stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <stdlib.h>
   4 #include <sys/time.h>
   5 #include <sys/stat.h>
   6 #include <math.h>
   7
   8 #include "fio.h"
   9 #include "diskutil.h"
  10 #include "lib/ieee754.h"
  11 #include "json.h"
  12 #include "lib/getrusage.h"
  13 #include "idletime.h"
  14 #include "lib/pow2.h"
  15 #include "lib/output_buffer.h"
  16 #include "helper_thread.h"
  17 #include "smalloc.h"
  18 #include "zbd.h"
  19 #include "oslib/asprintf.h"
  20
  21 #ifdef WIN32
  22 #define LOG_MSEC_SLACK  2
  23 #else
  24 #define LOG_MSEC_SLACK  1
  25 #endif
  26
  27 struct fio_sem *stat_sem;
  28
  29 void clear_rusage_stat(struct thread_data *td)
  30 {
  31         struct thread_stat *ts = &td->ts;
  32
  33         fio_getrusage(&td->ru_start);
  34         ts->usr_time = ts->sys_time = 0;
  35         ts->ctx = 0;
  36         ts->minf = ts->majf = 0;
  37 }
  38
  39 void update_rusage_stat(struct thread_data *td)
  40 {
  41         struct thread_stat *ts = &td->ts;
  42
  43         fio_getrusage(&td->ru_end);
  44         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
  45                                         &td->ru_end.ru_utime);
  46         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
  47                                         &td->ru_end.ru_stime);
  48         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  49                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  50         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  51         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  52
  53         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  54 }
  55
  56 /*
  57  * Given a latency, return the index of the corresponding bucket in
  58  * the structure tracking percentiles.
  59  *
  60  * (1) find the group (and error bits) that the value (latency)
  61  * belongs to by looking at its MSB. (2) find the bucket number in the
  62  * group by looking at the index bits.
  63  *
  64  */
  65 static unsigned int plat_val_to_idx(unsigned long long val)
  66 {
  67         unsigned int msb, error_bits, base, offset, idx;
  68
  69         /* Find MSB starting from bit 0 */
  70         if (val == 0)
  71                 msb = 0;
  72         else
  73                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
  74
  75         /*
  76          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  77          * all bits of the sample as index
  78          */
  79         if (msb <= FIO_IO_U_PLAT_BITS)
  80                 return val;
  81
  82         /* Compute the number of error bits to discard*/
  83         error_bits = msb - FIO_IO_U_PLAT_BITS;
  84
  85         /* Compute the number of buckets before the group */
  86         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  87
  88         /*
  89          * Discard the error bits and apply the mask to find the
  90          * index for the buckets in the group
  91          */
  92         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  93
  94         /* Make sure the index does not exceed (array size - 1) */
  95         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  96                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  97
  98         return idx;
  99 }
 100
 101 /*
 102  * Convert the given index of the bucket array to the value
 103  * represented by the bucket
 104  */
 105 static unsigned long long plat_idx_to_val(unsigned int idx)
 106 {
 107         unsigned int error_bits;
 108         unsigned long long k, base;
 109
 110         assert(idx < FIO_IO_U_PLAT_NR);
 111
 112         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 113          * all bits of the sample as index */
 114         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 115                 return idx;
 116
 117         /* Find the group and compute the minimum value of that group */
 118         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 119         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
 120
 121         /* Find its bucket number of the group */
 122         k = idx % FIO_IO_U_PLAT_VAL;
 123
 124         /* Return the mean of the range of the bucket */
 125         return base + ((k + 0.5) * (1 << error_bits));
 126 }
 127
 128 static int double_cmp(const void *a, const void *b)
 129 {
 130         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 131         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 132         int cmp = 0;
 133
 134         if (fa.u.f > fb.u.f)
 135                 cmp = 1;
 136         else if (fa.u.f < fb.u.f)
 137                 cmp = -1;
 138
 139         return cmp;
 140 }
 141
 142 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 143                                    fio_fp64_t *plist, unsigned long long **output,
 144                                    unsigned long long *maxv, unsigned long long *minv)
 145 {
 146         unsigned long long sum = 0;
 147         unsigned int len, i, j = 0;
 148         unsigned long long *ovals = NULL;
 149         bool is_last;
 150
 151         *minv = -1ULL;
 152         *maxv = 0;
 153
 154         len = 0;
 155         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 156                 len++;
 157
 158         if (!len)
 159                 return 0;
 160
 161         /*
 162          * Sort the percentile list. Note that it may already be sorted if
 163          * we are using the default values, but since it's a short list this
 164          * isn't a worry. Also note that this does not work for NaN values.
 165          */
 166         if (len > 1)
 167                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 168
 169         ovals = malloc(len * sizeof(*ovals));
 170         if (!ovals)
 171                 return 0;
 172
 173         /*
 174          * Calculate bucket values, note down max and min values
 175          */
 176         is_last = false;
 177         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 178                 sum += io_u_plat[i];
 179                 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
 180                         assert(plist[j].u.f <= 100.0);
 181
 182                         ovals[j] = plat_idx_to_val(i);
 183                         if (ovals[j] < *minv)
 184                                 *minv = ovals[j];
 185                         if (ovals[j] > *maxv)
 186                                 *maxv = ovals[j];
 187
 188                         is_last = (j == len - 1) != 0;
 189                         if (is_last)
 190                                 break;
 191
 192                         j++;
 193                 }
 194         }
 195
 196         if (!is_last)
 197                 log_err("fio: error calculating latency percentiles\n");
 198
 199         *output = ovals;
 200         return len;
 201 }
 202
 203 /*
 204  * Find and display the p-th percentile of clat
 205  */
 206 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
 207                                   fio_fp64_t *plist, unsigned int precision,
 208                                   const char *pre, struct buf_output *out)
 209 {
 210         unsigned int divisor, len, i, j = 0;
 211         unsigned long long minv, maxv;
 212         unsigned long long *ovals;
 213         int per_line, scale_down, time_width;
 214         bool is_last;
 215         char fmt[32];
 216
 217         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 218         if (!len || !ovals)
 219                 return;
 220
 221         /*
 222          * We default to nsecs, but if the value range is such that we
 223          * should scale down to usecs or msecs, do that.
 224          */
 225         if (minv > 2000000 && maxv > 99999999ULL) {
 226                 scale_down = 2;
 227                 divisor = 1000000;
 228                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 229         } else if (minv > 2000 && maxv > 99999) {
 230                 scale_down = 1;
 231                 divisor = 1000;
 232                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 233         } else {
 234                 scale_down = 0;
 235                 divisor = 1;
 236                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 237         }
 238
 239
 240         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 241         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 242                         precision, time_width);
 243         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 244         per_line = (80 - 7) / (precision + 10 + time_width);
 245
 246         for (j = 0; j < len; j++) {
 247                 /* for formatting */
 248                 if (j != 0 && (j % per_line) == 0)
 249                         log_buf(out, "     |");
 250
 251                 /* end of the list */
 252                 is_last = (j == len - 1) != 0;
 253
 254                 for (i = 0; i < scale_down; i++)
 255                         ovals[j] = (ovals[j] + 999) / 1000;
 256
 257                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 258
 259                 if (is_last)
 260                         break;
 261
 262                 if ((j % per_line) == per_line - 1)     /* for formatting */
 263                         log_buf(out, "\n");
 264         }
 265
 266         free(ovals);
 267 }
 268
 269 static int get_nr_prios_with_samples(struct thread_stat *ts, enum fio_ddir ddir)
 270 {
 271         int i, nr_prios_with_samples = 0;
 272
 273         for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
 274                 if (ts->clat_prio[ddir][i].clat_stat.samples)
 275                         nr_prios_with_samples++;
 276         }
 277
 278         return nr_prios_with_samples;
 279 }
 280
 281 bool calc_lat(struct io_stat *is, unsigned long long *min,
 282               unsigned long long *max, double *mean, double *dev)
 283 {
 284         double n = (double) is->samples;
 285
 286         if (n == 0)
 287                 return false;
 288
 289         *min = is->min_val;
 290         *max = is->max_val;
 291         *mean = is->mean.u.f;
 292
 293         if (n > 1.0)
 294                 *dev = sqrt(is->S.u.f / (n - 1.0));
 295         else
 296                 *dev = 0;
 297
 298         return true;
 299 }
 300
 301 void show_mixed_group_stats(struct group_run_stats *rs, struct buf_output *out)
 302 {
 303         char *io, *agg, *min, *max;
 304         char *ioalt, *aggalt, *minalt, *maxalt;
 305         uint64_t io_mix = 0, agg_mix = 0, min_mix = -1, max_mix = 0;
 306         uint64_t min_run = -1, max_run = 0;
 307         const int i2p = is_power_of_2(rs->kb_base);
 308         int i;
 309
 310         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 311                 if (!rs->max_run[i])
 312                         continue;
 313                 io_mix += rs->iobytes[i];
 314                 agg_mix += rs->agg[i];
 315                 min_mix = min_mix < rs->min_bw[i] ? min_mix : rs->min_bw[i];
 316                 max_mix = max_mix > rs->max_bw[i] ? max_mix : rs->max_bw[i];
 317                 min_run = min_run < rs->min_run[i] ? min_run : rs->min_run[i];
 318                 max_run = max_run > rs->max_run[i] ? max_run : rs->max_run[i];
 319         }
 320         io = num2str(io_mix, rs->sig_figs, 1, i2p, N2S_BYTE);
 321         ioalt = num2str(io_mix, rs->sig_figs, 1, !i2p, N2S_BYTE);
 322         agg = num2str(agg_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 323         aggalt = num2str(agg_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 324         min = num2str(min_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 325         minalt = num2str(min_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 326         max = num2str(max_mix, rs->sig_figs, 1, i2p, rs->unit_base);
 327         maxalt = num2str(max_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
 328         log_buf(out, "  MIXED: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 329                         agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 330                         (unsigned long long) min_run,
 331                         (unsigned long long) max_run);
 332         free(io);
 333         free(agg);
 334         free(min);
 335         free(max);
 336         free(ioalt);
 337         free(aggalt);
 338         free(minalt);
 339         free(maxalt);
 340 }
 341
 342 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 343 {
 344         char *io, *agg, *min, *max;
 345         char *ioalt, *aggalt, *minalt, *maxalt;
 346         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 347         int i;
 348
 349         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 350
 351         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 352                 const int i2p = is_power_of_2(rs->kb_base);
 353
 354                 if (!rs->max_run[i])
 355                         continue;
 356
 357                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
 358                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
 359                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
 360                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 361                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 362                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 363                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 364                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 365                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 366                                 (rs->unified_rw_rep == UNIFIED_MIXED) ? "  MIXED" : str[i],
 367                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 368                                 (unsigned long long) rs->min_run[i],
 369                                 (unsigned long long) rs->max_run[i]);
 370
 371                 free(io);
 372                 free(agg);
 373                 free(min);
 374                 free(max);
 375                 free(ioalt);
 376                 free(aggalt);
 377                 free(minalt);
 378                 free(maxalt);
 379         }
 380
 381         /* Need to aggregate statistics to show mixed values */
 382         if (rs->unified_rw_rep == UNIFIED_BOTH)
 383                 show_mixed_group_stats(rs, out);
 384 }
 385
 386 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
 387 {
 388         int i;
 389
 390         /*
 391          * Do depth distribution calculations
 392          */
 393         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 394                 if (total) {
 395                         io_u_dist[i] = (double) map[i] / (double) total;
 396                         io_u_dist[i] *= 100.0;
 397                         if (io_u_dist[i] < 0.1 && map[i])
 398                                 io_u_dist[i] = 0.1;
 399                 } else
 400                         io_u_dist[i] = 0.0;
 401         }
 402 }
 403
 404 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 405                           uint64_t *src, int nr)
 406 {
 407         unsigned long total = ddir_rw_sum(ts->total_io_u);
 408         int i;
 409
 410         /*
 411          * Do latency distribution calculations
 412          */
 413         for (i = 0; i < nr; i++) {
 414                 if (total) {
 415                         dst[i] = (double) src[i] / (double) total;
 416                         dst[i] *= 100.0;
 417                         if (dst[i] < 0.01 && src[i])
 418                                 dst[i] = 0.01;
 419                 } else
 420                         dst[i] = 0.0;
 421         }
 422 }
 423
 424 /*
 425  * To keep the terse format unaltered, add all of the ns latency
 426  * buckets to the first us latency bucket
 427  */
 428 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 429 {
 430         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 431         int i;
 432
 433         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 434
 435         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 436                 ntotal += ts->io_u_lat_n[i];
 437
 438         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 439 }
 440
 441 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 442 {
 443         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 444 }
 445
 446 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 447 {
 448         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 449 }
 450
 451 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 452 {
 453         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 454 }
 455
 456 static void display_lat(const char *name, unsigned long long min,
 457                         unsigned long long max, double mean, double dev,
 458                         struct buf_output *out)
 459 {
 460         const char *base = "(nsec)";
 461         char *minp, *maxp;
 462
 463         if (nsec_to_msec(&min, &max, &mean, &dev))
 464                 base = "(msec)";
 465         else if (nsec_to_usec(&min, &max, &mean, &dev))
 466                 base = "(usec)";
 467
 468         minp = num2str(min, 6, 1, 0, N2S_NONE);
 469         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 470
 471         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 472                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 473
 474         free(minp);
 475         free(maxp);
 476 }
 477
 478 static struct thread_stat *gen_mixed_ddir_stats_from_ts(struct thread_stat *ts)
 479 {
 480         struct thread_stat *ts_lcl;
 481
 482         /*
 483          * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
 484          * Trims (ddir = 2)
 485          */
 486         ts_lcl = malloc(sizeof(struct thread_stat));
 487         if (!ts_lcl) {
 488                 log_err("fio: failed to allocate local thread stat\n");
 489                 return NULL;
 490         }
 491
 492         init_thread_stat(ts_lcl);
 493
 494         /* calculate mixed stats  */
 495         ts_lcl->unified_rw_rep = UNIFIED_MIXED;
 496         ts_lcl->lat_percentiles = ts->lat_percentiles;
 497         ts_lcl->clat_percentiles = ts->clat_percentiles;
 498         ts_lcl->slat_percentiles = ts->slat_percentiles;
 499         ts_lcl->percentile_precision = ts->percentile_precision;
 500         memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
 501
 502         sum_thread_stats(ts_lcl, ts);
 503
 504         return ts_lcl;
 505 }
 506
 507 static double convert_agg_kbytes_percent(struct group_run_stats *rs,
 508                                          enum fio_ddir ddir, int mean)
 509 {
 510         double p_of_agg = 100.0;
 511         if (rs && rs->agg[ddir] > 1024) {
 512                 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
 513
 514                 if (p_of_agg > 100.0)
 515                         p_of_agg = 100.0;
 516         }
 517         return p_of_agg;
 518 }
 519
 520 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 521                              enum fio_ddir ddir, struct buf_output *out)
 522 {
 523         unsigned long runt;
 524         unsigned long long min, max, bw, iops;
 525         double mean, dev;
 526         char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
 527         int i2p, i;
 528         const char *clat_type = ts->lat_percentiles ? "lat" : "clat";
 529
 530         if (ddir_sync(ddir)) {
 531                 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
 532                         log_buf(out, "  %s:\n", "fsync/fdatasync/sync_file_range");
 533                         display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
 534                         show_clat_percentiles(ts->io_u_sync_plat,
 535                                                 ts->sync_stat.samples,
 536                                                 ts->percentile_list,
 537                                                 ts->percentile_precision,
 538                                                 io_ddir_name(ddir), out);
 539                 }
 540                 return;
 541         }
 542
 543         assert(ddir_rw(ddir));
 544
 545         if (!ts->runtime[ddir])
 546                 return;
 547
 548         i2p = is_power_of_2(rs->kb_base);
 549         runt = ts->runtime[ddir];
 550
 551         bw = (1000 * ts->io_bytes[ddir]) / runt;
 552         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
 553         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
 554         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
 555
 556         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 557         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
 558         if (ddir == DDIR_WRITE || ddir == DDIR_TRIM)
 559                 post_st = zbd_write_status(ts);
 560         else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
 561                 uint64_t total;
 562                 double hit;
 563
 564                 total = ts->cachehit + ts->cachemiss;
 565                 hit = (double) ts->cachehit / (double) total;
 566                 hit *= 100.0;
 567                 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
 568                         post_st = NULL;
 569         }
 570
 571         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
 572                         (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
 573                         iops_p, bw_p, bw_p_alt, io_p,
 574                         (unsigned long long) ts->runtime[ddir],
 575                         post_st ? : "");
 576
 577         free(post_st);
 578         free(io_p);
 579         free(bw_p);
 580         free(bw_p_alt);
 581         free(iops_p);
 582
 583         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 584                 display_lat("slat", min, max, mean, dev, out);
 585         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 586                 display_lat("clat", min, max, mean, dev, out);
 587         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 588                 display_lat(" lat", min, max, mean, dev, out);
 589
 590         /* Only print per prio stats if there are >= 2 prios with samples */
 591         if (get_nr_prios_with_samples(ts, ddir) >= 2) {
 592                 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
 593                         char buf[64];
 594
 595                         if (!calc_lat(&ts->clat_prio[ddir][i].clat_stat, &min,
 596                                       &max, &mean, &dev))
 597                                 continue;
 598
 599                         snprintf(buf, sizeof(buf),
 600                                  "%s prio %u/%u/%u",
 601                                  clat_type,
 602                                  ioprio_class(ts->clat_prio[ddir][i].ioprio),
 603                                  ioprio(ts->clat_prio[ddir][i].ioprio),
 604                                  ioprio_hint(ts->clat_prio[ddir][i].ioprio));
 605                         display_lat(buf, min, max, mean, dev, out);
 606                 }
 607         }
 608
 609         if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
 610                 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
 611                                         ts->slat_stat[ddir].samples,
 612                                         ts->percentile_list,
 613                                         ts->percentile_precision, "slat", out);
 614         if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
 615                 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
 616                                         ts->clat_stat[ddir].samples,
 617                                         ts->percentile_list,
 618                                         ts->percentile_precision, "clat", out);
 619         if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
 620                 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
 621                                         ts->lat_stat[ddir].samples,
 622                                         ts->percentile_list,
 623                                         ts->percentile_precision, "lat", out);
 624
 625         if (ts->clat_percentiles || ts->lat_percentiles) {
 626                 char prio_name[64];
 627                 uint64_t samples;
 628
 629                 if (ts->lat_percentiles)
 630                         samples = ts->lat_stat[ddir].samples;
 631                 else
 632                         samples = ts->clat_stat[ddir].samples;
 633
 634                 /* Only print per prio stats if there are >= 2 prios with samples */
 635                 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
 636                         for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
 637                                 uint64_t prio_samples =
 638                                         ts->clat_prio[ddir][i].clat_stat.samples;
 639
 640                                 if (!prio_samples)
 641                                         continue;
 642
 643                                 snprintf(prio_name, sizeof(prio_name),
 644                                          "%s prio %u/%u/%u (%.2f%% of IOs)",
 645                                          clat_type,
 646                                          ioprio_class(ts->clat_prio[ddir][i].ioprio),
 647                                          ioprio(ts->clat_prio[ddir][i].ioprio),
 648                                          ioprio_hint(ts->clat_prio[ddir][i].ioprio),
 649                                          100. * (double) prio_samples / (double) samples);
 650                                 show_clat_percentiles(ts->clat_prio[ddir][i].io_u_plat,
 651                                                 prio_samples, ts->percentile_list,
 652                                                 ts->percentile_precision,
 653                                                 prio_name, out);
 654                         }
 655                 }
 656         }
 657
 658         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 659                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 660                 const char *bw_str;
 661
 662                 if ((rs->unit_base == 1) && i2p)
 663                         bw_str = "Kibit";
 664                 else if (rs->unit_base == 1)
 665                         bw_str = "kbit";
 666                 else if (i2p)
 667                         bw_str = "KiB";
 668                 else
 669                         bw_str = "kB";
 670
 671                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
 672
 673                 if (rs->unit_base == 1) {
 674                         min *= 8.0;
 675                         max *= 8.0;
 676                         mean *= 8.0;
 677                         dev *= 8.0;
 678                 }
 679
 680                 if (mean > fkb_base * fkb_base) {
 681                         min /= fkb_base;
 682                         max /= fkb_base;
 683                         mean /= fkb_base;
 684                         dev /= fkb_base;
 685                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 686                 }
 687
 688                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 689                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 690                         bw_str, min, max, p_of_agg, mean, dev,
 691                         (&ts->bw_stat[ddir])->samples);
 692         }
 693         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 694                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 695                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 696                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 697         }
 698 }
 699
 700 static void show_mixed_ddir_status(struct group_run_stats *rs,
 701                                    struct thread_stat *ts,
 702                                    struct buf_output *out)
 703 {
 704         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
 705
 706         if (ts_lcl)
 707                 show_ddir_status(rs, ts_lcl, DDIR_READ, out);
 708
 709         free_clat_prio_stats(ts_lcl);
 710         free(ts_lcl);
 711 }
 712
 713 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
 714                      const char *msg, struct buf_output *out)
 715 {
 716         bool new_line = true, shown = false;
 717         int i, line = 0;
 718
 719         for (i = 0; i < nr; i++) {
 720                 if (io_u_lat[i] <= 0.0)
 721                         continue;
 722                 shown = true;
 723                 if (new_line) {
 724                         if (line)
 725                                 log_buf(out, "\n");
 726                         log_buf(out, "  lat (%s)   : ", msg);
 727                         new_line = false;
 728                         line = 0;
 729                 }
 730                 if (line)
 731                         log_buf(out, ", ");
 732                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 733                 line++;
 734                 if (line == 5)
 735                         new_line = true;
 736         }
 737
 738         if (shown)
 739                 log_buf(out, "\n");
 740
 741         return true;
 742 }
 743
 744 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 745 {
 746         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 747                                  "250=", "500=", "750=", "1000=", };
 748
 749         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 750 }
 751
 752 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 753 {
 754         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 755                                  "250=", "500=", "750=", "1000=", };
 756
 757         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 758 }
 759
 760 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 761 {
 762         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 763                                  "250=", "500=", "750=", "1000=", "2000=",
 764                                  ">=2000=", };
 765
 766         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 767 }
 768
 769 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 770 {
 771         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 772         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 773         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 774
 775         stat_calc_lat_n(ts, io_u_lat_n);
 776         stat_calc_lat_u(ts, io_u_lat_u);
 777         stat_calc_lat_m(ts, io_u_lat_m);
 778
 779         show_lat_n(io_u_lat_n, out);
 780         show_lat_u(io_u_lat_u, out);
 781         show_lat_m(io_u_lat_m, out);
 782 }
 783
 784 static int block_state_category(int block_state)
 785 {
 786         switch (block_state) {
 787         case BLOCK_STATE_UNINIT:
 788                 return 0;
 789         case BLOCK_STATE_TRIMMED:
 790         case BLOCK_STATE_WRITTEN:
 791                 return 1;
 792         case BLOCK_STATE_WRITE_FAILURE:
 793         case BLOCK_STATE_TRIM_FAILURE:
 794                 return 2;
 795         default:
 796                 /* Silence compile warning on some BSDs and have a return */
 797                 assert(0);
 798                 return -1;
 799         }
 800 }
 801
 802 static int compare_block_infos(const void *bs1, const void *bs2)
 803 {
 804         uint64_t block1 = *(uint64_t *)bs1;
 805         uint64_t block2 = *(uint64_t *)bs2;
 806         int state1 = BLOCK_INFO_STATE(block1);
 807         int state2 = BLOCK_INFO_STATE(block2);
 808         int bscat1 = block_state_category(state1);
 809         int bscat2 = block_state_category(state2);
 810         int cycles1 = BLOCK_INFO_TRIMS(block1);
 811         int cycles2 = BLOCK_INFO_TRIMS(block2);
 812
 813         if (bscat1 < bscat2)
 814                 return -1;
 815         if (bscat1 > bscat2)
 816                 return 1;
 817
 818         if (cycles1 < cycles2)
 819                 return -1;
 820         if (cycles1 > cycles2)
 821                 return 1;
 822
 823         if (state1 < state2)
 824                 return -1;
 825         if (state1 > state2)
 826                 return 1;
 827
 828         assert(block1 == block2);
 829         return 0;
 830 }
 831
 832 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 833                                   fio_fp64_t *plist, unsigned int **percentiles,
 834                                   unsigned int *types)
 835 {
 836         int len = 0;
 837         int i, nr_uninit;
 838
 839         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 840
 841         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 842                 len++;
 843
 844         if (!len)
 845                 return 0;
 846
 847         /*
 848          * Sort the percentile list. Note that it may already be sorted if
 849          * we are using the default values, but since it's a short list this
 850          * isn't a worry. Also note that this does not work for NaN values.
 851          */
 852         if (len > 1)
 853                 qsort(plist, len, sizeof(plist[0]), double_cmp);
 854
 855         /* Start only after the uninit entries end */
 856         for (nr_uninit = 0;
 857              nr_uninit < nr_block_infos
 858                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 859              nr_uninit ++)
 860                 ;
 861
 862         if (nr_uninit == nr_block_infos)
 863                 return 0;
 864
 865         *percentiles = calloc(len, sizeof(**percentiles));
 866
 867         for (i = 0; i < len; i++) {
 868                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 869                                 + nr_uninit;
 870                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 871         }
 872
 873         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 874         for (i = 0; i < nr_block_infos; i++)
 875                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 876
 877         return len;
 878 }
 879
 880 static const char *block_state_names[] = {
 881         [BLOCK_STATE_UNINIT] = "unwritten",
 882         [BLOCK_STATE_TRIMMED] = "trimmed",
 883         [BLOCK_STATE_WRITTEN] = "written",
 884         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 885         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 886 };
 887
 888 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 889                              fio_fp64_t *plist, struct buf_output *out)
 890 {
 891         int len, pos, i;
 892         unsigned int *percentiles = NULL;
 893         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 894
 895         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 896                                      &percentiles, block_state_counts);
 897
 898         log_buf(out, "  block lifetime percentiles :\n   |");
 899         pos = 0;
 900         for (i = 0; i < len; i++) {
 901                 uint32_t block_info = percentiles[i];
 902 #define LINE_LENGTH     75
 903                 char str[LINE_LENGTH];
 904                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 905                                      plist[i].u.f, block_info,
 906                                      i == len - 1 ? '\n' : ',');
 907                 assert(strln < LINE_LENGTH);
 908                 if (pos + strln > LINE_LENGTH) {
 909                         pos = 0;
 910                         log_buf(out, "\n   |");
 911                 }
 912                 log_buf(out, "%s", str);
 913                 pos += strln;
 914 #undef LINE_LENGTH
 915         }
 916         if (percentiles)
 917                 free(percentiles);
 918
 919         log_buf(out, "        states               :");
 920         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 921                 log_buf(out, " %s=%u%c",
 922                          block_state_names[i], block_state_counts[i],
 923                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 924 }
 925
 926 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 927 {
 928         char *p1, *p1alt, *p2;
 929         unsigned long long bw_mean, iops_mean;
 930         const int i2p = is_power_of_2(ts->kb_base);
 931
 932         if (!ts->ss_dur)
 933                 return;
 934
 935         bw_mean = steadystate_bw_mean(ts);
 936         iops_mean = steadystate_iops_mean(ts);
 937
 938         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
 939         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
 940         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
 941
 942         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 943                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
 944                 p1, p1alt, p2,
 945                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
 946                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
 947                 ts->ss_criterion.u.f,
 948                 ts->ss_state & FIO_SS_PCT ? "%" : "");
 949
 950         free(p1);
 951         free(p1alt);
 952         free(p2);
 953 }
 954
 955 static void show_agg_stats(struct disk_util_agg *agg, int terse,
 956                            struct buf_output *out)
 957 {
 958         if (!agg->slavecount)
 959                 return;
 960
 961         if (!terse) {
 962                 log_buf(out, ", aggrios=%llu/%llu, aggsectors=%llu/%llu, "
 963                          "aggrmerge=%llu/%llu, aggrticks=%llu/%llu, "
 964                          "aggrin_queue=%llu, aggrutil=%3.2f%%",
 965                         (unsigned long long) agg->ios[0] / agg->slavecount,
 966                         (unsigned long long) agg->ios[1] / agg->slavecount,
 967                         (unsigned long long) agg->sectors[0] / agg->slavecount,
 968                         (unsigned long long) agg->sectors[1] / agg->slavecount,
 969                         (unsigned long long) agg->merges[0] / agg->slavecount,
 970                         (unsigned long long) agg->merges[1] / agg->slavecount,
 971                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 972                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 973                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 974                         agg->max_util.u.f);
 975         } else {
 976                 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
 977                         (unsigned long long) agg->ios[0] / agg->slavecount,
 978                         (unsigned long long) agg->ios[1] / agg->slavecount,
 979                         (unsigned long long) agg->merges[0] / agg->slavecount,
 980                         (unsigned long long) agg->merges[1] / agg->slavecount,
 981                         (unsigned long long) agg->ticks[0] / agg->slavecount,
 982                         (unsigned long long) agg->ticks[1] / agg->slavecount,
 983                         (unsigned long long) agg->time_in_queue / agg->slavecount,
 984                         agg->max_util.u.f);
 985         }
 986 }
 987
 988 static void aggregate_slaves_stats(struct disk_util *masterdu)
 989 {
 990         struct disk_util_agg *agg = &masterdu->agg;
 991         struct disk_util_stat *dus;
 992         struct flist_head *entry;
 993         struct disk_util *slavedu;
 994         double util;
 995
 996         flist_for_each(entry, &masterdu->slaves) {
 997                 slavedu = flist_entry(entry, struct disk_util, slavelist);
 998                 dus = &slavedu->dus;
 999                 agg->ios[0] += dus->s.ios[0];
1000                 agg->ios[1] += dus->s.ios[1];
1001                 agg->merges[0] += dus->s.merges[0];
1002                 agg->merges[1] += dus->s.merges[1];
1003                 agg->sectors[0] += dus->s.sectors[0];
1004                 agg->sectors[1] += dus->s.sectors[1];
1005                 agg->ticks[0] += dus->s.ticks[0];
1006                 agg->ticks[1] += dus->s.ticks[1];
1007                 agg->time_in_queue += dus->s.time_in_queue;
1008                 agg->slavecount++;
1009
1010                 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
1011                 /* System utilization is the utilization of the
1012                  * component with the highest utilization.
1013                  */
1014                 if (util > agg->max_util.u.f)
1015                         agg->max_util.u.f = util;
1016
1017         }
1018
1019         if (agg->max_util.u.f > 100.0)
1020                 agg->max_util.u.f = 100.0;
1021 }
1022
1023 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
1024                      int terse, struct buf_output *out)
1025 {
1026         double util = 0;
1027
1028         if (dus->s.msec)
1029                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1030         if (util > 100.0)
1031                 util = 100.0;
1032
1033         if (!terse) {
1034                 if (agg->slavecount)
1035                         log_buf(out, "  ");
1036
1037                 log_buf(out, "  %s: ios=%llu/%llu, sectors=%llu/%llu, "
1038                         "merge=%llu/%llu, ticks=%llu/%llu, in_queue=%llu, "
1039                         "util=%3.2f%%",
1040                                 dus->name,
1041                                 (unsigned long long) dus->s.ios[0],
1042                                 (unsigned long long) dus->s.ios[1],
1043                                 (unsigned long long) dus->s.sectors[0],
1044                                 (unsigned long long) dus->s.sectors[1],
1045                                 (unsigned long long) dus->s.merges[0],
1046                                 (unsigned long long) dus->s.merges[1],
1047                                 (unsigned long long) dus->s.ticks[0],
1048                                 (unsigned long long) dus->s.ticks[1],
1049                                 (unsigned long long) dus->s.time_in_queue,
1050                                 util);
1051         } else {
1052                 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1053                                 dus->name,
1054                                 (unsigned long long) dus->s.ios[0],
1055                                 (unsigned long long) dus->s.ios[1],
1056                                 (unsigned long long) dus->s.merges[0],
1057                                 (unsigned long long) dus->s.merges[1],
1058                                 (unsigned long long) dus->s.ticks[0],
1059                                 (unsigned long long) dus->s.ticks[1],
1060                                 (unsigned long long) dus->s.time_in_queue,
1061                                 util);
1062         }
1063
1064         /*
1065          * If the device has slaves, aggregate the stats for
1066          * those slave devices also.
1067          */
1068         show_agg_stats(agg, terse, out);
1069
1070         if (!terse)
1071                 log_buf(out, "\n");
1072 }
1073
1074 void json_array_add_disk_util(struct disk_util_stat *dus,
1075                 struct disk_util_agg *agg, struct json_array *array)
1076 {
1077         struct json_object *obj;
1078         double util = 0;
1079
1080         if (dus->s.msec)
1081                 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1082         if (util > 100.0)
1083                 util = 100.0;
1084
1085         obj = json_create_object();
1086         json_array_add_value_object(array, obj);
1087
1088         json_object_add_value_string(obj, "name", (const char *)dus->name);
1089         json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1090         json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1091         json_object_add_value_int(obj, "read_sectors", dus->s.sectors[0]);
1092         json_object_add_value_int(obj, "write_sectors", dus->s.sectors[1]);
1093         json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1094         json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1095         json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1096         json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1097         json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1098         json_object_add_value_float(obj, "util", util);
1099
1100         /*
1101          * If the device has slaves, aggregate the stats for
1102          * those slave devices also.
1103          */
1104         if (!agg->slavecount)
1105                 return;
1106         json_object_add_value_int(obj, "aggr_read_ios",
1107                                 agg->ios[0] / agg->slavecount);
1108         json_object_add_value_int(obj, "aggr_write_ios",
1109                                 agg->ios[1] / agg->slavecount);
1110         json_object_add_value_int(obj, "aggr_read_sectors",
1111                                 agg->sectors[0] / agg->slavecount);
1112         json_object_add_value_int(obj, "aggr_write_sectors",
1113                                 agg->sectors[1] / agg->slavecount);
1114         json_object_add_value_int(obj, "aggr_read_merges",
1115                                 agg->merges[0] / agg->slavecount);
1116         json_object_add_value_int(obj, "aggr_write_merge",
1117                                 agg->merges[1] / agg->slavecount);
1118         json_object_add_value_int(obj, "aggr_read_ticks",
1119                                 agg->ticks[0] / agg->slavecount);
1120         json_object_add_value_int(obj, "aggr_write_ticks",
1121                                 agg->ticks[1] / agg->slavecount);
1122         json_object_add_value_int(obj, "aggr_in_queue",
1123                                 agg->time_in_queue / agg->slavecount);
1124         json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1125 }
1126
1127 static void json_object_add_disk_utils(struct json_object *obj,
1128                                        struct flist_head *head)
1129 {
1130         struct json_array *array = json_create_array();
1131         struct flist_head *entry;
1132         struct disk_util *du;
1133
1134         json_object_add_value_array(obj, "disk_util", array);
1135
1136         flist_for_each(entry, head) {
1137                 du = flist_entry(entry, struct disk_util, list);
1138
1139                 aggregate_slaves_stats(du);
1140                 json_array_add_disk_util(&du->dus, &du->agg, array);
1141         }
1142 }
1143
1144 void show_disk_util(int terse, struct json_object *parent,
1145                     struct buf_output *out)
1146 {
1147         struct flist_head *entry;
1148         struct disk_util *du;
1149         bool do_json;
1150
1151         if (!is_running_backend())
1152                 return;
1153
1154         if (flist_empty(&disk_list))
1155                 return;
1156
1157         if ((output_format & FIO_OUTPUT_JSON) && parent)
1158                 do_json = true;
1159         else
1160                 do_json = false;
1161
1162         if (!terse && !do_json)
1163                 log_buf(out, "\nDisk stats (read/write):\n");
1164
1165         if (do_json) {
1166                 json_object_add_disk_utils(parent, &disk_list);
1167         } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1168                 flist_for_each(entry, &disk_list) {
1169                         du = flist_entry(entry, struct disk_util, list);
1170
1171                         aggregate_slaves_stats(du);
1172                         print_disk_util(&du->dus, &du->agg, terse, out);
1173                 }
1174         }
1175 }
1176
1177 static void show_thread_status_normal(struct thread_stat *ts,
1178                                       struct group_run_stats *rs,
1179                                       struct buf_output *out)
1180 {
1181         double usr_cpu, sys_cpu;
1182         unsigned long runtime;
1183         double io_u_dist[FIO_IO_U_MAP_NR];
1184         time_t time_p;
1185         char time_buf[32];
1186
1187         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1188                 return;
1189
1190         memset(time_buf, 0, sizeof(time_buf));
1191
1192         time(&time_p);
1193         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1194
1195         if (!ts->error) {
1196                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1197                                         ts->name, ts->groupid, ts->members,
1198                                         ts->error, (int) ts->pid, time_buf);
1199         } else {
1200                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1201                                         ts->name, ts->groupid, ts->members,
1202                                         ts->error, ts->verror, (int) ts->pid,
1203                                         time_buf);
1204         }
1205
1206         if (strlen(ts->description))
1207                 log_buf(out, "  Description  : [%s]\n", ts->description);
1208
1209         for_each_rw_ddir(ddir) {
1210                 if (ts->io_bytes[ddir])
1211                         show_ddir_status(rs, ts, ddir, out);
1212         }
1213
1214         if (ts->unified_rw_rep == UNIFIED_BOTH)
1215                 show_mixed_ddir_status(rs, ts, out);
1216
1217         show_latencies(ts, out);
1218
1219         if (ts->sync_stat.samples)
1220                 show_ddir_status(rs, ts, DDIR_SYNC, out);
1221
1222         runtime = ts->total_run_time;
1223         if (runtime) {
1224                 double runt = (double) runtime;
1225
1226                 usr_cpu = (double) ts->usr_time * 100 / runt;
1227                 sys_cpu = (double) ts->sys_time * 100 / runt;
1228         } else {
1229                 usr_cpu = 0;
1230                 sys_cpu = 0;
1231         }
1232
1233         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1234                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1235                         (unsigned long long) ts->ctx,
1236                         (unsigned long long) ts->majf,
1237                         (unsigned long long) ts->minf);
1238
1239         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1240         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1241                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1242                                         io_u_dist[1], io_u_dist[2],
1243                                         io_u_dist[3], io_u_dist[4],
1244                                         io_u_dist[5], io_u_dist[6]);
1245
1246         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1247         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1248                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1249                                         io_u_dist[1], io_u_dist[2],
1250                                         io_u_dist[3], io_u_dist[4],
1251                                         io_u_dist[5], io_u_dist[6]);
1252         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1253         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1254                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1255                                         io_u_dist[1], io_u_dist[2],
1256                                         io_u_dist[3], io_u_dist[4],
1257                                         io_u_dist[5], io_u_dist[6]);
1258         log_buf(out, "     issued rwts: total=%llu,%llu,%llu,%llu"
1259                                  " short=%llu,%llu,%llu,0"
1260                                  " dropped=%llu,%llu,%llu,0\n",
1261                                         (unsigned long long) ts->total_io_u[0],
1262                                         (unsigned long long) ts->total_io_u[1],
1263                                         (unsigned long long) ts->total_io_u[2],
1264                                         (unsigned long long) ts->total_io_u[3],
1265                                         (unsigned long long) ts->short_io_u[0],
1266                                         (unsigned long long) ts->short_io_u[1],
1267                                         (unsigned long long) ts->short_io_u[2],
1268                                         (unsigned long long) ts->drop_io_u[0],
1269                                         (unsigned long long) ts->drop_io_u[1],
1270                                         (unsigned long long) ts->drop_io_u[2]);
1271         if (ts->continue_on_error) {
1272                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
1273                                         (unsigned long long)ts->total_err_count,
1274                                         ts->first_error,
1275                                         strerror(ts->first_error));
1276         }
1277         if (ts->latency_depth) {
1278                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1279                                         (unsigned long long)ts->latency_target,
1280                                         (unsigned long long)ts->latency_window,
1281                                         ts->latency_percentile.u.f,
1282                                         ts->latency_depth);
1283         }
1284
1285         if (ts->nr_block_infos)
1286                 show_block_infos(ts->nr_block_infos, ts->block_infos,
1287                                   ts->percentile_list, out);
1288
1289         if (ts->ss_dur)
1290                 show_ss_normal(ts, out);
1291 }
1292
1293 static void show_ddir_status_terse(struct thread_stat *ts,
1294                                    struct group_run_stats *rs,
1295                                    enum fio_ddir ddir, int ver,
1296                                    struct buf_output *out)
1297 {
1298         unsigned long long min, max, minv, maxv, bw, iops;
1299         unsigned long long *ovals = NULL;
1300         double mean, dev;
1301         unsigned int len;
1302         int i, bw_stat;
1303
1304         assert(ddir_rw(ddir));
1305
1306         iops = bw = 0;
1307         if (ts->runtime[ddir]) {
1308                 uint64_t runt = ts->runtime[ddir];
1309
1310                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1311                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1312         }
1313
1314         log_buf(out, ";%llu;%llu;%llu;%llu",
1315                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1316                                         (unsigned long long) ts->runtime[ddir]);
1317
1318         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1319                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1320         else
1321                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1322
1323         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1324                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1325         else
1326                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1327
1328         if (ts->lat_percentiles) {
1329                 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1330                                         ts->lat_stat[ddir].samples,
1331                                         ts->percentile_list, &ovals, &maxv,
1332                                         &minv);
1333         } else if (ts->clat_percentiles) {
1334                 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1335                                         ts->clat_stat[ddir].samples,
1336                                         ts->percentile_list, &ovals, &maxv,
1337                                         &minv);
1338         } else {
1339                 len = 0;
1340         }
1341
1342         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1343                 if (i >= len) {
1344                         log_buf(out, ";0%%=0");
1345                         continue;
1346                 }
1347                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1348         }
1349
1350         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1351                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1352         else
1353                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1354
1355         free(ovals);
1356
1357         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1358         if (bw_stat) {
1359                 double p_of_agg = 100.0;
1360
1361                 if (rs->agg[ddir]) {
1362                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1363                         if (p_of_agg > 100.0)
1364                                 p_of_agg = 100.0;
1365                 }
1366
1367                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1368         } else {
1369                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1370         }
1371
1372         if (ver == 5) {
1373                 if (bw_stat)
1374                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1375                 else
1376                         log_buf(out, ";%lu", 0UL);
1377
1378                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1379                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1380                                 mean, dev, (&ts->iops_stat[ddir])->samples);
1381                 else
1382                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1383         }
1384 }
1385
1386 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1387                                    struct group_run_stats *rs,
1388                                    int ver, struct buf_output *out)
1389 {
1390         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1391
1392         if (ts_lcl)
1393                 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1394
1395         free_clat_prio_stats(ts_lcl);
1396         free(ts_lcl);
1397 }
1398
1399 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1400                                              uint32_t percentiles,
1401                                              struct io_stat *lat_stat,
1402                                              uint64_t *io_u_plat)
1403 {
1404         char buf[120];
1405         double mean, dev;
1406         unsigned int i, len;
1407         struct json_object *lat_object, *percentile_object, *clat_bins_object;
1408         unsigned long long min, max, maxv, minv, *ovals = NULL;
1409
1410         if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1411                 min = max = 0;
1412                 mean = dev = 0.0;
1413         }
1414         lat_object = json_create_object();
1415         json_object_add_value_int(lat_object, "min", min);
1416         json_object_add_value_int(lat_object, "max", max);
1417         json_object_add_value_float(lat_object, "mean", mean);
1418         json_object_add_value_float(lat_object, "stddev", dev);
1419         json_object_add_value_int(lat_object, "N", lat_stat->samples);
1420
1421         if (percentiles && lat_stat->samples) {
1422                 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1423                                 ts->percentile_list, &ovals, &maxv, &minv);
1424
1425                 if (len > FIO_IO_U_LIST_MAX_LEN)
1426                         len = FIO_IO_U_LIST_MAX_LEN;
1427
1428                 percentile_object = json_create_object();
1429                 json_object_add_value_object(lat_object, "percentile", percentile_object);
1430                 for (i = 0; i < len; i++) {
1431                         snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1432                         json_object_add_value_int(percentile_object, buf, ovals[i]);
1433                 }
1434                 free(ovals);
1435
1436                 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1437                         clat_bins_object = json_create_object();
1438                         json_object_add_value_object(lat_object, "bins", clat_bins_object);
1439
1440                         for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1441                                 if (io_u_plat[i]) {
1442                                         snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1443                                         json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1444                                 }
1445                 }
1446         }
1447
1448         return lat_object;
1449 }
1450
1451 static void add_ddir_status_json(struct thread_stat *ts,
1452                                  struct group_run_stats *rs, enum fio_ddir ddir,
1453                                  struct json_object *parent)
1454 {
1455         unsigned long long min, max;
1456         unsigned long long bw_bytes, bw;
1457         double mean, dev, iops;
1458         struct json_object *dir_object, *tmp_object;
1459         double p_of_agg = 100.0;
1460
1461         assert(ddir_rw(ddir) || ddir_sync(ddir));
1462
1463         if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1464                 return;
1465
1466         dir_object = json_create_object();
1467         json_object_add_value_object(parent,
1468                 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1469
1470         if (ddir_rw(ddir)) {
1471                 bw_bytes = 0;
1472                 bw = 0;
1473                 iops = 0.0;
1474                 if (ts->runtime[ddir]) {
1475                         uint64_t runt = ts->runtime[ddir];
1476
1477                         bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1478                         bw = bw_bytes / 1024; /* KiB/s */
1479                         iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1480                 }
1481
1482                 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1483                 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1484                 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1485                 json_object_add_value_int(dir_object, "bw", bw);
1486                 json_object_add_value_float(dir_object, "iops", iops);
1487                 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1488                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1489                 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1490                 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1491
1492                 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1493                                 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1494                 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1495
1496                 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1497                                 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1498                 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1499
1500                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1501                                 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1502                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1503         } else {
1504                 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1505                 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1506                                 &ts->sync_stat, ts->io_u_sync_plat);
1507                 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1508         }
1509
1510         if (!ddir_rw(ddir))
1511                 return;
1512
1513         /* Only include per prio stats if there are >= 2 prios with samples */
1514         if (get_nr_prios_with_samples(ts, ddir) >= 2) {
1515                 struct json_array *array = json_create_array();
1516                 const char *obj_name;
1517                 int i;
1518
1519                 if (ts->lat_percentiles)
1520                         obj_name = "lat_ns";
1521                 else
1522                         obj_name = "clat_ns";
1523
1524                 json_object_add_value_array(dir_object, "prios", array);
1525
1526                 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
1527                         struct json_object *obj;
1528
1529                         if (!ts->clat_prio[ddir][i].clat_stat.samples)
1530                                 continue;
1531
1532                         obj = json_create_object();
1533
1534                         json_object_add_value_int(obj, "prioclass",
1535                                 ioprio_class(ts->clat_prio[ddir][i].ioprio));
1536                         json_object_add_value_int(obj, "prio",
1537                                 ioprio(ts->clat_prio[ddir][i].ioprio));
1538                         json_object_add_value_int(obj, "priohint",
1539                                 ioprio_hint(ts->clat_prio[ddir][i].ioprio));
1540
1541                         tmp_object = add_ddir_lat_json(ts,
1542                                         ts->clat_percentiles | ts->lat_percentiles,
1543                                         &ts->clat_prio[ddir][i].clat_stat,
1544                                         ts->clat_prio[ddir][i].io_u_plat);
1545                         json_object_add_value_object(obj, obj_name, tmp_object);
1546                         json_array_add_value_object(array, obj);
1547                 }
1548         }
1549
1550         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1551                 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1552         } else {
1553                 min = max = 0;
1554                 p_of_agg = mean = dev = 0.0;
1555         }
1556
1557         json_object_add_value_int(dir_object, "bw_min", min);
1558         json_object_add_value_int(dir_object, "bw_max", max);
1559         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1560         json_object_add_value_float(dir_object, "bw_mean", mean);
1561         json_object_add_value_float(dir_object, "bw_dev", dev);
1562         json_object_add_value_int(dir_object, "bw_samples",
1563                                 (&ts->bw_stat[ddir])->samples);
1564
1565         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1566                 min = max = 0;
1567                 mean = dev = 0.0;
1568         }
1569         json_object_add_value_int(dir_object, "iops_min", min);
1570         json_object_add_value_int(dir_object, "iops_max", max);
1571         json_object_add_value_float(dir_object, "iops_mean", mean);
1572         json_object_add_value_float(dir_object, "iops_stddev", dev);
1573         json_object_add_value_int(dir_object, "iops_samples",
1574                                 (&ts->iops_stat[ddir])->samples);
1575
1576         if (ts->cachehit + ts->cachemiss) {
1577                 uint64_t total;
1578                 double hit;
1579
1580                 total = ts->cachehit + ts->cachemiss;
1581                 hit = (double) ts->cachehit / (double) total;
1582                 hit *= 100.0;
1583                 json_object_add_value_float(dir_object, "cachehit", hit);
1584         }
1585 }
1586
1587 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1588                 struct group_run_stats *rs, struct json_object *parent)
1589 {
1590         struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1591
1592         /* add the aggregated stats to json parent */
1593         if (ts_lcl)
1594                 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1595
1596         free_clat_prio_stats(ts_lcl);
1597         free(ts_lcl);
1598 }
1599
1600 static void show_thread_status_terse_all(struct thread_stat *ts,
1601                                          struct group_run_stats *rs, int ver,
1602                                          struct buf_output *out)
1603 {
1604         double io_u_dist[FIO_IO_U_MAP_NR];
1605         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1606         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1607         double usr_cpu, sys_cpu;
1608         int i;
1609
1610         /* General Info */
1611         if (ver == 2)
1612                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1613         else
1614                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1615                         ts->name, ts->groupid, ts->error);
1616
1617         /* Log Read Status, or mixed if unified_rw_rep = 1 */
1618         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1619         if (ts->unified_rw_rep != UNIFIED_MIXED) {
1620                 /* Log Write Status */
1621                 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1622                 /* Log Trim Status */
1623                 if (ver == 2 || ver == 4 || ver == 5)
1624                         show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1625         }
1626         if (ts->unified_rw_rep == UNIFIED_BOTH)
1627                 show_mixed_ddir_status_terse(ts, rs, ver, out);
1628         /* CPU Usage */
1629         if (ts->total_run_time) {
1630                 double runt = (double) ts->total_run_time;
1631
1632                 usr_cpu = (double) ts->usr_time * 100 / runt;
1633                 sys_cpu = (double) ts->sys_time * 100 / runt;
1634         } else {
1635                 usr_cpu = 0;
1636                 sys_cpu = 0;
1637         }
1638
1639         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1640                                                 (unsigned long long) ts->ctx,
1641                                                 (unsigned long long) ts->majf,
1642                                                 (unsigned long long) ts->minf);
1643
1644         /* Calc % distribution of IO depths, usecond, msecond latency */
1645         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1646         stat_calc_lat_nu(ts, io_u_lat_u);
1647         stat_calc_lat_m(ts, io_u_lat_m);
1648
1649         /* Only show fixed 7 I/O depth levels*/
1650         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1651                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1652                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1653
1654         /* Microsecond latency */
1655         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1656                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1657         /* Millisecond latency */
1658         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1659                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1660
1661         /* disk util stats, if any */
1662         if (ver >= 3 && is_running_backend())
1663                 show_disk_util(1, NULL, out);
1664
1665         /* Additional output if continue_on_error set - default off*/
1666         if (ts->continue_on_error)
1667                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1668
1669         /* Additional output if description is set */
1670         if (strlen(ts->description)) {
1671                 if (ver == 2)
1672                         log_buf(out, "\n");
1673                 log_buf(out, ";%s", ts->description);
1674         }
1675
1676         log_buf(out, "\n");
1677 }
1678
1679 static void json_add_job_opts(struct json_object *root, const char *name,
1680                               struct flist_head *opt_list)
1681 {
1682         struct json_object *dir_object;
1683         struct flist_head *entry;
1684         struct print_option *p;
1685
1686         if (flist_empty(opt_list))
1687                 return;
1688
1689         dir_object = json_create_object();
1690         json_object_add_value_object(root, name, dir_object);
1691
1692         flist_for_each(entry, opt_list) {
1693                 p = flist_entry(entry, struct print_option, list);
1694                 json_object_add_value_string(dir_object, p->name, p->value);
1695         }
1696 }
1697
1698 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1699                                                    struct group_run_stats *rs,
1700                                                    struct flist_head *opt_list)
1701 {
1702         struct json_object *root, *tmp;
1703         struct jobs_eta *je;
1704         double io_u_dist[FIO_IO_U_MAP_NR];
1705         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1706         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1707         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1708         double usr_cpu, sys_cpu;
1709         int i;
1710         size_t size;
1711
1712         root = json_create_object();
1713         json_object_add_value_string(root, "jobname", ts->name);
1714         json_object_add_value_int(root, "groupid", ts->groupid);
1715         json_object_add_value_int(root, "error", ts->error);
1716
1717         /* ETA Info */
1718         je = get_jobs_eta(true, &size);
1719         if (je) {
1720                 json_object_add_value_int(root, "eta", je->eta_sec);
1721                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1722                 free(je);
1723         }
1724
1725         if (opt_list)
1726                 json_add_job_opts(root, "job options", opt_list);
1727
1728         add_ddir_status_json(ts, rs, DDIR_READ, root);
1729         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1730         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1731         add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1732
1733         if (ts->unified_rw_rep == UNIFIED_BOTH)
1734                 add_mixed_ddir_status_json(ts, rs, root);
1735
1736         /* CPU Usage */
1737         if (ts->total_run_time) {
1738                 double runt = (double) ts->total_run_time;
1739
1740                 usr_cpu = (double) ts->usr_time * 100 / runt;
1741                 sys_cpu = (double) ts->sys_time * 100 / runt;
1742         } else {
1743                 usr_cpu = 0;
1744                 sys_cpu = 0;
1745         }
1746         json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1747         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1748         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1749         json_object_add_value_int(root, "ctx", ts->ctx);
1750         json_object_add_value_int(root, "majf", ts->majf);
1751         json_object_add_value_int(root, "minf", ts->minf);
1752
1753         /* Calc % distribution of IO depths */
1754         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1755         tmp = json_create_object();
1756         json_object_add_value_object(root, "iodepth_level", tmp);
1757         /* Only show fixed 7 I/O depth levels*/
1758         for (i = 0; i < 7; i++) {
1759                 char name[20];
1760                 if (i < 6)
1761                         snprintf(name, 20, "%d", 1 << i);
1762                 else
1763                         snprintf(name, 20, ">=%d", 1 << i);
1764                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1765         }
1766
1767         /* Calc % distribution of submit IO depths */
1768         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1769         tmp = json_create_object();
1770         json_object_add_value_object(root, "iodepth_submit", tmp);
1771         /* Only show fixed 7 I/O depth levels*/
1772         for (i = 0; i < 7; i++) {
1773                 char name[20];
1774                 if (i == 0)
1775                         snprintf(name, 20, "0");
1776                 else if (i < 6)
1777                         snprintf(name, 20, "%d", 1 << (i+1));
1778                 else
1779                         snprintf(name, 20, ">=%d", 1 << i);
1780                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1781         }
1782
1783         /* Calc % distribution of completion IO depths */
1784         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1785         tmp = json_create_object();
1786         json_object_add_value_object(root, "iodepth_complete", tmp);
1787         /* Only show fixed 7 I/O depth levels*/
1788         for (i = 0; i < 7; i++) {
1789                 char name[20];
1790                 if (i == 0)
1791                         snprintf(name, 20, "0");
1792                 else if (i < 6)
1793                         snprintf(name, 20, "%d", 1 << (i+1));
1794                 else
1795                         snprintf(name, 20, ">=%d", 1 << i);
1796                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1797         }
1798
1799         /* Calc % distribution of nsecond, usecond, msecond latency */
1800         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1801         stat_calc_lat_n(ts, io_u_lat_n);
1802         stat_calc_lat_u(ts, io_u_lat_u);
1803         stat_calc_lat_m(ts, io_u_lat_m);
1804
1805         /* Nanosecond latency */
1806         tmp = json_create_object();
1807         json_object_add_value_object(root, "latency_ns", tmp);
1808         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1809                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1810                                  "250", "500", "750", "1000", };
1811                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1812         }
1813         /* Microsecond latency */
1814         tmp = json_create_object();
1815         json_object_add_value_object(root, "latency_us", tmp);
1816         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1817                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1818                                  "250", "500", "750", "1000", };
1819                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1820         }
1821         /* Millisecond latency */
1822         tmp = json_create_object();
1823         json_object_add_value_object(root, "latency_ms", tmp);
1824         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1825                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1826                                  "250", "500", "750", "1000", "2000",
1827                                  ">=2000", };
1828                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1829         }
1830
1831         /* Additional output if continue_on_error set - default off*/
1832         if (ts->continue_on_error) {
1833                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1834                 json_object_add_value_int(root, "first_error", ts->first_error);
1835         }
1836
1837         if (ts->latency_depth) {
1838                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1839                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1840                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1841                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1842         }
1843
1844         /* Additional output if description is set */
1845         if (strlen(ts->description))
1846                 json_object_add_value_string(root, "desc", ts->description);
1847
1848         if (ts->nr_block_infos) {
1849                 /* Block error histogram and types */
1850                 int len;
1851                 unsigned int *percentiles = NULL;
1852                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1853
1854                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1855                                              ts->percentile_list,
1856                                              &percentiles, block_state_counts);
1857
1858                 if (len) {
1859                         struct json_object *block, *percentile_object, *states;
1860                         int state;
1861                         block = json_create_object();
1862                         json_object_add_value_object(root, "block", block);
1863
1864                         percentile_object = json_create_object();
1865                         json_object_add_value_object(block, "percentiles",
1866                                                      percentile_object);
1867                         for (i = 0; i < len; i++) {
1868                                 char buf[20];
1869                                 snprintf(buf, sizeof(buf), "%f",
1870                                          ts->percentile_list[i].u.f);
1871                                 json_object_add_value_int(percentile_object,
1872                                                           buf,
1873                                                           percentiles[i]);
1874                         }
1875
1876                         states = json_create_object();
1877                         json_object_add_value_object(block, "states", states);
1878                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1879                                 json_object_add_value_int(states,
1880                                         block_state_names[state],
1881                                         block_state_counts[state]);
1882                         }
1883                         free(percentiles);
1884                 }
1885         }
1886
1887         if (ts->ss_dur) {
1888                 struct json_object *data;
1889                 struct json_array *iops, *bw;
1890                 int j, k, l;
1891                 char ss_buf[64];
1892                 int intervals = ts->ss_dur / (ss_check_interval / 1000L);
1893
1894                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1895                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1896                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1897                         (float) ts->ss_limit.u.f,
1898                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1899
1900                 tmp = json_create_object();
1901                 json_object_add_value_object(root, "steadystate", tmp);
1902                 json_object_add_value_string(tmp, "ss", ss_buf);
1903                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1904                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1905
1906                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1907                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1908                 json_object_add_value_string(tmp, "criterion", ss_buf);
1909                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1910                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1911
1912                 data = json_create_object();
1913                 json_object_add_value_object(tmp, "data", data);
1914                 bw = json_create_array();
1915                 iops = json_create_array();
1916
1917                 /*
1918                 ** if ss was attained or the buffer is not full,
1919                 ** ss->head points to the first element in the list.
1920                 ** otherwise it actually points to the second element
1921                 ** in the list
1922                 */
1923                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1924                         j = ts->ss_head;
1925                 else
1926                         j = ts->ss_head == 0 ? intervals - 1 : ts->ss_head - 1;
1927                 for (l = 0; l < intervals; l++) {
1928                         k = (j + l) % intervals;
1929                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1930                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1931                 }
1932                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1933                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1934                 json_object_add_value_array(data, "iops", iops);
1935                 json_object_add_value_array(data, "bw", bw);
1936         }
1937
1938         return root;
1939 }
1940
1941 static void show_thread_status_terse(struct thread_stat *ts,
1942                                      struct group_run_stats *rs,
1943                                      struct buf_output *out)
1944 {
1945         if (terse_version >= 2 && terse_version <= 5)
1946                 show_thread_status_terse_all(ts, rs, terse_version, out);
1947         else
1948                 log_err("fio: bad terse version!? %d\n", terse_version);
1949 }
1950
1951 struct json_object *show_thread_status(struct thread_stat *ts,
1952                                        struct group_run_stats *rs,
1953                                        struct flist_head *opt_list,
1954                                        struct buf_output *out)
1955 {
1956         struct json_object *ret = NULL;
1957
1958         if (output_format & FIO_OUTPUT_TERSE)
1959                 show_thread_status_terse(ts, rs,  out);
1960         if (output_format & FIO_OUTPUT_JSON)
1961                 ret = show_thread_status_json(ts, rs, opt_list);
1962         if (output_format & FIO_OUTPUT_NORMAL)
1963                 show_thread_status_normal(ts, rs,  out);
1964
1965         return ret;
1966 }
1967
1968 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1969 {
1970         double mean, S;
1971
1972         dst->min_val = min(dst->min_val, src->min_val);
1973         dst->max_val = max(dst->max_val, src->max_val);
1974
1975         /*
1976          * Compute new mean and S after the merge
1977          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1978          *  #Parallel_algorithm>
1979          */
1980         if (first) {
1981                 mean = src->mean.u.f;
1982                 S = src->S.u.f;
1983         } else {
1984                 double delta = src->mean.u.f - dst->mean.u.f;
1985
1986                 mean = ((src->mean.u.f * src->samples) +
1987                         (dst->mean.u.f * dst->samples)) /
1988                         (dst->samples + src->samples);
1989
1990                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1991                         (dst->samples * src->samples) /
1992                         (dst->samples + src->samples);
1993         }
1994
1995         dst->samples += src->samples;
1996         dst->mean.u.f = mean;
1997         dst->S.u.f = S;
1998
1999 }
2000
2001 /*
2002  * We sum two kinds of stats - one that is time based, in which case we
2003  * apply the proper summing technique, and then one that is iops/bw
2004  * numbers. For group_reporting, we should just add those up, not make
2005  * them the mean of everything.
2006  */
2007 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
2008 {
2009         bool first = dst->samples == 0;
2010
2011         if (src->samples == 0)
2012                 return;
2013
2014         if (!pure_sum) {
2015                 __sum_stat(dst, src, first);
2016                 return;
2017         }
2018
2019         if (first) {
2020                 dst->min_val = src->min_val;
2021                 dst->max_val = src->max_val;
2022                 dst->samples = src->samples;
2023                 dst->mean.u.f = src->mean.u.f;
2024                 dst->S.u.f = src->S.u.f;
2025         } else {
2026                 dst->min_val += src->min_val;
2027                 dst->max_val += src->max_val;
2028                 dst->samples += src->samples;
2029                 dst->mean.u.f += src->mean.u.f;
2030                 dst->S.u.f += src->S.u.f;
2031         }
2032 }
2033
2034 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
2035 {
2036         int i;
2037
2038         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2039                 if (dst->max_run[i] < src->max_run[i])
2040                         dst->max_run[i] = src->max_run[i];
2041                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
2042                         dst->min_run[i] = src->min_run[i];
2043                 if (dst->max_bw[i] < src->max_bw[i])
2044                         dst->max_bw[i] = src->max_bw[i];
2045                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
2046                         dst->min_bw[i] = src->min_bw[i];
2047
2048                 dst->iobytes[i] += src->iobytes[i];
2049                 dst->agg[i] += src->agg[i];
2050         }
2051
2052         if (!dst->kb_base)
2053                 dst->kb_base = src->kb_base;
2054         if (!dst->unit_base)
2055                 dst->unit_base = src->unit_base;
2056         if (!dst->sig_figs)
2057                 dst->sig_figs = src->sig_figs;
2058 }
2059
2060 /*
2061  * Free the clat_prio_stat arrays allocated by alloc_clat_prio_stat_ddir().
2062  */
2063 void free_clat_prio_stats(struct thread_stat *ts)
2064 {
2065         enum fio_ddir ddir;
2066
2067         if (!ts)
2068                 return;
2069
2070         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2071                 sfree(ts->clat_prio[ddir]);
2072                 ts->clat_prio[ddir] = NULL;
2073                 ts->nr_clat_prio[ddir] = 0;
2074         }
2075 }
2076
2077 /*
2078  * Allocate a clat_prio_stat array. The array has to be allocated/freed using
2079  * smalloc/sfree, so that it is accessible by the process/thread summing the
2080  * thread_stats.
2081  */
2082 int alloc_clat_prio_stat_ddir(struct thread_stat *ts, enum fio_ddir ddir,
2083                               int nr_prios)
2084 {
2085         struct clat_prio_stat *clat_prio;
2086         int i;
2087
2088         clat_prio = scalloc(nr_prios, sizeof(*ts->clat_prio[ddir]));
2089         if (!clat_prio) {
2090                 log_err("fio: failed to allocate ts clat data\n");
2091                 return 1;
2092         }
2093
2094         for (i = 0; i < nr_prios; i++)
2095                 clat_prio[i].clat_stat.min_val = ULONG_MAX;
2096
2097         ts->clat_prio[ddir] = clat_prio;
2098         ts->nr_clat_prio[ddir] = nr_prios;
2099
2100         return 0;
2101 }
2102
2103 static int grow_clat_prio_stat(struct thread_stat *dst, enum fio_ddir ddir)
2104 {
2105         int curr_len = dst->nr_clat_prio[ddir];
2106         void *new_arr;
2107
2108         new_arr = scalloc(curr_len + 1, sizeof(*dst->clat_prio[ddir]));
2109         if (!new_arr) {
2110                 log_err("fio: failed to grow clat prio array\n");
2111                 return 1;
2112         }
2113
2114         memcpy(new_arr, dst->clat_prio[ddir],
2115                curr_len * sizeof(*dst->clat_prio[ddir]));
2116         sfree(dst->clat_prio[ddir]);
2117
2118         dst->clat_prio[ddir] = new_arr;
2119         dst->clat_prio[ddir][curr_len].clat_stat.min_val = ULONG_MAX;
2120         dst->nr_clat_prio[ddir]++;
2121
2122         return 0;
2123 }
2124
2125 static int find_clat_prio_index(struct thread_stat *dst, enum fio_ddir ddir,
2126                                 uint32_t ioprio)
2127 {
2128         int i, nr_prios = dst->nr_clat_prio[ddir];
2129
2130         for (i = 0; i < nr_prios; i++) {
2131                 if (dst->clat_prio[ddir][i].ioprio == ioprio)
2132                         return i;
2133         }
2134
2135         return -1;
2136 }
2137
2138 static int alloc_or_get_clat_prio_index(struct thread_stat *dst,
2139                                         enum fio_ddir ddir, uint32_t ioprio,
2140                                         int *idx)
2141 {
2142         int index = find_clat_prio_index(dst, ddir, ioprio);
2143
2144         if (index == -1) {
2145                 index = dst->nr_clat_prio[ddir];
2146
2147                 if (grow_clat_prio_stat(dst, ddir))
2148                         return 1;
2149
2150                 dst->clat_prio[ddir][index].ioprio = ioprio;
2151         }
2152
2153         *idx = index;
2154
2155         return 0;
2156 }
2157
2158 static int clat_prio_stats_copy(struct thread_stat *dst, struct thread_stat *src,
2159                                 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2160 {
2161         size_t sz = sizeof(*src->clat_prio[src_ddir]) *
2162                 src->nr_clat_prio[src_ddir];
2163
2164         dst->clat_prio[dst_ddir] = smalloc(sz);
2165         if (!dst->clat_prio[dst_ddir]) {
2166                 log_err("fio: failed to alloc clat prio array\n");
2167                 return 1;
2168         }
2169
2170         memcpy(dst->clat_prio[dst_ddir], src->clat_prio[src_ddir], sz);
2171         dst->nr_clat_prio[dst_ddir] = src->nr_clat_prio[src_ddir];
2172
2173         return 0;
2174 }
2175
2176 static int clat_prio_stat_add_samples(struct thread_stat *dst,
2177                                       enum fio_ddir dst_ddir, uint32_t ioprio,
2178                                       struct io_stat *io_stat,
2179                                       uint64_t *io_u_plat)
2180 {
2181         int i, dst_index;
2182
2183         if (!io_stat->samples)
2184                 return 0;
2185
2186         if (alloc_or_get_clat_prio_index(dst, dst_ddir, ioprio, &dst_index))
2187                 return 1;
2188
2189         sum_stat(&dst->clat_prio[dst_ddir][dst_index].clat_stat, io_stat,
2190                  false);
2191
2192         for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
2193                 dst->clat_prio[dst_ddir][dst_index].io_u_plat[i] += io_u_plat[i];
2194
2195         return 0;
2196 }
2197
2198 static int sum_clat_prio_stats_src_single_prio(struct thread_stat *dst,
2199                                                struct thread_stat *src,
2200                                                enum fio_ddir dst_ddir,
2201                                                enum fio_ddir src_ddir)
2202 {
2203         struct io_stat *io_stat;
2204         uint64_t *io_u_plat;
2205
2206         /*
2207          * If src ts has no clat_prio_stat array, then all I/Os were submitted
2208          * using src->ioprio. Thus, the global samples in src->clat_stat (or
2209          * src->lat_stat) can be used as the 'per prio' samples for src->ioprio.
2210          */
2211         assert(!src->clat_prio[src_ddir]);
2212         assert(src->nr_clat_prio[src_ddir] == 0);
2213
2214         if (src->lat_percentiles) {
2215                 io_u_plat = src->io_u_plat[FIO_LAT][src_ddir];
2216                 io_stat = &src->lat_stat[src_ddir];
2217         } else {
2218                 io_u_plat = src->io_u_plat[FIO_CLAT][src_ddir];
2219                 io_stat = &src->clat_stat[src_ddir];
2220         }
2221
2222         return clat_prio_stat_add_samples(dst, dst_ddir, src->ioprio, io_stat,
2223                                           io_u_plat);
2224 }
2225
2226 static int sum_clat_prio_stats_src_multi_prio(struct thread_stat *dst,
2227                                               struct thread_stat *src,
2228                                               enum fio_ddir dst_ddir,
2229                                               enum fio_ddir src_ddir)
2230 {
2231         int i;
2232
2233         /*
2234          * If src ts has a clat_prio_stat array, then there are multiple prios
2235          * in use (i.e. src ts had cmdprio_percentage or cmdprio_bssplit set).
2236          * The samples for the default prio will exist in the src->clat_prio
2237          * array, just like the samples for any other prio.
2238          */
2239         assert(src->clat_prio[src_ddir]);
2240         assert(src->nr_clat_prio[src_ddir]);
2241
2242         /* If the dst ts doesn't yet have a clat_prio array, simply memcpy. */
2243         if (!dst->clat_prio[dst_ddir])
2244                 return clat_prio_stats_copy(dst, src, dst_ddir, src_ddir);
2245
2246         /* The dst ts already has a clat_prio_array, add src stats into it. */
2247         for (i = 0; i < src->nr_clat_prio[src_ddir]; i++) {
2248                 struct io_stat *io_stat = &src->clat_prio[src_ddir][i].clat_stat;
2249                 uint64_t *io_u_plat = src->clat_prio[src_ddir][i].io_u_plat;
2250                 uint32_t ioprio = src->clat_prio[src_ddir][i].ioprio;
2251
2252                 if (clat_prio_stat_add_samples(dst, dst_ddir, ioprio, io_stat, io_u_plat))
2253                         return 1;
2254         }
2255
2256         return 0;
2257 }
2258
2259 static int sum_clat_prio_stats(struct thread_stat *dst, struct thread_stat *src,
2260                                enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2261 {
2262         if (dst->disable_prio_stat)
2263                 return 0;
2264
2265         if (!src->clat_prio[src_ddir])
2266                 return sum_clat_prio_stats_src_single_prio(dst, src, dst_ddir,
2267                                                            src_ddir);
2268
2269         return sum_clat_prio_stats_src_multi_prio(dst, src, dst_ddir, src_ddir);
2270 }
2271
2272 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2273 {
2274         int k, l, m;
2275
2276         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2277                 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2278                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2279                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2280                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2281                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2282                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2283                         sum_clat_prio_stats(dst, src, l, l);
2284
2285                         dst->io_bytes[l] += src->io_bytes[l];
2286
2287                         if (dst->runtime[l] < src->runtime[l])
2288                                 dst->runtime[l] = src->runtime[l];
2289                 } else {
2290                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2291                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2292                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2293                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2294                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2295                         sum_clat_prio_stats(dst, src, 0, l);
2296
2297                         dst->io_bytes[0] += src->io_bytes[l];
2298
2299                         if (dst->runtime[0] < src->runtime[l])
2300                                 dst->runtime[0] = src->runtime[l];
2301                 }
2302         }
2303
2304         sum_stat(&dst->sync_stat, &src->sync_stat, false);
2305         dst->usr_time += src->usr_time;
2306         dst->sys_time += src->sys_time;
2307         dst->ctx += src->ctx;
2308         dst->majf += src->majf;
2309         dst->minf += src->minf;
2310
2311         for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2312                 dst->io_u_map[k] += src->io_u_map[k];
2313                 dst->io_u_submit[k] += src->io_u_submit[k];
2314                 dst->io_u_complete[k] += src->io_u_complete[k];
2315         }
2316
2317         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2318                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2319         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2320                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2321         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2322                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2323
2324         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2325                 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2326                         dst->total_io_u[k] += src->total_io_u[k];
2327                         dst->short_io_u[k] += src->short_io_u[k];
2328                         dst->drop_io_u[k] += src->drop_io_u[k];
2329                 } else {
2330                         dst->total_io_u[0] += src->total_io_u[k];
2331                         dst->short_io_u[0] += src->short_io_u[k];
2332                         dst->drop_io_u[0] += src->drop_io_u[k];
2333                 }
2334         }
2335
2336         dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2337
2338         for (k = 0; k < FIO_LAT_CNT; k++)
2339                 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2340                         for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2341                                 if (dst->unified_rw_rep != UNIFIED_MIXED)
2342                                         dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2343                                 else
2344                                         dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2345
2346         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2347                 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2348
2349         dst->total_run_time += src->total_run_time;
2350         dst->total_submit += src->total_submit;
2351         dst->total_complete += src->total_complete;
2352         dst->nr_zone_resets += src->nr_zone_resets;
2353         dst->cachehit += src->cachehit;
2354         dst->cachemiss += src->cachemiss;
2355 }
2356
2357 void init_group_run_stat(struct group_run_stats *gs)
2358 {
2359         int i;
2360         memset(gs, 0, sizeof(*gs));
2361
2362         for (i = 0; i < DDIR_RWDIR_CNT; i++)
2363                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2364 }
2365
2366 void init_thread_stat_min_vals(struct thread_stat *ts)
2367 {
2368         int i;
2369
2370         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2371                 ts->clat_stat[i].min_val = ULONG_MAX;
2372                 ts->slat_stat[i].min_val = ULONG_MAX;
2373                 ts->lat_stat[i].min_val = ULONG_MAX;
2374                 ts->bw_stat[i].min_val = ULONG_MAX;
2375                 ts->iops_stat[i].min_val = ULONG_MAX;
2376         }
2377         ts->sync_stat.min_val = ULONG_MAX;
2378 }
2379
2380 void init_thread_stat(struct thread_stat *ts)
2381 {
2382         memset(ts, 0, sizeof(*ts));
2383
2384         init_thread_stat_min_vals(ts);
2385         ts->groupid = -1;
2386 }
2387
2388 static void init_per_prio_stats(struct thread_stat *threadstats, int nr_ts)
2389 {
2390         struct thread_stat *ts;
2391         int i, j, last_ts, idx;
2392         enum fio_ddir ddir;
2393
2394         j = 0;
2395         last_ts = -1;
2396         idx = 0;
2397
2398         /*
2399          * Loop through all tds, if a td requires per prio stats, temporarily
2400          * store a 1 in ts->disable_prio_stat, and then do an additional
2401          * loop at the end where we invert the ts->disable_prio_stat values.
2402          */
2403         for_each_td(td) {
2404                 if (!td->o.stats)
2405                         continue;
2406                 if (idx &&
2407                     (!td->o.group_reporting ||
2408                      (td->o.group_reporting && last_ts != td->groupid))) {
2409                         idx = 0;
2410                         j++;
2411                 }
2412
2413                 last_ts = td->groupid;
2414                 ts = &threadstats[j];
2415
2416                 /* idx == 0 means first td in group, or td is not in a group. */
2417                 if (idx == 0)
2418                         ts->ioprio = td->ioprio;
2419                 else if (td->ioprio != ts->ioprio)
2420                         ts->disable_prio_stat = 1;
2421
2422                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2423                         if (td->ts.clat_prio[ddir]) {
2424                                 ts->disable_prio_stat = 1;
2425                                 break;
2426                         }
2427                 }
2428
2429                 idx++;
2430         } end_for_each();
2431
2432         /* Loop through all dst threadstats and fixup the values. */
2433         for (i = 0; i < nr_ts; i++) {
2434                 ts = &threadstats[i];
2435                 ts->disable_prio_stat = !ts->disable_prio_stat;
2436         }
2437 }
2438
2439 void __show_run_stats(void)
2440 {
2441         struct group_run_stats *runstats, *rs;
2442         struct thread_stat *threadstats, *ts;
2443         int i, j, k, nr_ts, last_ts, idx;
2444         bool kb_base_warned = false;
2445         bool unit_base_warned = false;
2446         struct json_object *root = NULL;
2447         struct json_array *array = NULL;
2448         struct buf_output output[FIO_OUTPUT_NR];
2449         struct flist_head **opt_lists;
2450
2451         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2452
2453         for (i = 0; i < groupid + 1; i++)
2454                 init_group_run_stat(&runstats[i]);
2455
2456         /*
2457          * find out how many threads stats we need. if group reporting isn't
2458          * enabled, it's one-per-td.
2459          */
2460         nr_ts = 0;
2461         last_ts = -1;
2462         for_each_td(td) {
2463                 if (!td->o.group_reporting) {
2464                         nr_ts++;
2465                         continue;
2466                 }
2467                 if (last_ts == td->groupid)
2468                         continue;
2469                 if (!td->o.stats)
2470                         continue;
2471
2472                 last_ts = td->groupid;
2473                 nr_ts++;
2474         } end_for_each();
2475
2476         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2477         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2478
2479         for (i = 0; i < nr_ts; i++) {
2480                 init_thread_stat(&threadstats[i]);
2481                 opt_lists[i] = NULL;
2482         }
2483
2484         init_per_prio_stats(threadstats, nr_ts);
2485
2486         j = 0;
2487         last_ts = -1;
2488         idx = 0;
2489         for_each_td(td) {
2490                 if (!td->o.stats)
2491                         continue;
2492                 if (idx && (!td->o.group_reporting ||
2493                     (td->o.group_reporting && last_ts != td->groupid))) {
2494                         idx = 0;
2495                         j++;
2496                 }
2497
2498                 last_ts = td->groupid;
2499
2500                 ts = &threadstats[j];
2501
2502                 ts->clat_percentiles = td->o.clat_percentiles;
2503                 ts->lat_percentiles = td->o.lat_percentiles;
2504                 ts->slat_percentiles = td->o.slat_percentiles;
2505                 ts->percentile_precision = td->o.percentile_precision;
2506                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2507                 opt_lists[j] = &td->opt_list;
2508
2509                 idx++;
2510
2511                 if (ts->groupid == -1) {
2512                         /*
2513                          * These are per-group shared already
2514                          */
2515                         snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2516                         if (td->o.description)
2517                                 snprintf(ts->description,
2518                                          sizeof(ts->description), "%s",
2519                                          td->o.description);
2520                         else
2521                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2522
2523                         /*
2524                          * If multiple entries in this group, this is
2525                          * the first member.
2526                          */
2527                         ts->thread_number = td->thread_number;
2528                         ts->groupid = td->groupid;
2529
2530                         /*
2531                          * first pid in group, not very useful...
2532                          */
2533                         ts->pid = td->pid;
2534
2535                         ts->kb_base = td->o.kb_base;
2536                         ts->unit_base = td->o.unit_base;
2537                         ts->sig_figs = td->o.sig_figs;
2538                         ts->unified_rw_rep = td->o.unified_rw_rep;
2539                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2540                         log_info("fio: kb_base differs for jobs in group, using"
2541                                  " %u as the base\n", ts->kb_base);
2542                         kb_base_warned = true;
2543                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2544                         log_info("fio: unit_base differs for jobs in group, using"
2545                                  " %u as the base\n", ts->unit_base);
2546                         unit_base_warned = true;
2547                 }
2548
2549                 ts->continue_on_error = td->o.continue_on_error;
2550                 ts->total_err_count += td->total_err_count;
2551                 ts->first_error = td->first_error;
2552                 if (!ts->error) {
2553                         if (!td->error && td->o.continue_on_error &&
2554                             td->first_error) {
2555                                 ts->error = td->first_error;
2556                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2557                                          td->verror);
2558                         } else  if (td->error) {
2559                                 ts->error = td->error;
2560                                 snprintf(ts->verror, sizeof(ts->verror), "%s",
2561                                          td->verror);
2562                         }
2563                 }
2564
2565                 ts->latency_depth = td->latency_qd;
2566                 ts->latency_target = td->o.latency_target;
2567                 ts->latency_percentile = td->o.latency_percentile;
2568                 ts->latency_window = td->o.latency_window;
2569
2570                 ts->nr_block_infos = td->ts.nr_block_infos;
2571                 for (k = 0; k < ts->nr_block_infos; k++)
2572                         ts->block_infos[k] = td->ts.block_infos[k];
2573
2574                 sum_thread_stats(ts, &td->ts);
2575
2576                 ts->members++;
2577
2578                 if (td->o.ss_dur) {
2579                         ts->ss_state = td->ss.state;
2580                         ts->ss_dur = td->ss.dur;
2581                         ts->ss_head = td->ss.head;
2582                         ts->ss_bw_data = td->ss.bw_data;
2583                         ts->ss_iops_data = td->ss.iops_data;
2584                         ts->ss_limit.u.f = td->ss.limit;
2585                         ts->ss_slope.u.f = td->ss.slope;
2586                         ts->ss_deviation.u.f = td->ss.deviation;
2587                         ts->ss_criterion.u.f = td->ss.criterion;
2588                 }
2589                 else
2590                         ts->ss_dur = ts->ss_state = 0;
2591         } end_for_each();
2592
2593         for (i = 0; i < nr_ts; i++) {
2594                 unsigned long long bw;
2595
2596                 ts = &threadstats[i];
2597                 if (ts->groupid == -1)
2598                         continue;
2599                 rs = &runstats[ts->groupid];
2600                 rs->kb_base = ts->kb_base;
2601                 rs->unit_base = ts->unit_base;
2602                 rs->sig_figs = ts->sig_figs;
2603                 rs->unified_rw_rep |= ts->unified_rw_rep;
2604
2605                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2606                         if (!ts->runtime[j])
2607                                 continue;
2608                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2609                                 rs->min_run[j] = ts->runtime[j];
2610                         if (ts->runtime[j] > rs->max_run[j])
2611                                 rs->max_run[j] = ts->runtime[j];
2612
2613                         bw = 0;
2614                         if (ts->runtime[j])
2615                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2616                         if (bw < rs->min_bw[j])
2617                                 rs->min_bw[j] = bw;
2618                         if (bw > rs->max_bw[j])
2619                                 rs->max_bw[j] = bw;
2620
2621                         rs->iobytes[j] += ts->io_bytes[j];
2622                 }
2623         }
2624
2625         for (i = 0; i < groupid + 1; i++) {
2626                 enum fio_ddir ddir;
2627
2628                 rs = &runstats[i];
2629
2630                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2631                         if (rs->max_run[ddir])
2632                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2633                                                 rs->max_run[ddir];
2634                 }
2635         }
2636
2637         for (i = 0; i < FIO_OUTPUT_NR; i++)
2638                 buf_output_init(&output[i]);
2639
2640         /*
2641          * don't overwrite last signal output
2642          */
2643         if (output_format & FIO_OUTPUT_NORMAL)
2644                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2645         if (output_format & FIO_OUTPUT_JSON) {
2646                 struct thread_data *global;
2647                 char time_buf[32];
2648                 struct timeval now;
2649                 unsigned long long ms_since_epoch;
2650                 time_t tv_sec;
2651
2652                 gettimeofday(&now, NULL);
2653                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2654                                  (unsigned long long)(now.tv_usec) / 1000;
2655
2656                 tv_sec = now.tv_sec;
2657                 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2658                 if (time_buf[strlen(time_buf) - 1] == '\n')
2659                         time_buf[strlen(time_buf) - 1] = '\0';
2660
2661                 root = json_create_object();
2662                 json_object_add_value_string(root, "fio version", fio_version_string);
2663                 json_object_add_value_int(root, "timestamp", now.tv_sec);
2664                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2665                 json_object_add_value_string(root, "time", time_buf);
2666                 global = get_global_options();
2667                 json_add_job_opts(root, "global options", &global->opt_list);
2668                 array = json_create_array();
2669                 json_object_add_value_array(root, "jobs", array);
2670         }
2671
2672         if (is_backend)
2673                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2674
2675         for (i = 0; i < nr_ts; i++) {
2676                 ts = &threadstats[i];
2677                 rs = &runstats[ts->groupid];
2678
2679                 if (is_backend) {
2680                         fio_server_send_job_options(opt_lists[i], i);
2681                         fio_server_send_ts(ts, rs);
2682                 } else {
2683                         if (output_format & FIO_OUTPUT_TERSE)
2684                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2685                         if (output_format & FIO_OUTPUT_JSON) {
2686                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2687                                 json_array_add_value_object(array, tmp);
2688                         }
2689                         if (output_format & FIO_OUTPUT_NORMAL)
2690                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2691                 }
2692         }
2693         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2694                 /* disk util stats, if any */
2695                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2696
2697                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2698
2699                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2700                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2701                 json_free_object(root);
2702         }
2703
2704         for (i = 0; i < groupid + 1; i++) {
2705                 rs = &runstats[i];
2706
2707                 rs->groupid = i;
2708                 if (is_backend)
2709                         fio_server_send_gs(rs);
2710                 else if (output_format & FIO_OUTPUT_NORMAL)
2711                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2712         }
2713
2714         if (is_backend)
2715                 fio_server_send_du();
2716         else if (output_format & FIO_OUTPUT_NORMAL) {
2717                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2718                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2719         }
2720
2721         for (i = 0; i < FIO_OUTPUT_NR; i++) {
2722                 struct buf_output *out = &output[i];
2723
2724                 log_info_buf(out->buf, out->buflen);
2725                 buf_output_free(out);
2726         }
2727
2728         fio_idle_prof_cleanup();
2729
2730         log_info_flush();
2731         free(runstats);
2732
2733         /* free arrays allocated by sum_thread_stats(), if any */
2734         for (i = 0; i < nr_ts; i++) {
2735                 ts = &threadstats[i];
2736                 free_clat_prio_stats(ts);
2737         }
2738         free(threadstats);
2739         free(opt_lists);
2740 }
2741
2742 int __show_running_run_stats(void)
2743 {
2744         unsigned long long *rt;
2745         struct timespec ts;
2746
2747         fio_sem_down(stat_sem);
2748
2749         rt = malloc(thread_number * sizeof(unsigned long long));
2750         fio_gettime(&ts, NULL);
2751
2752         for_each_td(td) {
2753                 if (td->runstate >= TD_EXITED)
2754                         continue;
2755
2756                 td->update_rusage = 1;
2757                 for_each_rw_ddir(ddir) {
2758                         td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2759                 }
2760                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2761
2762                 rt[__td_index] = mtime_since(&td->start, &ts);
2763                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2764                         td->ts.runtime[DDIR_READ] += rt[__td_index];
2765                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2766                         td->ts.runtime[DDIR_WRITE] += rt[__td_index];
2767                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2768                         td->ts.runtime[DDIR_TRIM] += rt[__td_index];
2769         } end_for_each();
2770
2771         for_each_td(td) {
2772                 if (td->runstate >= TD_EXITED)
2773                         continue;
2774                 if (td->rusage_sem) {
2775                         td->update_rusage = 1;
2776                         fio_sem_down(td->rusage_sem);
2777                 }
2778                 td->update_rusage = 0;
2779         } end_for_each();
2780
2781         __show_run_stats();
2782
2783         for_each_td(td) {
2784                 if (td->runstate >= TD_EXITED)
2785                         continue;
2786
2787                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2788                         td->ts.runtime[DDIR_READ] -= rt[__td_index];
2789                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2790                         td->ts.runtime[DDIR_WRITE] -= rt[__td_index];
2791                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2792                         td->ts.runtime[DDIR_TRIM] -= rt[__td_index];
2793         } end_for_each();
2794
2795         free(rt);
2796         fio_sem_up(stat_sem);
2797
2798         return 0;
2799 }
2800
2801 static bool status_file_disabled;
2802
2803 #define FIO_STATUS_FILE         "fio-dump-status"
2804
2805 static int check_status_file(void)
2806 {
2807         struct stat sb;
2808         const char *temp_dir;
2809         char fio_status_file_path[PATH_MAX];
2810
2811         if (status_file_disabled)
2812                 return 0;
2813
2814         temp_dir = getenv("TMPDIR");
2815         if (temp_dir == NULL) {
2816                 temp_dir = getenv("TEMP");
2817                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2818                         temp_dir = NULL;
2819         }
2820         if (temp_dir == NULL)
2821                 temp_dir = "/tmp";
2822 #ifdef __COVERITY__
2823         __coverity_tainted_data_sanitize__(temp_dir);
2824 #endif
2825
2826         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2827
2828         if (stat(fio_status_file_path, &sb))
2829                 return 0;
2830
2831         if (unlink(fio_status_file_path) < 0) {
2832                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2833                                                         strerror(errno));
2834                 log_err("fio: disabling status file updates\n");
2835                 status_file_disabled = true;
2836         }
2837
2838         return 1;
2839 }
2840
2841 void check_for_running_stats(void)
2842 {
2843         if (check_status_file()) {
2844                 show_running_run_stats();
2845                 return;
2846         }
2847 }
2848
2849 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2850 {
2851         double val = data;
2852         double delta;
2853
2854         if (data > is->max_val)
2855                 is->max_val = data;
2856         if (data < is->min_val)
2857                 is->min_val = data;
2858
2859         delta = val - is->mean.u.f;
2860         if (delta) {
2861                 is->mean.u.f += delta / (is->samples + 1.0);
2862                 is->S.u.f += delta * (val - is->mean.u.f);
2863         }
2864
2865         is->samples++;
2866 }
2867
2868 static inline void add_stat_prio_sample(struct clat_prio_stat *clat_prio,
2869                                         unsigned short clat_prio_index,
2870                                         unsigned long long nsec)
2871 {
2872         if (clat_prio)
2873                 add_stat_sample(&clat_prio[clat_prio_index].clat_stat, nsec);
2874 }
2875
2876 /*
2877  * Return a struct io_logs, which is added to the tail of the log
2878  * list for 'iolog'.
2879  */
2880 static struct io_logs *get_new_log(struct io_log *iolog)
2881 {
2882         size_t new_samples;
2883         struct io_logs *cur_log;
2884
2885         /*
2886          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2887          * forever
2888          */
2889         if (!iolog->cur_log_max) {
2890                 if (iolog->td)
2891                         new_samples = iolog->td->o.log_entries;
2892                 else
2893                         new_samples = DEF_LOG_ENTRIES;
2894         } else {
2895                 new_samples = iolog->cur_log_max * 2;
2896                 if (new_samples > MAX_LOG_ENTRIES)
2897                         new_samples = MAX_LOG_ENTRIES;
2898         }
2899
2900         cur_log = smalloc(sizeof(*cur_log));
2901         if (cur_log) {
2902                 INIT_FLIST_HEAD(&cur_log->list);
2903                 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2904                 if (cur_log->log) {
2905                         cur_log->nr_samples = 0;
2906                         cur_log->max_samples = new_samples;
2907                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2908                         iolog->cur_log_max = new_samples;
2909                         return cur_log;
2910                 }
2911                 sfree(cur_log);
2912         }
2913
2914         return NULL;
2915 }
2916
2917 /*
2918  * Add and return a new log chunk, or return current log if big enough
2919  */
2920 static struct io_logs *regrow_log(struct io_log *iolog)
2921 {
2922         struct io_logs *cur_log;
2923         int i;
2924
2925         if (!iolog || iolog->disabled)
2926                 goto disable;
2927
2928         cur_log = iolog_cur_log(iolog);
2929         if (!cur_log) {
2930                 cur_log = get_new_log(iolog);
2931                 if (!cur_log)
2932                         return NULL;
2933         }
2934
2935         if (cur_log->nr_samples < cur_log->max_samples)
2936                 return cur_log;
2937
2938         /*
2939          * No room for a new sample. If we're compressing on the fly, flush
2940          * out the current chunk
2941          */
2942         if (iolog->log_gz) {
2943                 if (iolog_cur_flush(iolog, cur_log)) {
2944                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2945                         return NULL;
2946                 }
2947         }
2948
2949         /*
2950          * Get a new log array, and add to our list
2951          */
2952         cur_log = get_new_log(iolog);
2953         if (!cur_log) {
2954                 log_err("fio: failed extending iolog! Will stop logging.\n");
2955                 return NULL;
2956         }
2957
2958         if (!iolog->pending || !iolog->pending->nr_samples)
2959                 return cur_log;
2960
2961         /*
2962          * Flush pending items to new log
2963          */
2964         for (i = 0; i < iolog->pending->nr_samples; i++) {
2965                 struct io_sample *src, *dst;
2966
2967                 src = get_sample(iolog, iolog->pending, i);
2968                 dst = get_sample(iolog, cur_log, i);
2969                 memcpy(dst, src, log_entry_sz(iolog));
2970         }
2971         cur_log->nr_samples = iolog->pending->nr_samples;
2972
2973         iolog->pending->nr_samples = 0;
2974         return cur_log;
2975 disable:
2976         if (iolog)
2977                 iolog->disabled = true;
2978         return NULL;
2979 }
2980
2981 void regrow_logs(struct thread_data *td)
2982 {
2983         regrow_log(td->slat_log);
2984         regrow_log(td->clat_log);
2985         regrow_log(td->clat_hist_log);
2986         regrow_log(td->lat_log);
2987         regrow_log(td->bw_log);
2988         regrow_log(td->iops_log);
2989         td->flags &= ~TD_F_REGROW_LOGS;
2990 }
2991
2992 void regrow_agg_logs(void)
2993 {
2994         enum fio_ddir ddir;
2995
2996         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2997                 regrow_log(agg_io_log[ddir]);
2998 }
2999
3000 static struct io_logs *get_cur_log(struct io_log *iolog)
3001 {
3002         struct io_logs *cur_log;
3003
3004         cur_log = iolog_cur_log(iolog);
3005         if (!cur_log) {
3006                 cur_log = get_new_log(iolog);
3007                 if (!cur_log)
3008                         return NULL;
3009         }
3010
3011         if (cur_log->nr_samples < cur_log->max_samples)
3012                 return cur_log;
3013
3014         /*
3015          * Out of space. If we're in IO offload mode, or we're not doing
3016          * per unit logging (hence logging happens outside of the IO thread
3017          * as well), add a new log chunk inline. If we're doing inline
3018          * submissions, flag 'td' as needing a log regrow and we'll take
3019          * care of it on the submission side.
3020          */
3021         if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
3022             !per_unit_log(iolog))
3023                 return regrow_log(iolog);
3024
3025         if (iolog->td)
3026                 iolog->td->flags |= TD_F_REGROW_LOGS;
3027         if (iolog->pending)
3028                 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
3029         return iolog->pending;
3030 }
3031
3032 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
3033                              enum fio_ddir ddir, unsigned long long bs,
3034                              unsigned long t, uint64_t offset,
3035                              unsigned int priority)
3036 {
3037         struct io_logs *cur_log;
3038
3039         if (iolog->disabled)
3040                 return;
3041         if (flist_empty(&iolog->io_logs))
3042                 iolog->avg_last[ddir] = t;
3043
3044         cur_log = get_cur_log(iolog);
3045         if (cur_log) {
3046                 struct io_sample *s;
3047
3048                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
3049
3050                 s->data = data;
3051                 s->time = t + (iolog->td ? iolog->td->alternate_epoch : 0);
3052                 io_sample_set_ddir(iolog, s, ddir);
3053                 s->bs = bs;
3054                 s->priority = priority;
3055
3056                 if (iolog->log_offset) {
3057                         struct io_sample_offset *so = (void *) s;
3058
3059                         so->offset = offset;
3060                 }
3061
3062                 cur_log->nr_samples++;
3063                 return;
3064         }
3065
3066         iolog->disabled = true;
3067 }
3068
3069 static inline void reset_io_stat(struct io_stat *ios)
3070 {
3071         ios->min_val = -1ULL;
3072         ios->max_val = ios->samples = 0;
3073         ios->mean.u.f = ios->S.u.f = 0;
3074 }
3075
3076 static inline void reset_io_u_plat(uint64_t *io_u_plat)
3077 {
3078         int i;
3079
3080         for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
3081                 io_u_plat[i] = 0;
3082 }
3083
3084 static inline void reset_clat_prio_stats(struct thread_stat *ts)
3085 {
3086         enum fio_ddir ddir;
3087         int i;
3088
3089         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3090                 if (!ts->clat_prio[ddir])
3091                         continue;
3092
3093                 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
3094                         reset_io_stat(&ts->clat_prio[ddir][i].clat_stat);
3095                         reset_io_u_plat(ts->clat_prio[ddir][i].io_u_plat);
3096                 }
3097         }
3098 }
3099
3100 void reset_io_stats(struct thread_data *td)
3101 {
3102         struct thread_stat *ts = &td->ts;
3103         int i, j;
3104
3105         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
3106                 reset_io_stat(&ts->clat_stat[i]);
3107                 reset_io_stat(&ts->slat_stat[i]);
3108                 reset_io_stat(&ts->lat_stat[i]);
3109                 reset_io_stat(&ts->bw_stat[i]);
3110                 reset_io_stat(&ts->iops_stat[i]);
3111
3112                 ts->io_bytes[i] = 0;
3113                 ts->runtime[i] = 0;
3114                 ts->total_io_u[i] = 0;
3115                 ts->short_io_u[i] = 0;
3116                 ts->drop_io_u[i] = 0;
3117         }
3118
3119         for (i = 0; i < FIO_LAT_CNT; i++)
3120                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
3121                         reset_io_u_plat(ts->io_u_plat[i][j]);
3122
3123         reset_clat_prio_stats(ts);
3124
3125         ts->total_io_u[DDIR_SYNC] = 0;
3126         reset_io_u_plat(ts->io_u_sync_plat);
3127
3128         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
3129                 ts->io_u_map[i] = 0;
3130                 ts->io_u_submit[i] = 0;
3131                 ts->io_u_complete[i] = 0;
3132         }
3133
3134         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
3135                 ts->io_u_lat_n[i] = 0;
3136         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
3137                 ts->io_u_lat_u[i] = 0;
3138         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
3139                 ts->io_u_lat_m[i] = 0;
3140
3141         ts->total_submit = 0;
3142         ts->total_complete = 0;
3143         ts->nr_zone_resets = 0;
3144         ts->cachehit = ts->cachemiss = 0;
3145 }
3146
3147 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
3148                               unsigned long elapsed, bool log_max)
3149 {
3150         /*
3151          * Note an entry in the log. Use the mean from the logged samples,
3152          * making sure to properly round up. Only write a log entry if we
3153          * had actual samples done.
3154          */
3155         if (iolog->avg_window[ddir].samples) {
3156                 union io_sample_data data;
3157
3158                 if (log_max)
3159                         data.val = iolog->avg_window[ddir].max_val;
3160                 else
3161                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
3162
3163                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
3164         }
3165
3166         reset_io_stat(&iolog->avg_window[ddir]);
3167 }
3168
3169 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
3170                              bool log_max)
3171 {
3172         enum fio_ddir ddir;
3173
3174         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
3175                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
3176 }
3177
3178 static unsigned long add_log_sample(struct thread_data *td,
3179                                     struct io_log *iolog,
3180                                     union io_sample_data data,
3181                                     enum fio_ddir ddir, unsigned long long bs,
3182                                     uint64_t offset, unsigned int ioprio)
3183 {
3184         unsigned long elapsed, this_window;
3185
3186         if (!ddir_rw(ddir))
3187                 return 0;
3188
3189         elapsed = mtime_since_now(&td->epoch);
3190
3191         /*
3192          * If no time averaging, just add the log sample.
3193          */
3194         if (!iolog->avg_msec) {
3195                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
3196                                  ioprio);
3197                 return 0;
3198         }
3199
3200         /*
3201          * Add the sample. If the time period has passed, then
3202          * add that entry to the log and clear.
3203          */
3204         add_stat_sample(&iolog->avg_window[ddir], data.val);
3205
3206         /*
3207          * If period hasn't passed, adding the above sample is all we
3208          * need to do.
3209          */
3210         this_window = elapsed - iolog->avg_last[ddir];
3211         if (elapsed < iolog->avg_last[ddir])
3212                 return iolog->avg_last[ddir] - elapsed;
3213         else if (this_window < iolog->avg_msec) {
3214                 unsigned long diff = iolog->avg_msec - this_window;
3215
3216                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
3217                         return diff;
3218         }
3219
3220         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
3221
3222         iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
3223
3224         return iolog->avg_msec;
3225 }
3226
3227 void finalize_logs(struct thread_data *td, bool unit_logs)
3228 {
3229         unsigned long elapsed;
3230
3231         elapsed = mtime_since_now(&td->epoch);
3232
3233         if (td->clat_log && unit_logs)
3234                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
3235         if (td->slat_log && unit_logs)
3236                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
3237         if (td->lat_log && unit_logs)
3238                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
3239         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
3240                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
3241         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
3242                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
3243 }
3244
3245 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
3246                     unsigned long long bs)
3247 {
3248         struct io_log *iolog;
3249
3250         if (!ddir_rw(ddir))
3251                 return;
3252
3253         iolog = agg_io_log[ddir];
3254         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
3255 }
3256
3257 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
3258 {
3259         unsigned int idx = plat_val_to_idx(nsec);
3260         assert(idx < FIO_IO_U_PLAT_NR);
3261
3262         ts->io_u_sync_plat[idx]++;
3263         add_stat_sample(&ts->sync_stat, nsec);
3264 }
3265
3266 static inline void add_lat_percentile_sample(struct thread_stat *ts,
3267                                              unsigned long long nsec,
3268                                              enum fio_ddir ddir,
3269                                              enum fio_lat lat)
3270 {
3271         unsigned int idx = plat_val_to_idx(nsec);
3272         assert(idx < FIO_IO_U_PLAT_NR);
3273
3274         ts->io_u_plat[lat][ddir][idx]++;
3275 }
3276
3277 static inline void
3278 add_lat_percentile_prio_sample(struct thread_stat *ts, unsigned long long nsec,
3279                                enum fio_ddir ddir,
3280                                unsigned short clat_prio_index)
3281 {
3282         unsigned int idx = plat_val_to_idx(nsec);
3283
3284         if (ts->clat_prio[ddir])
3285                 ts->clat_prio[ddir][clat_prio_index].io_u_plat[idx]++;
3286 }
3287
3288 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
3289                      unsigned long long nsec, unsigned long long bs,
3290                      uint64_t offset, unsigned int ioprio,
3291                      unsigned short clat_prio_index)
3292 {
3293         const bool needs_lock = td_async_processing(td);
3294         unsigned long elapsed, this_window;
3295         struct thread_stat *ts = &td->ts;
3296         struct io_log *iolog = td->clat_hist_log;
3297
3298         if (needs_lock)
3299                 __td_io_u_lock(td);
3300
3301         add_stat_sample(&ts->clat_stat[ddir], nsec);
3302
3303         /*
3304          * When lat_percentiles=1 (default 0), the reported per priority
3305          * percentiles and stats are used for describing total latency values,
3306          * even though the variable names themselves start with clat_.
3307          *
3308          * Because of the above definition, add a prio stat sample only when
3309          * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3310          * when lat_percentiles=1.
3311          */
3312         if (!ts->lat_percentiles)
3313                 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3314                                      nsec);
3315
3316         if (td->clat_log)
3317                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3318                                offset, ioprio);
3319
3320         if (ts->clat_percentiles) {
3321                 /*
3322                  * Because of the above definition, add a prio lat percentile
3323                  * sample only when lat_percentiles=0. add_lat_sample() will add
3324                  * the prio lat percentile sample when lat_percentiles=1.
3325                  */
3326                 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3327                 if (!ts->lat_percentiles)
3328                         add_lat_percentile_prio_sample(ts, nsec, ddir,
3329                                                        clat_prio_index);
3330         }
3331
3332         if (iolog && iolog->hist_msec) {
3333                 struct io_hist *hw = &iolog->hist_window[ddir];
3334
3335                 hw->samples++;
3336                 elapsed = mtime_since_now(&td->epoch);
3337                 if (!hw->hist_last)
3338                         hw->hist_last = elapsed;
3339                 this_window = elapsed - hw->hist_last;
3340
3341                 if (this_window >= iolog->hist_msec) {
3342                         uint64_t *io_u_plat;
3343                         struct io_u_plat_entry *dst;
3344
3345                         /*
3346                          * Make a byte-for-byte copy of the latency histogram
3347                          * stored in td->ts.io_u_plat[ddir], recording it in a
3348                          * log sample. Note that the matching call to free() is
3349                          * located in iolog.c after printing this sample to the
3350                          * log file.
3351                          */
3352                         io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3353                         dst = malloc(sizeof(struct io_u_plat_entry));
3354                         memcpy(&(dst->io_u_plat), io_u_plat,
3355                                 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3356                         flist_add(&dst->list, &hw->list);
3357                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3358                                          elapsed, offset, ioprio);
3359
3360                         /*
3361                          * Update the last time we recorded as being now, minus
3362                          * any drift in time we encountered before actually
3363                          * making the record.
3364                          */
3365                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3366                         hw->samples = 0;
3367                 }
3368         }
3369
3370         if (needs_lock)
3371                 __td_io_u_unlock(td);
3372 }
3373
3374 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3375                      unsigned long long nsec, unsigned long long bs,
3376                      uint64_t offset, unsigned int ioprio)
3377 {
3378         const bool needs_lock = td_async_processing(td);
3379         struct thread_stat *ts = &td->ts;
3380
3381         if (!ddir_rw(ddir))
3382                 return;
3383
3384         if (needs_lock)
3385                 __td_io_u_lock(td);
3386
3387         add_stat_sample(&ts->slat_stat[ddir], nsec);
3388
3389         if (td->slat_log)
3390                 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3391                                offset, ioprio);
3392
3393         if (ts->slat_percentiles)
3394                 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3395
3396         if (needs_lock)
3397                 __td_io_u_unlock(td);
3398 }
3399
3400 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3401                     unsigned long long nsec, unsigned long long bs,
3402                     uint64_t offset, unsigned int ioprio,
3403                     unsigned short clat_prio_index)
3404 {
3405         const bool needs_lock = td_async_processing(td);
3406         struct thread_stat *ts = &td->ts;
3407
3408         if (!ddir_rw(ddir))
3409                 return;
3410
3411         if (needs_lock)
3412                 __td_io_u_lock(td);
3413
3414         add_stat_sample(&ts->lat_stat[ddir], nsec);
3415
3416         if (td->lat_log)
3417                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3418                                offset, ioprio);
3419
3420         /*
3421          * When lat_percentiles=1 (default 0), the reported per priority
3422          * percentiles and stats are used for describing total latency values,
3423          * even though the variable names themselves start with clat_.
3424          *
3425          * Because of the above definition, add a prio stat and prio lat
3426          * percentile sample only when lat_percentiles=1. add_clat_sample() will
3427          * add the prio stat and prio lat percentile sample when
3428          * lat_percentiles=0.
3429          */
3430         if (ts->lat_percentiles) {
3431                 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3432                 add_lat_percentile_prio_sample(ts, nsec, ddir, clat_prio_index);
3433                 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3434                                      nsec);
3435         }
3436         if (needs_lock)
3437                 __td_io_u_unlock(td);
3438 }
3439
3440 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3441                    unsigned int bytes, unsigned long long spent)
3442 {
3443         const bool needs_lock = td_async_processing(td);
3444         struct thread_stat *ts = &td->ts;
3445         unsigned long rate;
3446
3447         if (spent)
3448                 rate = (unsigned long) (bytes * 1000000ULL / spent);
3449         else
3450                 rate = 0;
3451
3452         if (needs_lock)
3453                 __td_io_u_lock(td);
3454
3455         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3456
3457         if (td->bw_log)
3458                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3459                                bytes, io_u->offset, io_u->ioprio);
3460
3461         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3462
3463         if (needs_lock)
3464                 __td_io_u_unlock(td);
3465 }
3466
3467 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3468                          struct timespec *t, unsigned int avg_time,
3469                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3470                          struct io_stat *stat, struct io_log *log,
3471                          bool is_kb)
3472 {
3473         const bool needs_lock = td_async_processing(td);
3474         unsigned long spent, rate;
3475         enum fio_ddir ddir;
3476         unsigned long next, next_log;
3477
3478         next_log = avg_time;
3479
3480         spent = mtime_since(parent_tv, t);
3481         if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3482                 return avg_time - spent;
3483
3484         if (needs_lock)
3485                 __td_io_u_lock(td);
3486
3487         /*
3488          * Compute both read and write rates for the interval.
3489          */
3490         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3491                 uint64_t delta;
3492
3493                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3494                 if (!delta)
3495                         continue; /* No entries for interval */
3496
3497                 if (spent) {
3498                         if (is_kb)
3499                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
3500                         else
3501                                 rate = (delta * 1000) / spent;
3502                 } else
3503                         rate = 0;
3504
3505                 add_stat_sample(&stat[ddir], rate);
3506
3507                 if (log) {
3508                         unsigned long long bs = 0;
3509
3510                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3511                                 bs = td->o.min_bs[ddir];
3512
3513                         next = add_log_sample(td, log, sample_val(rate), ddir,
3514                                               bs, 0, 0);
3515                         next_log = min(next_log, next);
3516                 }
3517
3518                 stat_io_bytes[ddir] = this_io_bytes[ddir];
3519         }
3520
3521         *parent_tv = *t;
3522
3523         if (needs_lock)
3524                 __td_io_u_unlock(td);
3525
3526         if (spent <= avg_time)
3527                 next = avg_time;
3528         else
3529                 next = avg_time - (1 + spent - avg_time);
3530
3531         return min(next, next_log);
3532 }
3533
3534 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3535 {
3536         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3537                                 td->this_io_bytes, td->stat_io_bytes,
3538                                 td->ts.bw_stat, td->bw_log, true);
3539 }
3540
3541 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3542                      unsigned int bytes)
3543 {
3544         const bool needs_lock = td_async_processing(td);
3545         struct thread_stat *ts = &td->ts;
3546
3547         if (needs_lock)
3548                 __td_io_u_lock(td);
3549
3550         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3551
3552         if (td->iops_log)
3553                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3554                                bytes, io_u->offset, io_u->ioprio);
3555
3556         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3557
3558         if (needs_lock)
3559                 __td_io_u_unlock(td);
3560 }
3561
3562 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3563 {
3564         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3565                                 td->this_io_blocks, td->stat_io_blocks,
3566                                 td->ts.iops_stat, td->iops_log, false);
3567 }
3568
3569 /*
3570  * Returns msecs to next event
3571  */
3572 int calc_log_samples(void)
3573 {
3574         unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3575         struct timespec now;
3576         long elapsed_time = 0;
3577
3578         fio_gettime(&now, NULL);
3579
3580         for_each_td(td) {
3581                 elapsed_time = mtime_since_now(&td->epoch);
3582
3583                 if (!td->o.stats)
3584                         continue;
3585                 if (in_ramp_time(td) ||
3586                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3587                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3588                         continue;
3589                 }
3590                 if (!td->bw_log ||
3591                         (td->bw_log && !per_unit_log(td->bw_log))) {
3592                         tmp = add_bw_samples(td, &now);
3593
3594                         if (td->bw_log)
3595                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3596                 }
3597                 if (!td->iops_log ||
3598                         (td->iops_log && !per_unit_log(td->iops_log))) {
3599                         tmp = add_iops_samples(td, &now);
3600
3601                         if (td->iops_log)
3602                                 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3603                 }
3604
3605                 if (tmp < next)
3606                         next = tmp;
3607         } end_for_each();
3608
3609         /* if log_avg_msec_min has not been changed, set it to 0 */
3610         if (log_avg_msec_min == -1U)
3611                 log_avg_msec_min = 0;
3612
3613         if (log_avg_msec_min == 0)
3614                 next_mod = elapsed_time;
3615         else
3616                 next_mod = elapsed_time % log_avg_msec_min;
3617
3618         /* correction to keep the time on the log avg msec boundary */
3619         next = min(next, (log_avg_msec_min - next_mod));
3620
3621         return next == ~0U ? 0 : next;
3622 }
3623
3624 void stat_init(void)
3625 {
3626         stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3627 }
3628
3629 void stat_exit(void)
3630 {
3631         /*
3632          * When we have the mutex, we know out-of-band access to it
3633          * have ended.
3634          */
3635         fio_sem_down(stat_sem);
3636         fio_sem_remove(stat_sem);
3637 }
3638
3639 /*
3640  * Called from signal handler. Wake up status thread.
3641  */
3642 void show_running_run_stats(void)
3643 {
3644         helper_do_stat();
3645 }
3646
3647 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3648 {
3649         /* Ignore io_u's which span multiple blocks--they will just get
3650          * inaccurate counts. */
3651         int idx = (io_u->offset - io_u->file->file_offset)
3652                         / td->o.bs[DDIR_TRIM];
3653         uint32_t *info = &td->ts.block_infos[idx];
3654         assert(idx < td->ts.nr_block_infos);
3655         return info;
3656 }
3657