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 #ifdef WIN32
  21 #define LOG_MSEC_SLACK  2
  22 #else
  23 #define LOG_MSEC_SLACK  1
  24 #endif
  25
  26 struct fio_sem *stat_sem;
  27
  28 void clear_rusage_stat(struct thread_data *td)
  29 {
  30         struct thread_stat *ts = &td->ts;
  31
  32         fio_getrusage(&td->ru_start);
  33         ts->usr_time = ts->sys_time = 0;
  34         ts->ctx = 0;
  35         ts->minf = ts->majf = 0;
  36 }
  37
  38 void update_rusage_stat(struct thread_data *td)
  39 {
  40         struct thread_stat *ts = &td->ts;
  41
  42         fio_getrusage(&td->ru_end);
  43         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
  44                                         &td->ru_end.ru_utime);
  45         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
  46                                         &td->ru_end.ru_stime);
  47         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  48                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  49         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  50         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  51
  52         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  53 }
  54
  55 /*
  56  * Given a latency, return the index of the corresponding bucket in
  57  * the structure tracking percentiles.
  58  *
  59  * (1) find the group (and error bits) that the value (latency)
  60  * belongs to by looking at its MSB. (2) find the bucket number in the
  61  * group by looking at the index bits.
  62  *
  63  */
  64 static unsigned int plat_val_to_idx(unsigned long long val)
  65 {
  66         unsigned int msb, error_bits, base, offset, idx;
  67
  68         /* Find MSB starting from bit 0 */
  69         if (val == 0)
  70                 msb = 0;
  71         else
  72                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
  73
  74         /*
  75          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  76          * all bits of the sample as index
  77          */
  78         if (msb <= FIO_IO_U_PLAT_BITS)
  79                 return val;
  80
  81         /* Compute the number of error bits to discard*/
  82         error_bits = msb - FIO_IO_U_PLAT_BITS;
  83
  84         /* Compute the number of buckets before the group */
  85         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  86
  87         /*
  88          * Discard the error bits and apply the mask to find the
  89          * index for the buckets in the group
  90          */
  91         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  92
  93         /* Make sure the index does not exceed (array size - 1) */
  94         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  95                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  96
  97         return idx;
  98 }
  99
 100 /*
 101  * Convert the given index of the bucket array to the value
 102  * represented by the bucket
 103  */
 104 static unsigned long long plat_idx_to_val(unsigned int idx)
 105 {
 106         unsigned int error_bits;
 107         unsigned long long k, base;
 108
 109         assert(idx < FIO_IO_U_PLAT_NR);
 110
 111         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 112          * all bits of the sample as index */
 113         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 114                 return idx;
 115
 116         /* Find the group and compute the minimum value of that group */
 117         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 118         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
 119
 120         /* Find its bucket number of the group */
 121         k = idx % FIO_IO_U_PLAT_VAL;
 122
 123         /* Return the mean of the range of the bucket */
 124         return base + ((k + 0.5) * (1 << error_bits));
 125 }
 126
 127 static int double_cmp(const void *a, const void *b)
 128 {
 129         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 130         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 131         int cmp = 0;
 132
 133         if (fa.u.f > fb.u.f)
 134                 cmp = 1;
 135         else if (fa.u.f < fb.u.f)
 136                 cmp = -1;
 137
 138         return cmp;
 139 }
 140
 141 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 142                                    fio_fp64_t *plist, unsigned long long **output,
 143                                    unsigned long long *maxv, unsigned long long *minv)
 144 {
 145         unsigned long long sum = 0;
 146         unsigned int len, i, j = 0;
 147         unsigned long long *ovals = NULL;
 148         bool is_last;
 149
 150         *minv = -1ULL;
 151         *maxv = 0;
 152
 153         len = 0;
 154         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 155                 len++;
 156
 157         if (!len)
 158                 return 0;
 159
 160         /*
 161          * Sort the percentile list. Note that it may already be sorted if
 162          * we are using the default values, but since it's a short list this
 163          * isn't a worry. Also note that this does not work for NaN values.
 164          */
 165         if (len > 1)
 166                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 167
 168         ovals = malloc(len * sizeof(*ovals));
 169         if (!ovals)
 170                 return 0;
 171
 172         /*
 173          * Calculate bucket values, note down max and min values
 174          */
 175         is_last = false;
 176         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 177                 sum += io_u_plat[i];
 178                 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
 179                         assert(plist[j].u.f <= 100.0);
 180
 181                         ovals[j] = plat_idx_to_val(i);
 182                         if (ovals[j] < *minv)
 183                                 *minv = ovals[j];
 184                         if (ovals[j] > *maxv)
 185                                 *maxv = ovals[j];
 186
 187                         is_last = (j == len - 1) != 0;
 188                         if (is_last)
 189                                 break;
 190
 191                         j++;
 192                 }
 193         }
 194
 195         if (!is_last)
 196                 log_err("fio: error calculating latency percentiles\n");
 197
 198         *output = ovals;
 199         return len;
 200 }
 201
 202 /*
 203  * Find and display the p-th percentile of clat
 204  */
 205 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 206                                   fio_fp64_t *plist, unsigned int precision,
 207                                   const char *pre, struct buf_output *out)
 208 {
 209         unsigned int divisor, len, i, j = 0;
 210         unsigned long long minv, maxv;
 211         unsigned long long *ovals;
 212         int per_line, scale_down, time_width;
 213         bool is_last;
 214         char fmt[32];
 215
 216         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 217         if (!len || !ovals)
 218                 return;
 219
 220         /*
 221          * We default to nsecs, but if the value range is such that we
 222          * should scale down to usecs or msecs, do that.
 223          */
 224         if (minv > 2000000 && maxv > 99999999ULL) {
 225                 scale_down = 2;
 226                 divisor = 1000000;
 227                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 228         } else if (minv > 2000 && maxv > 99999) {
 229                 scale_down = 1;
 230                 divisor = 1000;
 231                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 232         } else {
 233                 scale_down = 0;
 234                 divisor = 1;
 235                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 236         }
 237
 238
 239         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 240         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 241                         precision, time_width);
 242         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 243         per_line = (80 - 7) / (precision + 10 + time_width);
 244
 245         for (j = 0; j < len; j++) {
 246                 /* for formatting */
 247                 if (j != 0 && (j % per_line) == 0)
 248                         log_buf(out, "     |");
 249
 250                 /* end of the list */
 251                 is_last = (j == len - 1) != 0;
 252
 253                 for (i = 0; i < scale_down; i++)
 254                         ovals[j] = (ovals[j] + 999) / 1000;
 255
 256                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 257
 258                 if (is_last)
 259                         break;
 260
 261                 if ((j % per_line) == per_line - 1)     /* for formatting */
 262                         log_buf(out, "\n");
 263         }
 264
 265         free(ovals);
 266 }
 267
 268 bool calc_lat(struct io_stat *is, unsigned long long *min,
 269               unsigned long long *max, double *mean, double *dev)
 270 {
 271         double n = (double) is->samples;
 272
 273         if (n == 0)
 274                 return false;
 275
 276         *min = is->min_val;
 277         *max = is->max_val;
 278         *mean = is->mean.u.f;
 279
 280         if (n > 1.0)
 281                 *dev = sqrt(is->S.u.f / (n - 1.0));
 282         else
 283                 *dev = 0;
 284
 285         return true;
 286 }
 287
 288 void show_mixed_group_stats(struct group_run_stats *rs, struct buf_output *out)
 289 {
 290         char *io, *agg, *min, *max;
 291         char *ioalt, *aggalt, *minalt, *maxalt;
 292         uint64_t io_mix = 0, agg_mix = 0, min_mix = -1, max_mix = 0;
 293         uint64_t min_run = -1, max_run = 0;
 294         const int i2p = is_power_of_2(rs->kb_base);
 295         int i;
 296
 297         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 298                 if (!rs->max_run[i])
 299                         continue;
 300                 io_mix += rs->iobytes[i];
 301                 agg_mix += rs->agg[i];
 302                 min_mix = min_mix < rs->min_bw[i] ? min_mix : rs->min_bw[i];
 303                 max_mix = max_mix > rs->max_bw[i] ? max_mix : rs->max_bw[i];
 304                 min_run = min_run < rs->min_run[i] ? min_run : rs->min_run[i];
 305                 max_run = max_run > rs->max_run[i] ? max_run : rs->max_run[i];
 306         }
 307         io = num2str(io_mix, rs->sig_figs, 1, i2p, N2S_BYTE);
 308         ioalt = num2str(io_mix, rs->sig_figs, 1, !i2p, N2S_BYTE);
 309         agg = num2str(agg_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 310         aggalt = num2str(agg_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 311         min = num2str(min_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 312         minalt = num2str(min_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 313         max = num2str(max_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 314         maxalt = num2str(max_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 315         log_buf(out, "  MIXED: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 316                         agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 317                         (unsigned long long) min_run,
 318                         (unsigned long long) max_run);
 319         free(io);
 320         free(agg);
 321         free(min);
 322         free(max);
 323         free(ioalt);
 324         free(aggalt);
 325         free(minalt);
 326         free(maxalt);
 327 }
 328
 329 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 330 {
 331         char *io, *agg, *min, *max;
 332         char *ioalt, *aggalt, *minalt, *maxalt;
 333         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 334         int i;
 335
 336         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 337
 338         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 339                 const int i2p = is_power_of_2(rs->kb_base);
 340
 341                 if (!rs->max_run[i])
 342                         continue;
 343
 344                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
 345                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
 346                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
 347                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 348                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 349                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 350                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 351                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 352                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 353                                 (rs->unified_rw_rep == UNIFIED_MIXED) ? "  MIXED" : str[i],
 354                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 355                                 (unsigned long long) rs->min_run[i],
 356                                 (unsigned long long) rs->max_run[i]);
 357
 358                 free(io);
 359                 free(agg);
 360                 free(min);
 361                 free(max);
 362                 free(ioalt);
 363                 free(aggalt);
 364                 free(minalt);
 365                 free(maxalt);
 366         }
 367
 368         /* Need to aggregate statisitics to show mixed values */
 369         if (rs->unified_rw_rep == UNIFIED_BOTH)
 370                 show_mixed_group_stats(rs, out);
 371 }
 372
 373 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
 374 {
 375         int i;
 376
 377         /*
 378          * Do depth distribution calculations
 379          */
 380         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 381                 if (total) {
 382                         io_u_dist[i] = (double) map[i] / (double) total;
 383                         io_u_dist[i] *= 100.0;
 384                         if (io_u_dist[i] < 0.1 && map[i])
 385                                 io_u_dist[i] = 0.1;
 386                 } else
 387                         io_u_dist[i] = 0.0;
 388         }
 389 }
 390
 391 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 392                           uint64_t *src, int nr)
 393 {
 394         unsigned long total = ddir_rw_sum(ts->total_io_u);
 395         int i;
 396
 397         /*
 398          * Do latency distribution calculations
 399          */
 400         for (i = 0; i < nr; i++) {
 401                 if (total) {
 402                         dst[i] = (double) src[i] / (double) total;
 403                         dst[i] *= 100.0;
 404                         if (dst[i] < 0.01 && src[i])
 405                                 dst[i] = 0.01;
 406                 } else
 407                         dst[i] = 0.0;
 408         }
 409 }
 410
 411 /*
 412  * To keep the terse format unaltered, add all of the ns latency
 413  * buckets to the first us latency bucket
 414  */
 415 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 416 {
 417         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 418         int i;
 419
 420         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 421
 422         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 423                 ntotal += ts->io_u_lat_n[i];
 424
 425         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 426 }
 427
 428 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 429 {
 430         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 431 }
 432
 433 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 434 {
 435         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 436 }
 437
 438 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 439 {
 440         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 441 }
 442
 443 static void display_lat(const char *name, unsigned long long min,
 444                         unsigned long long max, double mean, double dev,
 445                         struct buf_output *out)
 446 {
 447         const char *base = "(nsec)";
 448         char *minp, *maxp;
 449
 450         if (nsec_to_msec(&min, &max, &mean, &dev))
 451                 base = "(msec)";
 452         else if (nsec_to_usec(&min, &max, &mean, &dev))
 453                 base = "(usec)";
 454
 455         minp = num2str(min, 6, 1, 0, N2S_NONE);
 456         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 457
 458         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 459                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 460
 461         free(minp);
 462         free(maxp);
 463 }
 464
 465 static struct thread_stat *gen_mixed_ddir_stats_from_ts(struct thread_stat *ts)
 466 {
 467         struct thread_stat *ts_lcl;
 468
 469         /*
 470          * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
 471          * Trims (ddir = 2)
 472          */
 473         ts_lcl = malloc(sizeof(struct thread_stat));
 474         if (!ts_lcl) {
 475                 log_err("fio: failed to allocate local thread stat\n");
 476                 return NULL;
 477         }
 478
 479         init_thread_stat(ts_lcl);
 480
 481         /* calculate mixed stats  */
 482         ts_lcl->unified_rw_rep = UNIFIED_MIXED;
 483         ts_lcl->lat_percentiles = ts->lat_percentiles;
 484         ts_lcl->clat_percentiles = ts->clat_percentiles;
 485         ts_lcl->slat_percentiles = ts->slat_percentiles;
 486         ts_lcl->percentile_precision = ts->percentile_precision;
 487         memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
 488
 489         sum_thread_stats(ts_lcl, ts);
 490
 491         return ts_lcl;
 492 }
 493
 494 static double convert_agg_kbytes_percent(struct group_run_stats *rs, int ddir, int mean)
 495 {
 496         double p_of_agg = 100.0;
 497         if (rs && rs->agg[ddir] > 1024) {
 498                 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
 499
 500                 if (p_of_agg > 100.0)
 501                         p_of_agg = 100.0;
 502         }
 503         return p_of_agg;
 504 }
 505
 506 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 507                              int ddir, struct buf_output *out)
 508 {
 509         unsigned long runt;
 510         unsigned long long min, max, bw, iops;
 511         double mean, dev;
 512         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
 513         int i2p;
 514
 515         if (ddir_sync(ddir)) {
 516                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
 517                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
 518                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
 519                         show_clat_percentiles(ts->io_u_sync_plat,
 520                                                 ts->sync_stat.samples,
 521                                                 ts->percentile_list,
 522                                                 ts->percentile_precision,
 523                                                 io_ddir_name(ddir), out);
 524                 }
 525                 return;
 526         }
 527
 528         assert(ddir_rw(ddir));
 529
 530         if (!ts->runtime[ddir])
 531                 return;
 532
 533         i2p = is_power_of_2(rs->kb_base);
 534         runt = ts->runtime[ddir];
 535
 536         bw = (1000 * ts->io_bytes[ddir]) / runt;
 537         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
 538         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
 539         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
 540
 541         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 542         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
 543         if (ddir == DDIR_WRITE)
 544                 post_st = zbd_write_status(ts);
 545         else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
 546                 uint64_t total;
 547                 double hit;
 548
 549                 total = ts->cachehit + ts->cachemiss;
 550                 hit = (double) ts->cachehit / (double) total;
 551                 hit *= 100.0;
 552                 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
 553                         post_st = NULL;
 554         }
 555
 556         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
 557                         (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
 558                         iops_p, bw_p, bw_p_alt, io_p,
 559                         (unsigned long long) ts->runtime[ddir],
 560                         post_st ? : "");
 561
 562         free(post_st);
 563         free(io_p);
 564         free(bw_p);
 565         free(bw_p_alt);
 566         free(iops_p);
 567
 568         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 569                 display_lat("slat", min, max, mean, dev, out);
 570         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 571                 display_lat("clat", min, max, mean, dev, out);
 572         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 573                 display_lat(" lat", min, max, mean, dev, out);
 574         if (calc_lat(&ts->clat_high_prio_stat[ddir], &min, &max, &mean, &dev)) {
 575                 display_lat(ts->lat_percentiles ? "high prio_lat" : "high prio_clat",
 576                                 min, max, mean, dev, out);
 577                 if (calc_lat(&ts->clat_low_prio_stat[ddir], &min, &max, &mean, &dev))
 578                         display_lat(ts->lat_percentiles ? "low prio_lat" : "low prio_clat",
 579                                         min, max, mean, dev, out);
 580         }
 581
 582         if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
 583                 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
 584                                         ts->slat_stat[ddir].samples,
 585                                         ts->percentile_list,
 586                                         ts->percentile_precision, "slat", out);
 587         if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
 588                 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
 589                                         ts->clat_stat[ddir].samples,
 590                                         ts->percentile_list,
 591                                         ts->percentile_precision, "clat", out);
 592         if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
 593                 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
 594                                         ts->lat_stat[ddir].samples,
 595                                         ts->percentile_list,
 596                                         ts->percentile_precision, "lat", out);
 597
 598         if (ts->clat_percentiles || ts->lat_percentiles) {
 599                 const char *name = ts->lat_percentiles ? "lat" : "clat";
 600                 char prio_name[32];
 601                 uint64_t samples;
 602
 603                 if (ts->lat_percentiles)
 604                         samples = ts->lat_stat[ddir].samples;
 605                 else
 606                         samples = ts->clat_stat[ddir].samples;
 607
 608                 /* Only print this if some high and low priority stats were collected */
 609                 if (ts->clat_high_prio_stat[ddir].samples > 0 &&
 610                         ts->clat_low_prio_stat[ddir].samples > 0)
 611                 {
 612                         sprintf(prio_name, "high prio (%.2f%%) %s",
 613                                         100. * (double) ts->clat_high_prio_stat[ddir].samples / (double) samples,
 614                                         name);
 615                         show_clat_percentiles(ts->io_u_plat_high_prio[ddir],
 616                                                 ts->clat_high_prio_stat[ddir].samples,
 617                                                 ts->percentile_list,
 618                                                 ts->percentile_precision, prio_name, out);
 619
 620                         sprintf(prio_name, "low prio (%.2f%%) %s",
 621                                         100. * (double) ts->clat_low_prio_stat[ddir].samples / (double) samples,
 622                                         name);
 623                         show_clat_percentiles(ts->io_u_plat_low_prio[ddir],
 624                                                 ts->clat_low_prio_stat[ddir].samples,
 625                                                 ts->percentile_list,
 626                                                 ts->percentile_precision, prio_name, out);
 627                 }
 628         }
 629
 630         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 631                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 632                 const char *bw_str;
 633
 634                 if ((rs->unit_base == 1) && i2p)
 635                         bw_str = "Kibit";
 636                 else if (rs->unit_base == 1)
 637                         bw_str = "kbit";
 638                 else if (i2p)
 639                         bw_str = "KiB";
 640                 else
 641                         bw_str = "kB";
 642
 643                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
 644
 645                 if (rs->unit_base == 1) {
 646                         min *= 8.0;
 647                         max *= 8.0;
 648                         mean *= 8.0;
 649                         dev *= 8.0;
 650                 }
 651
 652                 if (mean > fkb_base * fkb_base) {
 653                         min /= fkb_base;
 654                         max /= fkb_base;
 655                         mean /= fkb_base;
 656                         dev /= fkb_base;
 657                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 658                 }
 659
 660                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 661                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 662                         bw_str, min, max, p_of_agg, mean, dev,
 663                         (&ts->bw_stat[ddir])->samples);
 664         }
 665         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 666                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 667                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 668                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 669         }
 670 }
 671
 672 static void show_mixed_ddir_status(struct group_run_stats *rs,
 673                                    struct thread_stat *ts,
 674                                    struct buf_output *out)
 675 {
 676         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
 677
 678         if (ts_lcl)
 679                 show_ddir_status(rs, ts_lcl, DDIR_READ, out);
 680
 681         free(ts_lcl);
 682 }
 683
 684 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
 685                      const char *msg, struct buf_output *out)
 686 {
 687         bool new_line = true, shown = false;
 688         int i, line = 0;
 689
 690         for (i = 0; i < nr; i++) {
 691                 if (io_u_lat[i] <= 0.0)
 692                         continue;
 693                 shown = true;
 694                 if (new_line) {
 695                         if (line)
 696                                 log_buf(out, "\n");
 697                         log_buf(out, "  lat (%s)   : ", msg);
 698                         new_line = false;
 699                         line = 0;
 700                 }
 701                 if (line)
 702                         log_buf(out, ", ");
 703                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 704                 line++;
 705                 if (line == 5)
 706                         new_line = true;
 707         }
 708
 709         if (shown)
 710                 log_buf(out, "\n");
 711
 712         return true;
 713 }
 714
 715 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 716 {
 717         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 718                                  "250=", "500=", "750=", "1000=", };
 719
 720         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 721 }
 722
 723 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 724 {
 725         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 726                                  "250=", "500=", "750=", "1000=", };
 727
 728         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 729 }
 730
 731 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 732 {
 733         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 734                                  "250=", "500=", "750=", "1000=", "2000=",
 735                                  ">=2000=", };
 736
 737         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 738 }
 739
 740 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 741 {
 742         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 743         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 744         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 745
 746         stat_calc_lat_n(ts, io_u_lat_n);
 747         stat_calc_lat_u(ts, io_u_lat_u);
 748         stat_calc_lat_m(ts, io_u_lat_m);
 749
 750         show_lat_n(io_u_lat_n, out);
 751         show_lat_u(io_u_lat_u, out);
 752         show_lat_m(io_u_lat_m, out);
 753 }
 754
 755 static int block_state_category(int block_state)
 756 {
 757         switch (block_state) {
 758         case BLOCK_STATE_UNINIT:
 759                 return 0;
 760         case BLOCK_STATE_TRIMMED:
 761         case BLOCK_STATE_WRITTEN:
 762                 return 1;
 763         case BLOCK_STATE_WRITE_FAILURE:
 764         case BLOCK_STATE_TRIM_FAILURE:
 765                 return 2;
 766         default:
 767                 /* Silence compile warning on some BSDs and have a return */
 768                 assert(0);
 769                 return -1;
 770         }
 771 }
 772
 773 static int compare_block_infos(const void *bs1, const void *bs2)
 774 {
 775         uint64_t block1 = *(uint64_t *)bs1;
 776         uint64_t block2 = *(uint64_t *)bs2;
 777         int state1 = BLOCK_INFO_STATE(block1);
 778         int state2 = BLOCK_INFO_STATE(block2);
 779         int bscat1 = block_state_category(state1);
 780         int bscat2 = block_state_category(state2);
 781         int cycles1 = BLOCK_INFO_TRIMS(block1);
 782         int cycles2 = BLOCK_INFO_TRIMS(block2);
 783
 784         if (bscat1 < bscat2)
 785                 return -1;
 786         if (bscat1 > bscat2)
 787                 return 1;
 788
 789         if (cycles1 < cycles2)
 790                 return -1;
 791         if (cycles1 > cycles2)
 792                 return 1;
 793
 794         if (state1 < state2)
 795                 return -1;
 796         if (state1 > state2)
 797                 return 1;
 798
 799         assert(block1 == block2);
 800         return 0;
 801 }
 802
 803 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 804                                   fio_fp64_t *plist, unsigned int **percentiles,
 805                                   unsigned int *types)
 806 {
 807         int len = 0;
 808         int i, nr_uninit;
 809
 810         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 811
 812         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 813                 len++;
 814
 815         if (!len)
 816                 return 0;
 817
 818         /*
 819          * Sort the percentile list. Note that it may already be sorted if
 820          * we are using the default values, but since it's a short list this
 821          * isn't a worry. Also note that this does not work for NaN values.
 822          */
 823         if (len > 1)
 824                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 825
 826         /* Start only after the uninit entries end */
 827         for (nr_uninit = 0;
 828              nr_uninit < nr_block_infos
 829                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 830              nr_uninit ++)
 831                 ;
 832
 833         if (nr_uninit == nr_block_infos)
 834                 return 0;
 835
 836         *percentiles = calloc(len, sizeof(**percentiles));
 837
 838         for (i = 0; i < len; i++) {
 839                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 840                                 + nr_uninit;
 841                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 842         }
 843
 844         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 845         for (i = 0; i < nr_block_infos; i++)
 846                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 847
 848         return len;
 849 }
 850
 851 static const char *block_state_names[] = {
 852         [BLOCK_STATE_UNINIT] = "unwritten",
 853         [BLOCK_STATE_TRIMMED] = "trimmed",
 854         [BLOCK_STATE_WRITTEN] = "written",
 855         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 856         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 857 };
 858
 859 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 860                              fio_fp64_t *plist, struct buf_output *out)
 861 {
 862         int len, pos, i;
 863         unsigned int *percentiles = NULL;
 864         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 865
 866         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 867                                      &percentiles, block_state_counts);
 868
 869         log_buf(out, "  block lifetime percentiles :\n   |");
 870         pos = 0;
 871         for (i = 0; i < len; i++) {
 872                 uint32_t block_info = percentiles[i];
 873 #define LINE_LENGTH     75
 874                 char str[LINE_LENGTH];
 875                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 876                                      plist[i].u.f, block_info,
 877                                      i == len - 1 ? '\n' : ',');
 878                 assert(strln < LINE_LENGTH);
 879                 if (pos + strln > LINE_LENGTH) {
 880                         pos = 0;
 881                         log_buf(out, "\n   |");
 882                 }
 883                 log_buf(out, "%s", str);
 884                 pos += strln;
 885 #undef LINE_LENGTH
 886         }
 887         if (percentiles)
 888                 free(percentiles);
 889
 890         log_buf(out, "        states               :");
 891         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 892                 log_buf(out, " %s=%u%c",
 893                          block_state_names[i], block_state_counts[i],
 894                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 895 }
 896
 897 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 898 {
 899         char *p1, *p1alt, *p2;
 900         unsigned long long bw_mean, iops_mean;
 901         const int i2p = is_power_of_2(ts->kb_base);
 902
 903         if (!ts->ss_dur)
 904                 return;
 905
 906         bw_mean = steadystate_bw_mean(ts);
 907         iops_mean = steadystate_iops_mean(ts);
 908
 909         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
 910         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
 911         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
 912
 913         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 914                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
 915                 p1, p1alt, p2,
 916                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
 917                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
 918                 ts->ss_criterion.u.f,
 919                 ts->ss_state & FIO_SS_PCT ? "%" : "");
 920
 921         free(p1);
 922         free(p1alt);
 923         free(p2);
 924 }
 925
 926 static void show_agg_stats(struct disk_util_agg *agg, int terse,
 927                            struct buf_output *out)
 928 {
 929         if (!agg->slavecount)
 930                 return;
 931
 932         if (!terse) {
 933                 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
 934                          "aggrticks=%llu/%llu, aggrin_queue=%llu, "
 935                          "aggrutil=%3.2f%%",
 936                         (unsigned long long) agg->ios[0] / agg->slavecount,
 937                         (unsigned long long) agg->ios[1] / agg->slavecount,
 938                         (unsigned long long) agg->merges[0] / agg->slavecount,
 939                         (unsigned long long) agg->merges[1] / agg->slavecount,
 940                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 941                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 942                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 943                         agg->max_util.u.f);
 944         } else {
 945                 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
 946                         (unsigned long long) agg->ios[0] / agg->slavecount,
 947                         (unsigned long long) agg->ios[1] / agg->slavecount,
 948                         (unsigned long long) agg->merges[0] / agg->slavecount,
 949                         (unsigned long long) agg->merges[1] / agg->slavecount,
 950                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 951                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 952                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 953                         agg->max_util.u.f);
 954         }
 955 }
 956
 957 static void aggregate_slaves_stats(struct disk_util *masterdu)
 958 {
 959         struct disk_util_agg *agg = &masterdu->agg;
 960         struct disk_util_stat *dus;
 961         struct flist_head *entry;
 962         struct disk_util *slavedu;
 963         double util;
 964
 965         flist_for_each(entry, &masterdu->slaves) {
 966                 slavedu = flist_entry(entry, struct disk_util, slavelist);
 967                 dus = &slavedu->dus;
 968                 agg->ios[0] += dus->s.ios[0];
 969                 agg->ios[1] += dus->s.ios[1];
 970                 agg->merges[0] += dus->s.merges[0];
 971                 agg->merges[1] += dus->s.merges[1];
 972                 agg->sectors[0] += dus->s.sectors[0];
 973                 agg->sectors[1] += dus->s.sectors[1];
 974                 agg->ticks[0] += dus->s.ticks[0];
 975                 agg->ticks[1] += dus->s.ticks[1];
 976                 agg->time_in_queue += dus->s.time_in_queue;
 977                 agg->slavecount++;
 978
 979                 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
 980                 /* System utilization is the utilization of the
 981                  * component with the highest utilization.
 982                  */
 983                 if (util > agg->max_util.u.f)
 984                         agg->max_util.u.f = util;
 985
 986         }
 987
 988         if (agg->max_util.u.f > 100.0)
 989                 agg->max_util.u.f = 100.0;
 990 }
 991
 992 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
 993                      int terse, struct buf_output *out)
 994 {
 995         double util = 0;
 996
 997         if (dus->s.msec)
 998                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
 999         if (util > 100.0)
1000                 util = 100.0;
1001
1002         if (!terse) {
1003                 if (agg->slavecount)
1004                         log_buf(out, "  ");
1005
1006                 log_buf(out, "  %s: ios=%llu/%llu, merge=%llu/%llu, "
1007                          "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
1008                                 dus->name,
1009                                 (unsigned long long) dus->s.ios[0],
1010                                 (unsigned long long) dus->s.ios[1],
1011                                 (unsigned long long) dus->s.merges[0],
1012                                 (unsigned long long) dus->s.merges[1],
1013                                 (unsigned long long) dus->s.ticks[0],
1014                                 (unsigned long long) dus->s.ticks[1],
1015                                 (unsigned long long) dus->s.time_in_queue,
1016                                 util);
1017         } else {
1018                 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1019                                 dus->name,
1020                                 (unsigned long long) dus->s.ios[0],
1021                                 (unsigned long long) dus->s.ios[1],
1022                                 (unsigned long long) dus->s.merges[0],
1023                                 (unsigned long long) dus->s.merges[1],
1024                                 (unsigned long long) dus->s.ticks[0],
1025                                 (unsigned long long) dus->s.ticks[1],
1026                                 (unsigned long long) dus->s.time_in_queue,
1027                                 util);
1028         }
1029
1030         /*
1031          * If the device has slaves, aggregate the stats for
1032          * those slave devices also.
1033          */
1034         show_agg_stats(agg, terse, out);
1035
1036         if (!terse)
1037                 log_buf(out, "\n");
1038 }
1039
1040 void json_array_add_disk_util(struct disk_util_stat *dus,
1041                 struct disk_util_agg *agg, struct json_array *array)
1042 {
1043         struct json_object *obj;
1044         double util = 0;
1045
1046         if (dus->s.msec)
1047                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1048         if (util > 100.0)
1049                 util = 100.0;
1050
1051         obj = json_create_object();
1052         json_array_add_value_object(array, obj);
1053
1054         json_object_add_value_string(obj, "name", (const char *)dus->name);
1055         json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1056         json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1057         json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1058         json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1059         json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1060         json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1061         json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1062         json_object_add_value_float(obj, "util", util);
1063
1064         /*
1065          * If the device has slaves, aggregate the stats for
1066          * those slave devices also.
1067          */
1068         if (!agg->slavecount)
1069                 return;
1070         json_object_add_value_int(obj, "aggr_read_ios",
1071                                 agg->ios[0] / agg->slavecount);
1072         json_object_add_value_int(obj, "aggr_write_ios",
1073                                 agg->ios[1] / agg->slavecount);
1074         json_object_add_value_int(obj, "aggr_read_merges",
1075                                 agg->merges[0] / agg->slavecount);
1076         json_object_add_value_int(obj, "aggr_write_merge",
1077                                 agg->merges[1] / agg->slavecount);
1078         json_object_add_value_int(obj, "aggr_read_ticks",
1079                                 agg->ticks[0] / agg->slavecount);
1080         json_object_add_value_int(obj, "aggr_write_ticks",
1081                                 agg->ticks[1] / agg->slavecount);
1082         json_object_add_value_int(obj, "aggr_in_queue",
1083                                 agg->time_in_queue / agg->slavecount);
1084         json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1085 }
1086
1087 static void json_object_add_disk_utils(struct json_object *obj,
1088                                        struct flist_head *head)
1089 {
1090         struct json_array *array = json_create_array();
1091         struct flist_head *entry;
1092         struct disk_util *du;
1093
1094         json_object_add_value_array(obj, "disk_util", array);
1095
1096         flist_for_each(entry, head) {
1097                 du = flist_entry(entry, struct disk_util, list);
1098
1099                 aggregate_slaves_stats(du);
1100                 json_array_add_disk_util(&du->dus, &du->agg, array);
1101         }
1102 }
1103
1104 void show_disk_util(int terse, struct json_object *parent,
1105                     struct buf_output *out)
1106 {
1107         struct flist_head *entry;
1108         struct disk_util *du;
1109         bool do_json;
1110
1111         if (!is_running_backend())
1112                 return;
1113
1114         if (flist_empty(&disk_list))
1115                 return;
1116
1117         if ((output_format & FIO_OUTPUT_JSON) && parent)
1118                 do_json = true;
1119         else
1120                 do_json = false;
1121
1122         if (!terse && !do_json)
1123                 log_buf(out, "\nDisk stats (read/write):\n");
1124
1125         if (do_json) {
1126                 json_object_add_disk_utils(parent, &disk_list);
1127         } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1128                 flist_for_each(entry, &disk_list) {
1129                         du = flist_entry(entry, struct disk_util, list);
1130
1131                         aggregate_slaves_stats(du);
1132                         print_disk_util(&du->dus, &du->agg, terse, out);
1133                 }
1134         }
1135 }
1136
1137 static void show_thread_status_normal(struct thread_stat *ts,
1138                                       struct group_run_stats *rs,
1139                                       struct buf_output *out)
1140 {
1141         double usr_cpu, sys_cpu;
1142         unsigned long runtime;
1143         double io_u_dist[FIO_IO_U_MAP_NR];
1144         time_t time_p;
1145         char time_buf[32];
1146
1147         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1148                 return;
1149
1150         memset(time_buf, 0, sizeof(time_buf));
1151
1152         time(&time_p);
1153         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1154
1155         if (!ts->error) {
1156                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1157                                         ts->name, ts->groupid, ts->members,
1158                                         ts->error, (int) ts->pid, time_buf);
1159         } else {
1160                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1161                                         ts->name, ts->groupid, ts->members,
1162                                         ts->error, ts->verror, (int) ts->pid,
1163                                         time_buf);
1164         }
1165
1166         if (strlen(ts->description))
1167                 log_buf(out, "  Description  : [%s]\n", ts->description);
1168
1169         for_each_rw_ddir(ddir) {
1170                 if (ts->io_bytes[ddir])
1171                         show_ddir_status(rs, ts, ddir, out);
1172         }
1173
1174         if (ts->unified_rw_rep == UNIFIED_BOTH)
1175                 show_mixed_ddir_status(rs, ts, out);
1176
1177         show_latencies(ts, out);
1178
1179         if (ts->sync_stat.samples)
1180                 show_ddir_status(rs, ts, DDIR_SYNC, out);
1181
1182         runtime = ts->total_run_time;
1183         if (runtime) {
1184                 double runt = (double) runtime;
1185
1186                 usr_cpu = (double) ts->usr_time * 100 / runt;
1187                 sys_cpu = (double) ts->sys_time * 100 / runt;
1188         } else {
1189                 usr_cpu = 0;
1190                 sys_cpu = 0;
1191         }
1192
1193         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1194                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1195                         (unsigned long long) ts->ctx,
1196                         (unsigned long long) ts->majf,
1197                         (unsigned long long) ts->minf);
1198
1199         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1200         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1201                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1202                                         io_u_dist[1], io_u_dist[2],
1203                                         io_u_dist[3], io_u_dist[4],
1204                                         io_u_dist[5], io_u_dist[6]);
1205
1206         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1207         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1208                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1209                                         io_u_dist[1], io_u_dist[2],
1210                                         io_u_dist[3], io_u_dist[4],
1211                                         io_u_dist[5], io_u_dist[6]);
1212         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1213         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1214                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1215                                         io_u_dist[1], io_u_dist[2],
1216                                         io_u_dist[3], io_u_dist[4],
1217                                         io_u_dist[5], io_u_dist[6]);
1218         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
1219                                  " short=%llu,%llu,%llu,0"
1220                                  " dropped=%llu,%llu,%llu,0\n",
1221                                         (unsigned long long) ts->total_io_u[0],
1222                                         (unsigned long long) ts->total_io_u[1],
1223                                         (unsigned long long) ts->total_io_u[2],
1224                                         (unsigned long long) ts->total_io_u[3],
1225                                         (unsigned long long) ts->short_io_u[0],
1226                                         (unsigned long long) ts->short_io_u[1],
1227                                         (unsigned long long) ts->short_io_u[2],
1228                                         (unsigned long long) ts->drop_io_u[0],
1229                                         (unsigned long long) ts->drop_io_u[1],
1230                                         (unsigned long long) ts->drop_io_u[2]);
1231         if (ts->continue_on_error) {
1232                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
1233                                         (unsigned long long)ts->total_err_count,
1234                                         ts->first_error,
1235                                         strerror(ts->first_error));
1236         }
1237         if (ts->latency_depth) {
1238                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1239                                         (unsigned long long)ts->latency_target,
1240                                         (unsigned long long)ts->latency_window,
1241                                         ts->latency_percentile.u.f,
1242                                         ts->latency_depth);
1243         }
1244
1245         if (ts->nr_block_infos)
1246                 show_block_infos(ts->nr_block_infos, ts->block_infos,
1247                                   ts->percentile_list, out);
1248
1249         if (ts->ss_dur)
1250                 show_ss_normal(ts, out);
1251 }
1252
1253 static void show_ddir_status_terse(struct thread_stat *ts,
1254                                    struct group_run_stats *rs, int ddir,
1255                                    int ver, struct buf_output *out)
1256 {
1257         unsigned long long min, max, minv, maxv, bw, iops;
1258         unsigned long long *ovals = NULL;
1259         double mean, dev;
1260         unsigned int len;
1261         int i, bw_stat;
1262
1263         assert(ddir_rw(ddir));
1264
1265         iops = bw = 0;
1266         if (ts->runtime[ddir]) {
1267                 uint64_t runt = ts->runtime[ddir];
1268
1269                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1270                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1271         }
1272
1273         log_buf(out, ";%llu;%llu;%llu;%llu",
1274                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1275                                         (unsigned long long) ts->runtime[ddir]);
1276
1277         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1278                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1279         else
1280                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1281
1282         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1283                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1284         else
1285                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1286
1287         if (ts->lat_percentiles) {
1288                 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1289                                         ts->lat_stat[ddir].samples,
1290                                         ts->percentile_list, &ovals, &maxv,
1291                                         &minv);
1292         } else if (ts->clat_percentiles) {
1293                 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1294                                         ts->clat_stat[ddir].samples,
1295                                         ts->percentile_list, &ovals, &maxv,
1296                                         &minv);
1297         } else {
1298                 len = 0;
1299         }
1300
1301         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1302                 if (i >= len) {
1303                         log_buf(out, ";0%%=0");
1304                         continue;
1305                 }
1306                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1307         }
1308
1309         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1310                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1311         else
1312                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1313
1314         free(ovals);
1315
1316         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1317         if (bw_stat) {
1318                 double p_of_agg = 100.0;
1319
1320                 if (rs->agg[ddir]) {
1321                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1322                         if (p_of_agg > 100.0)
1323                                 p_of_agg = 100.0;
1324                 }
1325
1326                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1327         } else {
1328                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1329         }
1330
1331         if (ver == 5) {
1332                 if (bw_stat)
1333                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1334                 else
1335                         log_buf(out, ";%lu", 0UL);
1336
1337                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1338                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1339                                 mean, dev, (&ts->iops_stat[ddir])->samples);
1340                 else
1341                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1342         }
1343 }
1344
1345 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1346                                    struct group_run_stats *rs,
1347                                    int ver, struct buf_output *out)
1348 {
1349         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1350
1351         if (ts_lcl)
1352                 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1353
1354         free(ts_lcl);
1355 }
1356
1357 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1358                                              uint32_t percentiles,
1359                                              struct io_stat *lat_stat,
1360                                              uint64_t *io_u_plat)
1361 {
1362         char buf[120];
1363         double mean, dev;
1364         unsigned int i, len;
1365         struct json_object *lat_object, *percentile_object, *clat_bins_object;
1366         unsigned long long min, max, maxv, minv, *ovals = NULL;
1367
1368         if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1369                 min = max = 0;
1370                 mean = dev = 0.0;
1371         }
1372         lat_object = json_create_object();
1373         json_object_add_value_int(lat_object, "min", min);
1374         json_object_add_value_int(lat_object, "max", max);
1375         json_object_add_value_float(lat_object, "mean", mean);
1376         json_object_add_value_float(lat_object, "stddev", dev);
1377         json_object_add_value_int(lat_object, "N", lat_stat->samples);
1378
1379         if (percentiles && lat_stat->samples) {
1380                 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1381                                 ts->percentile_list, &ovals, &maxv, &minv);
1382
1383                 if (len > FIO_IO_U_LIST_MAX_LEN)
1384                         len = FIO_IO_U_LIST_MAX_LEN;
1385
1386                 percentile_object = json_create_object();
1387                 json_object_add_value_object(lat_object, "percentile", percentile_object);
1388                 for (i = 0; i < len; i++) {
1389                         snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1390                         json_object_add_value_int(percentile_object, buf, ovals[i]);
1391                 }
1392                 free(ovals);
1393
1394                 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1395                         clat_bins_object = json_create_object();
1396                         json_object_add_value_object(lat_object, "bins", clat_bins_object);
1397
1398                         for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1399                                 if (io_u_plat[i]) {
1400                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1401                                         json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1402                                 }
1403                 }
1404         }
1405
1406         return lat_object;
1407 }
1408
1409 static void add_ddir_status_json(struct thread_stat *ts,
1410                 struct group_run_stats *rs, int ddir, struct json_object *parent)
1411 {
1412         unsigned long long min, max;
1413         unsigned long long bw_bytes, bw;
1414         double mean, dev, iops;
1415         struct json_object *dir_object, *tmp_object;
1416         double p_of_agg = 100.0;
1417
1418         assert(ddir_rw(ddir) || ddir_sync(ddir));
1419
1420         if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1421                 return;
1422
1423         dir_object = json_create_object();
1424         json_object_add_value_object(parent,
1425                 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1426
1427         if (ddir_rw(ddir)) {
1428                 bw_bytes = 0;
1429                 bw = 0;
1430                 iops = 0.0;
1431                 if (ts->runtime[ddir]) {
1432                         uint64_t runt = ts->runtime[ddir];
1433
1434                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1435                         bw = bw_bytes / 1024; /* KiB/s */
1436                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1437                 }
1438
1439                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1440                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1441                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1442                 json_object_add_value_int(dir_object, "bw", bw);
1443                 json_object_add_value_float(dir_object, "iops", iops);
1444                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1445                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1446                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1447                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1448
1449                 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1450                                 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1451                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1452
1453                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1454                                 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1455                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1456
1457                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1458                                 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1459                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1460         } else {
1461                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1462                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1463                                 &ts->sync_stat, ts->io_u_sync_plat);
1464                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1465         }
1466
1467         if (!ddir_rw(ddir))
1468                 return;
1469
1470         /* Only print PRIO latencies if some high priority samples were gathered */
1471         if (ts->clat_high_prio_stat[ddir].samples > 0) {
1472                 const char *high, *low;
1473
1474                 if (ts->lat_percentiles) {
1475                         high = "lat_high_prio";
1476                         low = "lat_low_prio";
1477                 } else {
1478                         high = "clat_high_prio";
1479                         low = "clat_low_prio";
1480                 }
1481
1482                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1483                                 &ts->clat_high_prio_stat[ddir], ts->io_u_plat_high_prio[ddir]);
1484                 json_object_add_value_object(dir_object, high, tmp_object);
1485
1486                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1487                                 &ts->clat_low_prio_stat[ddir], ts->io_u_plat_low_prio[ddir]);
1488                 json_object_add_value_object(dir_object, low, tmp_object);
1489         }
1490
1491         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1492                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1493         } else {
1494                 min = max = 0;
1495                 p_of_agg = mean = dev = 0.0;
1496         }
1497
1498         json_object_add_value_int(dir_object, "bw_min", min);
1499         json_object_add_value_int(dir_object, "bw_max", max);
1500         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1501         json_object_add_value_float(dir_object, "bw_mean", mean);
1502         json_object_add_value_float(dir_object, "bw_dev", dev);
1503         json_object_add_value_int(dir_object, "bw_samples",
1504                                 (&ts->bw_stat[ddir])->samples);
1505
1506         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1507                 min = max = 0;
1508                 mean = dev = 0.0;
1509         }
1510         json_object_add_value_int(dir_object, "iops_min", min);
1511         json_object_add_value_int(dir_object, "iops_max", max);
1512         json_object_add_value_float(dir_object, "iops_mean", mean);
1513         json_object_add_value_float(dir_object, "iops_stddev", dev);
1514         json_object_add_value_int(dir_object, "iops_samples",
1515                                 (&ts->iops_stat[ddir])->samples);
1516
1517         if (ts->cachehit + ts->cachemiss) {
1518                 uint64_t total;
1519                 double hit;
1520
1521                 total = ts->cachehit + ts->cachemiss;
1522                 hit = (double) ts->cachehit / (double) total;
1523                 hit *= 100.0;
1524                 json_object_add_value_float(dir_object, "cachehit", hit);
1525         }
1526 }
1527
1528 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1529                 struct group_run_stats *rs, struct json_object *parent)
1530 {
1531         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1532
1533         /* add the aggregated stats to json parent */
1534         if (ts_lcl)
1535                 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1536
1537         free(ts_lcl);
1538 }
1539
1540 static void show_thread_status_terse_all(struct thread_stat *ts,
1541                                          struct group_run_stats *rs, int ver,
1542                                          struct buf_output *out)
1543 {
1544         double io_u_dist[FIO_IO_U_MAP_NR];
1545         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1546         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1547         double usr_cpu, sys_cpu;
1548         int i;
1549
1550         /* General Info */
1551         if (ver == 2)
1552                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1553         else
1554                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1555                         ts->name, ts->groupid, ts->error);
1556
1557         /* Log Read Status, or mixed if unified_rw_rep = 1 */
1558         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1559         if (ts->unified_rw_rep != UNIFIED_MIXED) {
1560                 /* Log Write Status */
1561                 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1562                 /* Log Trim Status */
1563                 if (ver == 2 || ver == 4 || ver == 5)
1564                         show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1565         }
1566         if (ts->unified_rw_rep == UNIFIED_BOTH)
1567                 show_mixed_ddir_status_terse(ts, rs, ver, out);
1568         /* CPU Usage */
1569         if (ts->total_run_time) {
1570                 double runt = (double) ts->total_run_time;
1571
1572                 usr_cpu = (double) ts->usr_time * 100 / runt;
1573                 sys_cpu = (double) ts->sys_time * 100 / runt;
1574         } else {
1575                 usr_cpu = 0;
1576                 sys_cpu = 0;
1577         }
1578
1579         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1580                                                 (unsigned long long) ts->ctx,
1581                                                 (unsigned long long) ts->majf,
1582                                                 (unsigned long long) ts->minf);
1583
1584         /* Calc % distribution of IO depths, usecond, msecond latency */
1585         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1586         stat_calc_lat_nu(ts, io_u_lat_u);
1587         stat_calc_lat_m(ts, io_u_lat_m);
1588
1589         /* Only show fixed 7 I/O depth levels*/
1590         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1591                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1592                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1593
1594         /* Microsecond latency */
1595         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1596                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1597         /* Millisecond latency */
1598         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1599                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1600
1601         /* disk util stats, if any */
1602         if (ver >= 3 && is_running_backend())
1603                 show_disk_util(1, NULL, out);
1604
1605         /* Additional output if continue_on_error set - default off*/
1606         if (ts->continue_on_error)
1607                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1608
1609         /* Additional output if description is set */
1610         if (strlen(ts->description)) {
1611                 if (ver == 2)
1612                         log_buf(out, "\n");
1613                 log_buf(out, ";%s", ts->description);
1614         }
1615
1616         log_buf(out, "\n");
1617 }
1618
1619 static void json_add_job_opts(struct json_object *root, const char *name,
1620                               struct flist_head *opt_list)
1621 {
1622         struct json_object *dir_object;
1623         struct flist_head *entry;
1624         struct print_option *p;
1625
1626         if (flist_empty(opt_list))
1627                 return;
1628
1629         dir_object = json_create_object();
1630         json_object_add_value_object(root, name, dir_object);
1631
1632         flist_for_each(entry, opt_list) {
1633                 p = flist_entry(entry, struct print_option, list);
1634                 json_object_add_value_string(dir_object, p->name, p->value);
1635         }
1636 }
1637
1638 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1639                                                    struct group_run_stats *rs,
1640                                                    struct flist_head *opt_list)
1641 {
1642         struct json_object *root, *tmp;
1643         struct jobs_eta *je;
1644         double io_u_dist[FIO_IO_U_MAP_NR];
1645         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1646         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1647         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1648         double usr_cpu, sys_cpu;
1649         int i;
1650         size_t size;
1651
1652         root = json_create_object();
1653         json_object_add_value_string(root, "jobname", ts->name);
1654         json_object_add_value_int(root, "groupid", ts->groupid);
1655         json_object_add_value_int(root, "error", ts->error);
1656
1657         /* ETA Info */
1658         je = get_jobs_eta(true, &size);
1659         if (je) {
1660                 json_object_add_value_int(root, "eta", je->eta_sec);
1661                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1662         }
1663
1664         if (opt_list)
1665                 json_add_job_opts(root, "job options", opt_list);
1666
1667         add_ddir_status_json(ts, rs, DDIR_READ, root);
1668         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1669         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1670         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1671
1672         if (ts->unified_rw_rep == UNIFIED_BOTH)
1673                 add_mixed_ddir_status_json(ts, rs, root);
1674
1675         /* CPU Usage */
1676         if (ts->total_run_time) {
1677                 double runt = (double) ts->total_run_time;
1678
1679                 usr_cpu = (double) ts->usr_time * 100 / runt;
1680                 sys_cpu = (double) ts->sys_time * 100 / runt;
1681         } else {
1682                 usr_cpu = 0;
1683                 sys_cpu = 0;
1684         }
1685         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1686         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1687         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1688         json_object_add_value_int(root, "ctx", ts->ctx);
1689         json_object_add_value_int(root, "majf", ts->majf);
1690         json_object_add_value_int(root, "minf", ts->minf);
1691
1692         /* Calc % distribution of IO depths */
1693         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1694         tmp = json_create_object();
1695         json_object_add_value_object(root, "iodepth_level", tmp);
1696         /* Only show fixed 7 I/O depth levels*/
1697         for (i = 0; i < 7; i++) {
1698                 char name[20];
1699                 if (i < 6)
1700                         snprintf(name, 20, "%d", 1 << i);
1701                 else
1702                         snprintf(name, 20, ">=%d", 1 << i);
1703                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1704         }
1705
1706         /* Calc % distribution of submit IO depths */
1707         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1708         tmp = json_create_object();
1709         json_object_add_value_object(root, "iodepth_submit", tmp);
1710         /* Only show fixed 7 I/O depth levels*/
1711         for (i = 0; i < 7; i++) {
1712                 char name[20];
1713                 if (i == 0)
1714                         snprintf(name, 20, "0");
1715                 else if (i < 6)
1716                         snprintf(name, 20, "%d", 1 << (i+1));
1717                 else
1718                         snprintf(name, 20, ">=%d", 1 << i);
1719                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1720         }
1721
1722         /* Calc % distribution of completion IO depths */
1723         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1724         tmp = json_create_object();
1725         json_object_add_value_object(root, "iodepth_complete", tmp);
1726         /* Only show fixed 7 I/O depth levels*/
1727         for (i = 0; i < 7; i++) {
1728                 char name[20];
1729                 if (i == 0)
1730                         snprintf(name, 20, "0");
1731                 else if (i < 6)
1732                         snprintf(name, 20, "%d", 1 << (i+1));
1733                 else
1734                         snprintf(name, 20, ">=%d", 1 << i);
1735                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1736         }
1737
1738         /* Calc % distribution of nsecond, usecond, msecond latency */
1739         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1740         stat_calc_lat_n(ts, io_u_lat_n);
1741         stat_calc_lat_u(ts, io_u_lat_u);
1742         stat_calc_lat_m(ts, io_u_lat_m);
1743
1744         /* Nanosecond latency */
1745         tmp = json_create_object();
1746         json_object_add_value_object(root, "latency_ns", tmp);
1747         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1748                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1749                                  "250", "500", "750", "1000", };
1750                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1751         }
1752         /* Microsecond latency */
1753         tmp = json_create_object();
1754         json_object_add_value_object(root, "latency_us", tmp);
1755         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1756                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1757                                  "250", "500", "750", "1000", };
1758                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1759         }
1760         /* Millisecond latency */
1761         tmp = json_create_object();
1762         json_object_add_value_object(root, "latency_ms", tmp);
1763         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1764                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1765                                  "250", "500", "750", "1000", "2000",
1766                                  ">=2000", };
1767                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1768         }
1769
1770         /* Additional output if continue_on_error set - default off*/
1771         if (ts->continue_on_error) {
1772                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1773                 json_object_add_value_int(root, "first_error", ts->first_error);
1774         }
1775
1776         if (ts->latency_depth) {
1777                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1778                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1779                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1780                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1781         }
1782
1783         /* Additional output if description is set */
1784         if (strlen(ts->description))
1785                 json_object_add_value_string(root, "desc", ts->description);
1786
1787         if (ts->nr_block_infos) {
1788                 /* Block error histogram and types */
1789                 int len;
1790                 unsigned int *percentiles = NULL;
1791                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1792
1793                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1794                                              ts->percentile_list,
1795                                              &percentiles, block_state_counts);
1796
1797                 if (len) {
1798                         struct json_object *block, *percentile_object, *states;
1799                         int state;
1800                         block = json_create_object();
1801                         json_object_add_value_object(root, "block", block);
1802
1803                         percentile_object = json_create_object();
1804                         json_object_add_value_object(block, "percentiles",
1805                                                      percentile_object);
1806                         for (i = 0; i < len; i++) {
1807                                 char buf[20];
1808                                 snprintf(buf, sizeof(buf), "%f",
1809                                          ts->percentile_list[i].u.f);
1810                                 json_object_add_value_int(percentile_object,
1811                                                           buf,
1812                                                           percentiles[i]);
1813                         }
1814
1815                         states = json_create_object();
1816                         json_object_add_value_object(block, "states", states);
1817                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1818                                 json_object_add_value_int(states,
1819                                         block_state_names[state],
1820                                         block_state_counts[state]);
1821                         }
1822                         free(percentiles);
1823                 }
1824         }
1825
1826         if (ts->ss_dur) {
1827                 struct json_object *data;
1828                 struct json_array *iops, *bw;
1829                 int j, k, l;
1830                 char ss_buf[64];
1831
1832                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1833                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1834                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1835                         (float) ts->ss_limit.u.f,
1836                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1837
1838                 tmp = json_create_object();
1839                 json_object_add_value_object(root, "steadystate", tmp);
1840                 json_object_add_value_string(tmp, "ss", ss_buf);
1841                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1842                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1843
1844                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1845                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1846                 json_object_add_value_string(tmp, "criterion", ss_buf);
1847                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1848                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1849
1850                 data = json_create_object();
1851                 json_object_add_value_object(tmp, "data", data);
1852                 bw = json_create_array();
1853                 iops = json_create_array();
1854
1855                 /*
1856                 ** if ss was attained or the buffer is not full,
1857                 ** ss->head points to the first element in the list.
1858                 ** otherwise it actually points to the second element
1859                 ** in the list
1860                 */
1861                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1862                         j = ts->ss_head;
1863                 else
1864                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1865                 for (l = 0; l < ts->ss_dur; l++) {
1866                         k = (j + l) % ts->ss_dur;
1867                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1868                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1869                 }
1870                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1871                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1872                 json_object_add_value_array(data, "iops", iops);
1873                 json_object_add_value_array(data, "bw", bw);
1874         }
1875
1876         return root;
1877 }
1878
1879 static void show_thread_status_terse(struct thread_stat *ts,
1880                                      struct group_run_stats *rs,
1881                                      struct buf_output *out)
1882 {
1883         if (terse_version >= 2 && terse_version <= 5)
1884                 show_thread_status_terse_all(ts, rs, terse_version, out);
1885         else
1886                 log_err("fio: bad terse version!? %d\n", terse_version);
1887 }
1888
1889 struct json_object *show_thread_status(struct thread_stat *ts,
1890                                        struct group_run_stats *rs,
1891                                        struct flist_head *opt_list,
1892                                        struct buf_output *out)
1893 {
1894         struct json_object *ret = NULL;
1895
1896         if (output_format & FIO_OUTPUT_TERSE)
1897                 show_thread_status_terse(ts, rs,  out);
1898         if (output_format & FIO_OUTPUT_JSON)
1899                 ret = show_thread_status_json(ts, rs, opt_list);
1900         if (output_format & FIO_OUTPUT_NORMAL)
1901                 show_thread_status_normal(ts, rs,  out);
1902
1903         return ret;
1904 }
1905
1906 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1907 {
1908         double mean, S;
1909
1910         dst->min_val = min(dst->min_val, src->min_val);
1911         dst->max_val = max(dst->max_val, src->max_val);
1912
1913         /*
1914          * Compute new mean and S after the merge
1915          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1916          *  #Parallel_algorithm>
1917          */
1918         if (first) {
1919                 mean = src->mean.u.f;
1920                 S = src->S.u.f;
1921         } else {
1922                 double delta = src->mean.u.f - dst->mean.u.f;
1923
1924                 mean = ((src->mean.u.f * src->samples) +
1925                         (dst->mean.u.f * dst->samples)) /
1926                         (dst->samples + src->samples);
1927
1928                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1929                         (dst->samples * src->samples) /
1930                         (dst->samples + src->samples);
1931         }
1932
1933         dst->samples += src->samples;
1934         dst->mean.u.f = mean;
1935         dst->S.u.f = S;
1936
1937 }
1938
1939 /*
1940  * We sum two kinds of stats - one that is time based, in which case we
1941  * apply the proper summing technique, and then one that is iops/bw
1942  * numbers. For group_reporting, we should just add those up, not make
1943  * them the mean of everything.
1944  */
1945 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
1946 {
1947         bool first = dst->samples == 0;
1948
1949         if (src->samples == 0)
1950                 return;
1951
1952         if (!pure_sum) {
1953                 __sum_stat(dst, src, first);
1954                 return;
1955         }
1956
1957         if (first) {
1958                 dst->min_val = src->min_val;
1959                 dst->max_val = src->max_val;
1960                 dst->samples = src->samples;
1961                 dst->mean.u.f = src->mean.u.f;
1962                 dst->S.u.f = src->S.u.f;
1963         } else {
1964                 dst->min_val += src->min_val;
1965                 dst->max_val += src->max_val;
1966                 dst->samples += src->samples;
1967                 dst->mean.u.f += src->mean.u.f;
1968                 dst->S.u.f += src->S.u.f;
1969         }
1970 }
1971
1972 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1973 {
1974         int i;
1975
1976         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1977                 if (dst->max_run[i] < src->max_run[i])
1978                         dst->max_run[i] = src->max_run[i];
1979                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1980                         dst->min_run[i] = src->min_run[i];
1981                 if (dst->max_bw[i] < src->max_bw[i])
1982                         dst->max_bw[i] = src->max_bw[i];
1983                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1984                         dst->min_bw[i] = src->min_bw[i];
1985
1986                 dst->iobytes[i] += src->iobytes[i];
1987                 dst->agg[i] += src->agg[i];
1988         }
1989
1990         if (!dst->kb_base)
1991                 dst->kb_base = src->kb_base;
1992         if (!dst->unit_base)
1993                 dst->unit_base = src->unit_base;
1994         if (!dst->sig_figs)
1995                 dst->sig_figs = src->sig_figs;
1996 }
1997
1998 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
1999 {
2000         int k, l, m;
2001
2002         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2003                 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2004                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2005                         sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], false);
2006                         sum_stat(&dst->clat_low_prio_stat[l], &src->clat_low_prio_stat[l], false);
2007                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2008                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2009                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2010                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2011
2012                         dst->io_bytes[l] += src->io_bytes[l];
2013
2014                         if (dst->runtime[l] < src->runtime[l])
2015                                 dst->runtime[l] = src->runtime[l];
2016                 } else {
2017                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2018                         sum_stat(&dst->clat_high_prio_stat[0], &src->clat_high_prio_stat[l], false);
2019                         sum_stat(&dst->clat_low_prio_stat[0], &src->clat_low_prio_stat[l], false);
2020                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2021                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2022                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2023                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2024
2025                         dst->io_bytes[0] += src->io_bytes[l];
2026
2027                         if (dst->runtime[0] < src->runtime[l])
2028                                 dst->runtime[0] = src->runtime[l];
2029                 }
2030         }
2031
2032         sum_stat(&dst->sync_stat, &src->sync_stat, false);
2033         dst->usr_time += src->usr_time;
2034         dst->sys_time += src->sys_time;
2035         dst->ctx += src->ctx;
2036         dst->majf += src->majf;
2037         dst->minf += src->minf;
2038
2039         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2040                 dst->io_u_map[k] += src->io_u_map[k];
2041                 dst->io_u_submit[k] += src->io_u_submit[k];
2042                 dst->io_u_complete[k] += src->io_u_complete[k];
2043         }
2044
2045         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2046                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2047         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2048                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2049         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2050                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2051
2052         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2053                 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2054                         dst->total_io_u[k] += src->total_io_u[k];
2055                         dst->short_io_u[k] += src->short_io_u[k];
2056                         dst->drop_io_u[k] += src->drop_io_u[k];
2057                 } else {
2058                         dst->total_io_u[0] += src->total_io_u[k];
2059                         dst->short_io_u[0] += src->short_io_u[k];
2060                         dst->drop_io_u[0] += src->drop_io_u[k];
2061                 }
2062         }
2063
2064         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2065
2066         for (k = 0; k < FIO_LAT_CNT; k++)
2067                 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2068                         for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2069                                 if (dst->unified_rw_rep != UNIFIED_MIXED)
2070                                         dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2071                                 else
2072                                         dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2073
2074         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2075                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2076
2077         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2078                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
2079                         if (dst->unified_rw_rep != UNIFIED_MIXED) {
2080                                 dst->io_u_plat_high_prio[k][m] += src->io_u_plat_high_prio[k][m];
2081                                 dst->io_u_plat_low_prio[k][m] += src->io_u_plat_low_prio[k][m];
2082                         } else {
2083                                 dst->io_u_plat_high_prio[0][m] += src->io_u_plat_high_prio[k][m];
2084                                 dst->io_u_plat_low_prio[0][m] += src->io_u_plat_low_prio[k][m];
2085                         }
2086
2087                 }
2088         }
2089
2090         dst->total_run_time += src->total_run_time;
2091         dst->total_submit += src->total_submit;
2092         dst->total_complete += src->total_complete;
2093         dst->nr_zone_resets += src->nr_zone_resets;
2094         dst->cachehit += src->cachehit;
2095         dst->cachemiss += src->cachemiss;
2096 }
2097
2098 void init_group_run_stat(struct group_run_stats *gs)
2099 {
2100         int i;
2101         memset(gs, 0, sizeof(*gs));
2102
2103         for (i = 0; i < DDIR_RWDIR_CNT; i++)
2104                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2105 }
2106
2107 void init_thread_stat_min_vals(struct thread_stat *ts)
2108 {
2109         int i;
2110
2111         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2112                 ts->clat_stat[i].min_val = ULONG_MAX;
2113                 ts->slat_stat[i].min_val = ULONG_MAX;
2114                 ts->lat_stat[i].min_val = ULONG_MAX;
2115                 ts->bw_stat[i].min_val = ULONG_MAX;
2116                 ts->iops_stat[i].min_val = ULONG_MAX;
2117                 ts->clat_high_prio_stat[i].min_val = ULONG_MAX;
2118                 ts->clat_low_prio_stat[i].min_val = ULONG_MAX;
2119         }
2120         ts->sync_stat.min_val = ULONG_MAX;
2121 }
2122
2123 void init_thread_stat(struct thread_stat *ts)
2124 {
2125         memset(ts, 0, sizeof(*ts));
2126
2127         init_thread_stat_min_vals(ts);
2128         ts->groupid = -1;
2129 }
2130
2131 void __show_run_stats(void)
2132 {
2133         struct group_run_stats *runstats, *rs;
2134         struct thread_data *td;
2135         struct thread_stat *threadstats, *ts;
2136         int i, j, k, nr_ts, last_ts, idx;
2137         bool kb_base_warned = false;
2138         bool unit_base_warned = false;
2139         struct json_object *root = NULL;
2140         struct json_array *array = NULL;
2141         struct buf_output output[FIO_OUTPUT_NR];
2142         struct flist_head **opt_lists;
2143
2144         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2145
2146         for (i = 0; i < groupid + 1; i++)
2147                 init_group_run_stat(&runstats[i]);
2148
2149         /*
2150          * find out how many threads stats we need. if group reporting isn't
2151          * enabled, it's one-per-td.
2152          */
2153         nr_ts = 0;
2154         last_ts = -1;
2155         for_each_td(td, i) {
2156                 if (!td->o.group_reporting) {
2157                         nr_ts++;
2158                         continue;
2159                 }
2160                 if (last_ts == td->groupid)
2161                         continue;
2162                 if (!td->o.stats)
2163                         continue;
2164
2165                 last_ts = td->groupid;
2166                 nr_ts++;
2167         }
2168
2169         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2170         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2171
2172         for (i = 0; i < nr_ts; i++) {
2173                 init_thread_stat(&threadstats[i]);
2174                 opt_lists[i] = NULL;
2175         }
2176
2177         j = 0;
2178         last_ts = -1;
2179         idx = 0;
2180         for_each_td(td, i) {
2181                 if (!td->o.stats)
2182                         continue;
2183                 if (idx && (!td->o.group_reporting ||
2184                     (td->o.group_reporting && last_ts != td->groupid))) {
2185                         idx = 0;
2186                         j++;
2187                 }
2188
2189                 last_ts = td->groupid;
2190
2191                 ts = &threadstats[j];
2192
2193                 ts->clat_percentiles = td->o.clat_percentiles;
2194                 ts->lat_percentiles = td->o.lat_percentiles;
2195                 ts->slat_percentiles = td->o.slat_percentiles;
2196                 ts->percentile_precision = td->o.percentile_precision;
2197                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2198                 opt_lists[j] = &td->opt_list;
2199
2200                 idx++;
2201                 ts->members++;
2202
2203                 if (ts->groupid == -1) {
2204                         /*
2205                          * These are per-group shared already
2206                          */
2207                         snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2208                         if (td->o.description)
2209                                 snprintf(ts->description,
2210                                          sizeof(ts->description), "%s",
2211                                          td->o.description);
2212                         else
2213                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2214
2215                         /*
2216                          * If multiple entries in this group, this is
2217                          * the first member.
2218                          */
2219                         ts->thread_number = td->thread_number;
2220                         ts->groupid = td->groupid;
2221
2222                         /*
2223                          * first pid in group, not very useful...
2224                          */
2225                         ts->pid = td->pid;
2226
2227                         ts->kb_base = td->o.kb_base;
2228                         ts->unit_base = td->o.unit_base;
2229                         ts->sig_figs = td->o.sig_figs;
2230                         ts->unified_rw_rep = td->o.unified_rw_rep;
2231                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2232                         log_info("fio: kb_base differs for jobs in group, using"
2233                                  " %u as the base\n", ts->kb_base);
2234                         kb_base_warned = true;
2235                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2236                         log_info("fio: unit_base differs for jobs in group, using"
2237                                  " %u as the base\n", ts->unit_base);
2238                         unit_base_warned = true;
2239                 }
2240
2241                 ts->continue_on_error = td->o.continue_on_error;
2242                 ts->total_err_count += td->total_err_count;
2243                 ts->first_error = td->first_error;
2244                 if (!ts->error) {
2245                         if (!td->error && td->o.continue_on_error &&
2246                             td->first_error) {
2247                                 ts->error = td->first_error;
2248                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2249                                          td->verror);
2250                         } else  if (td->error) {
2251                                 ts->error = td->error;
2252                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2253                                          td->verror);
2254                         }
2255                 }
2256
2257                 ts->latency_depth = td->latency_qd;
2258                 ts->latency_target = td->o.latency_target;
2259                 ts->latency_percentile = td->o.latency_percentile;
2260                 ts->latency_window = td->o.latency_window;
2261
2262                 ts->nr_block_infos = td->ts.nr_block_infos;
2263                 for (k = 0; k < ts->nr_block_infos; k++)
2264                         ts->block_infos[k] = td->ts.block_infos[k];
2265
2266                 sum_thread_stats(ts, &td->ts);
2267
2268                 if (td->o.ss_dur) {
2269                         ts->ss_state = td->ss.state;
2270                         ts->ss_dur = td->ss.dur;
2271                         ts->ss_head = td->ss.head;
2272                         ts->ss_bw_data = td->ss.bw_data;
2273                         ts->ss_iops_data = td->ss.iops_data;
2274                         ts->ss_limit.u.f = td->ss.limit;
2275                         ts->ss_slope.u.f = td->ss.slope;
2276                         ts->ss_deviation.u.f = td->ss.deviation;
2277                         ts->ss_criterion.u.f = td->ss.criterion;
2278                 }
2279                 else
2280                         ts->ss_dur = ts->ss_state = 0;
2281         }
2282
2283         for (i = 0; i < nr_ts; i++) {
2284                 unsigned long long bw;
2285
2286                 ts = &threadstats[i];
2287                 if (ts->groupid == -1)
2288                         continue;
2289                 rs = &runstats[ts->groupid];
2290                 rs->kb_base = ts->kb_base;
2291                 rs->unit_base = ts->unit_base;
2292                 rs->sig_figs = ts->sig_figs;
2293                 rs->unified_rw_rep |= ts->unified_rw_rep;
2294
2295                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2296                         if (!ts->runtime[j])
2297                                 continue;
2298                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2299                                 rs->min_run[j] = ts->runtime[j];
2300                         if (ts->runtime[j] > rs->max_run[j])
2301                                 rs->max_run[j] = ts->runtime[j];
2302
2303                         bw = 0;
2304                         if (ts->runtime[j])
2305                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2306                         if (bw < rs->min_bw[j])
2307                                 rs->min_bw[j] = bw;
2308                         if (bw > rs->max_bw[j])
2309                                 rs->max_bw[j] = bw;
2310
2311                         rs->iobytes[j] += ts->io_bytes[j];
2312                 }
2313         }
2314
2315         for (i = 0; i < groupid + 1; i++) {
2316                 int ddir;
2317
2318                 rs = &runstats[i];
2319
2320                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2321                         if (rs->max_run[ddir])
2322                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2323                                                 rs->max_run[ddir];
2324                 }
2325         }
2326
2327         for (i = 0; i < FIO_OUTPUT_NR; i++)
2328                 buf_output_init(&output[i]);
2329
2330         /*
2331          * don't overwrite last signal output
2332          */
2333         if (output_format & FIO_OUTPUT_NORMAL)
2334                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2335         if (output_format & FIO_OUTPUT_JSON) {
2336                 struct thread_data *global;
2337                 char time_buf[32];
2338                 struct timeval now;
2339                 unsigned long long ms_since_epoch;
2340                 time_t tv_sec;
2341
2342                 gettimeofday(&now, NULL);
2343                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2344                                  (unsigned long long)(now.tv_usec) / 1000;
2345
2346                 tv_sec = now.tv_sec;
2347                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2348                 if (time_buf[strlen(time_buf) - 1] == '\n')
2349                         time_buf[strlen(time_buf) - 1] = '\0';
2350
2351                 root = json_create_object();
2352                 json_object_add_value_string(root, "fio version", fio_version_string);
2353                 json_object_add_value_int(root, "timestamp", now.tv_sec);
2354                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2355                 json_object_add_value_string(root, "time", time_buf);
2356                 global = get_global_options();
2357                 json_add_job_opts(root, "global options", &global->opt_list);
2358                 array = json_create_array();
2359                 json_object_add_value_array(root, "jobs", array);
2360         }
2361
2362         if (is_backend)
2363                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2364
2365         for (i = 0; i < nr_ts; i++) {
2366                 ts = &threadstats[i];
2367                 rs = &runstats[ts->groupid];
2368
2369                 if (is_backend) {
2370                         fio_server_send_job_options(opt_lists[i], i);
2371                         fio_server_send_ts(ts, rs);
2372                 } else {
2373                         if (output_format & FIO_OUTPUT_TERSE)
2374                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2375                         if (output_format & FIO_OUTPUT_JSON) {
2376                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2377                                 json_array_add_value_object(array, tmp);
2378                         }
2379                         if (output_format & FIO_OUTPUT_NORMAL)
2380                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2381                 }
2382         }
2383         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2384                 /* disk util stats, if any */
2385                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2386
2387                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2388
2389                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2390                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2391                 json_free_object(root);
2392         }
2393
2394         for (i = 0; i < groupid + 1; i++) {
2395                 rs = &runstats[i];
2396
2397                 rs->groupid = i;
2398                 if (is_backend)
2399                         fio_server_send_gs(rs);
2400                 else if (output_format & FIO_OUTPUT_NORMAL)
2401                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2402         }
2403
2404         if (is_backend)
2405                 fio_server_send_du();
2406         else if (output_format & FIO_OUTPUT_NORMAL) {
2407                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2408                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2409         }
2410
2411         for (i = 0; i < FIO_OUTPUT_NR; i++) {
2412                 struct buf_output *out = &output[i];
2413
2414                 log_info_buf(out->buf, out->buflen);
2415                 buf_output_free(out);
2416         }
2417
2418         fio_idle_prof_cleanup();
2419
2420         log_info_flush();
2421         free(runstats);
2422         free(threadstats);
2423         free(opt_lists);
2424 }
2425
2426 int __show_running_run_stats(void)
2427 {
2428         struct thread_data *td;
2429         unsigned long long *rt;
2430         struct timespec ts;
2431         int i;
2432
2433         fio_sem_down(stat_sem);
2434
2435         rt = malloc(thread_number * sizeof(unsigned long long));
2436         fio_gettime(&ts, NULL);
2437
2438         for_each_td(td, i) {
2439                 td->update_rusage = 1;
2440                 for_each_rw_ddir(ddir) {
2441                         td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2442                 }
2443                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2444
2445                 rt[i] = mtime_since(&td->start, &ts);
2446                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2447                         td->ts.runtime[DDIR_READ] += rt[i];
2448                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2449                         td->ts.runtime[DDIR_WRITE] += rt[i];
2450                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2451                         td->ts.runtime[DDIR_TRIM] += rt[i];
2452         }
2453
2454         for_each_td(td, i) {
2455                 if (td->runstate >= TD_EXITED)
2456                         continue;
2457                 if (td->rusage_sem) {
2458                         td->update_rusage = 1;
2459                         fio_sem_down(td->rusage_sem);
2460                 }
2461                 td->update_rusage = 0;
2462         }
2463
2464         __show_run_stats();
2465
2466         for_each_td(td, i) {
2467                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2468                         td->ts.runtime[DDIR_READ] -= rt[i];
2469                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2470                         td->ts.runtime[DDIR_WRITE] -= rt[i];
2471                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2472                         td->ts.runtime[DDIR_TRIM] -= rt[i];
2473         }
2474
2475         free(rt);
2476         fio_sem_up(stat_sem);
2477
2478         return 0;
2479 }
2480
2481 static bool status_file_disabled;
2482
2483 #define FIO_STATUS_FILE         "fio-dump-status"
2484
2485 static int check_status_file(void)
2486 {
2487         struct stat sb;
2488         const char *temp_dir;
2489         char fio_status_file_path[PATH_MAX];
2490
2491         if (status_file_disabled)
2492                 return 0;
2493
2494         temp_dir = getenv("TMPDIR");
2495         if (temp_dir == NULL) {
2496                 temp_dir = getenv("TEMP");
2497                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2498                         temp_dir = NULL;
2499         }
2500         if (temp_dir == NULL)
2501                 temp_dir = "/tmp";
2502 #ifdef __COVERITY__
2503         __coverity_tainted_data_sanitize__(temp_dir);
2504 #endif
2505
2506         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2507
2508         if (stat(fio_status_file_path, &sb))
2509                 return 0;
2510
2511         if (unlink(fio_status_file_path) < 0) {
2512                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2513                                                         strerror(errno));
2514                 log_err("fio: disabling status file updates\n");
2515                 status_file_disabled = true;
2516         }
2517
2518         return 1;
2519 }
2520
2521 void check_for_running_stats(void)
2522 {
2523         if (check_status_file()) {
2524                 show_running_run_stats();
2525                 return;
2526         }
2527 }
2528
2529 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2530 {
2531         double val = data;
2532         double delta;
2533
2534         if (data > is->max_val)
2535                 is->max_val = data;
2536         if (data < is->min_val)
2537                 is->min_val = data;
2538
2539         delta = val - is->mean.u.f;
2540         if (delta) {
2541                 is->mean.u.f += delta / (is->samples + 1.0);
2542                 is->S.u.f += delta * (val - is->mean.u.f);
2543         }
2544
2545         is->samples++;
2546 }
2547
2548 /*
2549  * Return a struct io_logs, which is added to the tail of the log
2550  * list for 'iolog'.
2551  */
2552 static struct io_logs *get_new_log(struct io_log *iolog)
2553 {
2554         size_t new_samples;
2555         struct io_logs *cur_log;
2556
2557         /*
2558          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2559          * forever
2560          */
2561         if (!iolog->cur_log_max) {
2562                 new_samples = iolog->td->o.log_entries;
2563         } else {
2564                 new_samples = iolog->cur_log_max * 2;
2565                 if (new_samples > MAX_LOG_ENTRIES)
2566                         new_samples = MAX_LOG_ENTRIES;
2567         }
2568
2569         cur_log = smalloc(sizeof(*cur_log));
2570         if (cur_log) {
2571                 INIT_FLIST_HEAD(&cur_log->list);
2572                 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2573                 if (cur_log->log) {
2574                         cur_log->nr_samples = 0;
2575                         cur_log->max_samples = new_samples;
2576                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2577                         iolog->cur_log_max = new_samples;
2578                         return cur_log;
2579                 }
2580                 sfree(cur_log);
2581         }
2582
2583         return NULL;
2584 }
2585
2586 /*
2587  * Add and return a new log chunk, or return current log if big enough
2588  */
2589 static struct io_logs *regrow_log(struct io_log *iolog)
2590 {
2591         struct io_logs *cur_log;
2592         int i;
2593
2594         if (!iolog || iolog->disabled)
2595                 goto disable;
2596
2597         cur_log = iolog_cur_log(iolog);
2598         if (!cur_log) {
2599                 cur_log = get_new_log(iolog);
2600                 if (!cur_log)
2601                         return NULL;
2602         }
2603
2604         if (cur_log->nr_samples < cur_log->max_samples)
2605                 return cur_log;
2606
2607         /*
2608          * No room for a new sample. If we're compressing on the fly, flush
2609          * out the current chunk
2610          */
2611         if (iolog->log_gz) {
2612                 if (iolog_cur_flush(iolog, cur_log)) {
2613                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2614                         return NULL;
2615                 }
2616         }
2617
2618         /*
2619          * Get a new log array, and add to our list
2620          */
2621         cur_log = get_new_log(iolog);
2622         if (!cur_log) {
2623                 log_err("fio: failed extending iolog! Will stop logging.\n");
2624                 return NULL;
2625         }
2626
2627         if (!iolog->pending || !iolog->pending->nr_samples)
2628                 return cur_log;
2629
2630         /*
2631          * Flush pending items to new log
2632          */
2633         for (i = 0; i < iolog->pending->nr_samples; i++) {
2634                 struct io_sample *src, *dst;
2635
2636                 src = get_sample(iolog, iolog->pending, i);
2637                 dst = get_sample(iolog, cur_log, i);
2638                 memcpy(dst, src, log_entry_sz(iolog));
2639         }
2640         cur_log->nr_samples = iolog->pending->nr_samples;
2641
2642         iolog->pending->nr_samples = 0;
2643         return cur_log;
2644 disable:
2645         if (iolog)
2646                 iolog->disabled = true;
2647         return NULL;
2648 }
2649
2650 void regrow_logs(struct thread_data *td)
2651 {
2652         regrow_log(td->slat_log);
2653         regrow_log(td->clat_log);
2654         regrow_log(td->clat_hist_log);
2655         regrow_log(td->lat_log);
2656         regrow_log(td->bw_log);
2657         regrow_log(td->iops_log);
2658         td->flags &= ~TD_F_REGROW_LOGS;
2659 }
2660
2661 void regrow_agg_logs(void)
2662 {
2663         enum fio_ddir ddir;
2664
2665         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2666                 regrow_log(agg_io_log[ddir]);
2667 }
2668
2669 static struct io_logs *get_cur_log(struct io_log *iolog)
2670 {
2671         struct io_logs *cur_log;
2672
2673         cur_log = iolog_cur_log(iolog);
2674         if (!cur_log) {
2675                 cur_log = get_new_log(iolog);
2676                 if (!cur_log)
2677                         return NULL;
2678         }
2679
2680         if (cur_log->nr_samples < cur_log->max_samples)
2681                 return cur_log;
2682
2683         /*
2684          * Out of space. If we're in IO offload mode, or we're not doing
2685          * per unit logging (hence logging happens outside of the IO thread
2686          * as well), add a new log chunk inline. If we're doing inline
2687          * submissions, flag 'td' as needing a log regrow and we'll take
2688          * care of it on the submission side.
2689          */
2690         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2691             !per_unit_log(iolog))
2692                 return regrow_log(iolog);
2693
2694         if (iolog->td)
2695                 iolog->td->flags |= TD_F_REGROW_LOGS;
2696         if (iolog->pending)
2697                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2698         return iolog->pending;
2699 }
2700
2701 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2702                              enum fio_ddir ddir, unsigned long long bs,
2703                              unsigned long t, uint64_t offset,
2704                              unsigned int priority)
2705 {
2706         struct io_logs *cur_log;
2707
2708         if (iolog->disabled)
2709                 return;
2710         if (flist_empty(&iolog->io_logs))
2711                 iolog->avg_last[ddir] = t;
2712
2713         cur_log = get_cur_log(iolog);
2714         if (cur_log) {
2715                 struct io_sample *s;
2716
2717                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2718
2719                 s->data = data;
2720                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2721                 io_sample_set_ddir(iolog, s, ddir);
2722                 s->bs = bs;
2723                 s->priority = priority;
2724
2725                 if (iolog->log_offset) {
2726                         struct io_sample_offset *so = (void *) s;
2727
2728                         so->offset = offset;
2729                 }
2730
2731                 cur_log->nr_samples++;
2732                 return;
2733         }
2734
2735         iolog->disabled = true;
2736 }
2737
2738 static inline void reset_io_stat(struct io_stat *ios)
2739 {
2740         ios->min_val = -1ULL;
2741         ios->max_val = ios->samples = 0;
2742         ios->mean.u.f = ios->S.u.f = 0;
2743 }
2744
2745 void reset_io_stats(struct thread_data *td)
2746 {
2747         struct thread_stat *ts = &td->ts;
2748         int i, j, k;
2749
2750         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2751                 reset_io_stat(&ts->clat_high_prio_stat[i]);
2752                 reset_io_stat(&ts->clat_low_prio_stat[i]);
2753                 reset_io_stat(&ts->clat_stat[i]);
2754                 reset_io_stat(&ts->slat_stat[i]);
2755                 reset_io_stat(&ts->lat_stat[i]);
2756                 reset_io_stat(&ts->bw_stat[i]);
2757                 reset_io_stat(&ts->iops_stat[i]);
2758
2759                 ts->io_bytes[i] = 0;
2760                 ts->runtime[i] = 0;
2761                 ts->total_io_u[i] = 0;
2762                 ts->short_io_u[i] = 0;
2763                 ts->drop_io_u[i] = 0;
2764
2765                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2766                         ts->io_u_plat_high_prio[i][j] = 0;
2767                         ts->io_u_plat_low_prio[i][j] = 0;
2768                         if (!i)
2769                                 ts->io_u_sync_plat[j] = 0;
2770                 }
2771         }
2772
2773         for (i = 0; i < FIO_LAT_CNT; i++)
2774                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
2775                         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2776                                 ts->io_u_plat[i][j][k] = 0;
2777
2778         ts->total_io_u[DDIR_SYNC] = 0;
2779
2780         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2781                 ts->io_u_map[i] = 0;
2782                 ts->io_u_submit[i] = 0;
2783                 ts->io_u_complete[i] = 0;
2784         }
2785
2786         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2787                 ts->io_u_lat_n[i] = 0;
2788         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2789                 ts->io_u_lat_u[i] = 0;
2790         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2791                 ts->io_u_lat_m[i] = 0;
2792
2793         ts->total_submit = 0;
2794         ts->total_complete = 0;
2795         ts->nr_zone_resets = 0;
2796         ts->cachehit = ts->cachemiss = 0;
2797 }
2798
2799 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2800                               unsigned long elapsed, bool log_max)
2801 {
2802         /*
2803          * Note an entry in the log. Use the mean from the logged samples,
2804          * making sure to properly round up. Only write a log entry if we
2805          * had actual samples done.
2806          */
2807         if (iolog->avg_window[ddir].samples) {
2808                 union io_sample_data data;
2809
2810                 if (log_max)
2811                         data.val = iolog->avg_window[ddir].max_val;
2812                 else
2813                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2814
2815                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
2816         }
2817
2818         reset_io_stat(&iolog->avg_window[ddir]);
2819 }
2820
2821 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2822                              bool log_max)
2823 {
2824         int ddir;
2825
2826         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2827                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2828 }
2829
2830 static unsigned long add_log_sample(struct thread_data *td,
2831                                     struct io_log *iolog,
2832                                     union io_sample_data data,
2833                                     enum fio_ddir ddir, unsigned long long bs,
2834                                     uint64_t offset, unsigned int ioprio)
2835 {
2836         unsigned long elapsed, this_window;
2837
2838         if (!ddir_rw(ddir))
2839                 return 0;
2840
2841         elapsed = mtime_since_now(&td->epoch);
2842
2843         /*
2844          * If no time averaging, just add the log sample.
2845          */
2846         if (!iolog->avg_msec) {
2847                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
2848                                  ioprio);
2849                 return 0;
2850         }
2851
2852         /*
2853          * Add the sample. If the time period has passed, then
2854          * add that entry to the log and clear.
2855          */
2856         add_stat_sample(&iolog->avg_window[ddir], data.val);
2857
2858         /*
2859          * If period hasn't passed, adding the above sample is all we
2860          * need to do.
2861          */
2862         this_window = elapsed - iolog->avg_last[ddir];
2863         if (elapsed < iolog->avg_last[ddir])
2864                 return iolog->avg_last[ddir] - elapsed;
2865         else if (this_window < iolog->avg_msec) {
2866                 unsigned long diff = iolog->avg_msec - this_window;
2867
2868                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2869                         return diff;
2870         }
2871
2872         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2873
2874         iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
2875
2876         return iolog->avg_msec;
2877 }
2878
2879 void finalize_logs(struct thread_data *td, bool unit_logs)
2880 {
2881         unsigned long elapsed;
2882
2883         elapsed = mtime_since_now(&td->epoch);
2884
2885         if (td->clat_log && unit_logs)
2886                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2887         if (td->slat_log && unit_logs)
2888                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2889         if (td->lat_log && unit_logs)
2890                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2891         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2892                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2893         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2894                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2895 }
2896
2897 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
2898                     unsigned long long bs)
2899 {
2900         struct io_log *iolog;
2901
2902         if (!ddir_rw(ddir))
2903                 return;
2904
2905         iolog = agg_io_log[ddir];
2906         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
2907 }
2908
2909 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2910 {
2911         unsigned int idx = plat_val_to_idx(nsec);
2912         assert(idx < FIO_IO_U_PLAT_NR);
2913
2914         ts->io_u_sync_plat[idx]++;
2915         add_stat_sample(&ts->sync_stat, nsec);
2916 }
2917
2918 static inline void add_lat_percentile_sample(struct thread_stat *ts,
2919                                              unsigned long long nsec,
2920                                              enum fio_ddir ddir,
2921                                              enum fio_lat lat)
2922 {
2923         unsigned int idx = plat_val_to_idx(nsec);
2924         assert(idx < FIO_IO_U_PLAT_NR);
2925
2926         ts->io_u_plat[lat][ddir][idx]++;
2927 }
2928
2929 static inline void add_lat_percentile_prio_sample(struct thread_stat *ts,
2930                                                   unsigned long long nsec,
2931                                                   enum fio_ddir ddir,
2932                                                   bool high_prio)
2933 {
2934         unsigned int idx = plat_val_to_idx(nsec);
2935
2936         if (!high_prio)
2937                 ts->io_u_plat_low_prio[ddir][idx]++;
2938         else
2939                 ts->io_u_plat_high_prio[ddir][idx]++;
2940 }
2941
2942 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2943                      unsigned long long nsec, unsigned long long bs,
2944                      uint64_t offset, unsigned int ioprio, bool high_prio)
2945 {
2946         const bool needs_lock = td_async_processing(td);
2947         unsigned long elapsed, this_window;
2948         struct thread_stat *ts = &td->ts;
2949         struct io_log *iolog = td->clat_hist_log;
2950
2951         if (needs_lock)
2952                 __td_io_u_lock(td);
2953
2954         add_stat_sample(&ts->clat_stat[ddir], nsec);
2955
2956         /*
2957          * When lat_percentiles=1 (default 0), the reported high/low priority
2958          * percentiles and stats are used for describing total latency values,
2959          * even though the variable names themselves start with clat_.
2960          *
2961          * Because of the above definition, add a prio stat sample only when
2962          * lat_percentiles=0. add_lat_sample() will add the prio stat sample
2963          * when lat_percentiles=1.
2964          */
2965         if (!ts->lat_percentiles) {
2966                 if (high_prio)
2967                         add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
2968                 else
2969                         add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
2970         }
2971
2972         if (td->clat_log)
2973                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2974                                offset, ioprio);
2975
2976         if (ts->clat_percentiles) {
2977                 /*
2978                  * Because of the above definition, add a prio lat percentile
2979                  * sample only when lat_percentiles=0. add_lat_sample() will add
2980                  * the prio lat percentile sample when lat_percentiles=1.
2981                  */
2982                 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
2983                 if (!ts->lat_percentiles)
2984                         add_lat_percentile_prio_sample(ts, nsec, ddir,
2985                                                        high_prio);
2986         }
2987
2988         if (iolog && iolog->hist_msec) {
2989                 struct io_hist *hw = &iolog->hist_window[ddir];
2990
2991                 hw->samples++;
2992                 elapsed = mtime_since_now(&td->epoch);
2993                 if (!hw->hist_last)
2994                         hw->hist_last = elapsed;
2995                 this_window = elapsed - hw->hist_last;
2996
2997                 if (this_window >= iolog->hist_msec) {
2998                         uint64_t *io_u_plat;
2999                         struct io_u_plat_entry *dst;
3000
3001                         /*
3002                          * Make a byte-for-byte copy of the latency histogram
3003                          * stored in td->ts.io_u_plat[ddir], recording it in a
3004                          * log sample. Note that the matching call to free() is
3005                          * located in iolog.c after printing this sample to the
3006                          * log file.
3007                          */
3008                         io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3009                         dst = malloc(sizeof(struct io_u_plat_entry));
3010                         memcpy(&(dst->io_u_plat), io_u_plat,
3011                                 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3012                         flist_add(&dst->list, &hw->list);
3013                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3014                                          elapsed, offset, ioprio);
3015
3016                         /*
3017                          * Update the last time we recorded as being now, minus
3018                          * any drift in time we encountered before actually
3019                          * making the record.
3020                          */
3021                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3022                         hw->samples = 0;
3023                 }
3024         }
3025
3026         if (needs_lock)
3027                 __td_io_u_unlock(td);
3028 }
3029
3030 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3031                      unsigned long long nsec, unsigned long long bs,
3032                      uint64_t offset, unsigned int ioprio)
3033 {
3034         const bool needs_lock = td_async_processing(td);
3035         struct thread_stat *ts = &td->ts;
3036
3037         if (!ddir_rw(ddir))
3038                 return;
3039
3040         if (needs_lock)
3041                 __td_io_u_lock(td);
3042
3043         add_stat_sample(&ts->slat_stat[ddir], nsec);
3044
3045         if (td->slat_log)
3046                 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3047                                offset, ioprio);
3048
3049         if (ts->slat_percentiles)
3050                 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3051
3052         if (needs_lock)
3053                 __td_io_u_unlock(td);
3054 }
3055
3056 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3057                     unsigned long long nsec, unsigned long long bs,
3058                     uint64_t offset, unsigned int ioprio, bool high_prio)
3059 {
3060         const bool needs_lock = td_async_processing(td);
3061         struct thread_stat *ts = &td->ts;
3062
3063         if (!ddir_rw(ddir))
3064                 return;
3065
3066         if (needs_lock)
3067                 __td_io_u_lock(td);
3068
3069         add_stat_sample(&ts->lat_stat[ddir], nsec);
3070
3071         if (td->lat_log)
3072                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3073                                offset, ioprio);
3074
3075         /*
3076          * When lat_percentiles=1 (default 0), the reported high/low priority
3077          * percentiles and stats are used for describing total latency values,
3078          * even though the variable names themselves start with clat_.
3079          *
3080          * Because of the above definition, add a prio stat and prio lat
3081          * percentile sample only when lat_percentiles=1. add_clat_sample() will
3082          * add the prio stat and prio lat percentile sample when
3083          * lat_percentiles=0.
3084          */
3085         if (ts->lat_percentiles) {
3086                 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3087                 add_lat_percentile_prio_sample(ts, nsec, ddir, high_prio);
3088                 if (high_prio)
3089                         add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
3090                 else
3091                         add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
3092
3093         }
3094         if (needs_lock)
3095                 __td_io_u_unlock(td);
3096 }
3097
3098 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3099                    unsigned int bytes, unsigned long long spent)
3100 {
3101         const bool needs_lock = td_async_processing(td);
3102         struct thread_stat *ts = &td->ts;
3103         unsigned long rate;
3104
3105         if (spent)
3106                 rate = (unsigned long) (bytes * 1000000ULL / spent);
3107         else
3108                 rate = 0;
3109
3110         if (needs_lock)
3111                 __td_io_u_lock(td);
3112
3113         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3114
3115         if (td->bw_log)
3116                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3117                                bytes, io_u->offset, io_u->ioprio);
3118
3119         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3120
3121         if (needs_lock)
3122                 __td_io_u_unlock(td);
3123 }
3124
3125 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3126                          struct timespec *t, unsigned int avg_time,
3127                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3128                          struct io_stat *stat, struct io_log *log,
3129                          bool is_kb)
3130 {
3131         const bool needs_lock = td_async_processing(td);
3132         unsigned long spent, rate;
3133         enum fio_ddir ddir;
3134         unsigned long next, next_log;
3135
3136         next_log = avg_time;
3137
3138         spent = mtime_since(parent_tv, t);
3139         if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3140                 return avg_time - spent;
3141
3142         if (needs_lock)
3143                 __td_io_u_lock(td);
3144
3145         /*
3146          * Compute both read and write rates for the interval.
3147          */
3148         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3149                 uint64_t delta;
3150
3151                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3152                 if (!delta)
3153                         continue; /* No entries for interval */
3154
3155                 if (spent) {
3156                         if (is_kb)
3157                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
3158                         else
3159                                 rate = (delta * 1000) / spent;
3160                 } else
3161                         rate = 0;
3162
3163                 add_stat_sample(&stat[ddir], rate);
3164
3165                 if (log) {
3166                         unsigned long long bs = 0;
3167
3168                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3169                                 bs = td->o.min_bs[ddir];
3170
3171                         next = add_log_sample(td, log, sample_val(rate), ddir,
3172                                               bs, 0, 0);
3173                         next_log = min(next_log, next);
3174                 }
3175
3176                 stat_io_bytes[ddir] = this_io_bytes[ddir];
3177         }
3178
3179         *parent_tv = *t;
3180
3181         if (needs_lock)
3182                 __td_io_u_unlock(td);
3183
3184         if (spent <= avg_time)
3185                 next = avg_time;
3186         else
3187                 next = avg_time - (1 + spent - avg_time);
3188
3189         return min(next, next_log);
3190 }
3191
3192 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3193 {
3194         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3195                                 td->this_io_bytes, td->stat_io_bytes,
3196                                 td->ts.bw_stat, td->bw_log, true);
3197 }
3198
3199 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3200                      unsigned int bytes)
3201 {
3202         const bool needs_lock = td_async_processing(td);
3203         struct thread_stat *ts = &td->ts;
3204
3205         if (needs_lock)
3206                 __td_io_u_lock(td);
3207
3208         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3209
3210         if (td->iops_log)
3211                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3212                                bytes, io_u->offset, io_u->ioprio);
3213
3214         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3215
3216         if (needs_lock)
3217                 __td_io_u_unlock(td);
3218 }
3219
3220 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3221 {
3222         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3223                                 td->this_io_blocks, td->stat_io_blocks,
3224                                 td->ts.iops_stat, td->iops_log, false);
3225 }
3226
3227 /*
3228  * Returns msecs to next event
3229  */
3230 int calc_log_samples(void)
3231 {
3232         struct thread_data *td;
3233         unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3234         struct timespec now;
3235         int i;
3236         long elapsed_time = 0;
3237
3238         fio_gettime(&now, NULL);
3239
3240         for_each_td(td, i) {
3241                 elapsed_time = mtime_since_now(&td->epoch);
3242
3243                 if (!td->o.stats)
3244                         continue;
3245                 if (in_ramp_time(td) ||
3246                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3247                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3248                         continue;
3249                 }
3250                 if (!td->bw_log ||
3251                         (td->bw_log && !per_unit_log(td->bw_log))) {
3252                         tmp = add_bw_samples(td, &now);
3253
3254                         if (td->bw_log)
3255                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3256                 }
3257                 if (!td->iops_log ||
3258                         (td->iops_log && !per_unit_log(td->iops_log))) {
3259                         tmp = add_iops_samples(td, &now);
3260
3261                         if (td->iops_log)
3262                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3263                 }
3264
3265                 if (tmp < next)
3266                         next = tmp;
3267         }
3268
3269         /* if log_avg_msec_min has not been changed, set it to 0 */
3270         if (log_avg_msec_min == -1U)
3271                 log_avg_msec_min = 0;
3272
3273         if (log_avg_msec_min == 0)
3274                 next_mod = elapsed_time;
3275         else
3276                 next_mod = elapsed_time % log_avg_msec_min;
3277
3278         /* correction to keep the time on the log avg msec boundary */
3279         next = min(next, (log_avg_msec_min - next_mod));
3280
3281         return next == ~0U ? 0 : next;
3282 }
3283
3284 void stat_init(void)
3285 {
3286         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3287 }
3288
3289 void stat_exit(void)
3290 {
3291         /*
3292          * When we have the mutex, we know out-of-band access to it
3293          * have ended.
3294          */
3295         fio_sem_down(stat_sem);
3296         fio_sem_remove(stat_sem);
3297 }
3298
3299 /*
3300  * Called from signal handler. Wake up status thread.
3301  */
3302 void show_running_run_stats(void)
3303 {
3304         helper_do_stat();
3305 }
3306
3307 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3308 {
3309         /* Ignore io_u's which span multiple blocks--they will just get
3310          * inaccurate counts. */
3311         int idx = (io_u->offset - io_u->file->file_offset)
3312                         / td->o.bs[DDIR_TRIM];
3313         uint32_t *info = &td->ts.block_infos[idx];
3314         assert(idx < td->ts.nr_block_infos);
3315         return info;
3316 }
3317