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 #include "lib/pow2.h"
  17 #include "lib/output_buffer.h"
  18 #include "helper_thread.h"
  19 #include "smalloc.h"
  20
  21 #define LOG_MSEC_SLACK  10
  22
  23 struct fio_mutex *stat_mutex;
  24
  25 void clear_rusage_stat(struct thread_data *td)
  26 {
  27         struct thread_stat *ts = &td->ts;
  28
  29         fio_getrusage(&td->ru_start);
  30         ts->usr_time = ts->sys_time = 0;
  31         ts->ctx = 0;
  32         ts->minf = ts->majf = 0;
  33 }
  34
  35 void update_rusage_stat(struct thread_data *td)
  36 {
  37         struct thread_stat *ts = &td->ts;
  38
  39         fio_getrusage(&td->ru_end);
  40         ts->usr_time += mtime_since(&td->ru_start.ru_utime,
  41                                         &td->ru_end.ru_utime);
  42         ts->sys_time += mtime_since(&td->ru_start.ru_stime,
  43                                         &td->ru_end.ru_stime);
  44         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  45                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  46         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  47         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  48
  49         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  50 }
  51
  52 /*
  53  * Given a latency, return the index of the corresponding bucket in
  54  * the structure tracking percentiles.
  55  *
  56  * (1) find the group (and error bits) that the value (latency)
  57  * belongs to by looking at its MSB. (2) find the bucket number in the
  58  * group by looking at the index bits.
  59  *
  60  */
  61 static unsigned int plat_val_to_idx(unsigned int val)
  62 {
  63         unsigned int msb, error_bits, base, offset, idx;
  64
  65         /* Find MSB starting from bit 0 */
  66         if (val == 0)
  67                 msb = 0;
  68         else
  69                 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
  70
  71         /*
  72          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  73          * all bits of the sample as index
  74          */
  75         if (msb <= FIO_IO_U_PLAT_BITS)
  76                 return val;
  77
  78         /* Compute the number of error bits to discard*/
  79         error_bits = msb - FIO_IO_U_PLAT_BITS;
  80
  81         /* Compute the number of buckets before the group */
  82         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  83
  84         /*
  85          * Discard the error bits and apply the mask to find the
  86          * index for the buckets in the group
  87          */
  88         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  89
  90         /* Make sure the index does not exceed (array size - 1) */
  91         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  92                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  93
  94         return idx;
  95 }
  96
  97 /*
  98  * Convert the given index of the bucket array to the value
  99  * represented by the bucket
 100  */
 101 static unsigned int plat_idx_to_val(unsigned int idx)
 102 {
 103         unsigned int error_bits, k, base;
 104
 105         assert(idx < FIO_IO_U_PLAT_NR);
 106
 107         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 108          * all bits of the sample as index */
 109         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 110                 return idx;
 111
 112         /* Find the group and compute the minimum value of that group */
 113         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 114         base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
 115
 116         /* Find its bucket number of the group */
 117         k = idx % FIO_IO_U_PLAT_VAL;
 118
 119         /* Return the mean of the range of the bucket */
 120         return base + ((k + 0.5) * (1 << error_bits));
 121 }
 122
 123 static int double_cmp(const void *a, const void *b)
 124 {
 125         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 126         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 127         int cmp = 0;
 128
 129         if (fa.u.f > fb.u.f)
 130                 cmp = 1;
 131         else if (fa.u.f < fb.u.f)
 132                 cmp = -1;
 133
 134         return cmp;
 135 }
 136
 137 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 138                                    fio_fp64_t *plist, unsigned int **output,
 139                                    unsigned int *maxv, unsigned int *minv)
 140 {
 141         unsigned long sum = 0;
 142         unsigned int len, i, j = 0;
 143         unsigned int oval_len = 0;
 144         unsigned int *ovals = NULL;
 145         int is_last;
 146
 147         *minv = -1U;
 148         *maxv = 0;
 149
 150         len = 0;
 151         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 152                 len++;
 153
 154         if (!len)
 155                 return 0;
 156
 157         /*
 158          * Sort the percentile list. Note that it may already be sorted if
 159          * we are using the default values, but since it's a short list this
 160          * isn't a worry. Also note that this does not work for NaN values.
 161          */
 162         if (len > 1)
 163                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 164
 165         /*
 166          * Calculate bucket values, note down max and min values
 167          */
 168         is_last = 0;
 169         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 170                 sum += io_u_plat[i];
 171                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 172                         assert(plist[j].u.f <= 100.0);
 173
 174                         if (j == oval_len) {
 175                                 oval_len += 100;
 176                                 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
 177                         }
 178
 179                         ovals[j] = plat_idx_to_val(i);
 180                         if (ovals[j] < *minv)
 181                                 *minv = ovals[j];
 182                         if (ovals[j] > *maxv)
 183                                 *maxv = ovals[j];
 184
 185                         is_last = (j == len - 1);
 186                         if (is_last)
 187                                 break;
 188
 189                         j++;
 190                 }
 191         }
 192
 193         *output = ovals;
 194         return len;
 195 }
 196
 197 /*
 198  * Find and display the p-th percentile of clat
 199  */
 200 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 201                                   fio_fp64_t *plist, unsigned int precision,
 202                                   struct buf_output *out)
 203 {
 204         unsigned int len, j = 0, minv, maxv;
 205         unsigned int *ovals;
 206         int is_last, per_line, scale_down;
 207         char fmt[32];
 208
 209         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 210         if (!len)
 211                 goto out;
 212
 213         /*
 214          * We default to usecs, but if the value range is such that we
 215          * should scale down to msecs, do that.
 216          */
 217         if (minv > 2000 && maxv > 99999) {
 218                 scale_down = 1;
 219                 log_buf(out, "    clat percentiles (msec):\n     |");
 220         } else {
 221                 scale_down = 0;
 222                 log_buf(out, "    clat percentiles (usec):\n     |");
 223         }
 224
 225         snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
 226         per_line = (80 - 7) / (precision + 14);
 227
 228         for (j = 0; j < len; j++) {
 229                 char fbuf[16], *ptr = fbuf;
 230
 231                 /* for formatting */
 232                 if (j != 0 && (j % per_line) == 0)
 233                         log_buf(out, "     |");
 234
 235                 /* end of the list */
 236                 is_last = (j == len - 1);
 237
 238                 if (plist[j].u.f < 10.0)
 239                         ptr += sprintf(fbuf, " ");
 240
 241                 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
 242
 243                 if (scale_down)
 244                         ovals[j] = (ovals[j] + 999) / 1000;
 245
 246                 log_buf(out, " %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
 247
 248                 if (is_last)
 249                         break;
 250
 251                 if ((j % per_line) == per_line - 1)     /* for formatting */
 252                         log_buf(out, "\n");
 253         }
 254
 255 out:
 256         if (ovals)
 257                 free(ovals);
 258 }
 259
 260 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 261              double *mean, double *dev)
 262 {
 263         double n = (double) is->samples;
 264
 265         if (n == 0)
 266                 return 0;
 267
 268         *min = is->min_val;
 269         *max = is->max_val;
 270         *mean = is->mean.u.f;
 271
 272         if (n > 1.0)
 273                 *dev = sqrt(is->S.u.f / (n - 1.0));
 274         else
 275                 *dev = 0;
 276
 277         return 1;
 278 }
 279
 280 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 281 {
 282         char *p1, *p2, *p3, *p4;
 283         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 284         int i;
 285
 286         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 287
 288         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 289                 const int i2p = is_power_of_2(rs->kb_base);
 290
 291                 if (!rs->max_run[i])
 292                         continue;
 293
 294                 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
 295                 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
 296                 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 297                 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 298
 299                 log_buf(out, "%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
 300                          " mint=%llumsec, maxt=%llumsec\n",
 301                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 302                                 p1, p2, p3, p4,
 303                                 (unsigned long long) rs->min_run[i],
 304                                 (unsigned long long) rs->max_run[i]);
 305
 306                 free(p1);
 307                 free(p2);
 308                 free(p3);
 309                 free(p4);
 310         }
 311 }
 312
 313 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 314 {
 315         int i;
 316
 317         /*
 318          * Do depth distribution calculations
 319          */
 320         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 321                 if (total) {
 322                         io_u_dist[i] = (double) map[i] / (double) total;
 323                         io_u_dist[i] *= 100.0;
 324                         if (io_u_dist[i] < 0.1 && map[i])
 325                                 io_u_dist[i] = 0.1;
 326                 } else
 327                         io_u_dist[i] = 0.0;
 328         }
 329 }
 330
 331 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 332                           unsigned int *src, int nr)
 333 {
 334         unsigned long total = ddir_rw_sum(ts->total_io_u);
 335         int i;
 336
 337         /*
 338          * Do latency distribution calculations
 339          */
 340         for (i = 0; i < nr; i++) {
 341                 if (total) {
 342                         dst[i] = (double) src[i] / (double) total;
 343                         dst[i] *= 100.0;
 344                         if (dst[i] < 0.01 && src[i])
 345                                 dst[i] = 0.01;
 346                 } else
 347                         dst[i] = 0.0;
 348         }
 349 }
 350
 351 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 352 {
 353         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 354 }
 355
 356 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 357 {
 358         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 359 }
 360
 361 static void display_lat(const char *name, unsigned long min, unsigned long max,
 362                         double mean, double dev, struct buf_output *out)
 363 {
 364         const char *base = "(usec)";
 365         char *minp, *maxp;
 366
 367         if (!usec_to_msec(&min, &max, &mean, &dev))
 368                 base = "(msec)";
 369
 370         minp = num2str(min, 6, 1, 0, 0);
 371         maxp = num2str(max, 6, 1, 0, 0);
 372
 373         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 374                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 375
 376         free(minp);
 377         free(maxp);
 378 }
 379
 380 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 381                              int ddir, struct buf_output *out)
 382 {
 383         const char *str[] = { "read ", "write", "trim" };
 384         unsigned long min, max, runt;
 385         unsigned long long bw, iops;
 386         double mean, dev;
 387         char *io_p, *bw_p, *iops_p;
 388         int i2p;
 389
 390         assert(ddir_rw(ddir));
 391
 392         if (!ts->runtime[ddir])
 393                 return;
 394
 395         i2p = is_power_of_2(rs->kb_base);
 396         runt = ts->runtime[ddir];
 397
 398         bw = (1000 * ts->io_bytes[ddir]) / runt;
 399         io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
 400         bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
 401
 402         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 403         iops_p = num2str(iops, 6, 1, 0, 0);
 404
 405         log_buf(out, "  %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
 406                                 rs->unified_rw_rep ? "mixed" : str[ddir],
 407                                 io_p, bw_p, iops_p,
 408                                 (unsigned long long) ts->runtime[ddir]);
 409
 410         free(io_p);
 411         free(bw_p);
 412         free(iops_p);
 413
 414         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 415                 display_lat("slat", min, max, mean, dev, out);
 416         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 417                 display_lat("clat", min, max, mean, dev, out);
 418         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 419                 display_lat(" lat", min, max, mean, dev, out);
 420
 421         if (ts->clat_percentiles) {
 422                 show_clat_percentiles(ts->io_u_plat[ddir],
 423                                         ts->clat_stat[ddir].samples,
 424                                         ts->percentile_list,
 425                                         ts->percentile_precision, out);
 426         }
 427         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 428                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 429                 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
 430
 431                 if (rs->unit_base == 1) {
 432                         min *= 8.0;
 433                         max *= 8.0;
 434                         mean *= 8.0;
 435                         dev *= 8.0;
 436                 }
 437
 438                 if (rs->agg[ddir]) {
 439                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 440                         if (p_of_agg > 100.0)
 441                                 p_of_agg = 100.0;
 442                 }
 443
 444                 if (mean > fkb_base * fkb_base) {
 445                         min /= fkb_base;
 446                         max /= fkb_base;
 447                         mean /= fkb_base;
 448                         dev /= fkb_base;
 449                         bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
 450                 }
 451
 452                 log_buf(out, "    bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
 453                          " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
 454                                                         p_of_agg, mean, dev);
 455         }
 456 }
 457
 458 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 459                     const char *msg, struct buf_output *out)
 460 {
 461         int new_line = 1, i, line = 0, shown = 0;
 462
 463         for (i = 0; i < nr; i++) {
 464                 if (io_u_lat[i] <= 0.0)
 465                         continue;
 466                 shown = 1;
 467                 if (new_line) {
 468                         if (line)
 469                                 log_buf(out, "\n");
 470                         log_buf(out, "    lat (%s) : ", msg);
 471                         new_line = 0;
 472                         line = 0;
 473                 }
 474                 if (line)
 475                         log_buf(out, ", ");
 476                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 477                 line++;
 478                 if (line == 5)
 479                         new_line = 1;
 480         }
 481
 482         if (shown)
 483                 log_buf(out, "\n");
 484
 485         return shown;
 486 }
 487
 488 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 489 {
 490         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 491                                  "250=", "500=", "750=", "1000=", };
 492
 493         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 494 }
 495
 496 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 497 {
 498         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 499                                  "250=", "500=", "750=", "1000=", "2000=",
 500                                  ">=2000=", };
 501
 502         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 503 }
 504
 505 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 506 {
 507         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 508         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 509
 510         stat_calc_lat_u(ts, io_u_lat_u);
 511         stat_calc_lat_m(ts, io_u_lat_m);
 512
 513         show_lat_u(io_u_lat_u, out);
 514         show_lat_m(io_u_lat_m, out);
 515 }
 516
 517 static int block_state_category(int block_state)
 518 {
 519         switch (block_state) {
 520         case BLOCK_STATE_UNINIT:
 521                 return 0;
 522         case BLOCK_STATE_TRIMMED:
 523         case BLOCK_STATE_WRITTEN:
 524                 return 1;
 525         case BLOCK_STATE_WRITE_FAILURE:
 526         case BLOCK_STATE_TRIM_FAILURE:
 527                 return 2;
 528         default:
 529                 /* Silence compile warning on some BSDs and have a return */
 530                 assert(0);
 531                 return -1;
 532         }
 533 }
 534
 535 static int compare_block_infos(const void *bs1, const void *bs2)
 536 {
 537         uint32_t block1 = *(uint32_t *)bs1;
 538         uint32_t block2 = *(uint32_t *)bs2;
 539         int state1 = BLOCK_INFO_STATE(block1);
 540         int state2 = BLOCK_INFO_STATE(block2);
 541         int bscat1 = block_state_category(state1);
 542         int bscat2 = block_state_category(state2);
 543         int cycles1 = BLOCK_INFO_TRIMS(block1);
 544         int cycles2 = BLOCK_INFO_TRIMS(block2);
 545
 546         if (bscat1 < bscat2)
 547                 return -1;
 548         if (bscat1 > bscat2)
 549                 return 1;
 550
 551         if (cycles1 < cycles2)
 552                 return -1;
 553         if (cycles1 > cycles2)
 554                 return 1;
 555
 556         if (state1 < state2)
 557                 return -1;
 558         if (state1 > state2)
 559                 return 1;
 560
 561         assert(block1 == block2);
 562         return 0;
 563 }
 564
 565 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 566                                   fio_fp64_t *plist, unsigned int **percentiles,
 567                                   unsigned int *types)
 568 {
 569         int len = 0;
 570         int i, nr_uninit;
 571
 572         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 573
 574         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 575                 len++;
 576
 577         if (!len)
 578                 return 0;
 579
 580         /*
 581          * Sort the percentile list. Note that it may already be sorted if
 582          * we are using the default values, but since it's a short list this
 583          * isn't a worry. Also note that this does not work for NaN values.
 584          */
 585         if (len > 1)
 586                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 587
 588         nr_uninit = 0;
 589         /* Start only after the uninit entries end */
 590         for (nr_uninit = 0;
 591              nr_uninit < nr_block_infos
 592                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 593              nr_uninit ++)
 594                 ;
 595
 596         if (nr_uninit == nr_block_infos)
 597                 return 0;
 598
 599         *percentiles = calloc(len, sizeof(**percentiles));
 600
 601         for (i = 0; i < len; i++) {
 602                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 603                                 + nr_uninit;
 604                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 605         }
 606
 607         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 608         for (i = 0; i < nr_block_infos; i++)
 609                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 610
 611         return len;
 612 }
 613
 614 static const char *block_state_names[] = {
 615         [BLOCK_STATE_UNINIT] = "unwritten",
 616         [BLOCK_STATE_TRIMMED] = "trimmed",
 617         [BLOCK_STATE_WRITTEN] = "written",
 618         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 619         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 620 };
 621
 622 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 623                              fio_fp64_t *plist, struct buf_output *out)
 624 {
 625         int len, pos, i;
 626         unsigned int *percentiles = NULL;
 627         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 628
 629         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 630                                      &percentiles, block_state_counts);
 631
 632         log_buf(out, "  block lifetime percentiles :\n   |");
 633         pos = 0;
 634         for (i = 0; i < len; i++) {
 635                 uint32_t block_info = percentiles[i];
 636 #define LINE_LENGTH     75
 637                 char str[LINE_LENGTH];
 638                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 639                                      plist[i].u.f, block_info,
 640                                      i == len - 1 ? '\n' : ',');
 641                 assert(strln < LINE_LENGTH);
 642                 if (pos + strln > LINE_LENGTH) {
 643                         pos = 0;
 644                         log_buf(out, "\n   |");
 645                 }
 646                 log_buf(out, "%s", str);
 647                 pos += strln;
 648 #undef LINE_LENGTH
 649         }
 650         if (percentiles)
 651                 free(percentiles);
 652
 653         log_buf(out, "        states               :");
 654         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 655                 log_buf(out, " %s=%u%c",
 656                          block_state_names[i], block_state_counts[i],
 657                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 658 }
 659
 660 static void show_thread_status_normal(struct thread_stat *ts,
 661                                       struct group_run_stats *rs,
 662                                       struct buf_output *out)
 663 {
 664         double usr_cpu, sys_cpu;
 665         unsigned long runtime;
 666         double io_u_dist[FIO_IO_U_MAP_NR];
 667         time_t time_p;
 668         char time_buf[32];
 669
 670         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 671                 return;
 672
 673         time(&time_p);
 674         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 675
 676         if (!ts->error) {
 677                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 678                                         ts->name, ts->groupid, ts->members,
 679                                         ts->error, (int) ts->pid, time_buf);
 680         } else {
 681                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 682                                         ts->name, ts->groupid, ts->members,
 683                                         ts->error, ts->verror, (int) ts->pid,
 684                                         time_buf);
 685         }
 686
 687         if (strlen(ts->description))
 688                 log_buf(out, "  Description  : [%s]\n", ts->description);
 689
 690         if (ts->io_bytes[DDIR_READ])
 691                 show_ddir_status(rs, ts, DDIR_READ, out);
 692         if (ts->io_bytes[DDIR_WRITE])
 693                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 694         if (ts->io_bytes[DDIR_TRIM])
 695                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 696
 697         show_latencies(ts, out);
 698
 699         runtime = ts->total_run_time;
 700         if (runtime) {
 701                 double runt = (double) runtime;
 702
 703                 usr_cpu = (double) ts->usr_time * 100 / runt;
 704                 sys_cpu = (double) ts->sys_time * 100 / runt;
 705         } else {
 706                 usr_cpu = 0;
 707                 sys_cpu = 0;
 708         }
 709
 710         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 711                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 712                         (unsigned long long) ts->ctx,
 713                         (unsigned long long) ts->majf,
 714                         (unsigned long long) ts->minf);
 715
 716         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 717         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 718                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 719                                         io_u_dist[1], io_u_dist[2],
 720                                         io_u_dist[3], io_u_dist[4],
 721                                         io_u_dist[5], io_u_dist[6]);
 722
 723         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 724         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 725                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 726                                         io_u_dist[1], io_u_dist[2],
 727                                         io_u_dist[3], io_u_dist[4],
 728                                         io_u_dist[5], io_u_dist[6]);
 729         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 730         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 731                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 732                                         io_u_dist[1], io_u_dist[2],
 733                                         io_u_dist[3], io_u_dist[4],
 734                                         io_u_dist[5], io_u_dist[6]);
 735         log_buf(out, "     issued    : total=r=%llu/w=%llu/d=%llu,"
 736                                  " short=r=%llu/w=%llu/d=%llu,"
 737                                  " drop=r=%llu/w=%llu/d=%llu\n",
 738                                         (unsigned long long) ts->total_io_u[0],
 739                                         (unsigned long long) ts->total_io_u[1],
 740                                         (unsigned long long) ts->total_io_u[2],
 741                                         (unsigned long long) ts->short_io_u[0],
 742                                         (unsigned long long) ts->short_io_u[1],
 743                                         (unsigned long long) ts->short_io_u[2],
 744                                         (unsigned long long) ts->drop_io_u[0],
 745                                         (unsigned long long) ts->drop_io_u[1],
 746                                         (unsigned long long) ts->drop_io_u[2]);
 747         if (ts->continue_on_error) {
 748                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 749                                         (unsigned long long)ts->total_err_count,
 750                                         ts->first_error,
 751                                         strerror(ts->first_error));
 752         }
 753         if (ts->latency_depth) {
 754                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 755                                         (unsigned long long)ts->latency_target,
 756                                         (unsigned long long)ts->latency_window,
 757                                         ts->latency_percentile.u.f,
 758                                         ts->latency_depth);
 759         }
 760
 761         if (ts->nr_block_infos)
 762                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 763                                   ts->percentile_list, out);
 764 }
 765
 766 static void show_ddir_status_terse(struct thread_stat *ts,
 767                                    struct group_run_stats *rs, int ddir,
 768                                    struct buf_output *out)
 769 {
 770         unsigned long min, max;
 771         unsigned long long bw, iops;
 772         unsigned int *ovals = NULL;
 773         double mean, dev;
 774         unsigned int len, minv, maxv;
 775         int i;
 776
 777         assert(ddir_rw(ddir));
 778
 779         iops = bw = 0;
 780         if (ts->runtime[ddir]) {
 781                 uint64_t runt = ts->runtime[ddir];
 782
 783                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 784                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 785         }
 786
 787         log_buf(out, ";%llu;%llu;%llu;%llu",
 788                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 789                                         (unsigned long long) ts->runtime[ddir]);
 790
 791         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 792                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 793         else
 794                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 795
 796         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 797                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 798         else
 799                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 800
 801         if (ts->clat_percentiles) {
 802                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 803                                         ts->clat_stat[ddir].samples,
 804                                         ts->percentile_list, &ovals, &maxv,
 805                                         &minv);
 806         } else
 807                 len = 0;
 808
 809         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 810                 if (i >= len) {
 811                         log_buf(out, ";0%%=0");
 812                         continue;
 813                 }
 814                 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
 815         }
 816
 817         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 818                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 819         else
 820                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 821
 822         if (ovals)
 823                 free(ovals);
 824
 825         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 826                 double p_of_agg = 100.0;
 827
 828                 if (rs->agg[ddir]) {
 829                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 830                         if (p_of_agg > 100.0)
 831                                 p_of_agg = 100.0;
 832                 }
 833
 834                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 835         } else
 836                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 837 }
 838
 839 static void add_ddir_status_json(struct thread_stat *ts,
 840                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 841 {
 842         unsigned long min, max;
 843         unsigned long long bw;
 844         unsigned int *ovals = NULL;
 845         double mean, dev, iops;
 846         unsigned int len, minv, maxv;
 847         int i;
 848         const char *ddirname[] = {"read", "write", "trim"};
 849         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 850         char buf[120];
 851         double p_of_agg = 100.0;
 852
 853         assert(ddir_rw(ddir));
 854
 855         if (ts->unified_rw_rep && ddir != DDIR_READ)
 856                 return;
 857
 858         dir_object = json_create_object();
 859         json_object_add_value_object(parent,
 860                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 861
 862         bw = 0;
 863         iops = 0.0;
 864         if (ts->runtime[ddir]) {
 865                 uint64_t runt = ts->runtime[ddir];
 866
 867                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 868                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 869         }
 870
 871         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
 872         json_object_add_value_int(dir_object, "bw", bw);
 873         json_object_add_value_float(dir_object, "iops", iops);
 874         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 875         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 876         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 877         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 878
 879         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 880                 min = max = 0;
 881                 mean = dev = 0.0;
 882         }
 883         tmp_object = json_create_object();
 884         json_object_add_value_object(dir_object, "slat", tmp_object);
 885         json_object_add_value_int(tmp_object, "min", min);
 886         json_object_add_value_int(tmp_object, "max", max);
 887         json_object_add_value_float(tmp_object, "mean", mean);
 888         json_object_add_value_float(tmp_object, "stddev", dev);
 889
 890         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 891                 min = max = 0;
 892                 mean = dev = 0.0;
 893         }
 894         tmp_object = json_create_object();
 895         json_object_add_value_object(dir_object, "clat", tmp_object);
 896         json_object_add_value_int(tmp_object, "min", min);
 897         json_object_add_value_int(tmp_object, "max", max);
 898         json_object_add_value_float(tmp_object, "mean", mean);
 899         json_object_add_value_float(tmp_object, "stddev", dev);
 900
 901         if (ts->clat_percentiles) {
 902                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 903                                         ts->clat_stat[ddir].samples,
 904                                         ts->percentile_list, &ovals, &maxv,
 905                                         &minv);
 906         } else
 907                 len = 0;
 908
 909         percentile_object = json_create_object();
 910         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 911         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 912                 if (i >= len) {
 913                         json_object_add_value_int(percentile_object, "0.00", 0);
 914                         continue;
 915                 }
 916                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 917                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 918         }
 919
 920         if (output_format & FIO_OUTPUT_JSON_PLUS) {
 921                 clat_bins_object = json_create_object();
 922                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
 923                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
 924                         snprintf(buf, sizeof(buf), "%d", i);
 925                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
 926                 }
 927                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS);
 928                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL);
 929                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR);
 930         }
 931
 932         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
 933                 min = max = 0;
 934                 mean = dev = 0.0;
 935         }
 936         tmp_object = json_create_object();
 937         json_object_add_value_object(dir_object, "lat", tmp_object);
 938         json_object_add_value_int(tmp_object, "min", min);
 939         json_object_add_value_int(tmp_object, "max", max);
 940         json_object_add_value_float(tmp_object, "mean", mean);
 941         json_object_add_value_float(tmp_object, "stddev", dev);
 942         if (ovals)
 943                 free(ovals);
 944
 945         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 946                 if (rs->agg[ddir]) {
 947                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 948                         if (p_of_agg > 100.0)
 949                                 p_of_agg = 100.0;
 950                 }
 951         } else {
 952                 min = max = 0;
 953                 p_of_agg = mean = dev = 0.0;
 954         }
 955         json_object_add_value_int(dir_object, "bw_min", min);
 956         json_object_add_value_int(dir_object, "bw_max", max);
 957         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
 958         json_object_add_value_float(dir_object, "bw_mean", mean);
 959         json_object_add_value_float(dir_object, "bw_dev", dev);
 960 }
 961
 962 static void show_thread_status_terse_v2(struct thread_stat *ts,
 963                                         struct group_run_stats *rs,
 964                                         struct buf_output *out)
 965 {
 966         double io_u_dist[FIO_IO_U_MAP_NR];
 967         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 968         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 969         double usr_cpu, sys_cpu;
 970         int i;
 971
 972         /* General Info */
 973         log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
 974         /* Log Read Status */
 975         show_ddir_status_terse(ts, rs, DDIR_READ, out);
 976         /* Log Write Status */
 977         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
 978         /* Log Trim Status */
 979         show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
 980
 981         /* CPU Usage */
 982         if (ts->total_run_time) {
 983                 double runt = (double) ts->total_run_time;
 984
 985                 usr_cpu = (double) ts->usr_time * 100 / runt;
 986                 sys_cpu = (double) ts->sys_time * 100 / runt;
 987         } else {
 988                 usr_cpu = 0;
 989                 sys_cpu = 0;
 990         }
 991
 992         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
 993                                                 (unsigned long long) ts->ctx,
 994                                                 (unsigned long long) ts->majf,
 995                                                 (unsigned long long) ts->minf);
 996
 997         /* Calc % distribution of IO depths, usecond, msecond latency */
 998         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 999         stat_calc_lat_u(ts, io_u_lat_u);
1000         stat_calc_lat_m(ts, io_u_lat_m);
1001
1002         /* Only show fixed 7 I/O depth levels*/
1003         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1004                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1005                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1006
1007         /* Microsecond latency */
1008         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1009                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1010         /* Millisecond latency */
1011         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1012                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1013         /* Additional output if continue_on_error set - default off*/
1014         if (ts->continue_on_error)
1015                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1016         log_buf(out, "\n");
1017
1018         /* Additional output if description is set */
1019         if (strlen(ts->description))
1020                 log_buf(out, ";%s", ts->description);
1021
1022         log_buf(out, "\n");
1023 }
1024
1025 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1026                                            struct group_run_stats *rs, int ver,
1027                                            struct buf_output *out)
1028 {
1029         double io_u_dist[FIO_IO_U_MAP_NR];
1030         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1031         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1032         double usr_cpu, sys_cpu;
1033         int i;
1034
1035         /* General Info */
1036         log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1037                                         ts->name, ts->groupid, ts->error);
1038         /* Log Read Status */
1039         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1040         /* Log Write Status */
1041         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1042         /* Log Trim Status */
1043         if (ver == 4)
1044                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1045
1046         /* CPU Usage */
1047         if (ts->total_run_time) {
1048                 double runt = (double) ts->total_run_time;
1049
1050                 usr_cpu = (double) ts->usr_time * 100 / runt;
1051                 sys_cpu = (double) ts->sys_time * 100 / runt;
1052         } else {
1053                 usr_cpu = 0;
1054                 sys_cpu = 0;
1055         }
1056
1057         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1058                                                 (unsigned long long) ts->ctx,
1059                                                 (unsigned long long) ts->majf,
1060                                                 (unsigned long long) ts->minf);
1061
1062         /* Calc % distribution of IO depths, usecond, msecond latency */
1063         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1064         stat_calc_lat_u(ts, io_u_lat_u);
1065         stat_calc_lat_m(ts, io_u_lat_m);
1066
1067         /* Only show fixed 7 I/O depth levels*/
1068         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1069                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1070                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1071
1072         /* Microsecond latency */
1073         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1074                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1075         /* Millisecond latency */
1076         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1077                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1078
1079         /* disk util stats, if any */
1080         show_disk_util(1, NULL, out);
1081
1082         /* Additional output if continue_on_error set - default off*/
1083         if (ts->continue_on_error)
1084                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1085
1086         /* Additional output if description is set */
1087         if (strlen(ts->description))
1088                 log_buf(out, ";%s", ts->description);
1089
1090         log_buf(out, "\n");
1091 }
1092
1093 void json_add_job_opts(struct json_object *root, const char *name,
1094                        struct flist_head *opt_list, bool num_jobs)
1095 {
1096         struct json_object *dir_object;
1097         struct flist_head *entry;
1098         struct print_option *p;
1099
1100         if (flist_empty(opt_list))
1101                 return;
1102
1103         dir_object = json_create_object();
1104         json_object_add_value_object(root, name, dir_object);
1105
1106         flist_for_each(entry, opt_list) {
1107                 const char *pos = "";
1108
1109                 p = flist_entry(entry, struct print_option, list);
1110                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1111                         continue;
1112                 if (p->value)
1113                         pos = p->value;
1114                 json_object_add_value_string(dir_object, p->name, pos);
1115         }
1116 }
1117
1118 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1119                                                    struct group_run_stats *rs,
1120                                                    struct flist_head *opt_list)
1121 {
1122         struct json_object *root, *tmp;
1123         struct jobs_eta *je;
1124         double io_u_dist[FIO_IO_U_MAP_NR];
1125         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1126         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1127         double usr_cpu, sys_cpu;
1128         int i;
1129         size_t size;
1130
1131         root = json_create_object();
1132         json_object_add_value_string(root, "jobname", ts->name);
1133         json_object_add_value_int(root, "groupid", ts->groupid);
1134         json_object_add_value_int(root, "error", ts->error);
1135
1136         /* ETA Info */
1137         je = get_jobs_eta(true, &size);
1138         if (je) {
1139                 json_object_add_value_int(root, "eta", je->eta_sec);
1140                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1141         }
1142
1143         if (opt_list)
1144                 json_add_job_opts(root, "job options", opt_list, true);
1145
1146         add_ddir_status_json(ts, rs, DDIR_READ, root);
1147         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1148         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1149
1150         /* CPU Usage */
1151         if (ts->total_run_time) {
1152                 double runt = (double) ts->total_run_time;
1153
1154                 usr_cpu = (double) ts->usr_time * 100 / runt;
1155                 sys_cpu = (double) ts->sys_time * 100 / runt;
1156         } else {
1157                 usr_cpu = 0;
1158                 sys_cpu = 0;
1159         }
1160         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1161         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1162         json_object_add_value_int(root, "ctx", ts->ctx);
1163         json_object_add_value_int(root, "majf", ts->majf);
1164         json_object_add_value_int(root, "minf", ts->minf);
1165
1166
1167         /* Calc % distribution of IO depths, usecond, msecond latency */
1168         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1169         stat_calc_lat_u(ts, io_u_lat_u);
1170         stat_calc_lat_m(ts, io_u_lat_m);
1171
1172         tmp = json_create_object();
1173         json_object_add_value_object(root, "iodepth_level", tmp);
1174         /* Only show fixed 7 I/O depth levels*/
1175         for (i = 0; i < 7; i++) {
1176                 char name[20];
1177                 if (i < 6)
1178                         snprintf(name, 20, "%d", 1 << i);
1179                 else
1180                         snprintf(name, 20, ">=%d", 1 << i);
1181                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1182         }
1183
1184         tmp = json_create_object();
1185         json_object_add_value_object(root, "latency_us", tmp);
1186         /* Microsecond latency */
1187         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1188                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1189                                  "250", "500", "750", "1000", };
1190                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1191         }
1192         /* Millisecond latency */
1193         tmp = json_create_object();
1194         json_object_add_value_object(root, "latency_ms", tmp);
1195         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1196                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1197                                  "250", "500", "750", "1000", "2000",
1198                                  ">=2000", };
1199                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1200         }
1201
1202         /* Additional output if continue_on_error set - default off*/
1203         if (ts->continue_on_error) {
1204                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1205                 json_object_add_value_int(root, "first_error", ts->first_error);
1206         }
1207
1208         if (ts->latency_depth) {
1209                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1210                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1211                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1212                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1213         }
1214
1215         /* Additional output if description is set */
1216         if (strlen(ts->description))
1217                 json_object_add_value_string(root, "desc", ts->description);
1218
1219         if (ts->nr_block_infos) {
1220                 /* Block error histogram and types */
1221                 int len;
1222                 unsigned int *percentiles = NULL;
1223                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1224
1225                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1226                                              ts->percentile_list,
1227                                              &percentiles, block_state_counts);
1228
1229                 if (len) {
1230                         struct json_object *block, *percentile_object, *states;
1231                         int state;
1232                         block = json_create_object();
1233                         json_object_add_value_object(root, "block", block);
1234
1235                         percentile_object = json_create_object();
1236                         json_object_add_value_object(block, "percentiles",
1237                                                      percentile_object);
1238                         for (i = 0; i < len; i++) {
1239                                 char buf[20];
1240                                 snprintf(buf, sizeof(buf), "%f",
1241                                          ts->percentile_list[i].u.f);
1242                                 json_object_add_value_int(percentile_object,
1243                                                           (const char *)buf,
1244                                                           percentiles[i]);
1245                         }
1246
1247                         states = json_create_object();
1248                         json_object_add_value_object(block, "states", states);
1249                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1250                                 json_object_add_value_int(states,
1251                                         block_state_names[state],
1252                                         block_state_counts[state]);
1253                         }
1254                         free(percentiles);
1255                 }
1256         }
1257
1258         /* steady state detection; move this behind json+? */
1259         if (ts->ss) {
1260                 struct json_object *data;
1261                 struct json_array *iops, *bw;
1262                 struct steadystate_data *ss = ts->ss;
1263                 unsigned long long sum_iops, sum_bw;
1264                 double mean_iops, mean_bw;
1265                 int i, x;
1266                 char ss_buf[64];
1267
1268                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1269                         ss->check_iops ? "iops" : "bw",
1270                         ss->check_slope ? "_slope" : "",
1271                         (float) ss->limit,
1272                         ss->pct ? "%" : "");
1273
1274                 tmp = json_create_object();
1275                 json_object_add_value_object(root, "steadystate", tmp);
1276                 json_object_add_value_string(tmp, "ss", ss_buf);
1277                 json_object_add_value_int(tmp, "duration", (int)ss->dur);
1278                 json_object_add_value_int(tmp, "steadystate_ramptime", ss->ramp_time / 1000000L);
1279                 json_object_add_value_int(tmp, "attained", ss->attained);
1280
1281                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ss->criterion, ss->pct ? "%" : "");
1282                 json_object_add_value_string(tmp, "criterion", ss_buf);
1283                 json_object_add_value_float(tmp, "max_deviation", ss->deviation);
1284                 json_object_add_value_float(tmp, "slope", ss->slope);
1285
1286                 data = json_create_object();
1287                 json_object_add_value_object(tmp, "data", data);
1288                 bw = json_create_array();
1289                 iops = json_create_array();
1290                 for (i = 0, sum_iops = 0, sum_bw = 0; i < ss->dur; i++) {
1291                         x = (ss->head + i) % ss->dur;
1292                         sum_bw += ss->bw_data[x];
1293                         sum_iops += ss->iops_data[x];
1294                         json_array_add_value_int(bw, ss->bw_data[x]);
1295                         json_array_add_value_int(iops, ss->iops_data[x]);
1296                 }
1297                 mean_bw = (double) sum_bw / ss->dur;
1298                 mean_iops = (double) sum_iops / ss->dur;
1299                 json_object_add_value_float(data, "bw_mean", mean_bw);
1300                 json_object_add_value_float(data, "iops_mean", mean_iops);
1301                 json_object_add_value_array(data, "iops", iops);
1302                 json_object_add_value_array(data, "bw", bw);
1303         }
1304
1305         return root;
1306 }
1307
1308 static void show_thread_status_terse(struct thread_stat *ts,
1309                                      struct group_run_stats *rs,
1310                                      struct buf_output *out)
1311 {
1312         if (terse_version == 2)
1313                 show_thread_status_terse_v2(ts, rs, out);
1314         else if (terse_version == 3 || terse_version == 4)
1315                 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1316         else
1317                 log_err("fio: bad terse version!? %d\n", terse_version);
1318 }
1319
1320 struct json_object *show_thread_status(struct thread_stat *ts,
1321                                        struct group_run_stats *rs,
1322                                        struct flist_head *opt_list,
1323                                        struct buf_output *out)
1324 {
1325         struct json_object *ret = NULL;
1326
1327         if (output_format & FIO_OUTPUT_TERSE)
1328                 show_thread_status_terse(ts, rs,  out);
1329         if (output_format & FIO_OUTPUT_JSON)
1330                 ret = show_thread_status_json(ts, rs, opt_list);
1331         if (output_format & FIO_OUTPUT_NORMAL)
1332                 show_thread_status_normal(ts, rs,  out);
1333
1334         return ret;
1335 }
1336
1337 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1338 {
1339         double mean, S;
1340
1341         if (src->samples == 0)
1342                 return;
1343
1344         dst->min_val = min(dst->min_val, src->min_val);
1345         dst->max_val = max(dst->max_val, src->max_val);
1346
1347         /*
1348          * Compute new mean and S after the merge
1349          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1350          *  #Parallel_algorithm>
1351          */
1352         if (first) {
1353                 mean = src->mean.u.f;
1354                 S = src->S.u.f;
1355         } else {
1356                 double delta = src->mean.u.f - dst->mean.u.f;
1357
1358                 mean = ((src->mean.u.f * src->samples) +
1359                         (dst->mean.u.f * dst->samples)) /
1360                         (dst->samples + src->samples);
1361
1362                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1363                         (dst->samples * src->samples) /
1364                         (dst->samples + src->samples);
1365         }
1366
1367         dst->samples += src->samples;
1368         dst->mean.u.f = mean;
1369         dst->S.u.f = S;
1370 }
1371
1372 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1373 {
1374         int i;
1375
1376         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1377                 if (dst->max_run[i] < src->max_run[i])
1378                         dst->max_run[i] = src->max_run[i];
1379                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1380                         dst->min_run[i] = src->min_run[i];
1381                 if (dst->max_bw[i] < src->max_bw[i])
1382                         dst->max_bw[i] = src->max_bw[i];
1383                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1384                         dst->min_bw[i] = src->min_bw[i];
1385
1386                 dst->io_kb[i] += src->io_kb[i];
1387                 dst->agg[i] += src->agg[i];
1388         }
1389
1390         if (!dst->kb_base)
1391                 dst->kb_base = src->kb_base;
1392         if (!dst->unit_base)
1393                 dst->unit_base = src->unit_base;
1394 }
1395
1396 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1397                       bool first)
1398 {
1399         int l, k;
1400
1401         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1402                 if (!dst->unified_rw_rep) {
1403                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1404                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1405                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1406                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1407
1408                         dst->io_bytes[l] += src->io_bytes[l];
1409
1410                         if (dst->runtime[l] < src->runtime[l])
1411                                 dst->runtime[l] = src->runtime[l];
1412                 } else {
1413                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1414                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1415                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1416                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1417
1418                         dst->io_bytes[0] += src->io_bytes[l];
1419
1420                         if (dst->runtime[0] < src->runtime[l])
1421                                 dst->runtime[0] = src->runtime[l];
1422
1423                         /*
1424                          * We're summing to the same destination, so override
1425                          * 'first' after the first iteration of the loop
1426                          */
1427                         first = false;
1428                 }
1429         }
1430
1431         dst->usr_time += src->usr_time;
1432         dst->sys_time += src->sys_time;
1433         dst->ctx += src->ctx;
1434         dst->majf += src->majf;
1435         dst->minf += src->minf;
1436
1437         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1438                 dst->io_u_map[k] += src->io_u_map[k];
1439         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1440                 dst->io_u_submit[k] += src->io_u_submit[k];
1441         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1442                 dst->io_u_complete[k] += src->io_u_complete[k];
1443         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1444                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1445         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1446                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1447
1448         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1449                 if (!dst->unified_rw_rep) {
1450                         dst->total_io_u[k] += src->total_io_u[k];
1451                         dst->short_io_u[k] += src->short_io_u[k];
1452                         dst->drop_io_u[k] += src->drop_io_u[k];
1453                 } else {
1454                         dst->total_io_u[0] += src->total_io_u[k];
1455                         dst->short_io_u[0] += src->short_io_u[k];
1456                         dst->drop_io_u[0] += src->drop_io_u[k];
1457                 }
1458         }
1459
1460         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1461                 int m;
1462
1463                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1464                         if (!dst->unified_rw_rep)
1465                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1466                         else
1467                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1468                 }
1469         }
1470
1471         dst->total_run_time += src->total_run_time;
1472         dst->total_submit += src->total_submit;
1473         dst->total_complete += src->total_complete;
1474 }
1475
1476 void init_group_run_stat(struct group_run_stats *gs)
1477 {
1478         int i;
1479         memset(gs, 0, sizeof(*gs));
1480
1481         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1482                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1483 }
1484
1485 void init_thread_stat(struct thread_stat *ts)
1486 {
1487         int j;
1488
1489         memset(ts, 0, sizeof(*ts));
1490
1491         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1492                 ts->lat_stat[j].min_val = -1UL;
1493                 ts->clat_stat[j].min_val = -1UL;
1494                 ts->slat_stat[j].min_val = -1UL;
1495                 ts->bw_stat[j].min_val = -1UL;
1496         }
1497         ts->groupid = -1;
1498 }
1499
1500 void __show_run_stats(void)
1501 {
1502         struct group_run_stats *runstats, *rs;
1503         struct thread_data *td;
1504         struct thread_stat *threadstats, *ts;
1505         int i, j, k, nr_ts, last_ts, idx;
1506         int kb_base_warned = 0;
1507         int unit_base_warned = 0;
1508         struct json_object *root = NULL;
1509         struct json_array *array = NULL;
1510         struct buf_output output[FIO_OUTPUT_NR];
1511         struct flist_head **opt_lists;
1512
1513         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1514
1515         for (i = 0; i < groupid + 1; i++)
1516                 init_group_run_stat(&runstats[i]);
1517
1518         /*
1519          * find out how many threads stats we need. if group reporting isn't
1520          * enabled, it's one-per-td.
1521          */
1522         nr_ts = 0;
1523         last_ts = -1;
1524         for_each_td(td, i) {
1525                 if (!td->o.group_reporting) {
1526                         nr_ts++;
1527                         continue;
1528                 }
1529                 if (last_ts == td->groupid)
1530                         continue;
1531
1532                 last_ts = td->groupid;
1533                 nr_ts++;
1534         }
1535
1536         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1537         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1538
1539         for (i = 0; i < nr_ts; i++) {
1540                 init_thread_stat(&threadstats[i]);
1541                 opt_lists[i] = NULL;
1542         }
1543
1544         j = 0;
1545         last_ts = -1;
1546         idx = 0;
1547         for_each_td(td, i) {
1548                 if (idx && (!td->o.group_reporting ||
1549                     (td->o.group_reporting && last_ts != td->groupid))) {
1550                         idx = 0;
1551                         j++;
1552                 }
1553
1554                 last_ts = td->groupid;
1555
1556                 ts = &threadstats[j];
1557
1558                 ts->clat_percentiles = td->o.clat_percentiles;
1559                 ts->percentile_precision = td->o.percentile_precision;
1560                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1561                 opt_lists[j] = &td->opt_list;
1562
1563                 idx++;
1564                 ts->members++;
1565
1566                 if (ts->groupid == -1) {
1567                         /*
1568                          * These are per-group shared already
1569                          */
1570                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1571                         if (td->o.description)
1572                                 strncpy(ts->description, td->o.description,
1573                                                 FIO_JOBDESC_SIZE - 1);
1574                         else
1575                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1576
1577                         /*
1578                          * If multiple entries in this group, this is
1579                          * the first member.
1580                          */
1581                         ts->thread_number = td->thread_number;
1582                         ts->groupid = td->groupid;
1583
1584                         /*
1585                          * first pid in group, not very useful...
1586                          */
1587                         ts->pid = td->pid;
1588
1589                         ts->kb_base = td->o.kb_base;
1590                         ts->unit_base = td->o.unit_base;
1591                         ts->unified_rw_rep = td->o.unified_rw_rep;
1592                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1593                         log_info("fio: kb_base differs for jobs in group, using"
1594                                  " %u as the base\n", ts->kb_base);
1595                         kb_base_warned = 1;
1596                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1597                         log_info("fio: unit_base differs for jobs in group, using"
1598                                  " %u as the base\n", ts->unit_base);
1599                         unit_base_warned = 1;
1600                 }
1601
1602                 ts->continue_on_error = td->o.continue_on_error;
1603                 ts->total_err_count += td->total_err_count;
1604                 ts->first_error = td->first_error;
1605                 if (!ts->error) {
1606                         if (!td->error && td->o.continue_on_error &&
1607                             td->first_error) {
1608                                 ts->error = td->first_error;
1609                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1610                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1611                         } else  if (td->error) {
1612                                 ts->error = td->error;
1613                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1614                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1615                         }
1616                 }
1617
1618                 ts->latency_depth = td->latency_qd;
1619                 ts->latency_target = td->o.latency_target;
1620                 ts->latency_percentile = td->o.latency_percentile;
1621                 ts->latency_window = td->o.latency_window;
1622
1623                 ts->nr_block_infos = td->ts.nr_block_infos;
1624                 for (k = 0; k < ts->nr_block_infos; k++)
1625                         ts->block_infos[k] = td->ts.block_infos[k];
1626
1627                 sum_thread_stats(ts, &td->ts, idx == 1);
1628
1629                 if (td->o.ss_dur)
1630                         ts->ss = &td->ss;
1631                 else
1632                         ts->ss = NULL;
1633         }
1634
1635         for (i = 0; i < nr_ts; i++) {
1636                 unsigned long long bw;
1637
1638                 ts = &threadstats[i];
1639                 if (ts->groupid == -1)
1640                         continue;
1641                 rs = &runstats[ts->groupid];
1642                 rs->kb_base = ts->kb_base;
1643                 rs->unit_base = ts->unit_base;
1644                 rs->unified_rw_rep += ts->unified_rw_rep;
1645
1646                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1647                         if (!ts->runtime[j])
1648                                 continue;
1649                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1650                                 rs->min_run[j] = ts->runtime[j];
1651                         if (ts->runtime[j] > rs->max_run[j])
1652                                 rs->max_run[j] = ts->runtime[j];
1653
1654                         bw = 0;
1655                         if (ts->runtime[j]) {
1656                                 unsigned long runt = ts->runtime[j];
1657                                 unsigned long long kb;
1658
1659                                 kb = ts->io_bytes[j] / rs->kb_base;
1660                                 bw = kb * 1000 / runt;
1661                         }
1662                         if (bw < rs->min_bw[j])
1663                                 rs->min_bw[j] = bw;
1664                         if (bw > rs->max_bw[j])
1665                                 rs->max_bw[j] = bw;
1666
1667                         rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1668                 }
1669         }
1670
1671         for (i = 0; i < groupid + 1; i++) {
1672                 int ddir;
1673
1674                 rs = &runstats[i];
1675
1676                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1677                         if (rs->max_run[ddir])
1678                                 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1679                                                 rs->max_run[ddir];
1680                 }
1681         }
1682
1683         for (i = 0; i < FIO_OUTPUT_NR; i++)
1684                 buf_output_init(&output[i]);
1685
1686         /*
1687          * don't overwrite last signal output
1688          */
1689         if (output_format & FIO_OUTPUT_NORMAL)
1690                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1691         if (output_format & FIO_OUTPUT_JSON) {
1692                 struct thread_data *global;
1693                 char time_buf[32];
1694                 struct timeval now;
1695                 unsigned long long ms_since_epoch;
1696
1697                 gettimeofday(&now, NULL);
1698                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1699                                  (unsigned long long)(now.tv_usec) / 1000;
1700
1701                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1702                                 sizeof(time_buf));
1703                 time_buf[strlen(time_buf) - 1] = '\0';
1704
1705                 root = json_create_object();
1706                 json_object_add_value_string(root, "fio version", fio_version_string);
1707                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1708                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1709                 json_object_add_value_string(root, "time", time_buf);
1710                 global = get_global_options();
1711                 json_add_job_opts(root, "global options", &global->opt_list, false);
1712                 array = json_create_array();
1713                 json_object_add_value_array(root, "jobs", array);
1714         }
1715
1716         if (is_backend)
1717                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1718
1719         for (i = 0; i < nr_ts; i++) {
1720                 ts = &threadstats[i];
1721                 rs = &runstats[ts->groupid];
1722
1723                 if (is_backend) {
1724                         fio_server_send_job_options(opt_lists[i], i);
1725                         fio_server_send_ts(ts, rs);
1726                 } else {
1727                         if (output_format & FIO_OUTPUT_TERSE)
1728                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1729                         if (output_format & FIO_OUTPUT_JSON) {
1730                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1731                                 json_array_add_value_object(array, tmp);
1732                         }
1733                         if (output_format & FIO_OUTPUT_NORMAL)
1734                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1735                 }
1736         }
1737         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1738                 /* disk util stats, if any */
1739                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1740
1741                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1742
1743                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1744                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1745                 json_free_object(root);
1746         }
1747
1748         for (i = 0; i < groupid + 1; i++) {
1749                 rs = &runstats[i];
1750
1751                 rs->groupid = i;
1752                 if (is_backend)
1753                         fio_server_send_gs(rs);
1754                 else if (output_format & FIO_OUTPUT_NORMAL)
1755                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1756         }
1757
1758         if (is_backend)
1759                 fio_server_send_du();
1760         else if (output_format & FIO_OUTPUT_NORMAL) {
1761                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1762                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1763         }
1764
1765         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1766                 buf_output_flush(&output[i]);
1767                 buf_output_free(&output[i]);
1768         }
1769
1770         log_info_flush();
1771         free(runstats);
1772         free(threadstats);
1773         free(opt_lists);
1774 }
1775
1776 void show_run_stats(void)
1777 {
1778         fio_mutex_down(stat_mutex);
1779         __show_run_stats();
1780         fio_mutex_up(stat_mutex);
1781 }
1782
1783 void __show_running_run_stats(void)
1784 {
1785         struct thread_data *td;
1786         unsigned long long *rt;
1787         struct timeval tv;
1788         int i;
1789
1790         fio_mutex_down(stat_mutex);
1791
1792         rt = malloc(thread_number * sizeof(unsigned long long));
1793         fio_gettime(&tv, NULL);
1794
1795         for_each_td(td, i) {
1796                 td->update_rusage = 1;
1797                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1798                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1799                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1800                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1801
1802                 rt[i] = mtime_since(&td->start, &tv);
1803                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1804                         td->ts.runtime[DDIR_READ] += rt[i];
1805                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1806                         td->ts.runtime[DDIR_WRITE] += rt[i];
1807                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1808                         td->ts.runtime[DDIR_TRIM] += rt[i];
1809         }
1810
1811         for_each_td(td, i) {
1812                 if (td->runstate >= TD_EXITED)
1813                         continue;
1814                 if (td->rusage_sem) {
1815                         td->update_rusage = 1;
1816                         fio_mutex_down(td->rusage_sem);
1817                 }
1818                 td->update_rusage = 0;
1819         }
1820
1821         __show_run_stats();
1822
1823         for_each_td(td, i) {
1824                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1825                         td->ts.runtime[DDIR_READ] -= rt[i];
1826                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1827                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1828                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1829                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1830         }
1831
1832         free(rt);
1833         fio_mutex_up(stat_mutex);
1834 }
1835
1836 static int status_interval_init;
1837 static struct timeval status_time;
1838 static int status_file_disabled;
1839
1840 #define FIO_STATUS_FILE         "fio-dump-status"
1841
1842 static int check_status_file(void)
1843 {
1844         struct stat sb;
1845         const char *temp_dir;
1846         char fio_status_file_path[PATH_MAX];
1847
1848         if (status_file_disabled)
1849                 return 0;
1850
1851         temp_dir = getenv("TMPDIR");
1852         if (temp_dir == NULL) {
1853                 temp_dir = getenv("TEMP");
1854                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1855                         temp_dir = NULL;
1856         }
1857         if (temp_dir == NULL)
1858                 temp_dir = "/tmp";
1859
1860         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1861
1862         if (stat(fio_status_file_path, &sb))
1863                 return 0;
1864
1865         if (unlink(fio_status_file_path) < 0) {
1866                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1867                                                         strerror(errno));
1868                 log_err("fio: disabling status file updates\n");
1869                 status_file_disabled = 1;
1870         }
1871
1872         return 1;
1873 }
1874
1875 void check_for_running_stats(void)
1876 {
1877         if (status_interval) {
1878                 if (!status_interval_init) {
1879                         fio_gettime(&status_time, NULL);
1880                         status_interval_init = 1;
1881                 } else if (mtime_since_now(&status_time) >= status_interval) {
1882                         show_running_run_stats();
1883                         fio_gettime(&status_time, NULL);
1884                         return;
1885                 }
1886         }
1887         if (check_status_file()) {
1888                 show_running_run_stats();
1889                 return;
1890         }
1891 }
1892
1893 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1894 {
1895         double val = data;
1896         double delta;
1897
1898         if (data > is->max_val)
1899                 is->max_val = data;
1900         if (data < is->min_val)
1901                 is->min_val = data;
1902
1903         delta = val - is->mean.u.f;
1904         if (delta) {
1905                 is->mean.u.f += delta / (is->samples + 1.0);
1906                 is->S.u.f += delta * (val - is->mean.u.f);
1907         }
1908
1909         is->samples++;
1910 }
1911
1912 /*
1913  * Return a struct io_logs, which is added to the tail of the log
1914  * list for 'iolog'.
1915  */
1916 static struct io_logs *get_new_log(struct io_log *iolog)
1917 {
1918         size_t new_size, new_samples;
1919         struct io_logs *cur_log;
1920
1921         /*
1922          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1923          * forever
1924          */
1925         if (!iolog->cur_log_max)
1926                 new_samples = DEF_LOG_ENTRIES;
1927         else {
1928                 new_samples = iolog->cur_log_max * 2;
1929                 if (new_samples > MAX_LOG_ENTRIES)
1930                         new_samples = MAX_LOG_ENTRIES;
1931         }
1932
1933         new_size = new_samples * log_entry_sz(iolog);
1934
1935         cur_log = smalloc(sizeof(*cur_log));
1936         if (cur_log) {
1937                 INIT_FLIST_HEAD(&cur_log->list);
1938                 cur_log->log = malloc(new_size);
1939                 if (cur_log->log) {
1940                         cur_log->nr_samples = 0;
1941                         cur_log->max_samples = new_samples;
1942                         flist_add_tail(&cur_log->list, &iolog->io_logs);
1943                         iolog->cur_log_max = new_samples;
1944                         return cur_log;
1945                 }
1946                 sfree(cur_log);
1947         }
1948
1949         return NULL;
1950 }
1951
1952 /*
1953  * Add and return a new log chunk, or return current log if big enough
1954  */
1955 static struct io_logs *regrow_log(struct io_log *iolog)
1956 {
1957         struct io_logs *cur_log;
1958         int i;
1959
1960         if (!iolog || iolog->disabled)
1961                 goto disable;
1962
1963         cur_log = iolog_cur_log(iolog);
1964         if (!cur_log) {
1965                 cur_log = get_new_log(iolog);
1966                 if (!cur_log)
1967                         return NULL;
1968         }
1969
1970         if (cur_log->nr_samples < cur_log->max_samples)
1971                 return cur_log;
1972
1973         /*
1974          * No room for a new sample. If we're compressing on the fly, flush
1975          * out the current chunk
1976          */
1977         if (iolog->log_gz) {
1978                 if (iolog_cur_flush(iolog, cur_log)) {
1979                         log_err("fio: failed flushing iolog! Will stop logging.\n");
1980                         return NULL;
1981                 }
1982         }
1983
1984         /*
1985          * Get a new log array, and add to our list
1986          */
1987         cur_log = get_new_log(iolog);
1988         if (!cur_log) {
1989                 log_err("fio: failed extending iolog! Will stop logging.\n");
1990                 return NULL;
1991         }
1992
1993         if (!iolog->pending || !iolog->pending->nr_samples)
1994                 return cur_log;
1995
1996         /*
1997          * Flush pending items to new log
1998          */
1999         for (i = 0; i < iolog->pending->nr_samples; i++) {
2000                 struct io_sample *src, *dst;
2001
2002                 src = get_sample(iolog, iolog->pending, i);
2003                 dst = get_sample(iolog, cur_log, i);
2004                 memcpy(dst, src, log_entry_sz(iolog));
2005         }
2006         cur_log->nr_samples = iolog->pending->nr_samples;
2007
2008         iolog->pending->nr_samples = 0;
2009         return cur_log;
2010 disable:
2011         if (iolog)
2012                 iolog->disabled = true;
2013         return NULL;
2014 }
2015
2016 void regrow_logs(struct thread_data *td)
2017 {
2018         regrow_log(td->slat_log);
2019         regrow_log(td->clat_log);
2020         regrow_log(td->clat_hist_log);
2021         regrow_log(td->lat_log);
2022         regrow_log(td->bw_log);
2023         regrow_log(td->iops_log);
2024         td->flags &= ~TD_F_REGROW_LOGS;
2025 }
2026
2027 static struct io_logs *get_cur_log(struct io_log *iolog)
2028 {
2029         struct io_logs *cur_log;
2030
2031         cur_log = iolog_cur_log(iolog);
2032         if (!cur_log) {
2033                 cur_log = get_new_log(iolog);
2034                 if (!cur_log)
2035                         return NULL;
2036         }
2037
2038         if (cur_log->nr_samples < cur_log->max_samples)
2039                 return cur_log;
2040
2041         /*
2042          * Out of space. If we're in IO offload mode, or we're not doing
2043          * per unit logging (hence logging happens outside of the IO thread
2044          * as well), add a new log chunk inline. If we're doing inline
2045          * submissions, flag 'td' as needing a log regrow and we'll take
2046          * care of it on the submission side.
2047          */
2048         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2049             !per_unit_log(iolog))
2050                 return regrow_log(iolog);
2051
2052         iolog->td->flags |= TD_F_REGROW_LOGS;
2053         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2054         return iolog->pending;
2055 }
2056
2057 static void __add_log_sample(struct io_log *iolog, unsigned long val,
2058                              enum fio_ddir ddir, unsigned int bs,
2059                              unsigned long t, uint64_t offset)
2060 {
2061         struct io_logs *cur_log;
2062
2063         if (iolog->disabled)
2064                 return;
2065         if (flist_empty(&iolog->io_logs))
2066                 iolog->avg_last = t;
2067
2068         cur_log = get_cur_log(iolog);
2069         if (cur_log) {
2070                 struct io_sample *s;
2071
2072                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2073
2074                 s->val = val;
2075                 s->time = t;
2076                 io_sample_set_ddir(iolog, s, ddir);
2077                 s->bs = bs;
2078
2079                 if (iolog->log_offset) {
2080                         struct io_sample_offset *so = (void *) s;
2081
2082                         so->offset = offset;
2083                 }
2084
2085                 cur_log->nr_samples++;
2086                 return;
2087         }
2088
2089         iolog->disabled = true;
2090 }
2091
2092 static inline void reset_io_stat(struct io_stat *ios)
2093 {
2094         ios->max_val = ios->min_val = ios->samples = 0;
2095         ios->mean.u.f = ios->S.u.f = 0;
2096 }
2097
2098 void reset_io_stats(struct thread_data *td)
2099 {
2100         struct thread_stat *ts = &td->ts;
2101         int i, j;
2102
2103         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2104                 reset_io_stat(&ts->clat_stat[i]);
2105                 reset_io_stat(&ts->slat_stat[i]);
2106                 reset_io_stat(&ts->lat_stat[i]);
2107                 reset_io_stat(&ts->bw_stat[i]);
2108                 reset_io_stat(&ts->iops_stat[i]);
2109
2110                 ts->io_bytes[i] = 0;
2111                 ts->runtime[i] = 0;
2112
2113                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2114                         ts->io_u_plat[i][j] = 0;
2115         }
2116
2117         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2118                 ts->io_u_map[i] = 0;
2119                 ts->io_u_submit[i] = 0;
2120                 ts->io_u_complete[i] = 0;
2121                 ts->io_u_lat_u[i] = 0;
2122                 ts->io_u_lat_m[i] = 0;
2123                 ts->total_submit = 0;
2124                 ts->total_complete = 0;
2125         }
2126
2127         for (i = 0; i < 3; i++) {
2128                 ts->total_io_u[i] = 0;
2129                 ts->short_io_u[i] = 0;
2130                 ts->drop_io_u[i] = 0;
2131         }
2132 }
2133
2134 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2135                               unsigned long elapsed, bool log_max)
2136 {
2137         /*
2138          * Note an entry in the log. Use the mean from the logged samples,
2139          * making sure to properly round up. Only write a log entry if we
2140          * had actual samples done.
2141          */
2142         if (iolog->avg_window[ddir].samples) {
2143                 unsigned long val;
2144
2145                 if (log_max)
2146                         val = iolog->avg_window[ddir].max_val;
2147                 else
2148                         val = iolog->avg_window[ddir].mean.u.f + 0.50;
2149
2150                 __add_log_sample(iolog, val, ddir, 0, elapsed, 0);
2151         }
2152
2153         reset_io_stat(&iolog->avg_window[ddir]);
2154 }
2155
2156 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2157                              bool log_max)
2158 {
2159         int ddir;
2160
2161         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2162                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2163 }
2164
2165 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2166                            unsigned long val, enum fio_ddir ddir,
2167                            unsigned int bs, uint64_t offset)
2168 {
2169         unsigned long elapsed, this_window;
2170
2171         if (!ddir_rw(ddir))
2172                 return 0;
2173
2174         elapsed = mtime_since_now(&td->epoch);
2175
2176         /*
2177          * If no time averaging, just add the log sample.
2178          */
2179         if (!iolog->avg_msec) {
2180                 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
2181                 return 0;
2182         }
2183
2184         /*
2185          * Add the sample. If the time period has passed, then
2186          * add that entry to the log and clear.
2187          */
2188         add_stat_sample(&iolog->avg_window[ddir], val);
2189
2190         /*
2191          * If period hasn't passed, adding the above sample is all we
2192          * need to do.
2193          */
2194         this_window = elapsed - iolog->avg_last;
2195         if (elapsed < iolog->avg_last)
2196                 return iolog->avg_last - elapsed;
2197         else if (this_window < iolog->avg_msec) {
2198                 int diff = iolog->avg_msec - this_window;
2199
2200                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2201                         return diff;
2202         }
2203
2204         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2205
2206         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2207         return iolog->avg_msec;
2208 }
2209
2210 void finalize_logs(struct thread_data *td, bool unit_logs)
2211 {
2212         unsigned long elapsed;
2213
2214         elapsed = mtime_since_now(&td->epoch);
2215
2216         if (td->clat_log && unit_logs)
2217                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2218         if (td->slat_log && unit_logs)
2219                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2220         if (td->lat_log && unit_logs)
2221                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2222         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2223                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2224         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2225                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2226 }
2227
2228 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
2229 {
2230         struct io_log *iolog;
2231
2232         if (!ddir_rw(ddir))
2233                 return;
2234
2235         iolog = agg_io_log[ddir];
2236         __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
2237 }
2238
2239 static void add_clat_percentile_sample(struct thread_stat *ts,
2240                                 unsigned long usec, enum fio_ddir ddir)
2241 {
2242         unsigned int idx = plat_val_to_idx(usec);
2243         assert(idx < FIO_IO_U_PLAT_NR);
2244
2245         ts->io_u_plat[ddir][idx]++;
2246 }
2247
2248 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2249                      unsigned long usec, unsigned int bs, uint64_t offset)
2250 {
2251         unsigned long elapsed, this_window;
2252         struct thread_stat *ts = &td->ts;
2253         struct io_log *iolog = td->clat_hist_log;
2254
2255         td_io_u_lock(td);
2256
2257         add_stat_sample(&ts->clat_stat[ddir], usec);
2258
2259         if (td->clat_log)
2260                 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
2261
2262         if (ts->clat_percentiles)
2263                 add_clat_percentile_sample(ts, usec, ddir);
2264
2265         if (iolog && iolog->hist_msec) {
2266                 struct io_hist *hw = &iolog->hist_window[ddir];
2267
2268                 hw->samples++;
2269                 elapsed = mtime_since_now(&td->epoch);
2270                 if (!hw->hist_last)
2271                         hw->hist_last = elapsed;
2272                 this_window = elapsed - hw->hist_last;
2273
2274                 if (this_window >= iolog->hist_msec) {
2275                         unsigned int *io_u_plat;
2276                         unsigned int *dst;
2277
2278                         /*
2279                          * Make a byte-for-byte copy of the latency histogram
2280                          * stored in td->ts.io_u_plat[ddir], recording it in a
2281                          * log sample. Note that the matching call to free() is
2282                          * located in iolog.c after printing this sample to the
2283                          * log file.
2284                          */
2285                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2286                         dst = malloc(FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2287                         memcpy(dst, io_u_plat,
2288                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2289                         __add_log_sample(iolog, (unsigned long )dst, ddir, bs,
2290                                                 elapsed, offset);
2291
2292                         /*
2293                          * Update the last time we recorded as being now, minus
2294                          * any drift in time we encountered before actually
2295                          * making the record.
2296                          */
2297                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2298                         hw->samples = 0;
2299                 }
2300         }
2301
2302         td_io_u_unlock(td);
2303 }
2304
2305 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2306                      unsigned long usec, unsigned int bs, uint64_t offset)
2307 {
2308         struct thread_stat *ts = &td->ts;
2309
2310         if (!ddir_rw(ddir))
2311                 return;
2312
2313         td_io_u_lock(td);
2314
2315         add_stat_sample(&ts->slat_stat[ddir], usec);
2316
2317         if (td->slat_log)
2318                 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
2319
2320         td_io_u_unlock(td);
2321 }
2322
2323 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2324                     unsigned long usec, unsigned int bs, uint64_t offset)
2325 {
2326         struct thread_stat *ts = &td->ts;
2327
2328         if (!ddir_rw(ddir))
2329                 return;
2330
2331         td_io_u_lock(td);
2332
2333         add_stat_sample(&ts->lat_stat[ddir], usec);
2334
2335         if (td->lat_log)
2336                 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2337
2338         td_io_u_unlock(td);
2339 }
2340
2341 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2342                    unsigned int bytes, unsigned long spent)
2343 {
2344         struct thread_stat *ts = &td->ts;
2345         unsigned long rate;
2346
2347         if (spent)
2348                 rate = bytes * 1000 / spent;
2349         else
2350                 rate = 0;
2351
2352         td_io_u_lock(td);
2353
2354         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2355
2356         if (td->bw_log)
2357                 add_log_sample(td, td->bw_log, rate, io_u->ddir, bytes, io_u->offset);
2358
2359         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2360         td_io_u_unlock(td);
2361 }
2362
2363 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2364 {
2365         struct thread_stat *ts = &td->ts;
2366         unsigned long spent, rate;
2367         enum fio_ddir ddir;
2368         unsigned int next, next_log;
2369
2370         next_log = td->o.bw_avg_time;
2371
2372         spent = mtime_since(&td->bw_sample_time, t);
2373         if (spent < td->o.bw_avg_time &&
2374             td->o.bw_avg_time - spent >= LOG_MSEC_SLACK)
2375                 return td->o.bw_avg_time - spent;
2376
2377         td_io_u_lock(td);
2378
2379         /*
2380          * Compute both read and write rates for the interval.
2381          */
2382         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2383                 uint64_t delta;
2384
2385                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2386                 if (!delta)
2387                         continue; /* No entries for interval */
2388
2389                 if (spent)
2390                         rate = delta * 1000 / spent / 1024;
2391                 else
2392                         rate = 0;
2393
2394                 add_stat_sample(&ts->bw_stat[ddir], rate);
2395
2396                 if (td->bw_log) {
2397                         unsigned int bs = 0;
2398
2399                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2400                                 bs = td->o.min_bs[ddir];
2401
2402                         next = add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2403                         next_log = min(next_log, next);
2404                 }
2405
2406                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2407         }
2408
2409         timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time);
2410
2411         td_io_u_unlock(td);
2412
2413         if (spent <= td->o.bw_avg_time)
2414                 return min(next_log, td->o.bw_avg_time);
2415
2416         next = td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time);
2417         return min(next, next_log);
2418 }
2419
2420 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2421                      unsigned int bytes)
2422 {
2423         struct thread_stat *ts = &td->ts;
2424
2425         td_io_u_lock(td);
2426
2427         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2428
2429         if (td->iops_log)
2430                 add_log_sample(td, td->iops_log, 1, io_u->ddir, bytes, io_u->offset);
2431
2432         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2433         td_io_u_unlock(td);
2434 }
2435
2436 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2437 {
2438         struct thread_stat *ts = &td->ts;
2439         unsigned long spent, iops;
2440         enum fio_ddir ddir;
2441         unsigned int next, next_log;
2442
2443         next_log = td->o.iops_avg_time;
2444
2445         spent = mtime_since(&td->iops_sample_time, t);
2446         if (spent < td->o.iops_avg_time &&
2447             td->o.iops_avg_time - spent >= LOG_MSEC_SLACK)
2448                 return td->o.iops_avg_time - spent;
2449
2450         td_io_u_lock(td);
2451
2452         /*
2453          * Compute both read and write rates for the interval.
2454          */
2455         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2456                 uint64_t delta;
2457
2458                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2459                 if (!delta)
2460                         continue; /* No entries for interval */
2461
2462                 if (spent)
2463                         iops = (delta * 1000) / spent;
2464                 else
2465                         iops = 0;
2466
2467                 add_stat_sample(&ts->iops_stat[ddir], iops);
2468
2469                 if (td->iops_log) {
2470                         unsigned int bs = 0;
2471
2472                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2473                                 bs = td->o.min_bs[ddir];
2474
2475                         next = add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2476                         next_log = min(next_log, next);
2477                 }
2478
2479                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2480         }
2481
2482         timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time);
2483
2484         td_io_u_unlock(td);
2485
2486         if (spent <= td->o.iops_avg_time)
2487                 return min(next_log, td->o.iops_avg_time);
2488
2489         next = td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
2490         return min(next, next_log);
2491 }
2492
2493 /*
2494  * Returns msecs to next event
2495  */
2496 int calc_log_samples(void)
2497 {
2498         struct thread_data *td;
2499         unsigned int next = ~0U, tmp;
2500         struct timeval now;
2501         int i;
2502
2503         fio_gettime(&now, NULL);
2504
2505         for_each_td(td, i) {
2506                 if (in_ramp_time(td) ||
2507                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2508                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2509                         continue;
2510                 }
2511                 if (!per_unit_log(td->bw_log)) {
2512                         tmp = add_bw_samples(td, &now);
2513                         if (tmp < next)
2514                                 next = tmp;
2515                 }
2516                 if (!per_unit_log(td->iops_log)) {
2517                         tmp = add_iops_samples(td, &now);
2518                         if (tmp < next)
2519                                 next = tmp;
2520                 }
2521         }
2522
2523         return next == ~0U ? 0 : next;
2524 }
2525
2526 void stat_init(void)
2527 {
2528         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2529 }
2530
2531 void stat_exit(void)
2532 {
2533         /*
2534          * When we have the mutex, we know out-of-band access to it
2535          * have ended.
2536          */
2537         fio_mutex_down(stat_mutex);
2538         fio_mutex_remove(stat_mutex);
2539 }
2540
2541 /*
2542  * Called from signal handler. Wake up status thread.
2543  */
2544 void show_running_run_stats(void)
2545 {
2546         helper_do_stat();
2547 }
2548
2549 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2550 {
2551         /* Ignore io_u's which span multiple blocks--they will just get
2552          * inaccurate counts. */
2553         int idx = (io_u->offset - io_u->file->file_offset)
2554                         / td->o.bs[DDIR_TRIM];
2555         uint32_t *info = &td->ts.block_infos[idx];
2556         assert(idx < td->ts.nr_block_infos);
2557         return info;
2558 }