                         samples = ts->lat_stat[ddir].samples;
 528                 else
 529                         samples = ts->clat_stat[ddir].samples;
 530
 531                 /* Only print this if some high and low priority stats were collected */
 532                 if (ts->clat_high_prio_stat[ddir].samples > 0 &&
 533                         ts->clat_low_prio_stat[ddir].samples > 0)
 534                 {
 535                         sprintf(prio_name, "high prio (%.2f%%) %s",
 536                                         100. * (double) ts->clat_high_prio_stat[ddir].samples / (double) samples,
 537                                         name);
 538                         show_clat_percentiles(ts->io_u_plat_high_prio[ddir],
 539                                                 ts->clat_high_prio_stat[ddir].samples,
 540                                                 ts->percentile_list,
 541                                                 ts->percentile_precision, prio_name, out);
 542
 543                         sprintf(prio_name, "low prio (%.2f%%) %s",
 544                                         100. * (double) ts->clat_low_prio_stat[ddir].samples / (double) samples,
 545                                         name);
 546                         show_clat_percentiles(ts->io_u_plat_low_prio[ddir],
 547                                                 ts->clat_low_prio_stat[ddir].samples,
 548                                                 ts->percentile_list,
 549                                                 ts->percentile_precision, prio_name, out);
 550                 }
 551         }
 552
 553         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 554                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 555                 const char *bw_str;
 556
 557                 if ((rs->unit_base == 1) && i2p)
 558                         bw_str = "Kibit";
 559                 else if (rs->unit_base == 1)
 560                         bw_str = "kbit";
 561                 else if (i2p)
 562                         bw_str = "KiB";
 563                 else
 564                         bw_str = "kB";
 565
 566                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
 567
 568                 if (rs->unit_base == 1) {
 569                         min *= 8.0;
 570                         max *= 8.0;
 571                         mean *= 8.0;
 572                         dev *= 8.0;
 573                 }
 574
 575                 if (mean > fkb_base * fkb_base) {
 576                         min /= fkb_base;
 577                         max /= fkb_base;
 578                         mean /= fkb_base;
 579                         dev /= fkb_base;
 580                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 581                 }
 582
 583                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 584                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 585                         bw_str, min, max, p_of_agg, mean, dev,
 586                         (&ts->bw_stat[ddir])->samples);
 587         }
 588         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 589                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 590                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 591                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 592         }
 593 }
 594
 595 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
 596                      const char *msg, struct buf_output *out)
 597 {
 598         bool new_line = true, shown = false;
 599         int i, line = 0;
 600
 601         for (i = 0; i < nr; i++) {
 602                 if (io_u_lat[i] <= 0.0)
 603                         continue;
 604                 shown = true;
 605                 if (new_line) {
 606                         if (line)
 607                                 log_buf(out, "\n");
 608                         log_buf(out, "  lat (%s)   : ", msg);
 609                         new_line = false;
 610                         line = 0;
 611                 }
 612                 if (line)
 613                         log_buf(out, ", ");
 614                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 615                 line++;
 616                 if (line == 5)
 617                         new_line = true;
 618         }
 619
 620         if (shown)
 621                 log_buf(out, "\n");
 622
 623         return true;
 624 }
 625
 626 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 627 {
 628         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 629                                  "250=", "500=", "750=", "1000=", };
 630
 631         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 632 }
 633
 634 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 635 {
 636         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 637                                  "250=", "500=", "750=", "1000=", };
 638
 639         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 640 }
 641
 642 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 643 {
 644         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 645                                  "250=", "500=", "750=", "1000=", "2000=",
 646                                  ">=2000=", };
 647
 648         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 649 }
 650
 651 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 652 {
 653         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 654         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 655         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 656
 657         stat_calc_lat_n(ts, io_u_lat_n);
 658         stat_calc_lat_u(ts, io_u_lat_u);
 659         stat_calc_lat_m(ts, io_u_lat_m);
 660
 661         show_lat_n(io_u_lat_n, out);
 662         show_lat_u(io_u_lat_u, out);
 663         show_lat_m(io_u_lat_m, out);
 664 }
 665
 666 static int block_state_category(int block_state)
 667 {
 668         switch (block_state) {
 669         case BLOCK_STATE_UNINIT:
 670                 return 0;
 671         case BLOCK_STATE_TRIMMED:
 672         case BLOCK_STATE_WRITTEN:
 673                 return 1;
 674         case BLOCK_STATE_WRITE_FAILURE:
 675         case BLOCK_STATE_TRIM_FAILURE:
 676                 return 2;
 677         default:
 678                 /* Silence compile warning on some BSDs and have a return */
 679                 assert(0);
 680                 return -1;
 681         }
 682 }
 683
 684 static int compare_block_infos(const void *bs1, const void *bs2)
 685 {
 686         uint64_t block1 = *(uint64_t *)bs1;
 687         uint64_t block2 = *(uint64_t *)bs2;
 688         int state1 = BLOCK_INFO_STATE(block1);
 689         int state2 = BLOCK_INFO_STATE(block2);
 690         int bscat1 = block_state_category(state1);
 691         int bscat2 = block_state_category(state2);
 692         int cycles1 = BLOCK_INFO_TRIMS(block1);
 693         int cycles2 = BLOCK_INFO_TRIMS(block2);
 694
 695         if (bscat1 < bscat2)
 696                 return -1;
 697         if (bscat1 > bscat2)
 698                 return 1;
 699
 700         if (cycles1 < cycles2)
 701                 return -1;
 702         if (cycles1 > cycles2)
 703                 return 1;
 704
 705         if (state1 < state2)
 706                 return -1;
 707         if (state1 > state2)
 708                 return 1;
 709
 710         assert(block1 == block2);
 711         return 0;
 712 }
 713
 714 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 715                                   fio_fp64_t *plist, unsigned int **percentiles,
 716                                   unsigned int *types)
 717 {
 718         int len = 0;
 719         int i, nr_uninit;
 720
 721         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 722
 723         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 724                 len++;
 725
 726         if (!len)
 727                 return 0;
 728
 729         /*
 730          * Sort the percentile list. Note that it may already be sorted if
 731          * we are using the default values, but since it's a short list this
 732          * isn't a worry. Also note that this does not work for NaN values.
 733          */
 734         if (len > 1)
 735                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 736
 737         /* Start only after the uninit entries end */
 738         for (nr_uninit = 0;
 739              nr_uninit < nr_block_infos
 740                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 741              nr_uninit ++)
 742                 ;
 743
 744         if (nr_uninit == nr_block_infos)
 745                 return 0;
 746
 747         *percentiles = calloc(len, sizeof(**percentiles));
 748
 749         for (i = 0; i < len; i++) {
 750                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 751                                 + nr_uninit;
 752                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 753         }
 754
 755         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 756         for (i = 0; i < nr_block_infos; i++)
 757                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 758
 759         return len;
 760 }
 761
 762 static const char *block_state_names[] = {
 763         [BLOCK_STATE_UNINIT] = "unwritten",
 764         [BLOCK_STATE_TRIMMED] = "trimmed",
 765         [BLOCK_STATE_WRITTEN] = "written",
 766         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 767         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 768 };
 769
 770 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 771                              fio_fp64_t *plist, struct buf_output *out)
 772 {
 773         int len, pos, i;
 774         unsigned int *percentiles = NULL;
 775         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 776
 777         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 778                                      &percentiles, block_state_counts);
 779
 780         log_buf(out, "  block lifetime percentiles :\n   |");
 781         pos = 0;
 782         for (i = 0; i < len; i++) {
 783                 uint32_t block_info = percentiles[i];
 784 #define LINE_LENGTH     75
 785                 char str[LINE_LENGTH];
 786                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 787                                      plist[i].u.f, block_info,
 788                                      i == len - 1 ? '\n' : ',');
 789                 assert(strln < LINE_LENGTH);
 790                 if (pos + strln > LINE_LENGTH) {
 791                         pos = 0;
 792                         log_buf(out, "\n   |");
 793                 }
 794                 log_buf(out, "%s", str);
 795                 pos += strln;
 796 #undef LINE_LENGTH
 797         }
 798         if (percentiles)
 799                 free(percentiles);
 800
 801         log_buf(out, "        states               :");
 802         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 803                 log_buf(out, " %s=%u%c",
 804                          block_state_names[i], block_state_counts[i],
 805                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 806 }
 807
 808 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 809 {
 810         char *p1, *p1alt, *p2;
 811         unsigned long long bw_mean, iops_mean;
 812         const int i2p = is_power_of_2(ts->kb_base);
 813
 814         if (!ts->ss_dur)
 815                 return;
 816
 817         bw_mean = steadystate_bw_mean(ts);
 818         iops_mean = steadystate_iops_mean(ts);
 819
 820         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
 821         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
 822         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
 823
 824         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 825                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
 826                 p1, p1alt, p2,
 827                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
 828                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
 829                 ts->ss_criterion.u.f,
 830                 ts->ss_state & FIO_SS_PCT ? "%" : "");
 831
 832         free(p1);
 833         free(p1alt);
 834         free(p2);
 835 }
 836
 837 static void show_agg_stats(struct disk_util_agg *agg, int terse,
 838                            struct buf_output *out)
 839 {
 840         if (!agg->slavecount)
 841                 return;
 842
 843         if (!terse) {
 844                 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
 845                          "aggrticks=%llu/%llu, aggrin_queue=%llu, "
 846                          "aggrutil=%3.2f%%",
 847                         (unsigned long long) agg->ios[0] / agg->slavecount,
 848                         (unsigned long long) agg->ios[1] / agg->slavecount,
 849                         (unsigned long long) agg->merges[0] / agg->slavecount,
 850                         (unsigned long long) agg->merges[1] / agg->slavecount,
 851                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 852                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 853                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 854                         agg->max_util.u.f);
 855         } else {
 856                 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
 857                         (unsigned long long) agg->ios[0] / agg->slavecount,
 858                         (unsigned long long) agg->ios[1] / agg->slavecount,
 859                         (unsigned long long) agg->merges[0] / agg->slavecount,
 860                         (unsigned long long) agg->merges[1] / agg->slavecount,
 861                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 862                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 863                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 864                         agg->max_util.u.f);
 865         }
 866 }
 867
 868 static void aggregate_slaves_stats(struct disk_util *masterdu)
 869 {
 870         struct disk_util_agg *agg = &masterdu->agg;
 871         struct disk_util_stat *dus;
 872         struct flist_head *entry;
 873         struct disk_util *slavedu;
 874         double util;
 875
 876         flist_for_each(entry, &masterdu->slaves) {
 877                 slavedu = flist_entry(entry, struct disk_util, slavelist);
 878                 dus = &slavedu->dus;
 879                 agg->ios[0] += dus->s.ios[0];
 880                 agg->ios[1] += dus->s.ios[1];
 881                 agg->merges[0] += dus->s.merges[0];
 882                 agg->merges[1] += dus->s.merges[1];
 883                 agg->sectors[0] += dus->s.sectors[0];
 884                 agg->sectors[1] += dus->s.sectors[1];
 885                 agg->ticks[0] += dus->s.ticks[0];
 886                 agg->ticks[1] += dus->s.ticks[1];
 887                 agg->time_in_queue += dus->s.time_in_queue;
 888                 agg->slavecount++;
 889
 890                 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
 891                 /* System utilization is the utilization of the
 892                  * component with the highest utilization.
 893                  */
 894                 if (util > agg->max_util.u.f)
 895                         agg->max_util.u.f = util;
 896
 897         }
 898
 899         if (agg->max_util.u.f > 100.0)
 900                 agg->max_util.u.f = 100.0;
 901 }
 902
 903 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
 904                      int terse, struct buf_output *out)
 905 {
 906         double util = 0;
 907
 908         if (dus->s.msec)
 909                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
 910         if (util > 100.0)
 911                 util = 100.0;
 912
 913         if (!terse) {
 914                 if (agg->slavecount)
 915                         log_buf(out, "  ");
 916
 917                 log_buf(out, "  %s: ios=%llu/%llu, merge=%llu/%llu, "
 918                          "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
 919                                 dus->name,
 920                                 (unsigned long long) dus->s.ios[0],
 921                                 (unsigned long long) dus->s.ios[1],
 922                                 (unsigned long long) dus->s.merges[0],
 923                                 (unsigned long long) dus->s.merges[1],
 924                                 (unsigned long long) dus->s.ticks[0],
 925                                 (unsigned long long) dus->s.ticks[1],
 926                                 (unsigned long long) dus->s.time_in_queue,
 927                                 util);
 928         } else {
 929                 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
 930                                 dus->name,
 931                                 (unsigned long long) dus->s.ios[0],
 932                                 (unsigned long long) dus->s.ios[1],
 933                                 (unsigned long long) dus->s.merges[0],
 934                                 (unsigned long long) dus->s.merges[1],
 935                                 (unsigned long long) dus->s.ticks[0],
 936                                 (unsigned long long) dus->s.ticks[1],
 937                                 (unsigned long long) dus->s.time_in_queue,
 938                                 util);
 939         }
 940
 941         /*
 942          * If the device has slaves, aggregate the stats for
 943          * those slave devices also.
 944          */
 945         show_agg_stats(agg, terse, out);
 946
 947         if (!terse)
 948                 log_buf(out, "\n");
 949 }
 950
 951 void json_array_add_disk_util(struct disk_util_stat *dus,
 952                 struct disk_util_agg *agg, struct json_array *array)
 953 {
 954         struct json_object *obj;
 955         double util = 0;
 956
 957         if (dus->s.msec)
 958                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
 959         if (util > 100.0)
 960                 util = 100.0;
 961
 962         obj = json_create_object();
 963         json_array_add_value_object(array, obj);
 964
 965         json_object_add_value_string(obj, "name", (const char *)dus->name);
 966         json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
 967         json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
 968         json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
 969         json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
 970         json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
 971         json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
 972         json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
 973         json_object_add_value_float(obj, "util", util);
 974
 975         /*
 976          * If the device has slaves, aggregate the stats for
 977          * those slave devices also.
 978          */
 979         if (!agg->slavecount)
 980                 return;
 981         json_object_add_value_int(obj, "aggr_read_ios",
 982                                 agg->ios[0] / agg->slavecount);
 983         json_object_add_value_int(obj, "aggr_write_ios",
 984                                 agg->ios[1] / agg->slavecount);
 985         json_object_add_value_int(obj, "aggr_read_merges",
 986                                 agg->merges[0] / agg->slavecount);
 987         json_object_add_value_int(obj, "aggr_write_merge",
 988                                 agg->merges[1] / agg->slavecount);
 989         json_object_add_value_int(obj, "aggr_read_ticks",
 990                                 agg->ticks[0] / agg->slavecount);
 991         json_object_add_value_int(obj, "aggr_write_ticks",
 992                                 agg->ticks[1] / agg->slavecount);
 993         json_object_add_value_int(obj, "aggr_in_queue",
 994                                 agg->time_in_queue / agg->slavecount);
 995         json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
 996 }
 997
 998 static void json_object_add_disk_utils(struct json_object *obj,
 999                                        struct flist_head *head)
1000 {
1001         struct json_array *array = json_create_array();
1002         struct flist_head *entry;
1003         struct disk_util *du;
1004
1005         json_object_add_value_array(obj, "disk_util", array);
1006
1007         flist_for_each(entry, head) {
1008                 du = flist_entry(entry, struct disk_util, list);
1009
1010                 aggregate_slaves_stats(du);
1011                 json_array_add_disk_util(&du->dus, &du->agg, array);
1012         }
1013 }
1014
1015 void show_disk_util(int terse, struct json_object *parent,
1016                     struct buf_output *out)
1017 {
1018         struct flist_head *entry;
1019         struct disk_util *du;
1020         bool do_json;
1021
1022         if (!is_running_backend())
1023                 return;
1024
1025         if (flist_empty(&disk_list)) {
1026                 return;
1027         }
1028
1029         if ((output_format & FIO_OUTPUT_JSON) && parent)
1030                 do_json = true;
1031         else
1032                 do_json = false;
1033
1034         if (!terse && !do_json)
1035                 log_buf(out, "\nDisk stats (read/write):\n");
1036
1037         if (do_json)
1038                 json_object_add_disk_utils(parent, &disk_list);
1039         else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1040                 flist_for_each(entry, &disk_list) {
1041                         du = flist_entry(entry, struct disk_util, list);
1042
1043                         aggregate_slaves_stats(du);
1044                         print_disk_util(&du->dus, &du->agg, terse, out);
1045                 }
1046         }
1047 }
1048
1049 static void show_thread_status_normal(struct thread_stat *ts,
1050                                       struct group_run_stats *rs,
1051                                       struct buf_output *out)
1052 {
1053         double usr_cpu, sys_cpu;
1054         unsigned long runtime;
1055         double io_u_dist[FIO_IO_U_MAP_NR];
1056         time_t time_p;
1057         char time_buf[32];
1058
1059         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1060                 return;
1061
1062         memset(time_buf, 0, sizeof(time_buf));
1063
1064         time(&time_p);
1065         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1066
1067         if (!ts->error) {
1068                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1069                                         ts->name, ts->groupid, ts->members,
1070                                         ts->error, (int) ts->pid, time_buf);
1071         } else {
1072                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1073                                         ts->name, ts->groupid, ts->members,
1074                                         ts->error, ts->verror, (int) ts->pid,
1075                                         time_buf);
1076         }
1077
1078         if (strlen(ts->description))
1079                 log_buf(out, "  Description  : [%s]\n", ts->description);
1080
1081         for_each_rw_ddir(ddir) {
1082                 if (ts->io_bytes[ddir])
1083                         show_ddir_status(rs, ts, ddir, out);
1084         }
1085
1086         show_latencies(ts, out);
1087
1088         if (ts->sync_stat.samples)
1089                 show_ddir_status(rs, ts, DDIR_SYNC, out);
1090
1091         runtime = ts->total_run_time;
1092         if (runtime) {
1093                 double runt = (double) runtime;
1094
1095                 usr_cpu = (double) ts->usr_time * 100 / runt;
1096                 sys_cpu = (double) ts->sys_time * 100 / runt;
1097         } else {
1098                 usr_cpu = 0;
1099                 sys_cpu = 0;
1100         }
1101
1102         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1103                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1104                         (unsigned long long) ts->ctx,
1105                         (unsigned long long) ts->majf,
1106                         (unsigned long long) ts->minf);
1107
1108         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1109         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1110                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1111                                         io_u_dist[1], io_u_dist[2],
1112                                         io_u_dist[3], io_u_dist[4],
1113                                         io_u_dist[5], io_u_dist[6]);
1114
1115         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1116         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1117                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1118                                         io_u_dist[1], io_u_dist[2],
1119                                         io_u_dist[3], io_u_dist[4],
1120                                         io_u_dist[5], io_u_dist[6]);
1121         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1122         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1123                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1124                                         io_u_dist[1], io_u_dist[2],
1125                                         io_u_dist[3], io_u_dist[4],
1126                                         io_u_dist[5], io_u_dist[6]);
1127         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
1128                                  " short=%llu,%llu,%llu,0"
1129                                  " dropped=%llu,%llu,%llu,0\n",
1130                                         (unsigned long long) ts->total_io_u[0],
1131                                         (unsigned long long) ts->total_io_u[1],
1132                                         (unsigned long long) ts->total_io_u[2],
1133                                         (unsigned long long) ts->total_io_u[3],
1134                                         (unsigned long long) ts->short_io_u[0],
1135                                         (unsigned long long) ts->short_io_u[1],
1136                                         (unsigned long long) ts->short_io_u[2],
1137                                         (unsigned long long) ts->drop_io_u[0],
1138                                         (unsigned long long) ts->drop_io_u[1],
1139                                         (unsigned long long) ts->drop_io_u[2]);
1140         if (ts->continue_on_error) {
1141                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
1142                                         (unsigned long long)ts->total_err_count,
1143                                         ts->first_error,
1144                                         strerror(ts->first_error));
1145         }
1146         if (ts->latency_depth) {
1147                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1148                                         (unsigned long long)ts->latency_target,
1149                                         (unsigned long long)ts->latency_window,
1150                                         ts->latency_percentile.u.f,
1151                                         ts->latency_depth);
1152         }
1153
1154         if (ts->nr_block_infos)
1155                 show_block_infos(ts->nr_block_infos, ts->block_infos,
1156                                   ts->percentile_list, out);
1157
1158         if (ts->ss_dur)
1159                 show_ss_normal(ts, out);
1160 }
1161
1162 static void show_ddir_status_terse(struct thread_stat *ts,
1163                                    struct group_run_stats *rs, int ddir,
1164                                    int ver, struct buf_output *out)
1165 {
1166         unsigned long long min, max, minv, maxv, bw, iops;
1167         unsigned long long *ovals = NULL;
1168         double mean, dev;
1169         unsigned int len;
1170         int i, bw_stat;
1171
1172         assert(ddir_rw(ddir));
1173
1174         iops = bw = 0;
1175         if (ts->runtime[ddir]) {
1176                 uint64_t runt = ts->runtime[ddir];
1177
1178                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1179                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1180         }
1181
1182         log_buf(out, ";%llu;%llu;%llu;%llu",
1183                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1184                                         (unsigned long long) ts->runtime[ddir]);
1185
1186         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1187                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1188         else
1189                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1190
1191         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1192                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1193         else
1194                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1195
1196         if (ts->lat_percentiles)
1197                 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1198                                         ts->lat_stat[ddir].samples,
1199                                         ts->percentile_list, &ovals, &maxv,
1200                                         &minv);
1201         else if (ts->clat_percentiles)
1202                 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1203                                         ts->clat_stat[ddir].samples,
1204                                         ts->percentile_list, &ovals, &maxv,
1205                                         &minv);
1206         else
1207                 len = 0;
1208
1209         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1210                 if (i >= len) {
1211                         log_buf(out, ";0%%=0");
1212                         continue;
1213                 }
1214                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1215         }
1216
1217         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1218                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1219         else
1220                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1221
1222         free(ovals);
1223
1224         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1225         if (bw_stat) {
1226                 double p_of_agg = 100.0;
1227
1228                 if (rs->agg[ddir]) {
1229                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1230                         if (p_of_agg > 100.0)
1231                                 p_of_agg = 100.0;
1232                 }
1233
1234                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1235         } else
1236                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1237
1238         if (ver == 5) {
1239                 if (bw_stat)
1240                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1241                 else
1242                         log_buf(out, ";%lu", 0UL);
1243
1244                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1245                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1246                                 mean, dev, (&ts->iops_stat[ddir])->samples);
1247                 else
1248                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1249         }
1250 }
1251
1252 static struct json_object *add_ddir_lat_json(struct thread_stat *ts, uint32_t percentiles,
1253                 struct io_stat *lat_stat, uint64_t *io_u_plat)
1254 {
1255         char buf[120];
1256         double mean, dev;
1257         unsigned int i, len;
1258         struct json_object *lat_object, *percentile_object, *clat_bins_object;
1259         unsigned long long min, max, maxv, minv, *ovals = NULL;
1260
1261         if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1262                 min = max = 0;
1263                 mean = dev = 0.0;
1264         }
1265         lat_object = json_create_object();
1266         json_object_add_value_int(lat_object, "min", min);
1267         json_object_add_value_int(lat_object, "max", max);
1268         json_object_add_value_float(lat_object, "mean", mean);
1269         json_object_add_value_float(lat_object, "stddev", dev);
1270         json_object_add_value_int(lat_object, "N", lat_stat->samples);
1271
1272         if (percentiles && lat_stat->samples) {
1273                 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1274                                 ts->percentile_list, &ovals, &maxv, &minv);
1275
1276                 if (len > FIO_IO_U_LIST_MAX_LEN)
1277                         len = FIO_IO_U_LIST_MAX_LEN;
1278
1279                 percentile_object = json_create_object();
1280                 json_object_add_value_object(lat_object, "percentile", percentile_object);
1281                 for (i = 0; i < len; i++) {
1282                         snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1283                         json_object_add_value_int(percentile_object, buf, ovals[i]);
1284                 }
1285                 free(ovals);
1286
1287                 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1288                         clat_bins_object = json_create_object();
1289                         json_object_add_value_object(lat_object, "bins", clat_bins_object);
1290
1291                         for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1292                                 if (io_u_plat[i]) {
1293                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1294                                         json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1295                                 }
1296                 }
1297         }
1298
1299         return lat_object;
1300 }
1301
1302 static void add_ddir_status_json(struct thread_stat *ts,
1303                 struct group_run_stats *rs, int ddir, struct json_object *parent)
1304 {
1305         unsigned long long min, max;
1306         unsigned long long bw_bytes, bw;
1307         double mean, dev, iops;
1308         struct json_object *dir_object, *tmp_object;
1309         double p_of_agg = 100.0;
1310
1311         assert(ddir_rw(ddir) || ddir_sync(ddir));
1312
1313         if (ts->unified_rw_rep && ddir != DDIR_READ)
1314                 return;
1315
1316         dir_object = json_create_object();
1317         json_object_add_value_object(parent,
1318                 ts->unified_rw_rep ? "mixed" : io_ddir_name(ddir), dir_object);
1319
1320         if (ddir_rw(ddir)) {
1321                 bw_bytes = 0;
1322                 bw = 0;
1323                 iops = 0.0;
1324                 if (ts->runtime[ddir]) {
1325                         uint64_t runt = ts->runtime[ddir];
1326
1327                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1328                         bw = bw_bytes / 1024; /* KiB/s */
1329                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1330                 }
1331
1332                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1333                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1334                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1335                 json_object_add_value_int(dir_object, "bw", bw);
1336                 json_object_add_value_float(dir_object, "iops", iops);
1337                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1338                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1339                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1340                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1341
1342                 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1343                                 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1344                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1345
1346                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1347                                 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1348                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1349
1350                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1351                                 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1352                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1353         } else {
1354                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1355                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1356                                 &ts->sync_stat, ts->io_u_sync_plat);
1357                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1358         }
1359
1360         if (!ddir_rw(ddir))
1361                 return;
1362
1363         /* Only print PRIO latencies if some high priority samples were gathered */
1364         if (ts->clat_high_prio_stat[ddir].samples > 0) {
1365                 const char *high, *low;
1366
1367                 if (ts->lat_percentiles) {
1368                         high = "lat_high_prio";
1369                         low = "lat_low_prio";
1370                 } else {
1371                         high = "clat_high_prio";
1372                         low = "clat_low_prio";
1373                 }
1374
1375                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1376                                 &ts->clat_high_prio_stat[ddir], ts->io_u_plat_high_prio[ddir]);
1377                 json_object_add_value_object(dir_object, high, tmp_object);
1378
1379                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1380                                 &ts->clat_low_prio_stat[ddir], ts->io_u_plat_low_prio[ddir]);
1381                 json_object_add_value_object(dir_object, low, tmp_object);
1382         }
1383
1384         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1385                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1386         } else {
1387                 min = max = 0;
1388                 p_of_agg = mean = dev = 0.0;
1389         }
1390
1391         json_object_add_value_int(dir_object, "bw_min", min);
1392         json_object_add_value_int(dir_object, "bw_max", max);
1393         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1394         json_object_add_value_float(dir_object, "bw_mean", mean);
1395         json_object_add_value_float(dir_object, "bw_dev", dev);
1396         json_object_add_value_int(dir_object, "bw_samples",
1397                                 (&ts->bw_stat[ddir])->samples);
1398
1399         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1400                 min = max = 0;
1401                 mean = dev = 0.0;
1402         }
1403         json_object_add_value_int(dir_object, "iops_min", min);
1404         json_object_add_value_int(dir_object, "iops_max", max);
1405         json_object_add_value_float(dir_object, "iops_mean", mean);
1406         json_object_add_value_float(dir_object, "iops_stddev", dev);
1407         json_object_add_value_int(dir_object, "iops_samples",
1408                                 (&ts->iops_stat[ddir])->samples);
1409
1410         if (ts->cachehit + ts->cachemiss) {
1411                 uint64_t total;
1412                 double hit;
1413
1414                 total = ts->cachehit + ts->cachemiss;
1415                 hit = (double) ts->cachehit / (double) total;
1416                 hit *= 100.0;
1417                 json_object_add_value_float(dir_object, "cachehit", hit);
1418         }
1419 }
1420
1421 static void show_thread_status_terse_all(struct thread_stat *ts,
1422                                          struct group_run_stats *rs, int ver,
1423                                          struct buf_output *out)
1424 {
1425         double io_u_dist[FIO_IO_U_MAP_NR];
1426         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1427         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1428         double usr_cpu, sys_cpu;
1429         int i;
1430
1431         /* General Info */
1432         if (ver == 2)
1433                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1434         else
1435                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1436                         ts->name, ts->groupid, ts->error);
1437
1438         /* Log Read Status */
1439         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1440         /* Log Write Status */
1441         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1442         /* Log Trim Status */
1443         if (ver == 2 || ver == 4 || ver == 5)
1444                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1445
1446         /* CPU Usage */
1447         if (ts->total_run_time) {
1448                 double runt = (double) ts->total_run_time;
1449
1450                 usr_cpu = (double) ts->usr_time * 100 / runt;
1451                 sys_cpu = (double) ts->sys_time * 100 / runt;
1452         } else {
1453                 usr_cpu = 0;
1454                 sys_cpu = 0;
1455         }
1456
1457         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1458                                                 (unsigned long long) ts->ctx,
1459                                                 (unsigned long long) ts->majf,
1460                                                 (unsigned long long) ts->minf);
1461
1462         /* Calc % distribution of IO depths, usecond, msecond latency */
1463         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1464         stat_calc_lat_nu(ts, io_u_lat_u);
1465         stat_calc_lat_m(ts, io_u_lat_m);
1466
1467         /* Only show fixed 7 I/O depth levels*/
1468         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1469                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1470                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1471
1472         /* Microsecond latency */
1473         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1474                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1475         /* Millisecond latency */
1476         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1477                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1478
1479         /* disk util stats, if any */
1480         if (ver >= 3 && is_running_backend())
1481                 show_disk_util(1, NULL, out);
1482
1483         /* Additional output if continue_on_error set - default off*/
1484         if (ts->continue_on_error)
1485                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1486
1487         /* Additional output if description is set */
1488         if (strlen(ts->description)) {
1489                 if (ver == 2)
1490                         log_buf(out, "\n");
1491                 log_buf(out, ";%s", ts->description);
1492         }
1493
1494         log_buf(out, "\n");
1495 }
1496
1497 static void json_add_job_opts(struct json_object *root, const char *name,
1498                               struct flist_head *opt_list)
1499 {
1500         struct json_object *dir_object;
1501         struct flist_head *entry;
1502         struct print_option *p;
1503
1504         if (flist_empty(opt_list))
1505                 return;
1506
1507         dir_object = json_create_object();
1508         json_object_add_value_object(root, name, dir_object);
1509
1510         flist_for_each(entry, opt_list) {
1511                 p = flist_entry(entry, struct print_option, list);
1512                 json_object_add_value_string(dir_object, p->name, p->value);
1513         }
1514 }
1515
1516 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1517                                                    struct group_run_stats *rs,
1518                                                    struct flist_head *opt_list)
1519 {
1520         struct json_object *root, *tmp;
1521         struct jobs_eta *je;
1522         double io_u_dist[FIO_IO_U_MAP_NR];
1523         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1524         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1525         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1526         double usr_cpu, sys_cpu;
1527         int i;
1528         size_t size;
1529
1530         root = json_create_object();
1531         json_object_add_value_string(root, "jobname", ts->name);
1532         json_object_add_value_int(root, "groupid", ts->groupid);
1533         json_object_add_value_int(root, "error", ts->error);
1534
1535         /* ETA Info */
1536         je = get_jobs_eta(true, &size);
1537         if (je) {
1538                 json_object_add_value_int(root, "eta", je->eta_sec);
1539                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1540         }
1541
1542         if (opt_list)
1543                 json_add_job_opts(root, "job options", opt_list);
1544
1545         add_ddir_status_json(ts, rs, DDIR_READ, root);
1546         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1547         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1548         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1549
1550         /* CPU Usage */
1551         if (ts->total_run_time) {
1552                 double runt = (double) ts->total_run_time;
1553
1554                 usr_cpu = (double) ts->usr_time * 100 / runt;
1555                 sys_cpu = (double) ts->sys_time * 100 / runt;
1556         } else {
1557                 usr_cpu = 0;
1558                 sys_cpu = 0;
1559         }
1560         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1561         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1562         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1563         json_object_add_value_int(root, "ctx", ts->ctx);
1564         json_object_add_value_int(root, "majf", ts->majf);
1565         json_object_add_value_int(root, "minf", ts->minf);
1566
1567         /* Calc % distribution of IO depths */
1568         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1569         tmp = json_create_object();
1570         json_object_add_value_object(root, "iodepth_level", tmp);
1571         /* Only show fixed 7 I/O depth levels*/
1572         for (i = 0; i < 7; i++) {
1573                 char name[20];
1574                 if (i < 6)
1575                         snprintf(name, 20, "%d", 1 << i);
1576                 else
1577                         snprintf(name, 20, ">=%d", 1 << i);
1578                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1579         }
1580
1581         /* Calc % distribution of submit IO depths */
1582         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1583         tmp = json_create_object();
1584         json_object_add_value_object(root, "iodepth_submit", tmp);
1585         /* Only show fixed 7 I/O depth levels*/
1586         for (i = 0; i < 7; i++) {
1587                 char name[20];
1588                 if (i == 0)
1589                         snprintf(name, 20, "0");
1590                 else if (i < 6)
1591                         snprintf(name, 20, "%d", 1 << (i+1));
1592                 else
1593                         snprintf(name, 20, ">=%d", 1 << i);
1594                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1595         }
1596
1597         /* Calc % distribution of completion IO depths */
1598         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1599         tmp = json_create_object();
1600         json_object_add_value_object(root, "iodepth_complete", tmp);
1601         /* Only show fixed 7 I/O depth levels*/
1602         for (i = 0; i < 7; i++) {
1603                 char name[20];
1604                 if (i == 0)
1605                         snprintf(name, 20, "0");
1606                 else if (i < 6)
1607                         snprintf(name, 20, "%d", 1 << (i+1));
1608                 else
1609                         snprintf(name, 20, ">=%d", 1 << i);
1610                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1611         }
1612
1613         /* Calc % distribution of nsecond, usecond, msecond latency */
1614         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1615         stat_calc_lat_n(ts, io_u_lat_n);
1616         stat_calc_lat_u(ts, io_u_lat_u);
1617         stat_calc_lat_m(ts, io_u_lat_m);
1618
1619         /* Nanosecond latency */
1620         tmp = json_create_object();
1621         json_object_add_value_object(root, "latency_ns", tmp);
1622         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1623                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1624                                  "250", "500", "750", "1000", };
1625                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1626         }
1627         /* Microsecond latency */
1628         tmp = json_create_object();
1629         json_object_add_value_object(root, "latency_us", tmp);
1630         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1631                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1632                                  "250", "500", "750", "1000", };
1633                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1634         }
1635         /* Millisecond latency */
1636         tmp = json_create_object();
1637         json_object_add_value_object(root, "latency_ms", tmp);
1638         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1639                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1640                                  "250", "500", "750", "1000", "2000",
1641                                  ">=2000", };
1642                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1643         }
1644
1645         /* Additional output if continue_on_error set - default off*/
1646         if (ts->continue_on_error) {
1647                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1648                 json_object_add_value_int(root, "first_error", ts->first_error);
1649         }
1650
1651         if (ts->latency_depth) {
1652                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1653                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1654                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1655                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1656         }
1657
1658         /* Additional output if description is set */
1659         if (strlen(ts->description))
1660                 json_object_add_value_string(root, "desc", ts->description);
1661
1662         if (ts->nr_block_infos) {
1663                 /* Block error histogram and types */
1664                 int len;
1665                 unsigned int *percentiles = NULL;
1666                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1667
1668                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1669                                              ts->percentile_list,
1670                                              &percentiles, block_state_counts);
1671
1672                 if (len) {
1673                         struct json_object *block, *percentile_object, *states;
1674                         int state;
1675                         block = json_create_object();
1676                         json_object_add_value_object(root, "block", block);
1677
1678                         percentile_object = json_create_object();
1679                         json_object_add_value_object(block, "percentiles",
1680                                                      percentile_object);
1681                         for (i = 0; i < len; i++) {
1682                                 char buf[20];
1683                                 snprintf(buf, sizeof(buf), "%f",
1684                                          ts->percentile_list[i].u.f);
1685                                 json_object_add_value_int(percentile_object,
1686                                                           buf,
1687                                                           percentiles[i]);
1688                         }
1689
1690                         states = json_create_object();
1691                         json_object_add_value_object(block, "states", states);
1692                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1693                                 json_object_add_value_int(states,
1694                                         block_state_names[state],
1695                                         block_state_counts[state]);
1696                         }
1697                         free(percentiles);
1698                 }
1699         }
1700
1701         if (ts->ss_dur) {
1702                 struct json_object *data;
1703                 struct json_array *iops, *bw;
1704                 int j, k, l;
1705                 char ss_buf[64];
1706
1707                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1708                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1709                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1710                         (float) ts->ss_limit.u.f,
1711                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1712
1713                 tmp = json_create_object();
1714                 json_object_add_value_object(root, "steadystate", tmp);
1715                 json_object_add_value_string(tmp, "ss", ss_buf);
1716                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1717                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1718
1719                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1720                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1721                 json_object_add_value_string(tmp, "criterion", ss_buf);
1722                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1723                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1724
1725                 data = json_create_object();
1726                 json_object_add_value_object(tmp, "data", data);
1727                 bw = json_create_array();
1728                 iops = json_create_array();
1729
1730                 /*
1731                 ** if ss was attained or the buffer is not full,
1732                 ** ss->head points to the first element in the list.
1733                 ** otherwise it actually points to the second element
1734                 ** in the list
1735                 */
1736                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1737                         j = ts->ss_head;
1738                 else
1739                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1740                 for (l = 0; l < ts->ss_dur; l++) {
1741                         k = (j + l) % ts->ss_dur;
1742                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1743                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1744                 }
1745                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1746                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1747                 json_object_add_value_array(data, "iops", iops);
1748                 json_object_add_value_array(data, "bw", bw);
1749         }
1750
1751         return root;
1752 }
1753
1754 static void show_thread_status_terse(struct thread_stat *ts,
1755                                      struct group_run_stats *rs,
1756                                      struct buf_output *out)
1757 {
1758         if (terse_version >= 2 && terse_version <= 5)
1759                 show_thread_status_terse_all(ts, rs, terse_version, out);
1760         else
1761                 log_err("fio: bad terse version!? %d\n", terse_version);
1762 }
1763
1764 struct json_object *show_thread_status(struct thread_stat *ts,
1765                                        struct group_run_stats *rs,
1766                                        struct flist_head *opt_list,
1767                                        struct buf_output *out)
1768 {
1769         struct json_object *ret = NULL;
1770
1771         if (output_format & FIO_OUTPUT_TERSE)
1772                 show_thread_status_terse(ts, rs,  out);
1773         if (output_format & FIO_OUTPUT_JSON)
1774                 ret = show_thread_status_json(ts, rs, opt_list);
1775         if (output_format & FIO_OUTPUT_NORMAL)
1776                 show_thread_status_normal(ts, rs,  out);
1777
1778         return ret;
1779 }
1780
1781 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1782 {
1783         double mean, S;
1784
1785         dst->min_val = min(dst->min_val, src->min_val);
1786         dst->max_val = max(dst->max_val, src->max_val);
1787
1788         /*
1789          * Compute new mean and S after the merge
1790          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1791          *  #Parallel_algorithm>
1792          */
1793         if (first) {
1794                 mean = src->mean.u.f;
1795                 S = src->S.u.f;
1796         } else {
1797                 double delta = src->mean.u.f - dst->mean.u.f;
1798
1799                 mean = ((src->mean.u.f * src->samples) +
1800                         (dst->mean.u.f * dst->samples)) /
1801                         (dst->samples + src->samples);
1802
1803                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1804                         (dst->samples * src->samples) /
1805                         (dst->samples + src->samples);
1806         }
1807
1808         dst->samples += src->samples;
1809         dst->mean.u.f = mean;
1810         dst->S.u.f = S;
1811
1812 }
1813
1814 /*
1815  * We sum two kinds of stats - one that is time based, in which case we
1816  * apply the proper summing technique, and then one that is iops/bw
1817  * numbers. For group_reporting, we should just add those up, not make
1818  * them the mean of everything.
1819  */
1820 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first,
1821                      bool pure_sum)
1822 {
1823         if (src->samples == 0)
1824                 return;
1825
1826         if (!pure_sum) {
1827                 __sum_stat(dst, src, first);
1828                 return;
1829         }
1830
1831         if (first) {
1832                 dst->min_val = src->min_val;
1833                 dst->max_val = src->max_val;
1834                 dst->samples = src->samples;
1835                 dst->mean.u.f = src->mean.u.f;
1836                 dst->S.u.f = src->S.u.f;
1837         } else {
1838                 dst->min_val += src->min_val;
1839                 dst->max_val += src->max_val;
1840                 dst->samples += src->samples;
1841                 dst->mean.u.f += src->mean.u.f;
1842                 dst->S.u.f += src->S.u.f;
1843         }
1844 }
1845
1846 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1847 {
1848         int i;
1849
1850         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1851                 if (dst->max_run[i] < src->max_run[i])
1852                         dst->max_run[i] = src->max_run[i];
1853                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1854                         dst->min_run[i] = src->min_run[i];
1855                 if (dst->max_bw[i] < src->max_bw[i])
1856                         dst->max_bw[i] = src->max_bw[i];
1857                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1858                         dst->min_bw[i] = src->min_bw[i];
1859
1860                 dst->iobytes[i] += src->iobytes[i];
1861                 dst->agg[i] += src->agg[i];
1862         }
1863
1864         if (!dst->kb_base)
1865                 dst->kb_base = src->kb_base;
1866         if (!dst->unit_base)
1867                 dst->unit_base = src->unit_base;
1868         if (!dst->sig_figs)
1869                 dst->sig_figs = src->sig_figs;
1870 }
1871
1872 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1873                       bool first)
1874 {
1875         int k, l, m;
1876
1877         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1878                 if (!dst->unified_rw_rep) {
1879                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first, false);
1880                         sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], first, false);
1881                         sum_stat(&dst->clat_low_prio_stat[l], &src->clat_low_prio_stat[l], first, false);
1882                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first, false);
1883                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first, false);
1884                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first, true);
1885                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first, true);
1886
1887                         dst->io_bytes[l] += src->io_bytes[l];
1888
1889                         if (dst->runtime[l] < src->runtime[l])
1890                                 dst->runtime[l] = src->runtime[l];
1891                 } else {
1892                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first, false);
1893                         sum_stat(&dst->clat_high_prio_stat[0], &src->clat_high_prio_stat[l], first, false);
1894                         sum_stat(&dst->clat_low_prio_stat[0], &src->clat_low_prio_stat[l], first, false);
1895                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first, false);
1896                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first, false);
1897                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first, true);
1898                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first, true);
1899
1900                         dst->io_bytes[0] += src->io_bytes[l];
1901
1902                         if (dst->runtime[0] < src->runtime[l])
1903                                 dst->runtime[0] = src->runtime[l];
1904
1905                         /*
1906                          * We're summing to the same destination, so override
1907                          * 'first' after the first iteration of the loop
1908                          */
1909                         first = false;
1910                 }
1911         }
1912
1913         sum_stat(&dst->sync_stat, &src->sync_stat, first, false);
1914         dst->usr_time += src->usr_time;
1915         dst->sys_time += src->sys_time;
1916         dst->ctx += src->ctx;
1917         dst->majf += src->majf;
1918         dst->minf += src->minf;
1919
1920         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
1921                 dst->io_u_map[k] += src->io_u_map[k];
1922                 dst->io_u_submit[k] += src->io_u_submit[k];
1923                 dst->io_u_complete[k] += src->io_u_complete[k];
1924         }
1925
1926         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1927                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1928         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1929                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1930         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1931                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1932
1933         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1934                 if (!dst->unified_rw_rep) {
1935                         dst->total_io_u[k] += src->total_io_u[k];
1936                         dst->short_io_u[k] += src->short_io_u[k];
1937                         dst->drop_io_u[k] += src->drop_io_u[k];
1938                 } else {
1939                         dst->total_io_u[0] += src->total_io_u[k];
1940                         dst->short_io_u[0] += src->short_io_u[k];
1941                         dst->drop_io_u[0] += src->drop_io_u[k];
1942                 }
1943         }
1944
1945         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
1946
1947         for (k = 0; k < FIO_LAT_CNT; k++)
1948                 for (l = 0; l < DDIR_RWDIR_CNT; l++)
1949                         for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
1950                                 if (!dst->unified_rw_rep)
1951                                         dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
1952                                 else
1953                                         dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
1954
1955         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1956                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
1957
1958         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1959                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1960                         if (!dst->unified_rw_rep) {
1961                                 dst->io_u_plat_high_prio[k][m] += src->io_u_plat_high_prio[k][m];
1962                                 dst->io_u_plat_low_prio[k][m] += src->io_u_plat_low_prio[k][m];
1963                         } else {
1964                                 dst->io_u_plat_high_prio[0][m] += src->io_u_plat_high_prio[k][m];
1965                                 dst->io_u_plat_low_prio[0][m] += src->io_u_plat_low_prio[k][m];
1966                         }
1967
1968                 }
1969         }
1970
1971         dst->total_run_time += src->total_run_time;
1972         dst->total_submit += src->total_submit;
1973         dst->total_complete += src->total_complete;
1974         dst->nr_zone_resets += src->nr_zone_resets;
1975         dst->cachehit += src->cachehit;
1976         dst->cachemiss += src->cachemiss;
1977 }
1978
1979 void init_group_run_stat(struct group_run_stats *gs)
1980 {
1981         int i;
1982         memset(gs, 0, sizeof(*gs));
1983
1984         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1985                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1986 }
1987
1988 void init_thread_stat(struct thread_stat *ts)
1989 {
1990         int j;
1991
1992         memset(ts, 0, sizeof(*ts));
1993
1994         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1995                 ts->lat_stat[j].min_val = -1UL;
1996                 ts->clat_stat[j].min_val = -1UL;
1997                 ts->slat_stat[j].min_val = -1UL;
1998                 ts->bw_stat[j].min_val = -1UL;
1999                 ts->iops_stat[j].min_val = -1UL;
2000                 ts->clat_high_prio_stat[j].min_val = -1UL;
2001                 ts->clat_low_prio_stat[j].min_val = -1UL;
2002         }
2003         ts->sync_stat.min_val = -1UL;
2004         ts->groupid = -1;
2005 }
2006
2007 void __show_run_stats(void)
2008 {
2009         struct group_run_stats *runstats, *rs;
2010         struct thread_data *td;
2011         struct thread_stat *threadstats, *ts;
2012         int i, j, k, nr_ts, last_ts, idx;
2013         bool kb_base_warned = false;
2014         bool unit_base_warned = false;
2015         struct json_object *root = NULL;
2016         struct json_array *array = NULL;
2017         struct buf_output output[FIO_OUTPUT_NR];
2018         struct flist_head **opt_lists;
2019
2020         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2021
2022         for (i = 0; i < groupid + 1; i++)
2023                 init_group_run_stat(&runstats[i]);
2024
2025         /*
2026          * find out how many threads stats we need. if group reporting isn't
2027          * enabled, it's one-per-td.
2028          */
2029         nr_ts = 0;
2030         last_ts = -1;
2031         for_each_td(td, i) {
2032                 if (!td->o.group_reporting) {
2033                         nr_ts++;
2034                         continue;
2035                 }
2036                 if (last_ts == td->groupid)
2037                         continue;
2038                 if (!td->o.stats)
2039                         continue;
2040
2041                 last_ts = td->groupid;
2042                 nr_ts++;
2043         }
2044
2045         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2046         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2047
2048         for (i = 0; i < nr_ts; i++) {
2049                 init_thread_stat(&threadstats[i]);
2050                 opt_lists[i] = NULL;
2051         }
2052
2053         j = 0;
2054         last_ts = -1;
2055         idx = 0;
2056         for_each_td(td, i) {
2057                 if (!td->o.stats)
2058                         continue;
2059                 if (idx && (!td->o.group_reporting ||
2060                     (td->o.group_reporting && last_ts != td->groupid))) {
2061                         idx = 0;
2062                         j++;
2063                 }
2064
2065                 last_ts = td->groupid;
2066
2067                 ts = &threadstats[j];
2068
2069                 ts->clat_percentiles = td->o.clat_percentiles;
2070                 ts->lat_percentiles = td->o.lat_percentiles;
2071                 ts->slat_percentiles = td->o.slat_percentiles;
2072                 ts->percentile_precision = td->o.percentile_precision;
2073                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2074                 opt_lists[j] = &td->opt_list;
2075
2076                 idx++;
2077                 ts->members++;
2078
2079                 if (ts->groupid == -1) {
2080                         /*
2081                          * These are per-group shared already
2082                          */
2083                         snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2084                         if (td->o.description)
2085                                 snprintf(ts->description,
2086                                          sizeof(ts->description), "%s",
2087                                          td->o.description);
2088                         else
2089                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2090
2091                         /*
2092                          * If multiple entries in this group, this is
2093                          * the first member.
2094                          */
2095                         ts->thread_number = td->thread_number;
2096                         ts->groupid = td->groupid;
2097
2098                         /*
2099                          * first pid in group, not very useful...
2100                          */
2101                         ts->pid = td->pid;
2102
2103                         ts->kb_base = td->o.kb_base;
2104                         ts->unit_base = td->o.unit_base;
2105                         ts->sig_figs = td->o.sig_figs;
2106                         ts->unified_rw_rep = td->o.unified_rw_rep;
2107                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2108                         log_info("fio: kb_base differs for jobs in group, using"
2109                                  " %u as the base\n", ts->kb_base);
2110                         kb_base_warned = true;
2111                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2112                         log_info("fio: unit_base differs for jobs in group, using"
2113                                  " %u as the base\n", ts->unit_base);
2114                         unit_base_warned = true;
2115                 }
2116
2117                 ts->continue_on_error = td->o.continue_on_error;
2118                 ts->total_err_count += td->total_err_count;
2119                 ts->first_error = td->first_error;
2120                 if (!ts->error) {
2121                         if (!td->error && td->o.continue_on_error &&
2122                             td->first_error) {
2123                                 ts->error = td->first_error;
2124                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2125                                          td->verror);
2126                         } else  if (td->error) {
2127                                 ts->error = td->error;
2128                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2129                                          td->verror);
2130                         }
2131                 }
2132
2133                 ts->latency_depth = td->latency_qd;
2134                 ts->latency_target = td->o.latency_target;
2135                 ts->latency_percentile = td->o.latency_percentile;
2136                 ts->latency_window = td->o.latency_window;
2137
2138                 ts->nr_block_infos = td->ts.nr_block_infos;
2139                 for (k = 0; k < ts->nr_block_infos; k++)
2140                         ts->block_infos[k] = td->ts.block_infos[k];
2141
2142                 sum_thread_stats(ts, &td->ts, idx == 1);
2143
2144                 if (td->o.ss_dur) {
2145                         ts->ss_state = td->ss.state;
2146                         ts->ss_dur = td->ss.dur;
2147                         ts->ss_head = td->ss.head;
2148                         ts->ss_bw_data = td->ss.bw_data;
2149                         ts->ss_iops_data = td->ss.iops_data;
2150                         ts->ss_limit.u.f = td->ss.limit;
2151                         ts->ss_slope.u.f = td->ss.slope;
2152                         ts->ss_deviation.u.f = td->ss.deviation;
2153                         ts->ss_criterion.u.f = td->ss.criterion;
2154                 }
2155                 else
2156                         ts->ss_dur = ts->ss_state = 0;
2157         }
2158
2159         for (i = 0; i < nr_ts; i++) {
2160                 unsigned long long bw;
2161
2162                 ts = &threadstats[i];
2163                 if (ts->groupid == -1)
2164                         continue;
2165                 rs = &runstats[ts->groupid];
2166                 rs->kb_base = ts->kb_base;
2167                 rs->unit_base = ts->unit_base;
2168                 rs->sig_figs = ts->sig_figs;
2169                 rs->unified_rw_rep += ts->unified_rw_rep;
2170
2171                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2172                         if (!ts->runtime[j])
2173                                 continue;
2174                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2175                                 rs->min_run[j] = ts->runtime[j];
2176                         if (ts->runtime[j] > rs->max_run[j])
2177                                 rs->max_run[j] = ts->runtime[j];
2178
2179                         bw = 0;
2180                         if (ts->runtime[j])
2181                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2182                         if (bw < rs->min_bw[j])
2183                                 rs->min_bw[j] = bw;
2184                         if (bw > rs->max_bw[j])
2185                                 rs->max_bw[j] = bw;
2186
2187                         rs->iobytes[j] += ts->io_bytes[j];
2188                 }
2189         }
2190
2191         for (i = 0; i < groupid + 1; i++) {
2192                 int ddir;
2193
2194                 rs = &runstats[i];
2195
2196                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2197                         if (rs->max_run[ddir])
2198                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2199                                                 rs->max_run[ddir];
2200                 }
2201         }
2202
2203         for (i = 0; i < FIO_OUTPUT_NR; i++)
2204                 buf_output_init(&output[i]);
2205
2206         /*
2207          * don't overwrite last signal output
2208          */
2209         if (output_format & FIO_OUTPUT_NORMAL)
2210                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2211         if (output_format & FIO_OUTPUT_JSON) {
2212                 struct thread_data *global;
2213                 char time_buf[32];
2214                 struct timeval now;
2215                 unsigned long long ms_since_epoch;
2216                 time_t tv_sec;
2217
2218                 gettimeofday(&now, NULL);
2219                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2220                                  (unsigned long long)(now.tv_usec) / 1000;
2221
2222                 tv_sec = now.tv_sec;
2223                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2224                 if (time_buf[strlen(time_buf) - 1] == '\n')
2225                         time_buf[strlen(time_buf) - 1] = '\0';
2226
2227                 root = json_create_object();
2228                 json_object_add_value_string(root, "fio version", fio_version_string);
2229                 json_object_add_value_int(root, "timestamp", now.tv_sec);
2230                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2231                 json_object_add_value_string(root, "time", time_buf);
2232                 global = get_global_options();
2233                 json_add_job_opts(root, "global options", &global->opt_list);
2234                 array = json_create_array();
2235                 json_object_add_value_array(root, "jobs", array);
2236         }
2237
2238         if (is_backend)
2239                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2240
2241         for (i = 0; i < nr_ts; i++) {
2242                 ts = &threadstats[i];
2243                 rs = &runstats[ts->groupid];
2244
2245                 if (is_backend) {
2246                         fio_server_send_job_options(opt_lists[i], i);
2247                         fio_server_send_ts(ts, rs);
2248                 } else {
2249                         if (output_format & FIO_OUTPUT_TERSE)
2250                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2251                         if (output_format & FIO_OUTPUT_JSON) {
2252                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2253                                 json_array_add_value_object(array, tmp);
2254                         }
2255                         if (output_format & FIO_OUTPUT_NORMAL)
2256                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2257                 }
2258         }
2259         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2260                 /* disk util stats, if any */
2261                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2262
2263                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2264
2265                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2266                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2267                 json_free_object(root);
2268         }
2269
2270         for (i = 0; i < groupid + 1; i++) {
2271                 rs = &runstats[i];
2272
2273                 rs->groupid = i;
2274                 if (is_backend)
2275                         fio_server_send_gs(rs);
2276                 else if (output_format & FIO_OUTPUT_NORMAL)
2277                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2278         }
2279
2280         if (is_backend)
2281                 fio_server_send_du();
2282         else if (output_format & FIO_OUTPUT_NORMAL) {
2283                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2284                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2285         }
2286
2287         for (i = 0; i < FIO_OUTPUT_NR; i++) {
2288                 struct buf_output *out = &output[i];
2289
2290                 log_info_buf(out->buf, out->buflen);
2291                 buf_output_free(out);
2292         }
2293
2294         fio_idle_prof_cleanup();
2295
2296         log_info_flush();
2297         free(runstats);
2298         free(threadstats);
2299         free(opt_lists);
2300 }
2301
2302 int __show_running_run_stats(void)
2303 {
2304         struct thread_data *td;
2305         unsigned long long *rt;
2306         struct timespec ts;
2307         int i;
2308
2309         fio_sem_down(stat_sem);
2310
2311         rt = malloc(thread_number * sizeof(unsigned long long));
2312         fio_gettime(&ts, NULL);
2313
2314         for_each_td(td, i) {
2315                 td->update_rusage = 1;
2316                 for_each_rw_ddir(ddir) {
2317                         td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2318                 }
2319                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2320
2321                 rt[i] = mtime_since(&td->start, &ts);
2322                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2323                         td->ts.runtime[DDIR_READ] += rt[i];
2324                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2325                         td->ts.runtime[DDIR_WRITE] += rt[i];
2326                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2327                         td->ts.runtime[DDIR_TRIM] += rt[i];
2328         }
2329
2330         for_each_td(td, i) {
2331                 if (td->runstate >= TD_EXITED)
2332                         continue;
2333                 if (td->rusage_sem) {
2334                         td->update_rusage = 1;
2335                         fio_sem_down(td->rusage_sem);
2336                 }
2337                 td->update_rusage = 0;
2338         }
2339
2340         __show_run_stats();
2341
2342         for_each_td(td, i) {
2343                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2344                         td->ts.runtime[DDIR_READ] -= rt[i];
2345                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2346                         td->ts.runtime[DDIR_WRITE] -= rt[i];
2347                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2348                         td->ts.runtime[DDIR_TRIM] -= rt[i];
2349         }
2350
2351         free(rt);
2352         fio_sem_up(stat_sem);
2353
2354         return 0;
2355 }
2356
2357 static bool status_file_disabled;
2358
2359 #define FIO_STATUS_FILE         "fio-dump-status"
2360
2361 static int check_status_file(void)
2362 {
2363         struct stat sb;
2364         const char *temp_dir;
2365         char fio_status_file_path[PATH_MAX];
2366
2367         if (status_file_disabled)
2368                 return 0;
2369
2370         temp_dir = getenv("TMPDIR");
2371         if (temp_dir == NULL) {
2372                 temp_dir = getenv("TEMP");
2373                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2374                         temp_dir = NULL;
2375         }
2376         if (temp_dir == NULL)
2377                 temp_dir = "/tmp";
2378 #ifdef __COVERITY__
2379         __coverity_tainted_data_sanitize__(temp_dir);
2380 #endif
2381
2382         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2383
2384         if (stat(fio_status_file_path, &sb))
2385                 return 0;
2386
2387         if (unlink(fio_status_file_path) < 0) {
2388                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2389                                                         strerror(errno));
2390                 log_err("fio: disabling status file updates\n");
2391                 status_file_disabled = true;
2392         }
2393
2394         return 1;
2395 }
2396
2397 void check_for_running_stats(void)
2398 {
2399         if (check_status_file()) {
2400                 show_running_run_stats();
2401                 return;
2402         }
2403 }
2404
2405 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2406 {
2407         double val = data;
2408         double delta;
2409
2410         if (data > is->max_val)
2411                 is->max_val = data;
2412         if (data < is->min_val)
2413                 is->min_val = data;
2414
2415         delta = val - is->mean.u.f;
2416         if (delta) {
2417                 is->mean.u.f += delta / (is->samples + 1.0);
2418                 is->S.u.f += delta * (val - is->mean.u.f);
2419         }
2420
2421         is->samples++;
2422 }
2423
2424 /*
2425  * Return a struct io_logs, which is added to the tail of the log
2426  * list for 'iolog'.
2427  */
2428 static struct io_logs *get_new_log(struct io_log *iolog)
2429 {
2430         size_t new_size, new_samples;
2431         struct io_logs *cur_log;
2432
2433         /*
2434          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2435          * forever
2436          */
2437         if (!iolog->cur_log_max)
2438                 new_samples = DEF_LOG_ENTRIES;
2439         else {
2440                 new_samples = iolog->cur_log_max * 2;
2441                 if (new_samples > MAX_LOG_ENTRIES)
2442                         new_samples = MAX_LOG_ENTRIES;
2443         }
2444
2445         new_size = new_samples * log_entry_sz(iolog);
2446
2447         cur_log = smalloc(sizeof(*cur_log));
2448         if (cur_log) {
2449                 INIT_FLIST_HEAD(&cur_log->list);
2450                 cur_log->log = malloc(new_size);
2451                 if (cur_log->log) {
2452                         cur_log->nr_samples = 0;
2453                         cur_log->max_samples = new_samples;
2454                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2455                         iolog->cur_log_max = new_samples;
2456                         return cur_log;
2457                 }
2458                 sfree(cur_log);
2459         }
2460
2461         return NULL;
2462 }
2463
2464 /*
2465  * Add and return a new log chunk, or return current log if big enough
2466  */
2467 static struct io_logs *regrow_log(struct io_log *iolog)
2468 {
2469         struct io_logs *cur_log;
2470         int i;
2471
2472         if (!iolog || iolog->disabled)
2473                 goto disable;
2474
2475         cur_log = iolog_cur_log(iolog);
2476         if (!cur_log) {
2477                 cur_log = get_new_log(iolog);
2478                 if (!cur_log)
2479                         return NULL;
2480         }
2481
2482         if (cur_log->nr_samples < cur_log->max_samples)
2483                 return cur_log;
2484
2485         /*
2486          * No room for a new sample. If we're compressing on the fly, flush
2487          * out the current chunk
2488          */
2489         if (iolog->log_gz) {
2490                 if (iolog_cur_flush(iolog, cur_log)) {
2491                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2492                         return NULL;
2493                 }
2494         }
2495
2496         /*
2497          * Get a new log array, and add to our list
2498          */
2499         cur_log = get_new_log(iolog);
2500         if (!cur_log) {
2501                 log_err("fio: failed extending iolog! Will stop logging.\n");
2502                 return NULL;
2503         }
2504
2505         if (!iolog->pending || !iolog->pending->nr_samples)
2506                 return cur_log;
2507
2508         /*
2509          * Flush pending items to new log
2510          */
2511         for (i = 0; i < iolog->pending->nr_samples; i++) {
2512                 struct io_sample *src, *dst;
2513
2514                 src = get_sample(iolog, iolog->pending, i);
2515                 dst = get_sample(iolog, cur_log, i);
2516                 memcpy(dst, src, log_entry_sz(iolog));
2517         }
2518         cur_log->nr_samples = iolog->pending->nr_samples;
2519
2520         iolog->pending->nr_samples = 0;
2521         return cur_log;
2522 disable:
2523         if (iolog)
2524                 iolog->disabled = true;
2525         return NULL;
2526 }
2527
2528 void regrow_logs(struct thread_data *td)
2529 {
2530         regrow_log(td->slat_log);
2531         regrow_log(td->clat_log);
2532         regrow_log(td->clat_hist_log);
2533         regrow_log(td->lat_log);
2534         regrow_log(td->bw_log);
2535         regrow_log(td->iops_log);
2536         td->flags &= ~TD_F_REGROW_LOGS;
2537 }
2538
2539 void regrow_agg_logs(void)
2540 {
2541         enum fio_ddir ddir;
2542
2543         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2544                 regrow_log(agg_io_log[ddir]);
2545 }
2546
2547 static struct io_logs *get_cur_log(struct io_log *iolog)
2548 {
2549         struct io_logs *cur_log;
2550
2551         cur_log = iolog_cur_log(iolog);
2552         if (!cur_log) {
2553                 cur_log = get_new_log(iolog);
2554                 if (!cur_log)
2555                         return NULL;
2556         }
2557
2558         if (cur_log->nr_samples < cur_log->max_samples)
2559                 return cur_log;
2560
2561         /*
2562          * Out of space. If we're in IO offload mode, or we're not doing
2563          * per unit logging (hence logging happens outside of the IO thread
2564          * as well), add a new log chunk inline. If we're doing inline
2565          * submissions, flag 'td' as needing a log regrow and we'll take
2566          * care of it on the submission side.
2567          */
2568         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2569             !per_unit_log(iolog))
2570                 return regrow_log(iolog);
2571
2572         if (iolog->td)
2573                 iolog->td->flags |= TD_F_REGROW_LOGS;
2574         if (iolog->pending)
2575                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2576         return iolog->pending;
2577 }
2578
2579 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2580                              enum fio_ddir ddir, unsigned long long bs,
2581                              unsigned long t, uint64_t offset, uint8_t priority_bit)
2582 {
2583         struct io_logs *cur_log;
2584
2585         if (iolog->disabled)
2586                 return;
2587         if (flist_empty(&iolog->io_logs))
2588                 iolog->avg_last[ddir] = t;
2589
2590         cur_log = get_cur_log(iolog);
2591         if (cur_log) {
2592                 struct io_sample *s;
2593
2594                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2595
2596                 s->data = data;
2597                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2598                 io_sample_set_ddir(iolog, s, ddir);
2599                 s->bs = bs;
2600                 s->priority_bit = priority_bit;
2601
2602                 if (iolog->log_offset) {
2603                         struct io_sample_offset *so = (void *) s;
2604
2605                         so->offset = offset;
2606                 }
2607
2608                 cur_log->nr_samples++;
2609                 return;
2610         }
2611
2612         iolog->disabled = true;
2613 }
2614
2615 static inline void reset_io_stat(struct io_stat *ios)
2616 {
2617         ios->min_val = -1ULL;
2618         ios->max_val = ios->samples = 0;
2619         ios->mean.u.f = ios->S.u.f = 0;
2620 }
2621
2622 void reset_io_stats(struct thread_data *td)
2623 {
2624         struct thread_stat *ts = &td->ts;
2625         int i, j, k;
2626
2627         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2628                 reset_io_stat(&ts->clat_high_prio_stat[i]);
2629                 reset_io_stat(&ts->clat_low_prio_stat[i]);
2630                 reset_io_stat(&ts->clat_stat[i]);
2631                 reset_io_stat(&ts->slat_stat[i]);
2632                 reset_io_stat(&ts->lat_stat[i]);
2633                 reset_io_stat(&ts->bw_stat[i]);
2634                 reset_io_stat(&ts->iops_stat[i]);
2635
2636                 ts->io_bytes[i] = 0;
2637                 ts->runtime[i] = 0;
2638                 ts->total_io_u[i] = 0;
2639                 ts->short_io_u[i] = 0;
2640                 ts->drop_io_u[i] = 0;
2641
2642                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2643                         ts->io_u_plat_high_prio[i][j] = 0;
2644                         ts->io_u_plat_low_prio[i][j] = 0;
2645                         if (!i)
2646                                 ts->io_u_sync_plat[j] = 0;
2647                 }
2648         }
2649
2650         for (i = 0; i < FIO_LAT_CNT; i++)
2651                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
2652                         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2653                                 ts->io_u_plat[i][j][k] = 0;
2654
2655         ts->total_io_u[DDIR_SYNC] = 0;
2656
2657         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2658                 ts->io_u_map[i] = 0;
2659                 ts->io_u_submit[i] = 0;
2660                 ts->io_u_complete[i] = 0;
2661         }
2662
2663         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2664                 ts->io_u_lat_n[i] = 0;
2665         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2666                 ts->io_u_lat_u[i] = 0;
2667         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2668                 ts->io_u_lat_m[i] = 0;
2669
2670         ts->total_submit = 0;
2671         ts->total_complete = 0;
2672         ts->nr_zone_resets = 0;
2673         ts->cachehit = ts->cachemiss = 0;
2674 }
2675
2676 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2677                               unsigned long elapsed, bool log_max, uint8_t priority_bit)
2678 {
2679         /*
2680          * Note an entry in the log. Use the mean from the logged samples,
2681          * making sure to properly round up. Only write a log entry if we
2682          * had actual samples done.
2683          */
2684         if (iolog->avg_window[ddir].samples) {
2685                 union io_sample_data data;
2686
2687                 if (log_max)
2688                         data.val = iolog->avg_window[ddir].max_val;
2689                 else
2690                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2691
2692                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, priority_bit);
2693         }
2694
2695         reset_io_stat(&iolog->avg_window[ddir]);
2696 }
2697
2698 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2699                              bool log_max, uint8_t priority_bit)
2700 {
2701         int ddir;
2702
2703         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2704                 __add_stat_to_log(iolog, ddir, elapsed, log_max, priority_bit);
2705 }
2706
2707 static unsigned long add_log_sample(struct thread_data *td,
2708                                     struct io_log *iolog,
2709                                     union io_sample_data data,
2710                                     enum fio_ddir ddir, unsigned long long bs,
2711                                     uint64_t offset, uint8_t priority_bit)
2712 {
2713         unsigned long elapsed, this_window;
2714
2715         if (!ddir_rw(ddir))
2716                 return 0;
2717
2718         elapsed = mtime_since_now(&td->epoch);
2719
2720         /*
2721          * If no time averaging, just add the log sample.
2722          */
2723         if (!iolog->avg_msec) {
2724                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset, priority_bit);
2725                 return 0;
2726         }
2727
2728         /*
2729          * Add the sample. If the time period has passed, then
2730          * add that entry to the log and clear.
2731          */
2732         add_stat_sample(&iolog->avg_window[ddir], data.val);
2733
2734         /*
2735          * If period hasn't passed, adding the above sample is all we
2736          * need to do.
2737          */
2738         this_window = elapsed - iolog->avg_last[ddir];
2739         if (elapsed < iolog->avg_last[ddir])
2740                 return iolog->avg_last[ddir] - elapsed;
2741         else if (this_window < iolog->avg_msec) {
2742                 unsigned long diff = iolog->avg_msec - this_window;
2743
2744                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2745                         return diff;
2746         }
2747
2748         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0, priority_bit);
2749
2750         iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
2751
2752         return iolog->avg_msec;
2753 }
2754
2755 void finalize_logs(struct thread_data *td, bool unit_logs)
2756 {
2757         unsigned long elapsed;
2758
2759         elapsed = mtime_since_now(&td->epoch);
2760
2761         if (td->clat_log && unit_logs)
2762                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0, 0);
2763         if (td->slat_log && unit_logs)
2764                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0, 0);
2765         if (td->lat_log && unit_logs)
2766                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0, 0);
2767         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2768                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0, 0);
2769         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2770                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0, 0);
2771 }
2772
2773 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned long long bs,
2774                                         uint8_t priority_bit)
2775 {
2776         struct io_log *iolog;
2777
2778         if (!ddir_rw(ddir))
2779                 return;
2780
2781         iolog = agg_io_log[ddir];
2782         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, priority_bit);
2783 }
2784
2785 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2786 {
2787         unsigned int idx = plat_val_to_idx(nsec);
2788         assert(idx < FIO_IO_U_PLAT_NR);
2789
2790         ts->io_u_sync_plat[idx]++;
2791         add_stat_sample(&ts->sync_stat, nsec);
2792 }
2793
2794 static void add_lat_percentile_sample_noprio(struct thread_stat *ts,
2795                                 unsigned long long nsec, enum fio_ddir ddir, enum fio_lat lat)
2796 {
2797         unsigned int idx = plat_val_to_idx(nsec);
2798         assert(idx < FIO_IO_U_PLAT_NR);
2799
2800         ts->io_u_plat[lat][ddir][idx]++;
2801 }
2802
2803 static void add_lat_percentile_sample(struct thread_stat *ts,
2804                                 unsigned long long nsec, enum fio_ddir ddir, uint8_t priority_bit,
2805                                 enum fio_lat lat)
2806 {
2807         unsigned int idx = plat_val_to_idx(nsec);
2808
2809         add_lat_percentile_sample_noprio(ts, nsec, ddir, lat);
2810
2811         if (!priority_bit)
2812                 ts->io_u_plat_low_prio[ddir][idx]++;
2813         else
2814                 ts->io_u_plat_high_prio[ddir][idx]++;
2815 }
2816
2817 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2818                      unsigned long long nsec, unsigned long long bs,
2819                      uint64_t offset, uint8_t priority_bit)
2820 {
2821         const bool needs_lock = td_async_processing(td);
2822         unsigned long elapsed, this_window;
2823         struct thread_stat *ts = &td->ts;
2824         struct io_log *iolog = td->clat_hist_log;
2825
2826         if (needs_lock)
2827                 __td_io_u_lock(td);
2828
2829         add_stat_sample(&ts->clat_stat[ddir], nsec);
2830
2831         if (!ts->lat_percentiles) {
2832                 if (priority_bit)
2833                         add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
2834                 else
2835                         add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
2836         }
2837
2838         if (td->clat_log)
2839                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2840                                offset, priority_bit);
2841
2842         if (ts->clat_percentiles) {
2843                 if (ts->lat_percentiles)
2844                         add_lat_percentile_sample_noprio(ts, nsec, ddir, FIO_CLAT);
2845                 else
2846                         add_lat_percentile_sample(ts, nsec, ddir, priority_bit, FIO_CLAT);
2847         }
2848
2849         if (iolog && iolog->hist_msec) {
2850                 struct io_hist *hw = &iolog->hist_window[ddir];
2851
2852                 hw->samples++;
2853                 elapsed = mtime_since_now(&td->epoch);
2854                 if (!hw->hist_last)
2855                         hw->hist_last = elapsed;
2856                 this_window = elapsed - hw->hist_last;
2857
2858                 if (this_window >= iolog->hist_msec) {
2859                         uint64_t *io_u_plat;
2860                         struct io_u_plat_entry *dst;
2861
2862                         /*
2863                          * Make a byte-for-byte copy of the latency histogram
2864                          * stored in td->ts.io_u_plat[ddir], recording it in a
2865                          * log sample. Note that the matching call to free() is
2866                          * located in iolog.c after printing this sample to the
2867                          * log file.
2868                          */
2869                         io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
2870                         dst = malloc(sizeof(struct io_u_plat_entry));
2871                         memcpy(&(dst->io_u_plat), io_u_plat,
2872                                 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
2873                         flist_add(&dst->list, &hw->list);
2874                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2875                                                 elapsed, offset, priority_bit);
2876
2877                         /*
2878                          * Update the last time we recorded as being now, minus
2879                          * any drift in time we encountered before actually
2880                          * making the record.
2881                          */
2882                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2883                         hw->samples = 0;
2884                 }
2885         }
2886
2887         if (needs_lock)
2888                 __td_io_u_unlock(td);
2889 }
2890
2891 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2892                         unsigned long long nsec, unsigned long long bs, uint64_t offset,
2893                         uint8_t priority_bit)
2894 {
2895         const bool needs_lock = td_async_processing(td);
2896         struct thread_stat *ts = &td->ts;
2897
2898         if (!ddir_rw(ddir))
2899                 return;
2900
2901         if (needs_lock)
2902                 __td_io_u_lock(td);
2903
2904         add_stat_sample(&ts->slat_stat[ddir], nsec);
2905
2906         if (td->slat_log)
2907                 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs, offset,
2908                         priority_bit);
2909
2910         if (ts->slat_percentiles)
2911                 add_lat_percentile_sample_noprio(ts, nsec, ddir, FIO_SLAT);
2912
2913         if (needs_lock)
2914                 __td_io_u_unlock(td);
2915 }
2916
2917 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2918                     unsigned long long nsec, unsigned long long bs,
2919                     uint64_t offset, uint8_t priority_bit)
2920 {
2921         const bool needs_lock = td_async_processing(td);
2922         struct thread_stat *ts = &td->ts;
2923
2924         if (!ddir_rw(ddir))
2925                 return;
2926
2927         if (needs_lock)
2928                 __td_io_u_lock(td);
2929
2930         add_stat_sample(&ts->lat_stat[ddir], nsec);
2931
2932         if (td->lat_log)
2933                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2934                                offset, priority_bit);
2935
2936         if (ts->lat_percentiles) {
2937                 add_lat_percentile_sample(ts, nsec, ddir, priority_bit, FIO_LAT);
2938                 if (priority_bit)
2939                         add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
2940                 else
2941                         add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
2942
2943         }
2944         if (needs_lock)
2945                 __td_io_u_unlock(td);
2946 }
2947
2948 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2949                    unsigned int bytes, unsigned long long spent)
2950 {
2951         const bool needs_lock = td_async_processing(td);
2952         struct thread_stat *ts = &td->ts;
2953         unsigned long rate;
2954
2955         if (spent)
2956                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2957         else
2958                 rate = 0;
2959
2960         if (needs_lock)
2961                 __td_io_u_lock(td);
2962
2963         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2964
2965         if (td->bw_log)
2966                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2967                                bytes, io_u->offset, io_u_is_prio(io_u));
2968
2969         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2970
2971         if (needs_lock)
2972                 __td_io_u_unlock(td);
2973 }
2974
2975 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2976                          struct timespec *t, unsigned int avg_time,
2977                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2978                          struct io_stat *stat, struct io_log *log,
2979                          bool is_kb)
2980 {
2981         const bool needs_lock = td_async_processing(td);
2982         unsigned long spent, rate;
2983         enum fio_ddir ddir;
2984         unsigned long next, next_log;
2985
2986         next_log = avg_time;
2987
2988         spent = mtime_since(parent_tv, t);
2989         if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
2990                 return avg_time - spent;
2991
2992         if (needs_lock)
2993                 __td_io_u_lock(td);
2994
2995         /*
2996          * Compute both read and write rates for the interval.
2997          */
2998         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2999                 uint64_t delta;
3000
3001                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3002                 if (!delta)
3003                         continue; /* No entries for interval */
3004
3005                 if (spent) {
3006                         if (is_kb)
3007                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
3008                         else
3009                                 rate = (delta * 1000) / spent;
3010                 } else
3011                         rate = 0;
3012
3013                 add_stat_sample(&stat[ddir], rate);
3014
3015                 if (log) {
3016                         unsigned long long bs = 0;
3017
3018                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3019                                 bs = td->o.min_bs[ddir];
3020
3021                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0, 0);
3022                         next_log = min(next_log, next);
3023                 }
3024
3025                 stat_io_bytes[ddir] = this_io_bytes[ddir];
3026         }
3027
3028         *parent_tv = *t;
3029
3030         if (needs_lock)
3031                 __td_io_u_unlock(td);
3032
3033         if (spent <= avg_time)
3034                 next = avg_time;
3035         else
3036                 next = avg_time - (1 + spent - avg_time);
3037
3038         return min(next, next_log);
3039 }
3040
3041 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3042 {
3043         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3044                                 td->this_io_bytes, td->stat_io_bytes,
3045                                 td->ts.bw_stat, td->bw_log, true);
3046 }
3047
3048 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3049                      unsigned int bytes)
3050 {
3051         const bool needs_lock = td_async_processing(td);
3052         struct thread_stat *ts = &td->ts;
3053
3054         if (needs_lock)
3055                 __td_io_u_lock(td);
3056
3057         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3058
3059         if (td->iops_log)
3060                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3061                                bytes, io_u->offset, io_u_is_prio(io_u));
3062
3063         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3064
3065         if (needs_lock)
3066                 __td_io_u_unlock(td);
3067 }
3068
3069 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3070 {
3071         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3072                                 td->this_io_blocks, td->stat_io_blocks,
3073                                 td->ts.iops_stat, td->iops_log, false);
3074 }
3075
3076 /*
3077  * Returns msecs to next event
3078  */
3079 int calc_log_samples(void)
3080 {
3081         struct thread_data *td;
3082         unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3083         struct timespec now;
3084         int i;
3085         long elapsed_time = 0;
3086
3087         fio_gettime(&now, NULL);
3088
3089         for_each_td(td, i) {
3090                 elapsed_time = mtime_since_now(&td->epoch);
3091
3092                 if (!td->o.stats)
3093                         continue;
3094                 if (in_ramp_time(td) ||
3095                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3096                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3097                         continue;
3098                 }
3099                 if (!td->bw_log ||
3100                         (td->bw_log && !per_unit_log(td->bw_log))) {
3101                         tmp = add_bw_samples(td, &now);
3102
3103                         if (td->bw_log)
3104                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3105                 }
3106                 if (!td->iops_log ||
3107                         (td->iops_log && !per_unit_log(td->iops_log))) {
3108                         tmp = add_iops_samples(td, &now);
3109
3110                         if (td->iops_log)
3111                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3112                 }
3113
3114                 if (tmp < next)
3115                         next = tmp;
3116         }
3117
3118         /* if log_avg_msec_min has not been changed, set it to 0 */
3119         if (log_avg_msec_min == -1U)
3120                 log_avg_msec_min = 0;
3121
3122         if (log_avg_msec_min == 0)
3123                 next_mod = elapsed_time;
3124         else
3125                 next_mod = elapsed_time % log_avg_msec_min;
3126
3127         /* correction to keep the time on the log avg msec boundary */
3128         next = min(next, (log_avg_msec_min - next_mod));
3129
3130         return next == ~0U ? 0 : next;
3131 }
3132
3133 void stat_init(void)
3134 {
3135         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3136 }
3137
3138 void stat_exit(void)
3139 {
3140         /*
3141          * When we have the mutex, we know out-of-band access to it
3142          * have ended.
3143          */
3144         fio_sem_down(stat_sem);
3145         fio_sem_remove(stat_sem);
3146 }
3147
3148 /*
3149  * Called from signal handler. Wake up status thread.
3150  */
3151 void show_running_run_stats(void)
3152 {
3153         helper_do_stat();
3154 }
3155
3156 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3157 {
3158         /* Ignore io_u's which span multiple blocks--they will just get
3159          * inaccurate counts. */
3160         int idx = (io_u->offset - io_u->file->file_offset)
3161                         / td->o.bs[DDIR_TRIM];
3162         uint32_t *info = &td->ts.block_infos[idx];
3163         assert(idx < td->ts.nr_block_infos);
3164         return info;
3165 }
3166