stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/types.h>
   5 #include <sys/stat.h>
   6 #include <dirent.h>
   7 #include <libgen.h>
   8 #include <math.h>
   9
  10 #include "fio.h"
  11 #include "diskutil.h"
  12 #include "lib/ieee754.h"
  13 #include "json.h"
  14 #include "lib/getrusage.h"
  15 #include "idletime.h"
  16
  17 struct fio_mutex *stat_mutex;
  18
  19 void update_rusage_stat(struct thread_data *td)
  20 {
  21         struct thread_stat *ts = &td->ts;
  22
  23         fio_getrusage(&td->ru_end);
  24         ts->usr_time += mtime_since(&td->ru_start.ru_utime,
  25                                         &td->ru_end.ru_utime);
  26         ts->sys_time += mtime_since(&td->ru_start.ru_stime,
  27                                         &td->ru_end.ru_stime);
  28         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  29                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  30         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  31         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  32
  33         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  34 }
  35
  36 /*
  37  * Given a latency, return the index of the corresponding bucket in
  38  * the structure tracking percentiles.
  39  *
  40  * (1) find the group (and error bits) that the value (latency)
  41  * belongs to by looking at its MSB. (2) find the bucket number in the
  42  * group by looking at the index bits.
  43  *
  44  */
  45 static unsigned int plat_val_to_idx(unsigned int val)
  46 {
  47         unsigned int msb, error_bits, base, offset, idx;
  48
  49         /* Find MSB starting from bit 0 */
  50         if (val == 0)
  51                 msb = 0;
  52         else
  53                 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
  54
  55         /*
  56          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  57          * all bits of the sample as index
  58          */
  59         if (msb <= FIO_IO_U_PLAT_BITS)
  60                 return val;
  61
  62         /* Compute the number of error bits to discard*/
  63         error_bits = msb - FIO_IO_U_PLAT_BITS;
  64
  65         /* Compute the number of buckets before the group */
  66         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  67
  68         /*
  69          * Discard the error bits and apply the mask to find the
  70          * index for the buckets in the group
  71          */
  72         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  73
  74         /* Make sure the index does not exceed (array size - 1) */
  75         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  76                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  77
  78         return idx;
  79 }
  80
  81 /*
  82  * Convert the given index of the bucket array to the value
  83  * represented by the bucket
  84  */
  85 static unsigned int plat_idx_to_val(unsigned int idx)
  86 {
  87         unsigned int error_bits, k, base;
  88
  89         assert(idx < FIO_IO_U_PLAT_NR);
  90
  91         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  92          * all bits of the sample as index */
  93         if (idx < (FIO_IO_U_PLAT_VAL << 1))
  94                 return idx;
  95
  96         /* Find the group and compute the minimum value of that group */
  97         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
  98         base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
  99
 100         /* Find its bucket number of the group */
 101         k = idx % FIO_IO_U_PLAT_VAL;
 102
 103         /* Return the mean of the range of the bucket */
 104         return base + ((k + 0.5) * (1 << error_bits));
 105 }
 106
 107 static int double_cmp(const void *a, const void *b)
 108 {
 109         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 110         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 111         int cmp = 0;
 112
 113         if (fa.u.f > fb.u.f)
 114                 cmp = 1;
 115         else if (fa.u.f < fb.u.f)
 116                 cmp = -1;
 117
 118         return cmp;
 119 }
 120
 121 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 122                                    fio_fp64_t *plist, unsigned int **output,
 123                                    unsigned int *maxv, unsigned int *minv)
 124 {
 125         unsigned long sum = 0;
 126         unsigned int len, i, j = 0;
 127         unsigned int oval_len = 0;
 128         unsigned int *ovals = NULL;
 129         int is_last;
 130
 131         *minv = -1U;
 132         *maxv = 0;
 133
 134         len = 0;
 135         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 136                 len++;
 137
 138         if (!len)
 139                 return 0;
 140
 141         /*
 142          * Sort the percentile list. Note that it may already be sorted if
 143          * we are using the default values, but since it's a short list this
 144          * isn't a worry. Also note that this does not work for NaN values.
 145          */
 146         if (len > 1)
 147                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 148
 149         /*
 150          * Calculate bucket values, note down max and min values
 151          */
 152         is_last = 0;
 153         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 154                 sum += io_u_plat[i];
 155                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 156                         assert(plist[j].u.f <= 100.0);
 157
 158                         if (j == oval_len) {
 159                                 oval_len += 100;
 160                                 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
 161                         }
 162
 163                         ovals[j] = plat_idx_to_val(i);
 164                         if (ovals[j] < *minv)
 165                                 *minv = ovals[j];
 166                         if (ovals[j] > *maxv)
 167                                 *maxv = ovals[j];
 168
 169                         is_last = (j == len - 1);
 170                         if (is_last)
 171                                 break;
 172
 173                         j++;
 174                 }
 175         }
 176
 177         *output = ovals;
 178         return len;
 179 }
 180
 181 /*
 182  * Find and display the p-th percentile of clat
 183  */
 184 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 185                                   fio_fp64_t *plist, unsigned int precision)
 186 {
 187         unsigned int len, j = 0, minv, maxv;
 188         unsigned int *ovals;
 189         int is_last, per_line, scale_down;
 190         char fmt[32];
 191
 192         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 193         if (!len)
 194                 goto out;
 195
 196         /*
 197          * We default to usecs, but if the value range is such that we
 198          * should scale down to msecs, do that.
 199          */
 200         if (minv > 2000 && maxv > 99999) {
 201                 scale_down = 1;
 202                 log_info("    clat percentiles (msec):\n     |");
 203         } else {
 204                 scale_down = 0;
 205                 log_info("    clat percentiles (usec):\n     |");
 206         }
 207
 208         snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
 209         per_line = (80 - 7) / (precision + 14);
 210
 211         for (j = 0; j < len; j++) {
 212                 char fbuf[16], *ptr = fbuf;
 213
 214                 /* for formatting */
 215                 if (j != 0 && (j % per_line) == 0)
 216                         log_info("     |");
 217
 218                 /* end of the list */
 219                 is_last = (j == len - 1);
 220
 221                 if (plist[j].u.f < 10.0)
 222                         ptr += sprintf(fbuf, " ");
 223
 224                 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
 225
 226                 if (scale_down)
 227                         ovals[j] = (ovals[j] + 999) / 1000;
 228
 229                 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
 230
 231                 if (is_last)
 232                         break;
 233
 234                 if ((j % per_line) == per_line - 1)     /* for formatting */
 235                         log_info("\n");
 236         }
 237
 238 out:
 239         if (ovals)
 240                 free(ovals);
 241 }
 242
 243 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 244              double *mean, double *dev)
 245 {
 246         double n = (double) is->samples;
 247
 248         if (n == 0)
 249                 return 0;
 250
 251         *min = is->min_val;
 252         *max = is->max_val;
 253         *mean = is->mean.u.f;
 254
 255         if (n > 1.0)
 256                 *dev = sqrt(is->S.u.f / (n - 1.0));
 257         else
 258                 *dev = 0;
 259
 260         return 1;
 261 }
 262
 263 void show_group_stats(struct group_run_stats *rs)
 264 {
 265         char *p1, *p2, *p3, *p4;
 266         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 267         int i;
 268
 269         log_info("\nRun status group %d (all jobs):\n", rs->groupid);
 270
 271         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 272                 const int i2p = is_power_of_2(rs->kb_base);
 273
 274                 if (!rs->max_run[i])
 275                         continue;
 276
 277                 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
 278                 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
 279                 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 280                 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 281
 282                 log_info("%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
 283                          " mint=%llumsec, maxt=%llumsec\n",
 284                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 285                                 p1, p2, p3, p4,
 286                                 (unsigned long long) rs->min_run[i],
 287                                 (unsigned long long) rs->max_run[i]);
 288
 289                 free(p1);
 290                 free(p2);
 291                 free(p3);
 292                 free(p4);
 293         }
 294 }
 295
 296 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 297 {
 298         int i;
 299
 300         /*
 301          * Do depth distribution calculations
 302          */
 303         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 304                 if (total) {
 305                         io_u_dist[i] = (double) map[i] / (double) total;
 306                         io_u_dist[i] *= 100.0;
 307                         if (io_u_dist[i] < 0.1 && map[i])
 308                                 io_u_dist[i] = 0.1;
 309                 } else
 310                         io_u_dist[i] = 0.0;
 311         }
 312 }
 313
 314 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 315                           unsigned int *src, int nr)
 316 {
 317         unsigned long total = ddir_rw_sum(ts->total_io_u);
 318         int i;
 319
 320         /*
 321          * Do latency distribution calculations
 322          */
 323         for (i = 0; i < nr; i++) {
 324                 if (total) {
 325                         dst[i] = (double) src[i] / (double) total;
 326                         dst[i] *= 100.0;
 327                         if (dst[i] < 0.01 && src[i])
 328                                 dst[i] = 0.01;
 329                 } else
 330                         dst[i] = 0.0;
 331         }
 332 }
 333
 334 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 335 {
 336         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 337 }
 338
 339 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 340 {
 341         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 342 }
 343
 344 static void display_lat(const char *name, unsigned long min, unsigned long max,
 345                         double mean, double dev)
 346 {
 347         const char *base = "(usec)";
 348         char *minp, *maxp;
 349
 350         if (!usec_to_msec(&min, &max, &mean, &dev))
 351                 base = "(msec)";
 352
 353         minp = num2str(min, 6, 1, 0, 0);
 354         maxp = num2str(max, 6, 1, 0, 0);
 355
 356         log_info("    %s %s: min=%s, max=%s, avg=%5.02f,"
 357                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 358
 359         free(minp);
 360         free(maxp);
 361 }
 362
 363 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 364                              int ddir)
 365 {
 366         const char *str[] = { "read ", "write", "trim" };
 367         unsigned long min, max, runt;
 368         unsigned long long bw, iops;
 369         double mean, dev;
 370         char *io_p, *bw_p, *iops_p;
 371         int i2p;
 372
 373         assert(ddir_rw(ddir));
 374
 375         if (!ts->runtime[ddir])
 376                 return;
 377
 378         i2p = is_power_of_2(rs->kb_base);
 379         runt = ts->runtime[ddir];
 380
 381         bw = (1000 * ts->io_bytes[ddir]) / runt;
 382         io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
 383         bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
 384
 385         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 386         iops_p = num2str(iops, 6, 1, 0, 0);
 387
 388         log_info("  %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
 389                                 rs->unified_rw_rep ? "mixed" : str[ddir],
 390                                 io_p, bw_p, iops_p,
 391                                 (unsigned long long) ts->runtime[ddir]);
 392
 393         free(io_p);
 394         free(bw_p);
 395         free(iops_p);
 396
 397         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 398                 display_lat("slat", min, max, mean, dev);
 399         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 400                 display_lat("clat", min, max, mean, dev);
 401         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 402                 display_lat(" lat", min, max, mean, dev);
 403
 404         if (ts->clat_percentiles) {
 405                 show_clat_percentiles(ts->io_u_plat[ddir],
 406                                         ts->clat_stat[ddir].samples,
 407                                         ts->percentile_list,
 408                                         ts->percentile_precision);
 409         }
 410         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 411                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 412                 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
 413
 414                 if (rs->unit_base == 1) {
 415                         min *= 8.0;
 416                         max *= 8.0;
 417                         mean *= 8.0;
 418                         dev *= 8.0;
 419                 }
 420
 421                 if (rs->agg[ddir]) {
 422                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 423                         if (p_of_agg > 100.0)
 424                                 p_of_agg = 100.0;
 425                 }
 426
 427                 if (mean > fkb_base * fkb_base) {
 428                         min /= fkb_base;
 429                         max /= fkb_base;
 430                         mean /= fkb_base;
 431                         dev /= fkb_base;
 432                         bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
 433                 }
 434
 435                 log_info("    bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
 436                          " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
 437                                                         p_of_agg, mean, dev);
 438         }
 439 }
 440
 441 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 442                     const char *msg)
 443 {
 444         int new_line = 1, i, line = 0, shown = 0;
 445
 446         for (i = 0; i < nr; i++) {
 447                 if (io_u_lat[i] <= 0.0)
 448                         continue;
 449                 shown = 1;
 450                 if (new_line) {
 451                         if (line)
 452                                 log_info("\n");
 453                         log_info("    lat (%s) : ", msg);
 454                         new_line = 0;
 455                         line = 0;
 456                 }
 457                 if (line)
 458                         log_info(", ");
 459                 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
 460                 line++;
 461                 if (line == 5)
 462                         new_line = 1;
 463         }
 464
 465         if (shown)
 466                 log_info("\n");
 467
 468         return shown;
 469 }
 470
 471 static void show_lat_u(double *io_u_lat_u)
 472 {
 473         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 474                                  "250=", "500=", "750=", "1000=", };
 475
 476         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
 477 }
 478
 479 static void show_lat_m(double *io_u_lat_m)
 480 {
 481         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 482                                  "250=", "500=", "750=", "1000=", "2000=",
 483                                  ">=2000=", };
 484
 485         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
 486 }
 487
 488 static void show_latencies(struct thread_stat *ts)
 489 {
 490         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 491         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 492
 493         stat_calc_lat_u(ts, io_u_lat_u);
 494         stat_calc_lat_m(ts, io_u_lat_m);
 495
 496         show_lat_u(io_u_lat_u);
 497         show_lat_m(io_u_lat_m);
 498 }
 499
 500 static void show_thread_status_normal(struct thread_stat *ts,
 501                                       struct group_run_stats *rs)
 502 {
 503         double usr_cpu, sys_cpu;
 504         unsigned long runtime;
 505         double io_u_dist[FIO_IO_U_MAP_NR];
 506         time_t time_p;
 507         char time_buf[64];
 508
 509         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 510                 return;
 511
 512         time(&time_p);
 513         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 514
 515         if (!ts->error) {
 516                 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 517                                         ts->name, ts->groupid, ts->members,
 518                                         ts->error, (int) ts->pid, time_buf);
 519         } else {
 520                 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 521                                         ts->name, ts->groupid, ts->members,
 522                                         ts->error, ts->verror, (int) ts->pid,
 523                                         time_buf);
 524         }
 525
 526         if (strlen(ts->description))
 527                 log_info("  Description  : [%s]\n", ts->description);
 528
 529         if (ts->io_bytes[DDIR_READ])
 530                 show_ddir_status(rs, ts, DDIR_READ);
 531         if (ts->io_bytes[DDIR_WRITE])
 532                 show_ddir_status(rs, ts, DDIR_WRITE);
 533         if (ts->io_bytes[DDIR_TRIM])
 534                 show_ddir_status(rs, ts, DDIR_TRIM);
 535
 536         show_latencies(ts);
 537
 538         runtime = ts->total_run_time;
 539         if (runtime) {
 540                 double runt = (double) runtime;
 541
 542                 usr_cpu = (double) ts->usr_time * 100 / runt;
 543                 sys_cpu = (double) ts->sys_time * 100 / runt;
 544         } else {
 545                 usr_cpu = 0;
 546                 sys_cpu = 0;
 547         }
 548
 549         log_info("  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 550                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 551                         (unsigned long long) ts->ctx,
 552                         (unsigned long long) ts->majf,
 553                         (unsigned long long) ts->minf);
 554
 555         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 556         log_info("  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 557                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 558                                         io_u_dist[1], io_u_dist[2],
 559                                         io_u_dist[3], io_u_dist[4],
 560                                         io_u_dist[5], io_u_dist[6]);
 561
 562         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 563         log_info("     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 564                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 565                                         io_u_dist[1], io_u_dist[2],
 566                                         io_u_dist[3], io_u_dist[4],
 567                                         io_u_dist[5], io_u_dist[6]);
 568         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 569         log_info("     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 570                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 571                                         io_u_dist[1], io_u_dist[2],
 572                                         io_u_dist[3], io_u_dist[4],
 573                                         io_u_dist[5], io_u_dist[6]);
 574         log_info("     issued    : total=r=%llu/w=%llu/d=%llu,"
 575                                  " short=r=%llu/w=%llu/d=%llu,"
 576                                  " drop=r=%llu/w=%llu/d=%llu\n",
 577                                         (unsigned long long) ts->total_io_u[0],
 578                                         (unsigned long long) ts->total_io_u[1],
 579                                         (unsigned long long) ts->total_io_u[2],
 580                                         (unsigned long long) ts->short_io_u[0],
 581                                         (unsigned long long) ts->short_io_u[1],
 582                                         (unsigned long long) ts->short_io_u[2],
 583                                         (unsigned long long) ts->drop_io_u[0],
 584                                         (unsigned long long) ts->drop_io_u[1],
 585                                         (unsigned long long) ts->drop_io_u[2]);
 586         if (ts->continue_on_error) {
 587                 log_info("     errors    : total=%llu, first_error=%d/<%s>\n",
 588                                         (unsigned long long)ts->total_err_count,
 589                                         ts->first_error,
 590                                         strerror(ts->first_error));
 591         }
 592         if (ts->latency_depth) {
 593                 log_info("     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 594                                         (unsigned long long)ts->latency_target,
 595                                         (unsigned long long)ts->latency_window,
 596                                         ts->latency_percentile.u.f,
 597                                         ts->latency_depth);
 598         }
 599 }
 600
 601 static void show_ddir_status_terse(struct thread_stat *ts,
 602                                    struct group_run_stats *rs, int ddir)
 603 {
 604         unsigned long min, max;
 605         unsigned long long bw, iops;
 606         unsigned int *ovals = NULL;
 607         double mean, dev;
 608         unsigned int len, minv, maxv;
 609         int i;
 610
 611         assert(ddir_rw(ddir));
 612
 613         iops = bw = 0;
 614         if (ts->runtime[ddir]) {
 615                 uint64_t runt = ts->runtime[ddir];
 616
 617                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 618                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 619         }
 620
 621         log_info(";%llu;%llu;%llu;%llu",
 622                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 623                                         (unsigned long long) ts->runtime[ddir]);
 624
 625         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 626                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 627         else
 628                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 629
 630         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 631                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 632         else
 633                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 634
 635         if (ts->clat_percentiles) {
 636                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 637                                         ts->clat_stat[ddir].samples,
 638                                         ts->percentile_list, &ovals, &maxv,
 639                                         &minv);
 640         } else
 641                 len = 0;
 642
 643         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 644                 if (i >= len) {
 645                         log_info(";0%%=0");
 646                         continue;
 647                 }
 648                 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
 649         }
 650
 651         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 652                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 653         else
 654                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 655
 656         if (ovals)
 657                 free(ovals);
 658
 659         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 660                 double p_of_agg = 100.0;
 661
 662                 if (rs->agg[ddir]) {
 663                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 664                         if (p_of_agg > 100.0)
 665                                 p_of_agg = 100.0;
 666                 }
 667
 668                 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 669         } else
 670                 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 671 }
 672
 673 static void add_ddir_status_json(struct thread_stat *ts,
 674                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 675 {
 676         unsigned long min, max;
 677         unsigned long long bw, iops;
 678         unsigned int *ovals = NULL;
 679         double mean, dev;
 680         unsigned int len, minv, maxv;
 681         int i;
 682         const char *ddirname[] = {"read", "write", "trim"};
 683         struct json_object *dir_object, *tmp_object, *percentile_object;
 684         char buf[120];
 685         double p_of_agg = 100.0;
 686
 687         assert(ddir_rw(ddir));
 688
 689         if (ts->unified_rw_rep && ddir != DDIR_READ)
 690                 return;
 691
 692         dir_object = json_create_object();
 693         json_object_add_value_object(parent,
 694                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 695
 696         iops = bw = 0;
 697         if (ts->runtime[ddir]) {
 698                 uint64_t runt = ts->runtime[ddir];
 699
 700                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 701                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 702         }
 703
 704         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
 705         json_object_add_value_int(dir_object, "bw", bw);
 706         json_object_add_value_int(dir_object, "iops", iops);
 707         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 708         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 709         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 710         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 711
 712         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 713                 min = max = 0;
 714                 mean = dev = 0.0;
 715         }
 716         tmp_object = json_create_object();
 717         json_object_add_value_object(dir_object, "slat", tmp_object);
 718         json_object_add_value_int(tmp_object, "min", min);
 719         json_object_add_value_int(tmp_object, "max", max);
 720         json_object_add_value_float(tmp_object, "mean", mean);
 721         json_object_add_value_float(tmp_object, "stddev", dev);
 722
 723         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 724                 min = max = 0;
 725                 mean = dev = 0.0;
 726         }
 727         tmp_object = json_create_object();
 728         json_object_add_value_object(dir_object, "clat", tmp_object);
 729         json_object_add_value_int(tmp_object, "min", min);
 730         json_object_add_value_int(tmp_object, "max", max);
 731         json_object_add_value_float(tmp_object, "mean", mean);
 732         json_object_add_value_float(tmp_object, "stddev", dev);
 733
 734         if (ts->clat_percentiles) {
 735                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 736                                         ts->clat_stat[ddir].samples,
 737                                         ts->percentile_list, &ovals, &maxv,
 738                                         &minv);
 739         } else
 740                 len = 0;
 741
 742         percentile_object = json_create_object();
 743         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 744         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 745                 if (i >= len) {
 746                         json_object_add_value_int(percentile_object, "0.00", 0);
 747                         continue;
 748                 }
 749                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 750                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 751         }
 752
 753         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
 754                 min = max = 0;
 755                 mean = dev = 0.0;
 756         }
 757         tmp_object = json_create_object();
 758         json_object_add_value_object(dir_object, "lat", tmp_object);
 759         json_object_add_value_int(tmp_object, "min", min);
 760         json_object_add_value_int(tmp_object, "max", max);
 761         json_object_add_value_float(tmp_object, "mean", mean);
 762         json_object_add_value_float(tmp_object, "stddev", dev);
 763         if (ovals)
 764                 free(ovals);
 765
 766         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 767                 if (rs->agg[ddir]) {
 768                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 769                         if (p_of_agg > 100.0)
 770                                 p_of_agg = 100.0;
 771                 }
 772         } else {
 773                 min = max = 0;
 774                 p_of_agg = mean = dev = 0.0;
 775         }
 776         json_object_add_value_int(dir_object, "bw_min", min);
 777         json_object_add_value_int(dir_object, "bw_max", max);
 778         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
 779         json_object_add_value_float(dir_object, "bw_mean", mean);
 780         json_object_add_value_float(dir_object, "bw_dev", dev);
 781 }
 782
 783 static void show_thread_status_terse_v2(struct thread_stat *ts,
 784                                         struct group_run_stats *rs)
 785 {
 786         double io_u_dist[FIO_IO_U_MAP_NR];
 787         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 788         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 789         double usr_cpu, sys_cpu;
 790         int i;
 791
 792         /* General Info */
 793         log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
 794         /* Log Read Status */
 795         show_ddir_status_terse(ts, rs, DDIR_READ);
 796         /* Log Write Status */
 797         show_ddir_status_terse(ts, rs, DDIR_WRITE);
 798         /* Log Trim Status */
 799         show_ddir_status_terse(ts, rs, DDIR_TRIM);
 800
 801         /* CPU Usage */
 802         if (ts->total_run_time) {
 803                 double runt = (double) ts->total_run_time;
 804
 805                 usr_cpu = (double) ts->usr_time * 100 / runt;
 806                 sys_cpu = (double) ts->sys_time * 100 / runt;
 807         } else {
 808                 usr_cpu = 0;
 809                 sys_cpu = 0;
 810         }
 811
 812         log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
 813                                                 (unsigned long long) ts->ctx,
 814                                                 (unsigned long long) ts->majf,
 815                                                 (unsigned long long) ts->minf);
 816
 817         /* Calc % distribution of IO depths, usecond, msecond latency */
 818         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 819         stat_calc_lat_u(ts, io_u_lat_u);
 820         stat_calc_lat_m(ts, io_u_lat_m);
 821
 822         /* Only show fixed 7 I/O depth levels*/
 823         log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
 824                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
 825                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
 826
 827         /* Microsecond latency */
 828         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
 829                 log_info(";%3.2f%%", io_u_lat_u[i]);
 830         /* Millisecond latency */
 831         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
 832                 log_info(";%3.2f%%", io_u_lat_m[i]);
 833         /* Additional output if continue_on_error set - default off*/
 834         if (ts->continue_on_error)
 835                 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
 836         log_info("\n");
 837
 838         /* Additional output if description is set */
 839         if (strlen(ts->description))
 840                 log_info(";%s", ts->description);
 841
 842         log_info("\n");
 843 }
 844
 845 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
 846                                            struct group_run_stats *rs, int ver)
 847 {
 848         double io_u_dist[FIO_IO_U_MAP_NR];
 849         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 850         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 851         double usr_cpu, sys_cpu;
 852         int i;
 853
 854         /* General Info */
 855         log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
 856                                         ts->name, ts->groupid, ts->error);
 857         /* Log Read Status */
 858         show_ddir_status_terse(ts, rs, DDIR_READ);
 859         /* Log Write Status */
 860         show_ddir_status_terse(ts, rs, DDIR_WRITE);
 861         /* Log Trim Status */
 862         if (ver == 4)
 863                 show_ddir_status_terse(ts, rs, DDIR_TRIM);
 864
 865         /* CPU Usage */
 866         if (ts->total_run_time) {
 867                 double runt = (double) ts->total_run_time;
 868
 869                 usr_cpu = (double) ts->usr_time * 100 / runt;
 870                 sys_cpu = (double) ts->sys_time * 100 / runt;
 871         } else {
 872                 usr_cpu = 0;
 873                 sys_cpu = 0;
 874         }
 875
 876         log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
 877                                                 (unsigned long long) ts->ctx,
 878                                                 (unsigned long long) ts->majf,
 879                                                 (unsigned long long) ts->minf);
 880
 881         /* Calc % distribution of IO depths, usecond, msecond latency */
 882         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 883         stat_calc_lat_u(ts, io_u_lat_u);
 884         stat_calc_lat_m(ts, io_u_lat_m);
 885
 886         /* Only show fixed 7 I/O depth levels*/
 887         log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
 888                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
 889                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
 890
 891         /* Microsecond latency */
 892         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
 893                 log_info(";%3.2f%%", io_u_lat_u[i]);
 894         /* Millisecond latency */
 895         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
 896                 log_info(";%3.2f%%", io_u_lat_m[i]);
 897
 898         /* disk util stats, if any */
 899         show_disk_util(1, NULL);
 900
 901         /* Additional output if continue_on_error set - default off*/
 902         if (ts->continue_on_error)
 903                 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
 904
 905         /* Additional output if description is set */
 906         if (strlen(ts->description))
 907                 log_info(";%s", ts->description);
 908
 909         log_info("\n");
 910 }
 911
 912 static struct json_object *show_thread_status_json(struct thread_stat *ts,
 913                                     struct group_run_stats *rs)
 914 {
 915         struct json_object *root, *tmp;
 916         double io_u_dist[FIO_IO_U_MAP_NR];
 917         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 918         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 919         double usr_cpu, sys_cpu;
 920         int i;
 921
 922         root = json_create_object();
 923         json_object_add_value_string(root, "jobname", ts->name);
 924         json_object_add_value_int(root, "groupid", ts->groupid);
 925         json_object_add_value_int(root, "error", ts->error);
 926
 927         add_ddir_status_json(ts, rs, DDIR_READ, root);
 928         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
 929         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
 930
 931         /* CPU Usage */
 932         if (ts->total_run_time) {
 933                 double runt = (double) ts->total_run_time;
 934
 935                 usr_cpu = (double) ts->usr_time * 100 / runt;
 936                 sys_cpu = (double) ts->sys_time * 100 / runt;
 937         } else {
 938                 usr_cpu = 0;
 939                 sys_cpu = 0;
 940         }
 941         json_object_add_value_float(root, "usr_cpu", usr_cpu);
 942         json_object_add_value_float(root, "sys_cpu", sys_cpu);
 943         json_object_add_value_int(root, "ctx", ts->ctx);
 944         json_object_add_value_int(root, "majf", ts->majf);
 945         json_object_add_value_int(root, "minf", ts->minf);
 946
 947
 948         /* Calc % distribution of IO depths, usecond, msecond latency */
 949         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 950         stat_calc_lat_u(ts, io_u_lat_u);
 951         stat_calc_lat_m(ts, io_u_lat_m);
 952
 953         tmp = json_create_object();
 954         json_object_add_value_object(root, "iodepth_level", tmp);
 955         /* Only show fixed 7 I/O depth levels*/
 956         for (i = 0; i < 7; i++) {
 957                 char name[20];
 958                 if (i < 6)
 959                         snprintf(name, 20, "%d", 1 << i);
 960                 else
 961                         snprintf(name, 20, ">=%d", 1 << i);
 962                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
 963         }
 964
 965         tmp = json_create_object();
 966         json_object_add_value_object(root, "latency_us", tmp);
 967         /* Microsecond latency */
 968         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
 969                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
 970                                  "250", "500", "750", "1000", };
 971                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
 972         }
 973         /* Millisecond latency */
 974         tmp = json_create_object();
 975         json_object_add_value_object(root, "latency_ms", tmp);
 976         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
 977                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
 978                                  "250", "500", "750", "1000", "2000",
 979                                  ">=2000", };
 980                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
 981         }
 982
 983         /* Additional output if continue_on_error set - default off*/
 984         if (ts->continue_on_error) {
 985                 json_object_add_value_int(root, "total_err", ts->total_err_count);
 986                 json_object_add_value_int(root, "first_error", ts->first_error);
 987         }
 988
 989         if (ts->latency_depth) {
 990                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
 991                 json_object_add_value_int(root, "latency_target", ts->latency_target);
 992                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
 993                 json_object_add_value_int(root, "latency_window", ts->latency_window);
 994         }
 995
 996         /* Additional output if description is set */
 997         if (strlen(ts->description))
 998                 json_object_add_value_string(root, "desc", ts->description);
 999
1000         return root;
1001 }
1002
1003 static void show_thread_status_terse(struct thread_stat *ts,
1004                                      struct group_run_stats *rs)
1005 {
1006         if (terse_version == 2)
1007                 show_thread_status_terse_v2(ts, rs);
1008         else if (terse_version == 3 || terse_version == 4)
1009                 show_thread_status_terse_v3_v4(ts, rs, terse_version);
1010         else
1011                 log_err("fio: bad terse version!? %d\n", terse_version);
1012 }
1013
1014 struct json_object *show_thread_status(struct thread_stat *ts,
1015                                        struct group_run_stats *rs)
1016 {
1017         if (output_format == FIO_OUTPUT_TERSE)
1018                 show_thread_status_terse(ts, rs);
1019         else if (output_format == FIO_OUTPUT_JSON)
1020                 return show_thread_status_json(ts, rs);
1021         else
1022                 show_thread_status_normal(ts, rs);
1023         return NULL;
1024 }
1025
1026 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1027 {
1028         double mean, S;
1029
1030         if (src->samples == 0)
1031                 return;
1032
1033         dst->min_val = min(dst->min_val, src->min_val);
1034         dst->max_val = max(dst->max_val, src->max_val);
1035
1036         /*
1037          * Compute new mean and S after the merge
1038          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1039          *  #Parallel_algorithm>
1040          */
1041         if (nr == 1) {
1042                 mean = src->mean.u.f;
1043                 S = src->S.u.f;
1044         } else {
1045                 double delta = src->mean.u.f - dst->mean.u.f;
1046
1047                 mean = ((src->mean.u.f * src->samples) +
1048                         (dst->mean.u.f * dst->samples)) /
1049                         (dst->samples + src->samples);
1050
1051                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1052                         (dst->samples * src->samples) /
1053                         (dst->samples + src->samples);
1054         }
1055
1056         dst->samples += src->samples;
1057         dst->mean.u.f = mean;
1058         dst->S.u.f = S;
1059 }
1060
1061 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1062 {
1063         int i;
1064
1065         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1066                 if (dst->max_run[i] < src->max_run[i])
1067                         dst->max_run[i] = src->max_run[i];
1068                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1069                         dst->min_run[i] = src->min_run[i];
1070                 if (dst->max_bw[i] < src->max_bw[i])
1071                         dst->max_bw[i] = src->max_bw[i];
1072                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1073                         dst->min_bw[i] = src->min_bw[i];
1074
1075                 dst->io_kb[i] += src->io_kb[i];
1076                 dst->agg[i] += src->agg[i];
1077         }
1078
1079         if (!dst->kb_base)
1080                 dst->kb_base = src->kb_base;
1081         if (!dst->unit_base)
1082                 dst->unit_base = src->unit_base;
1083 }
1084
1085 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1086 {
1087         int l, k;
1088
1089         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1090                 if (!dst->unified_rw_rep) {
1091                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1092                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1093                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1094                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1095
1096                         dst->io_bytes[l] += src->io_bytes[l];
1097
1098                         if (dst->runtime[l] < src->runtime[l])
1099                                 dst->runtime[l] = src->runtime[l];
1100                 } else {
1101                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1102                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1103                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1104                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1105
1106                         dst->io_bytes[0] += src->io_bytes[l];
1107
1108                         if (dst->runtime[0] < src->runtime[l])
1109                                 dst->runtime[0] = src->runtime[l];
1110                 }
1111         }
1112
1113         dst->usr_time += src->usr_time;
1114         dst->sys_time += src->sys_time;
1115         dst->ctx += src->ctx;
1116         dst->majf += src->majf;
1117         dst->minf += src->minf;
1118
1119         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1120                 dst->io_u_map[k] += src->io_u_map[k];
1121         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1122                 dst->io_u_submit[k] += src->io_u_submit[k];
1123         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1124                 dst->io_u_complete[k] += src->io_u_complete[k];
1125         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1126                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1127         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1128                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1129
1130         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1131                 if (!dst->unified_rw_rep) {
1132                         dst->total_io_u[k] += src->total_io_u[k];
1133                         dst->short_io_u[k] += src->short_io_u[k];
1134                         dst->drop_io_u[k] += src->drop_io_u[k];
1135                 } else {
1136                         dst->total_io_u[0] += src->total_io_u[k];
1137                         dst->short_io_u[0] += src->short_io_u[k];
1138                         dst->drop_io_u[0] += src->drop_io_u[k];
1139                 }
1140         }
1141
1142         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1143                 int m;
1144
1145                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1146                         if (!dst->unified_rw_rep)
1147                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1148                         else
1149                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1150                 }
1151         }
1152
1153         dst->total_run_time += src->total_run_time;
1154         dst->total_submit += src->total_submit;
1155         dst->total_complete += src->total_complete;
1156 }
1157
1158 void init_group_run_stat(struct group_run_stats *gs)
1159 {
1160         int i;
1161         memset(gs, 0, sizeof(*gs));
1162
1163         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1164                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1165 }
1166
1167 void init_thread_stat(struct thread_stat *ts)
1168 {
1169         int j;
1170
1171         memset(ts, 0, sizeof(*ts));
1172
1173         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1174                 ts->lat_stat[j].min_val = -1UL;
1175                 ts->clat_stat[j].min_val = -1UL;
1176                 ts->slat_stat[j].min_val = -1UL;
1177                 ts->bw_stat[j].min_val = -1UL;
1178         }
1179         ts->groupid = -1;
1180 }
1181
1182 void __show_run_stats(void)
1183 {
1184         struct group_run_stats *runstats, *rs;
1185         struct thread_data *td;
1186         struct thread_stat *threadstats, *ts;
1187         int i, j, nr_ts, last_ts, idx;
1188         int kb_base_warned = 0;
1189         int unit_base_warned = 0;
1190         struct json_object *root = NULL;
1191         struct json_array *array = NULL;
1192
1193         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1194
1195         for (i = 0; i < groupid + 1; i++)
1196                 init_group_run_stat(&runstats[i]);
1197
1198         /*
1199          * find out how many threads stats we need. if group reporting isn't
1200          * enabled, it's one-per-td.
1201          */
1202         nr_ts = 0;
1203         last_ts = -1;
1204         for_each_td(td, i) {
1205                 if (!td->o.group_reporting) {
1206                         nr_ts++;
1207                         continue;
1208                 }
1209                 if (last_ts == td->groupid)
1210                         continue;
1211
1212                 last_ts = td->groupid;
1213                 nr_ts++;
1214         }
1215
1216         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1217
1218         for (i = 0; i < nr_ts; i++)
1219                 init_thread_stat(&threadstats[i]);
1220
1221         j = 0;
1222         last_ts = -1;
1223         idx = 0;
1224         for_each_td(td, i) {
1225                 if (idx && (!td->o.group_reporting ||
1226                     (td->o.group_reporting && last_ts != td->groupid))) {
1227                         idx = 0;
1228                         j++;
1229                 }
1230
1231                 last_ts = td->groupid;
1232
1233                 ts = &threadstats[j];
1234
1235                 ts->clat_percentiles = td->o.clat_percentiles;
1236                 ts->percentile_precision = td->o.percentile_precision;
1237                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1238
1239                 idx++;
1240                 ts->members++;
1241
1242                 if (ts->groupid == -1) {
1243                         /*
1244                          * These are per-group shared already
1245                          */
1246                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1247                         if (td->o.description)
1248                                 strncpy(ts->description, td->o.description,
1249                                                 FIO_JOBDESC_SIZE - 1);
1250                         else
1251                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1252
1253                         /*
1254                          * If multiple entries in this group, this is
1255                          * the first member.
1256                          */
1257                         ts->thread_number = td->thread_number;
1258                         ts->groupid = td->groupid;
1259
1260                         /*
1261                          * first pid in group, not very useful...
1262                          */
1263                         ts->pid = td->pid;
1264
1265                         ts->kb_base = td->o.kb_base;
1266                         ts->unit_base = td->o.unit_base;
1267                         ts->unified_rw_rep = td->o.unified_rw_rep;
1268                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1269                         log_info("fio: kb_base differs for jobs in group, using"
1270                                  " %u as the base\n", ts->kb_base);
1271                         kb_base_warned = 1;
1272                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1273                         log_info("fio: unit_base differs for jobs in group, using"
1274                                  " %u as the base\n", ts->unit_base);
1275                         unit_base_warned = 1;
1276                 }
1277
1278                 ts->continue_on_error = td->o.continue_on_error;
1279                 ts->total_err_count += td->total_err_count;
1280                 ts->first_error = td->first_error;
1281                 if (!ts->error) {
1282                         if (!td->error && td->o.continue_on_error &&
1283                             td->first_error) {
1284                                 ts->error = td->first_error;
1285                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1286                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1287                         } else  if (td->error) {
1288                                 ts->error = td->error;
1289                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1290                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1291                         }
1292                 }
1293
1294                 ts->latency_depth = td->latency_qd;
1295                 ts->latency_target = td->o.latency_target;
1296                 ts->latency_percentile = td->o.latency_percentile;
1297                 ts->latency_window = td->o.latency_window;
1298
1299                 sum_thread_stats(ts, &td->ts, idx);
1300         }
1301
1302         for (i = 0; i < nr_ts; i++) {
1303                 unsigned long long bw;
1304
1305                 ts = &threadstats[i];
1306                 rs = &runstats[ts->groupid];
1307                 rs->kb_base = ts->kb_base;
1308                 rs->unit_base = ts->unit_base;
1309                 rs->unified_rw_rep += ts->unified_rw_rep;
1310
1311                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1312                         if (!ts->runtime[j])
1313                                 continue;
1314                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1315                                 rs->min_run[j] = ts->runtime[j];
1316                         if (ts->runtime[j] > rs->max_run[j])
1317                                 rs->max_run[j] = ts->runtime[j];
1318
1319                         bw = 0;
1320                         if (ts->runtime[j]) {
1321                                 unsigned long runt = ts->runtime[j];
1322                                 unsigned long long kb;
1323
1324                                 kb = ts->io_bytes[j] / rs->kb_base;
1325                                 bw = kb * 1000 / runt;
1326                         }
1327                         if (bw < rs->min_bw[j])
1328                                 rs->min_bw[j] = bw;
1329                         if (bw > rs->max_bw[j])
1330                                 rs->max_bw[j] = bw;
1331
1332                         rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1333                 }
1334         }
1335
1336         for (i = 0; i < groupid + 1; i++) {
1337                 int ddir;
1338
1339                 rs = &runstats[i];
1340
1341                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1342                         if (rs->max_run[ddir])
1343                                 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1344                                                 rs->max_run[ddir];
1345                 }
1346         }
1347
1348         /*
1349          * don't overwrite last signal output
1350          */
1351         if (output_format == FIO_OUTPUT_NORMAL)
1352                 log_info("\n");
1353         else if (output_format == FIO_OUTPUT_JSON) {
1354                 root = json_create_object();
1355                 json_object_add_value_string(root, "fio version", fio_version_string);
1356                 array = json_create_array();
1357                 json_object_add_value_array(root, "jobs", array);
1358         }
1359
1360         for (i = 0; i < nr_ts; i++) {
1361                 ts = &threadstats[i];
1362                 rs = &runstats[ts->groupid];
1363
1364                 if (is_backend)
1365                         fio_server_send_ts(ts, rs);
1366                 else if (output_format == FIO_OUTPUT_TERSE)
1367                         show_thread_status_terse(ts, rs);
1368                 else if (output_format == FIO_OUTPUT_JSON) {
1369                         struct json_object *tmp = show_thread_status_json(ts, rs);
1370                         json_array_add_value_object(array, tmp);
1371                 } else
1372                         show_thread_status_normal(ts, rs);
1373         }
1374         if (output_format == FIO_OUTPUT_JSON) {
1375                 /* disk util stats, if any */
1376                 show_disk_util(1, root);
1377
1378                 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1379
1380                 json_print_object(root);
1381                 log_info("\n");
1382                 json_free_object(root);
1383         }
1384
1385         for (i = 0; i < groupid + 1; i++) {
1386                 rs = &runstats[i];
1387
1388                 rs->groupid = i;
1389                 if (is_backend)
1390                         fio_server_send_gs(rs);
1391                 else if (output_format == FIO_OUTPUT_NORMAL)
1392                         show_group_stats(rs);
1393         }
1394
1395         if (is_backend)
1396                 fio_server_send_du();
1397         else if (output_format == FIO_OUTPUT_NORMAL) {
1398                 show_disk_util(0, NULL);
1399                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1400         }
1401
1402         if ( !(output_format == FIO_OUTPUT_TERSE) && append_terse_output) {
1403                 log_info("\nAdditional Terse Output:\n");
1404
1405                 for (i = 0; i < nr_ts; i++) {
1406                         ts = &threadstats[i];
1407                         rs = &runstats[ts->groupid];
1408                         show_thread_status_terse(ts, rs);
1409                 }
1410         }
1411
1412         log_info_flush();
1413         free(runstats);
1414         free(threadstats);
1415 }
1416
1417 void show_run_stats(void)
1418 {
1419         fio_mutex_down(stat_mutex);
1420         __show_run_stats();
1421         fio_mutex_up(stat_mutex);
1422 }
1423
1424 void __show_running_run_stats(void)
1425 {
1426         struct thread_data *td;
1427         unsigned long long *rt;
1428         struct timeval tv;
1429         int i;
1430
1431         fio_mutex_down(stat_mutex);
1432
1433         rt = malloc(thread_number * sizeof(unsigned long long));
1434         fio_gettime(&tv, NULL);
1435
1436         for_each_td(td, i) {
1437                 rt[i] = mtime_since(&td->start, &tv);
1438                 if (td_read(td) && td->io_bytes[DDIR_READ])
1439                         td->ts.runtime[DDIR_READ] += rt[i];
1440                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1441                         td->ts.runtime[DDIR_WRITE] += rt[i];
1442                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1443                         td->ts.runtime[DDIR_TRIM] += rt[i];
1444
1445                 td->update_rusage = 1;
1446                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1447                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1448                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1449                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1450         }
1451
1452         for_each_td(td, i) {
1453                 if (td->runstate >= TD_EXITED)
1454                         continue;
1455                 if (td->rusage_sem) {
1456                         td->update_rusage = 1;
1457                         fio_mutex_down(td->rusage_sem);
1458                 }
1459                 td->update_rusage = 0;
1460         }
1461
1462         __show_run_stats();
1463
1464         for_each_td(td, i) {
1465                 if (td_read(td) && td->io_bytes[DDIR_READ])
1466                         td->ts.runtime[DDIR_READ] -= rt[i];
1467                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1468                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1469                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1470                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1471         }
1472
1473         free(rt);
1474         fio_mutex_up(stat_mutex);
1475 }
1476
1477 static int status_interval_init;
1478 static struct timeval status_time;
1479 static int status_file_disabled;
1480
1481 #define FIO_STATUS_FILE         "fio-dump-status"
1482
1483 static int check_status_file(void)
1484 {
1485         struct stat sb;
1486         const char *temp_dir;
1487         char fio_status_file_path[PATH_MAX];
1488
1489         if (status_file_disabled)
1490                 return 0;
1491
1492         temp_dir = getenv("TMPDIR");
1493         if (temp_dir == NULL) {
1494                 temp_dir = getenv("TEMP");
1495                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1496                         temp_dir = NULL;
1497         }
1498         if (temp_dir == NULL)
1499                 temp_dir = "/tmp";
1500
1501         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1502
1503         if (stat(fio_status_file_path, &sb))
1504                 return 0;
1505
1506         if (unlink(fio_status_file_path) < 0) {
1507                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1508                                                         strerror(errno));
1509                 log_err("fio: disabling status file updates\n");
1510                 status_file_disabled = 1;
1511         }
1512
1513         return 1;
1514 }
1515
1516 void check_for_running_stats(void)
1517 {
1518         if (status_interval) {
1519                 if (!status_interval_init) {
1520                         fio_gettime(&status_time, NULL);
1521                         status_interval_init = 1;
1522                 } else if (mtime_since_now(&status_time) >= status_interval) {
1523                         show_running_run_stats();
1524                         fio_gettime(&status_time, NULL);
1525                         return;
1526                 }
1527         }
1528         if (check_status_file()) {
1529                 show_running_run_stats();
1530                 return;
1531         }
1532 }
1533
1534 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1535 {
1536         double val = data;
1537         double delta;
1538
1539         if (data > is->max_val)
1540                 is->max_val = data;
1541         if (data < is->min_val)
1542                 is->min_val = data;
1543
1544         delta = val - is->mean.u.f;
1545         if (delta) {
1546                 is->mean.u.f += delta / (is->samples + 1.0);
1547                 is->S.u.f += delta * (val - is->mean.u.f);
1548         }
1549
1550         is->samples++;
1551 }
1552
1553 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1554                              enum fio_ddir ddir, unsigned int bs,
1555                              unsigned long t, uint64_t offset)
1556 {
1557         uint64_t nr_samples = iolog->nr_samples;
1558         struct io_sample *s;
1559
1560         if (iolog->disabled)
1561                 return;
1562
1563         if (!iolog->nr_samples)
1564                 iolog->avg_last = t;
1565
1566         if (iolog->nr_samples == iolog->max_samples) {
1567                 size_t new_size;
1568                 void *new_log;
1569
1570                 new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
1571
1572                 if (iolog->log_gz && (new_size > iolog->log_gz)) {
1573                         if (iolog_flush(iolog, 0)) {
1574                                 log_err("fio: failed flushing iolog! Will stop logging.\n");
1575                                 iolog->disabled = 1;
1576                                 return;
1577                         }
1578                         nr_samples = iolog->nr_samples;
1579                 } else {
1580                         new_log = realloc(iolog->log, new_size);
1581                         if (!new_log) {
1582                                 log_err("fio: failed extending iolog! Will stop logging.\n");
1583                                 iolog->disabled = 1;
1584                                 return;
1585                         }
1586                         iolog->log = new_log;
1587                         iolog->max_samples <<= 1;
1588                 }
1589         }
1590
1591         s = get_sample(iolog, nr_samples);
1592
1593         s->val = val;
1594         s->time = t;
1595         io_sample_set_ddir(iolog, s, ddir);
1596         s->bs = bs;
1597
1598         if (iolog->log_offset) {
1599                 struct io_sample_offset *so = (void *) s;
1600
1601                 so->offset = offset;
1602         }
1603
1604         iolog->nr_samples++;
1605 }
1606
1607 static inline void reset_io_stat(struct io_stat *ios)
1608 {
1609         ios->max_val = ios->min_val = ios->samples = 0;
1610         ios->mean.u.f = ios->S.u.f = 0;
1611 }
1612
1613 void reset_io_stats(struct thread_data *td)
1614 {
1615         struct thread_stat *ts = &td->ts;
1616         int i, j;
1617
1618         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1619                 reset_io_stat(&ts->clat_stat[i]);
1620                 reset_io_stat(&ts->slat_stat[i]);
1621                 reset_io_stat(&ts->lat_stat[i]);
1622                 reset_io_stat(&ts->bw_stat[i]);
1623                 reset_io_stat(&ts->iops_stat[i]);
1624
1625                 ts->io_bytes[i] = 0;
1626                 ts->runtime[i] = 0;
1627
1628                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1629                         ts->io_u_plat[i][j] = 0;
1630         }
1631
1632         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1633                 ts->io_u_map[i] = 0;
1634                 ts->io_u_submit[i] = 0;
1635                 ts->io_u_complete[i] = 0;
1636                 ts->io_u_lat_u[i] = 0;
1637                 ts->io_u_lat_m[i] = 0;
1638                 ts->total_submit = 0;
1639                 ts->total_complete = 0;
1640         }
1641
1642         for (i = 0; i < 3; i++) {
1643                 ts->total_io_u[i] = 0;
1644                 ts->short_io_u[i] = 0;
1645                 ts->drop_io_u[i] = 0;
1646         }
1647 }
1648
1649 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
1650 {
1651         /*
1652          * Note an entry in the log. Use the mean from the logged samples,
1653          * making sure to properly round up. Only write a log entry if we
1654          * had actual samples done.
1655          */
1656         if (iolog->avg_window[DDIR_READ].samples) {
1657                 unsigned long mr;
1658
1659                 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1660                 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed, 0);
1661         }
1662         if (iolog->avg_window[DDIR_WRITE].samples) {
1663                 unsigned long mw;
1664
1665                 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1666                 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
1667         }
1668         if (iolog->avg_window[DDIR_TRIM].samples) {
1669                 unsigned long mw;
1670
1671                 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1672                 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed, 0);
1673         }
1674
1675         reset_io_stat(&iolog->avg_window[DDIR_READ]);
1676         reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1677         reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1678 }
1679
1680 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1681                            unsigned long val, enum fio_ddir ddir,
1682                            unsigned int bs, uint64_t offset)
1683 {
1684         unsigned long elapsed, this_window;
1685
1686         if (!ddir_rw(ddir))
1687                 return;
1688
1689         elapsed = mtime_since_now(&td->epoch);
1690
1691         /*
1692          * If no time averaging, just add the log sample.
1693          */
1694         if (!iolog->avg_msec) {
1695                 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
1696                 return;
1697         }
1698
1699         /*
1700          * Add the sample. If the time period has passed, then
1701          * add that entry to the log and clear.
1702          */
1703         add_stat_sample(&iolog->avg_window[ddir], val);
1704
1705         /*
1706          * If period hasn't passed, adding the above sample is all we
1707          * need to do.
1708          */
1709         this_window = elapsed - iolog->avg_last;
1710         if (this_window < iolog->avg_msec)
1711                 return;
1712
1713         _add_stat_to_log(iolog, elapsed);
1714
1715         iolog->avg_last = elapsed;
1716 }
1717
1718 void finalize_logs(struct thread_data *td)
1719 {
1720         unsigned long elapsed;
1721
1722         elapsed = mtime_since_now(&td->epoch);
1723
1724         if (td->clat_log)
1725                 _add_stat_to_log(td->clat_log, elapsed);
1726         if (td->slat_log)
1727                 _add_stat_to_log(td->slat_log, elapsed);
1728         if (td->lat_log)
1729                 _add_stat_to_log(td->lat_log, elapsed);
1730         if (td->bw_log)
1731                 _add_stat_to_log(td->bw_log, elapsed);
1732         if (td->iops_log)
1733                 _add_stat_to_log(td->iops_log, elapsed);
1734 }
1735
1736 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1737 {
1738         struct io_log *iolog;
1739
1740         if (!ddir_rw(ddir))
1741                 return;
1742
1743         iolog = agg_io_log[ddir];
1744         __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
1745 }
1746
1747 static void add_clat_percentile_sample(struct thread_stat *ts,
1748                                 unsigned long usec, enum fio_ddir ddir)
1749 {
1750         unsigned int idx = plat_val_to_idx(usec);
1751         assert(idx < FIO_IO_U_PLAT_NR);
1752
1753         ts->io_u_plat[ddir][idx]++;
1754 }
1755
1756 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1757                      unsigned long usec, unsigned int bs, uint64_t offset)
1758 {
1759         struct thread_stat *ts = &td->ts;
1760
1761         if (!ddir_rw(ddir))
1762                 return;
1763
1764         add_stat_sample(&ts->clat_stat[ddir], usec);
1765
1766         if (td->clat_log)
1767                 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
1768
1769         if (ts->clat_percentiles)
1770                 add_clat_percentile_sample(ts, usec, ddir);
1771 }
1772
1773 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1774                      unsigned long usec, unsigned int bs, uint64_t offset)
1775 {
1776         struct thread_stat *ts = &td->ts;
1777
1778         if (!ddir_rw(ddir))
1779                 return;
1780
1781         add_stat_sample(&ts->slat_stat[ddir], usec);
1782
1783         if (td->slat_log)
1784                 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
1785 }
1786
1787 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1788                     unsigned long usec, unsigned int bs, uint64_t offset)
1789 {
1790         struct thread_stat *ts = &td->ts;
1791
1792         if (!ddir_rw(ddir))
1793                 return;
1794
1795         add_stat_sample(&ts->lat_stat[ddir], usec);
1796
1797         if (td->lat_log)
1798                 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
1799 }
1800
1801 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1802                    struct timeval *t)
1803 {
1804         struct thread_stat *ts = &td->ts;
1805         unsigned long spent, rate;
1806
1807         if (!ddir_rw(ddir))
1808                 return;
1809
1810         spent = mtime_since(&td->bw_sample_time, t);
1811         if (spent < td->o.bw_avg_time)
1812                 return;
1813
1814         /*
1815          * Compute both read and write rates for the interval.
1816          */
1817         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1818                 uint64_t delta;
1819
1820                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1821                 if (!delta)
1822                         continue; /* No entries for interval */
1823
1824                 if (spent)
1825                         rate = delta * 1000 / spent / 1024;
1826                 else
1827                         rate = 0;
1828
1829                 add_stat_sample(&ts->bw_stat[ddir], rate);
1830
1831                 if (td->bw_log)
1832                         add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
1833
1834                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1835         }
1836
1837         fio_gettime(&td->bw_sample_time, NULL);
1838 }
1839
1840 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1841                      struct timeval *t)
1842 {
1843         struct thread_stat *ts = &td->ts;
1844         unsigned long spent, iops;
1845
1846         if (!ddir_rw(ddir))
1847                 return;
1848
1849         spent = mtime_since(&td->iops_sample_time, t);
1850         if (spent < td->o.iops_avg_time)
1851                 return;
1852
1853         /*
1854          * Compute both read and write rates for the interval.
1855          */
1856         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1857                 uint64_t delta;
1858
1859                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1860                 if (!delta)
1861                         continue; /* No entries for interval */
1862
1863                 if (spent)
1864                         iops = (delta * 1000) / spent;
1865                 else
1866                         iops = 0;
1867
1868                 add_stat_sample(&ts->iops_stat[ddir], iops);
1869
1870                 if (td->iops_log)
1871                         add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
1872
1873                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1874         }
1875
1876         fio_gettime(&td->iops_sample_time, NULL);
1877 }
1878
1879 void stat_init(void)
1880 {
1881         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1882 }
1883
1884 void stat_exit(void)
1885 {
1886         /*
1887          * When we have the mutex, we know out-of-band access to it
1888          * have ended.
1889          */
1890         fio_mutex_down(stat_mutex);
1891         fio_mutex_remove(stat_mutex);
1892 }
1893
1894 /*
1895  * Called from signal handler. Wake up status thread.
1896  */
1897 void show_running_run_stats(void)
1898 {
1899         helper_do_stat = 1;
1900         pthread_cond_signal(&helper_cond);
1901 }