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