stat.c

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