stat.c

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