stat.c

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