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