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 static 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 *ddir_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" : ddir_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 *ddir_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" : ddir_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 (!(ts->io_bytes[DDIR_READ] + ts->io_bytes[DDIR_WRITE] +
 510             ts->io_bytes[DDIR_TRIM]) && !(ts->total_io_u[DDIR_READ] +
 511             ts->total_io_u[DDIR_WRITE] + ts->total_io_u[DDIR_TRIM]))
 512                 return;
 513
 514         time(&time_p);
 515         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 516
 517         if (!ts->error) {
 518                 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 519                                         ts->name, ts->groupid, ts->members,
 520                                         ts->error, (int) ts->pid, time_buf);
 521         } else {
 522                 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 523                                         ts->name, ts->groupid, ts->members,
 524                                         ts->error, ts->verror, (int) ts->pid,
 525                                         time_buf);
 526         }
 527
 528         if (strlen(ts->description))
 529                 log_info("  Description  : [%s]\n", ts->description);
 530
 531         if (ts->io_bytes[DDIR_READ])
 532                 show_ddir_status(rs, ts, DDIR_READ);
 533         if (ts->io_bytes[DDIR_WRITE])
 534                 show_ddir_status(rs, ts, DDIR_WRITE);
 535         if (ts->io_bytes[DDIR_TRIM])
 536                 show_ddir_status(rs, ts, DDIR_TRIM);
 537
 538         show_latencies(ts);
 539
 540         runtime = ts->total_run_time;
 541         if (runtime) {
 542                 double runt = (double) runtime;
 543
 544                 usr_cpu = (double) ts->usr_time * 100 / runt;
 545                 sys_cpu = (double) ts->sys_time * 100 / runt;
 546         } else {
 547                 usr_cpu = 0;
 548                 sys_cpu = 0;
 549         }
 550
 551         log_info("  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 552                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 553                         (unsigned long long) ts->ctx,
 554                         (unsigned long long) ts->majf,
 555                         (unsigned long long) ts->minf);
 556
 557         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 558         log_info("  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 559                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 560                                         io_u_dist[1], io_u_dist[2],
 561                                         io_u_dist[3], io_u_dist[4],
 562                                         io_u_dist[5], io_u_dist[6]);
 563
 564         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 565         log_info("     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 566                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 567                                         io_u_dist[1], io_u_dist[2],
 568                                         io_u_dist[3], io_u_dist[4],
 569                                         io_u_dist[5], io_u_dist[6]);
 570         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 571         log_info("     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 572                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 573                                         io_u_dist[1], io_u_dist[2],
 574                                         io_u_dist[3], io_u_dist[4],
 575                                         io_u_dist[5], io_u_dist[6]);
 576         log_info("     issued    : total=r=%llu/w=%llu/d=%llu,"
 577                                  " short=r=%llu/w=%llu/d=%llu,"
 578                                  " drop=r=%llu/w=%llu/d=%llu\n",
 579                                         (unsigned long long) ts->total_io_u[0],
 580                                         (unsigned long long) ts->total_io_u[1],
 581                                         (unsigned long long) ts->total_io_u[2],
 582                                         (unsigned long long) ts->short_io_u[0],
 583                                         (unsigned long long) ts->short_io_u[1],
 584                                         (unsigned long long) ts->short_io_u[2],
 585                                         (unsigned long long) ts->drop_io_u[0],
 586                                         (unsigned long long) ts->drop_io_u[1],
 587                                         (unsigned long long) ts->drop_io_u[2]);
 588         if (ts->continue_on_error) {
 589                 log_info("     errors    : total=%llu, first_error=%d/<%s>\n",
 590                                         (unsigned long long)ts->total_err_count,
 591                                         ts->first_error,
 592                                         strerror(ts->first_error));
 593         }
 594         if (ts->latency_depth) {
 595                 log_info("     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 596                                         (unsigned long long)ts->latency_target,
 597                                         (unsigned long long)ts->latency_window,
 598                                         ts->latency_percentile.u.f,
 599                                         ts->latency_depth);
 600         }
 601 }
 602
 603 static void show_ddir_status_terse(struct thread_stat *ts,
 604                                    struct group_run_stats *rs, int ddir)
 605 {
 606         unsigned long min, max;
 607         unsigned long long bw, iops;
 608         unsigned int *ovals = NULL;
 609         double mean, dev;
 610         unsigned int len, minv, maxv;
 611         int i;
 612
 613         assert(ddir_rw(ddir));
 614
 615         iops = bw = 0;
 616         if (ts->runtime[ddir]) {
 617                 uint64_t runt = ts->runtime[ddir];
 618
 619                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 620                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 621         }
 622
 623         log_info(";%llu;%llu;%llu;%llu",
 624                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 625                                         (unsigned long long) ts->runtime[ddir]);
 626
 627         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 628                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 629         else
 630                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 631
 632         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 633                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 634         else
 635                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 636
 637         if (ts->clat_percentiles) {
 638                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 639                                         ts->clat_stat[ddir].samples,
 640                                         ts->percentile_list, &ovals, &maxv,
 641                                         &minv);
 642         } else
 643                 len = 0;
 644
 645         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 646                 if (i >= len) {
 647                         log_info(";0%%=0");
 648                         continue;
 649                 }
 650                 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
 651         }
 652
 653         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 654                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 655         else
 656                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 657
 658         if (ovals)
 659                 free(ovals);
 660
 661         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 662                 double p_of_agg = 100.0;
 663
 664                 if (rs->agg[ddir]) {
 665                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 666                         if (p_of_agg > 100.0)
 667                                 p_of_agg = 100.0;
 668                 }
 669
 670                 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 671         } else
 672                 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 673 }
 674
 675 static void add_ddir_status_json(struct thread_stat *ts,
 676                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 677 {
 678         unsigned long min, max;
 679         unsigned long long bw, iops;
 680         unsigned int *ovals = NULL;
 681         double mean, dev;
 682         unsigned int len, minv, maxv;
 683         int i;
 684         const char *ddirname[] = {"read", "write", "trim"};
 685         struct json_object *dir_object, *tmp_object, *percentile_object;
 686         char buf[120];
 687         double p_of_agg = 100.0;
 688
 689         assert(ddir_rw(ddir));
 690
 691         if (ts->unified_rw_rep && ddir != DDIR_READ)
 692                 return;
 693
 694         dir_object = json_create_object();
 695         json_object_add_value_object(parent,
 696                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 697
 698         iops = bw = 0;
 699         if (ts->runtime[ddir]) {
 700                 uint64_t runt = ts->runtime[ddir];
 701
 702                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 703                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 704         }
 705
 706         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
 707         json_object_add_value_int(dir_object, "bw", bw);
 708         json_object_add_value_int(dir_object, "iops", iops);
 709         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 710
 711         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 712                 min = max = 0;
 713                 mean = dev = 0.0;
 714         }
 715         tmp_object = json_create_object();
 716         json_object_add_value_object(dir_object, "slat", tmp_object);
 717         json_object_add_value_int(tmp_object, "min", min);
 718         json_object_add_value_int(tmp_object, "max", max);
 719         json_object_add_value_float(tmp_object, "mean", mean);
 720         json_object_add_value_float(tmp_object, "stddev", dev);
 721
 722         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 723                 min = max = 0;
 724                 mean = dev = 0.0;
 725         }
 726         tmp_object = json_create_object();
 727         json_object_add_value_object(dir_object, "clat", tmp_object);
 728         json_object_add_value_int(tmp_object, "min", min);
 729         json_object_add_value_int(tmp_object, "max", max);
 730         json_object_add_value_float(tmp_object, "mean", mean);
 731         json_object_add_value_float(tmp_object, "stddev", dev);
 732
 733         if (ts->clat_percentiles) {
 734                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 735                                         ts->clat_stat[ddir].samples,
 736                                         ts->percentile_list, &ovals, &maxv,
 737                                         &minv);
 738         } else
 739                 len = 0;
 740
 741         percentile_object = json_create_object();
 742         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 743         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 744                 if (i >= len) {
 745                         json_object_add_value_int(percentile_object, "0.00", 0);
 746                         continue;
 747                 }
 748                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 749                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 750         }
 751
 752         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
 753                 min = max = 0;
 754                 mean = dev = 0.0;
 755         }
 756         tmp_object = json_create_object();
 757         json_object_add_value_object(dir_object, "lat", tmp_object);
 758         json_object_add_value_int(tmp_object, "min", min);
 759         json_object_add_value_int(tmp_object, "max", max);
 760         json_object_add_value_float(tmp_object, "mean", mean);
 761         json_object_add_value_float(tmp_object, "stddev", dev);
 762         if (ovals)
 763                 free(ovals);
 764
 765         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 766                 if (rs->agg[ddir]) {
 767                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 768                         if (p_of_agg > 100.0)
 769                                 p_of_agg = 100.0;
 770                 }
 771         } else {
 772                 min = max = 0;
 773                 p_of_agg = mean = dev = 0.0;
 774         }
 775         json_object_add_value_int(dir_object, "bw_min", min);
 776         json_object_add_value_int(dir_object, "bw_max", max);
 777         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
 778         json_object_add_value_float(dir_object, "bw_mean", mean);
 779         json_object_add_value_float(dir_object, "bw_dev", dev);
 780 }
 781
 782 static void show_thread_status_terse_v2(struct thread_stat *ts,
 783                                         struct group_run_stats *rs)
 784 {
 785         double io_u_dist[FIO_IO_U_MAP_NR];
 786         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 787         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 788         double usr_cpu, sys_cpu;
 789         int i;
 790
 791         /* General Info */
 792         log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
 793         /* Log Read Status */
 794         show_ddir_status_terse(ts, rs, DDIR_READ);
 795         /* Log Write Status */
 796         show_ddir_status_terse(ts, rs, DDIR_WRITE);
 797         /* Log Trim Status */
 798         show_ddir_status_terse(ts, rs, DDIR_TRIM);
 799
 800         /* CPU Usage */
 801         if (ts->total_run_time) {
 802                 double runt = (double) ts->total_run_time;
 803
 804                 usr_cpu = (double) ts->usr_time * 100 / runt;
 805                 sys_cpu = (double) ts->sys_time * 100 / runt;
 806         } else {
 807                 usr_cpu = 0;
 808                 sys_cpu = 0;
 809         }
 810
 811         log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
 812                                                 (unsigned long long) ts->ctx,
 813                                                 (unsigned long long) ts->majf,
 814                                                 (unsigned long long) ts->minf);
 815
 816         /* Calc % distribution of IO depths, usecond, msecond latency */
 817         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 818         stat_calc_lat_u(ts, io_u_lat_u);
 819         stat_calc_lat_m(ts, io_u_lat_m);
 820
 821         /* Only show fixed 7 I/O depth levels*/
 822         log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
 823                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
 824                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
 825
 826         /* Microsecond latency */
 827         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
 828                 log_info(";%3.2f%%", io_u_lat_u[i]);
 829         /* Millisecond latency */
 830         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
 831                 log_info(";%3.2f%%", io_u_lat_m[i]);
 832         /* Additional output if continue_on_error set - default off*/
 833         if (ts->continue_on_error)
 834                 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
 835         log_info("\n");
 836
 837         /* Additional output if description is set */
 838         if (strlen(ts->description))
 839                 log_info(";%s", ts->description);
 840
 841         log_info("\n");
 842 }
 843
 844 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
 845                                            struct group_run_stats *rs, int ver)
 846 {
 847         double io_u_dist[FIO_IO_U_MAP_NR];
 848         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 849         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 850         double usr_cpu, sys_cpu;
 851         int i;
 852
 853         /* General Info */
 854         log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
 855                                         ts->name, ts->groupid, ts->error);
 856         /* Log Read Status */
 857         show_ddir_status_terse(ts, rs, DDIR_READ);
 858         /* Log Write Status */
 859         show_ddir_status_terse(ts, rs, DDIR_WRITE);
 860         /* Log Trim Status */
 861         if (ver == 4)
 862                 show_ddir_status_terse(ts, rs, DDIR_TRIM);
 863
 864         /* CPU Usage */
 865         if (ts->total_run_time) {
 866                 double runt = (double) ts->total_run_time;
 867
 868                 usr_cpu = (double) ts->usr_time * 100 / runt;
 869                 sys_cpu = (double) ts->sys_time * 100 / runt;
 870         } else {
 871                 usr_cpu = 0;
 872                 sys_cpu = 0;
 873         }
 874
 875         log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
 876                                                 (unsigned long long) ts->ctx,
 877                                                 (unsigned long long) ts->majf,
 878                                                 (unsigned long long) ts->minf);
 879
 880         /* Calc % distribution of IO depths, usecond, msecond latency */
 881         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 882         stat_calc_lat_u(ts, io_u_lat_u);
 883         stat_calc_lat_m(ts, io_u_lat_m);
 884
 885         /* Only show fixed 7 I/O depth levels*/
 886         log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
 887                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
 888                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
 889
 890         /* Microsecond latency */
 891         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
 892                 log_info(";%3.2f%%", io_u_lat_u[i]);
 893         /* Millisecond latency */
 894         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
 895                 log_info(";%3.2f%%", io_u_lat_m[i]);
 896
 897         /* disk util stats, if any */
 898         if (is_backend)
 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 }
1080
1081 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1082 {
1083         int l, k;
1084
1085         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1086                 if (!dst->unified_rw_rep) {
1087                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1088                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1089                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1090                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1091
1092                         dst->io_bytes[l] += src->io_bytes[l];
1093
1094                         if (dst->runtime[l] < src->runtime[l])
1095                                 dst->runtime[l] = src->runtime[l];
1096                 } else {
1097                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1098                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1099                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1100                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1101
1102                         dst->io_bytes[0] += src->io_bytes[l];
1103
1104                         if (dst->runtime[0] < src->runtime[l])
1105                                 dst->runtime[0] = src->runtime[l];
1106                 }
1107         }
1108
1109         dst->usr_time += src->usr_time;
1110         dst->sys_time += src->sys_time;
1111         dst->ctx += src->ctx;
1112         dst->majf += src->majf;
1113         dst->minf += src->minf;
1114
1115         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1116                 dst->io_u_map[k] += src->io_u_map[k];
1117         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1118                 dst->io_u_submit[k] += src->io_u_submit[k];
1119         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1120                 dst->io_u_complete[k] += src->io_u_complete[k];
1121         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1122                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1123         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1124                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1125
1126         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1127                 if (!dst->unified_rw_rep) {
1128                         dst->total_io_u[k] += src->total_io_u[k];
1129                         dst->short_io_u[k] += src->short_io_u[k];
1130                         dst->drop_io_u[k] += src->drop_io_u[k];
1131                 } else {
1132                         dst->total_io_u[0] += src->total_io_u[k];
1133                         dst->short_io_u[0] += src->short_io_u[k];
1134                         dst->drop_io_u[0] += src->drop_io_u[k];
1135                 }
1136         }
1137
1138         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1139                 int m;
1140
1141                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1142                         if (!dst->unified_rw_rep)
1143                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1144                         else
1145                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1146                 }
1147         }
1148
1149         dst->total_run_time += src->total_run_time;
1150         dst->total_submit += src->total_submit;
1151         dst->total_complete += src->total_complete;
1152 }
1153
1154 void init_group_run_stat(struct group_run_stats *gs)
1155 {
1156         int i;
1157         memset(gs, 0, sizeof(*gs));
1158
1159         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1160                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1161 }
1162
1163 void init_thread_stat(struct thread_stat *ts)
1164 {
1165         int j;
1166
1167         memset(ts, 0, sizeof(*ts));
1168
1169         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1170                 ts->lat_stat[j].min_val = -1UL;
1171                 ts->clat_stat[j].min_val = -1UL;
1172                 ts->slat_stat[j].min_val = -1UL;
1173                 ts->bw_stat[j].min_val = -1UL;
1174         }
1175         ts->groupid = -1;
1176 }
1177
1178 void __show_run_stats(void)
1179 {
1180         struct group_run_stats *runstats, *rs;
1181         struct thread_data *td;
1182         struct thread_stat *threadstats, *ts;
1183         int i, j, nr_ts, last_ts, idx;
1184         int kb_base_warned = 0;
1185         int unit_base_warned = 0;
1186         struct json_object *root = NULL;
1187         struct json_array *array = NULL;
1188
1189         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1190
1191         for (i = 0; i < groupid + 1; i++)
1192                 init_group_run_stat(&runstats[i]);
1193
1194         /*
1195          * find out how many threads stats we need. if group reporting isn't
1196          * enabled, it's one-per-td.
1197          */
1198         nr_ts = 0;
1199         last_ts = -1;
1200         for_each_td(td, i) {
1201                 if (!td->o.group_reporting) {
1202                         nr_ts++;
1203                         continue;
1204                 }
1205                 if (last_ts == td->groupid)
1206                         continue;
1207
1208                 last_ts = td->groupid;
1209                 nr_ts++;
1210         }
1211
1212         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1213
1214         for (i = 0; i < nr_ts; i++)
1215                 init_thread_stat(&threadstats[i]);
1216
1217         j = 0;
1218         last_ts = -1;
1219         idx = 0;
1220         for_each_td(td, i) {
1221                 if (idx && (!td->o.group_reporting ||
1222                     (td->o.group_reporting && last_ts != td->groupid))) {
1223                         idx = 0;
1224                         j++;
1225                 }
1226
1227                 last_ts = td->groupid;
1228
1229                 ts = &threadstats[j];
1230
1231                 ts->clat_percentiles = td->o.clat_percentiles;
1232                 ts->percentile_precision = td->o.percentile_precision;
1233                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1234
1235                 idx++;
1236                 ts->members++;
1237
1238                 if (ts->groupid == -1) {
1239                         /*
1240                          * These are per-group shared already
1241                          */
1242                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1243                         if (td->o.description)
1244                                 strncpy(ts->description, td->o.description,
1245                                                 FIO_JOBDESC_SIZE - 1);
1246                         else
1247                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1248
1249                         /*
1250                          * If multiple entries in this group, this is
1251                          * the first member.
1252                          */
1253                         ts->thread_number = td->thread_number;
1254                         ts->groupid = td->groupid;
1255
1256                         /*
1257                          * first pid in group, not very useful...
1258                          */
1259                         ts->pid = td->pid;
1260
1261                         ts->kb_base = td->o.kb_base;
1262                         ts->unit_base = td->o.unit_base;
1263                         ts->unified_rw_rep = td->o.unified_rw_rep;
1264                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1265                         log_info("fio: kb_base differs for jobs in group, using"
1266                                  " %u as the base\n", ts->kb_base);
1267                         kb_base_warned = 1;
1268                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1269                         log_info("fio: unit_base differs for jobs in group, using"
1270                                  " %u as the base\n", ts->unit_base);
1271                         unit_base_warned = 1;
1272                 }
1273
1274                 ts->continue_on_error = td->o.continue_on_error;
1275                 ts->total_err_count += td->total_err_count;
1276                 ts->first_error = td->first_error;
1277                 if (!ts->error) {
1278                         if (!td->error && td->o.continue_on_error &&
1279                             td->first_error) {
1280                                 ts->error = td->first_error;
1281                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1282                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1283                         } else  if (td->error) {
1284                                 ts->error = td->error;
1285                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1286                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1287                         }
1288                 }
1289
1290                 ts->latency_depth = td->latency_qd;
1291                 ts->latency_target = td->o.latency_target;
1292                 ts->latency_percentile = td->o.latency_percentile;
1293                 ts->latency_window = td->o.latency_window;
1294
1295                 sum_thread_stats(ts, &td->ts, idx);
1296         }
1297
1298         for (i = 0; i < nr_ts; i++) {
1299                 unsigned long long bw;
1300
1301                 ts = &threadstats[i];
1302                 rs = &runstats[ts->groupid];
1303                 rs->kb_base = ts->kb_base;
1304                 rs->unit_base = ts->unit_base;
1305                 rs->unified_rw_rep += ts->unified_rw_rep;
1306
1307                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1308                         if (!ts->runtime[j])
1309                                 continue;
1310                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1311                                 rs->min_run[j] = ts->runtime[j];
1312                         if (ts->runtime[j] > rs->max_run[j])
1313                                 rs->max_run[j] = ts->runtime[j];
1314
1315                         bw = 0;
1316                         if (ts->runtime[j]) {
1317                                 unsigned long runt = ts->runtime[j];
1318                                 unsigned long long kb;
1319
1320                                 kb = ts->io_bytes[j] / rs->kb_base;
1321                                 bw = kb * 1000 / runt;
1322                         }
1323                         if (bw < rs->min_bw[j])
1324                                 rs->min_bw[j] = bw;
1325                         if (bw > rs->max_bw[j])
1326                                 rs->max_bw[j] = bw;
1327
1328                         rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1329                 }
1330         }
1331
1332         for (i = 0; i < groupid + 1; i++) {
1333                 int ddir;
1334
1335                 rs = &runstats[i];
1336
1337                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1338                         if (rs->max_run[ddir])
1339                                 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1340                                                 rs->max_run[ddir];
1341                 }
1342         }
1343
1344         /*
1345          * don't overwrite last signal output
1346          */
1347         if (output_format == FIO_OUTPUT_NORMAL)
1348                 log_info("\n");
1349         else if (output_format == FIO_OUTPUT_JSON) {
1350                 root = json_create_object();
1351                 json_object_add_value_string(root, "fio version", fio_version_string);
1352                 array = json_create_array();
1353                 json_object_add_value_array(root, "jobs", array);
1354         }
1355
1356         for (i = 0; i < nr_ts; i++) {
1357                 ts = &threadstats[i];
1358                 rs = &runstats[ts->groupid];
1359
1360                 if (is_backend)
1361                         fio_server_send_ts(ts, rs);
1362                 else if (output_format == FIO_OUTPUT_TERSE)
1363                         show_thread_status_terse(ts, rs);
1364                 else if (output_format == FIO_OUTPUT_JSON) {
1365                         struct json_object *tmp = show_thread_status_json(ts, rs);
1366                         json_array_add_value_object(array, tmp);
1367                 } else
1368                         show_thread_status_normal(ts, rs);
1369         }
1370         if (output_format == FIO_OUTPUT_JSON) {
1371                 /* disk util stats, if any */
1372                 show_disk_util(1, root);
1373
1374                 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1375
1376                 json_print_object(root);
1377                 log_info("\n");
1378                 json_free_object(root);
1379         }
1380
1381         for (i = 0; i < groupid + 1; i++) {
1382                 rs = &runstats[i];
1383
1384                 rs->groupid = i;
1385                 if (is_backend)
1386                         fio_server_send_gs(rs);
1387                 else if (output_format == FIO_OUTPUT_NORMAL)
1388                         show_group_stats(rs);
1389         }
1390
1391         if (is_backend)
1392                 fio_server_send_du();
1393         else if (output_format == FIO_OUTPUT_NORMAL) {
1394                 show_disk_util(0, NULL);
1395                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1396         }
1397
1398         if ( !(output_format == FIO_OUTPUT_TERSE) && append_terse_output) {
1399                 log_info("\nAdditional Terse Output:\n");
1400
1401                 for (i = 0; i < nr_ts; i++) {
1402                         ts = &threadstats[i];
1403                         rs = &runstats[ts->groupid];
1404                         show_thread_status_terse(ts, rs);
1405                 }
1406         }
1407
1408         log_info_flush();
1409         free(runstats);
1410         free(threadstats);
1411 }
1412
1413 void show_run_stats(void)
1414 {
1415         fio_mutex_down(stat_mutex);
1416         __show_run_stats();
1417         fio_mutex_up(stat_mutex);
1418 }
1419
1420 static void *__show_running_run_stats(void *arg)
1421 {
1422         struct thread_data *td;
1423         unsigned long long *rt;
1424         struct timeval tv;
1425         int i;
1426
1427         fio_mutex_down(stat_mutex);
1428
1429         rt = malloc(thread_number * sizeof(unsigned long long));
1430         fio_gettime(&tv, NULL);
1431
1432         for_each_td(td, i) {
1433                 rt[i] = mtime_since(&td->start, &tv);
1434                 if (td_read(td) && td->io_bytes[DDIR_READ])
1435                         td->ts.runtime[DDIR_READ] += rt[i];
1436                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1437                         td->ts.runtime[DDIR_WRITE] += rt[i];
1438                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1439                         td->ts.runtime[DDIR_TRIM] += rt[i];
1440
1441                 td->update_rusage = 1;
1442                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1443                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1444                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1445                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1446         }
1447
1448         for_each_td(td, i) {
1449                 if (td->rusage_sem) {
1450                         td->update_rusage = 1;
1451                         fio_mutex_down(td->rusage_sem);
1452                 }
1453                 td->update_rusage = 0;
1454         }
1455
1456         __show_run_stats();
1457
1458         for_each_td(td, i) {
1459                 if (td_read(td) && td->io_bytes[DDIR_READ])
1460                         td->ts.runtime[DDIR_READ] -= rt[i];
1461                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1462                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1463                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1464                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1465         }
1466
1467         free(rt);
1468         fio_mutex_up(stat_mutex);
1469         free(arg);
1470         return NULL;
1471 }
1472
1473 /*
1474  * Called from signal handler. It _should_ be safe to just run this inline
1475  * in the sig handler, but we should be disturbing the system less by just
1476  * creating a thread to do it.
1477  */
1478 void show_running_run_stats(void)
1479 {
1480         pthread_t *thread;
1481
1482         thread = calloc(1, sizeof(*thread));
1483         if (!thread)
1484                 return;
1485
1486         if (!pthread_create(thread, NULL, __show_running_run_stats, thread)) {
1487                 int err;
1488
1489                 err = pthread_detach(*thread);
1490                 if (err)
1491                         log_err("fio: DU thread detach failed: %s\n", strerror(err));
1492
1493                 return;
1494         }
1495
1496         free(thread);
1497 }
1498
1499 static int status_interval_init;
1500 static struct timeval status_time;
1501 static int status_file_disabled;
1502
1503 #define FIO_STATUS_FILE         "fio-dump-status"
1504
1505 static int check_status_file(void)
1506 {
1507         struct stat sb;
1508         const char *temp_dir;
1509         char fio_status_file_path[PATH_MAX];
1510
1511         if (status_file_disabled)
1512                 return 0;
1513
1514         temp_dir = getenv("TMPDIR");
1515         if (temp_dir == NULL) {
1516                 temp_dir = getenv("TEMP");
1517                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1518                         temp_dir = NULL;
1519         }
1520         if (temp_dir == NULL)
1521                 temp_dir = "/tmp";
1522
1523         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1524
1525         if (stat(fio_status_file_path, &sb))
1526                 return 0;
1527
1528         if (unlink(fio_status_file_path) < 0) {
1529                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1530                                                         strerror(errno));
1531                 log_err("fio: disabling status file updates\n");
1532                 status_file_disabled = 1;
1533         }
1534
1535         return 1;
1536 }
1537
1538 void check_for_running_stats(void)
1539 {
1540         if (status_interval) {
1541                 if (!status_interval_init) {
1542                         fio_gettime(&status_time, NULL);
1543                         status_interval_init = 1;
1544                 } else if (mtime_since_now(&status_time) >= status_interval) {
1545                         show_running_run_stats();
1546                         fio_gettime(&status_time, NULL);
1547                         return;
1548                 }
1549         }
1550         if (check_status_file()) {
1551                 show_running_run_stats();
1552                 return;
1553         }
1554 }
1555
1556 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1557 {
1558         double val = data;
1559         double delta;
1560
1561         if (data > is->max_val)
1562                 is->max_val = data;
1563         if (data < is->min_val)
1564                 is->min_val = data;
1565
1566         delta = val - is->mean.u.f;
1567         if (delta) {
1568                 is->mean.u.f += delta / (is->samples + 1.0);
1569                 is->S.u.f += delta * (val - is->mean.u.f);
1570         }
1571
1572         is->samples++;
1573 }
1574
1575 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1576                              enum fio_ddir ddir, unsigned int bs,
1577                              unsigned long t, uint64_t offset)
1578 {
1579         uint64_t nr_samples = iolog->nr_samples;
1580         struct io_sample *s;
1581
1582         if (iolog->disabled)
1583                 return;
1584
1585         if (!iolog->nr_samples)
1586                 iolog->avg_last = t;
1587
1588         if (iolog->nr_samples == iolog->max_samples) {
1589                 size_t new_size;
1590                 void *new_log;
1591
1592                 new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
1593
1594                 if (iolog->log_gz && (new_size > iolog->log_gz)) {
1595                         if (iolog_flush(iolog, 0)) {
1596                                 log_err("fio: failed flushing iolog! Will stop logging.\n");
1597                                 iolog->disabled = 1;
1598                                 return;
1599                         }
1600                         nr_samples = iolog->nr_samples;
1601                 } else {
1602                         new_log = realloc(iolog->log, new_size);
1603                         if (!new_log) {
1604                                 log_err("fio: failed extending iolog! Will stop logging.\n");
1605                                 iolog->disabled = 1;
1606                                 return;
1607                         }
1608                         iolog->log = new_log;
1609                         iolog->max_samples <<= 1;
1610                 }
1611         }
1612
1613         s = get_sample(iolog, nr_samples);
1614
1615         s->val = val;
1616         s->time = t;
1617         io_sample_set_ddir(iolog, s, ddir);
1618         s->bs = bs;
1619
1620         if (iolog->log_offset) {
1621                 struct io_sample_offset *so = (void *) s;
1622
1623                 so->offset = offset;
1624         }
1625
1626         iolog->nr_samples++;
1627 }
1628
1629 static inline void reset_io_stat(struct io_stat *ios)
1630 {
1631         ios->max_val = ios->min_val = ios->samples = 0;
1632         ios->mean.u.f = ios->S.u.f = 0;
1633 }
1634
1635 void reset_io_stats(struct thread_data *td)
1636 {
1637         struct thread_stat *ts = &td->ts;
1638         int i, j;
1639
1640         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1641                 reset_io_stat(&ts->clat_stat[i]);
1642                 reset_io_stat(&ts->slat_stat[i]);
1643                 reset_io_stat(&ts->lat_stat[i]);
1644                 reset_io_stat(&ts->bw_stat[i]);
1645                 reset_io_stat(&ts->iops_stat[i]);
1646
1647                 ts->io_bytes[i] = 0;
1648                 ts->runtime[i] = 0;
1649
1650                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1651                         ts->io_u_plat[i][j] = 0;
1652         }
1653
1654         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1655                 ts->io_u_map[i] = 0;
1656                 ts->io_u_submit[i] = 0;
1657                 ts->io_u_complete[i] = 0;
1658                 ts->io_u_lat_u[i] = 0;
1659                 ts->io_u_lat_m[i] = 0;
1660                 ts->total_submit = 0;
1661                 ts->total_complete = 0;
1662         }
1663
1664         for (i = 0; i < 3; i++) {
1665                 ts->total_io_u[i] = 0;
1666                 ts->short_io_u[i] = 0;
1667                 ts->drop_io_u[i] = 0;
1668         }
1669 }
1670
1671 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
1672 {
1673         /*
1674          * Note an entry in the log. Use the mean from the logged samples,
1675          * making sure to properly round up. Only write a log entry if we
1676          * had actual samples done.
1677          */
1678         if (iolog->avg_window[DDIR_READ].samples) {
1679                 unsigned long mr;
1680
1681                 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1682                 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed, 0);
1683         }
1684         if (iolog->avg_window[DDIR_WRITE].samples) {
1685                 unsigned long mw;
1686
1687                 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1688                 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
1689         }
1690         if (iolog->avg_window[DDIR_TRIM].samples) {
1691                 unsigned long mw;
1692
1693                 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1694                 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed, 0);
1695         }
1696
1697         reset_io_stat(&iolog->avg_window[DDIR_READ]);
1698         reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1699         reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1700 }
1701
1702 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1703                            unsigned long val, enum fio_ddir ddir,
1704                            unsigned int bs, uint64_t offset)
1705 {
1706         unsigned long elapsed, this_window;
1707
1708         if (!ddir_rw(ddir))
1709                 return;
1710
1711         elapsed = mtime_since_now(&td->epoch);
1712
1713         /*
1714          * If no time averaging, just add the log sample.
1715          */
1716         if (!iolog->avg_msec) {
1717                 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
1718                 return;
1719         }
1720
1721         /*
1722          * Add the sample. If the time period has passed, then
1723          * add that entry to the log and clear.
1724          */
1725         add_stat_sample(&iolog->avg_window[ddir], val);
1726
1727         /*
1728          * If period hasn't passed, adding the above sample is all we
1729          * need to do.
1730          */
1731         this_window = elapsed - iolog->avg_last;
1732         if (this_window < iolog->avg_msec)
1733                 return;
1734
1735         _add_stat_to_log(iolog, elapsed);
1736
1737         iolog->avg_last = elapsed;
1738 }
1739
1740 void finalize_logs(struct thread_data *td)
1741 {
1742         unsigned long elapsed;
1743
1744         elapsed = mtime_since_now(&td->epoch);
1745
1746         if (td->clat_log)
1747                 _add_stat_to_log(td->clat_log, elapsed);
1748         if (td->slat_log)
1749                 _add_stat_to_log(td->slat_log, elapsed);
1750         if (td->lat_log)
1751                 _add_stat_to_log(td->lat_log, elapsed);
1752         if (td->bw_log)
1753                 _add_stat_to_log(td->bw_log, elapsed);
1754         if (td->iops_log)
1755                 _add_stat_to_log(td->iops_log, elapsed);
1756 }
1757
1758 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1759 {
1760         struct io_log *iolog;
1761
1762         if (!ddir_rw(ddir))
1763                 return;
1764
1765         iolog = agg_io_log[ddir];
1766         __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
1767 }
1768
1769 static void add_clat_percentile_sample(struct thread_stat *ts,
1770                                 unsigned long usec, enum fio_ddir ddir)
1771 {
1772         unsigned int idx = plat_val_to_idx(usec);
1773         assert(idx < FIO_IO_U_PLAT_NR);
1774
1775         ts->io_u_plat[ddir][idx]++;
1776 }
1777
1778 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1779                      unsigned long usec, unsigned int bs, uint64_t offset)
1780 {
1781         struct thread_stat *ts = &td->ts;
1782
1783         if (!ddir_rw(ddir))
1784                 return;
1785
1786         add_stat_sample(&ts->clat_stat[ddir], usec);
1787
1788         if (td->clat_log)
1789                 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
1790
1791         if (ts->clat_percentiles)
1792                 add_clat_percentile_sample(ts, usec, ddir);
1793 }
1794
1795 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1796                      unsigned long usec, unsigned int bs, uint64_t offset)
1797 {
1798         struct thread_stat *ts = &td->ts;
1799
1800         if (!ddir_rw(ddir))
1801                 return;
1802
1803         add_stat_sample(&ts->slat_stat[ddir], usec);
1804
1805         if (td->slat_log)
1806                 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
1807 }
1808
1809 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1810                     unsigned long usec, unsigned int bs, uint64_t offset)
1811 {
1812         struct thread_stat *ts = &td->ts;
1813
1814         if (!ddir_rw(ddir))
1815                 return;
1816
1817         add_stat_sample(&ts->lat_stat[ddir], usec);
1818
1819         if (td->lat_log)
1820                 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
1821 }
1822
1823 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1824                    struct timeval *t)
1825 {
1826         struct thread_stat *ts = &td->ts;
1827         unsigned long spent, rate;
1828
1829         if (!ddir_rw(ddir))
1830                 return;
1831
1832         spent = mtime_since(&td->bw_sample_time, t);
1833         if (spent < td->o.bw_avg_time)
1834                 return;
1835
1836         /*
1837          * Compute both read and write rates for the interval.
1838          */
1839         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1840                 uint64_t delta;
1841
1842                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1843                 if (!delta)
1844                         continue; /* No entries for interval */
1845
1846                 if (spent)
1847                         rate = delta * 1000 / spent / 1024;
1848                 else
1849                         rate = 0;
1850
1851                 add_stat_sample(&ts->bw_stat[ddir], rate);
1852
1853                 if (td->bw_log)
1854                         add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
1855
1856                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1857         }
1858
1859         fio_gettime(&td->bw_sample_time, NULL);
1860 }
1861
1862 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1863                      struct timeval *t)
1864 {
1865         struct thread_stat *ts = &td->ts;
1866         unsigned long spent, iops;
1867
1868         if (!ddir_rw(ddir))
1869                 return;
1870
1871         spent = mtime_since(&td->iops_sample_time, t);
1872         if (spent < td->o.iops_avg_time)
1873                 return;
1874
1875         /*
1876          * Compute both read and write rates for the interval.
1877          */
1878         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1879                 uint64_t delta;
1880
1881                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1882                 if (!delta)
1883                         continue; /* No entries for interval */
1884
1885                 if (spent)
1886                         iops = (delta * 1000) / spent;
1887                 else
1888                         iops = 0;
1889
1890                 add_stat_sample(&ts->iops_stat[ddir], iops);
1891
1892                 if (td->iops_log)
1893                         add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
1894
1895                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1896         }
1897
1898         fio_gettime(&td->iops_sample_time, NULL);
1899 }
1900
1901 void stat_init(void)
1902 {
1903         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1904 }
1905
1906 void stat_exit(void)
1907 {
1908         /*
1909          * When we have the mutex, we know out-of-band access to it
1910          * have ended.
1911          */
1912         fio_mutex_down(stat_mutex);
1913         fio_mutex_remove(stat_mutex);
1914 }