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_ss_normal(struct thread_stat *ts, struct buf_output *out)
 661 {
 662         char *p1, *p2;
 663         unsigned long long bw_mean, iops_mean;
 664         const int i2p = is_power_of_2(ts->kb_base);
 665
 666         if (!ts->ss_state)
 667                 return;
 668
 669         bw_mean = steadystate_bw_mean(ts);
 670         iops_mean = steadystate_iops_mean(ts);
 671
 672         p1 = num2str(bw_mean / ts->kb_base, 6, ts->kb_base, i2p, ts->unit_base);
 673         p2 = num2str(iops_mean, 6, 1, 0, 0);
 674
 675         log_buf(out, "  steadystate  : attained=%s, bw=%s/s, iops=%s, %s%s=%.3f%s\n",
 676                 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
 677                 p1, p2,
 678                 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
 679                 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
 680                 ts->ss_criterion.u.f,
 681                 ts->ss_state & __FIO_SS_PCT ? "%" : "");
 682
 683         free(p1);
 684         free(p2);
 685 }
 686
 687 static void show_thread_status_normal(struct thread_stat *ts,
 688                                       struct group_run_stats *rs,
 689                                       struct buf_output *out)
 690 {
 691         double usr_cpu, sys_cpu;
 692         unsigned long runtime;
 693         double io_u_dist[FIO_IO_U_MAP_NR];
 694         time_t time_p;
 695         char time_buf[32];
 696
 697         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 698                 return;
 699
 700         time(&time_p);
 701         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 702
 703         if (!ts->error) {
 704                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 705                                         ts->name, ts->groupid, ts->members,
 706                                         ts->error, (int) ts->pid, time_buf);
 707         } else {
 708                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 709                                         ts->name, ts->groupid, ts->members,
 710                                         ts->error, ts->verror, (int) ts->pid,
 711                                         time_buf);
 712         }
 713
 714         if (strlen(ts->description))
 715                 log_buf(out, "  Description  : [%s]\n", ts->description);
 716
 717         if (ts->io_bytes[DDIR_READ])
 718                 show_ddir_status(rs, ts, DDIR_READ, out);
 719         if (ts->io_bytes[DDIR_WRITE])
 720                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 721         if (ts->io_bytes[DDIR_TRIM])
 722                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 723
 724         show_latencies(ts, out);
 725
 726         runtime = ts->total_run_time;
 727         if (runtime) {
 728                 double runt = (double) runtime;
 729
 730                 usr_cpu = (double) ts->usr_time * 100 / runt;
 731                 sys_cpu = (double) ts->sys_time * 100 / runt;
 732         } else {
 733                 usr_cpu = 0;
 734                 sys_cpu = 0;
 735         }
 736
 737         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 738                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 739                         (unsigned long long) ts->ctx,
 740                         (unsigned long long) ts->majf,
 741                         (unsigned long long) ts->minf);
 742
 743         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 744         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 745                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 746                                         io_u_dist[1], io_u_dist[2],
 747                                         io_u_dist[3], io_u_dist[4],
 748                                         io_u_dist[5], io_u_dist[6]);
 749
 750         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 751         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 752                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 753                                         io_u_dist[1], io_u_dist[2],
 754                                         io_u_dist[3], io_u_dist[4],
 755                                         io_u_dist[5], io_u_dist[6]);
 756         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 757         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 758                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 759                                         io_u_dist[1], io_u_dist[2],
 760                                         io_u_dist[3], io_u_dist[4],
 761                                         io_u_dist[5], io_u_dist[6]);
 762         log_buf(out, "     issued    : total=r=%llu/w=%llu/d=%llu,"
 763                                  " short=r=%llu/w=%llu/d=%llu,"
 764                                  " drop=r=%llu/w=%llu/d=%llu\n",
 765                                         (unsigned long long) ts->total_io_u[0],
 766                                         (unsigned long long) ts->total_io_u[1],
 767                                         (unsigned long long) ts->total_io_u[2],
 768                                         (unsigned long long) ts->short_io_u[0],
 769                                         (unsigned long long) ts->short_io_u[1],
 770                                         (unsigned long long) ts->short_io_u[2],
 771                                         (unsigned long long) ts->drop_io_u[0],
 772                                         (unsigned long long) ts->drop_io_u[1],
 773                                         (unsigned long long) ts->drop_io_u[2]);
 774         if (ts->continue_on_error) {
 775                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 776                                         (unsigned long long)ts->total_err_count,
 777                                         ts->first_error,
 778                                         strerror(ts->first_error));
 779         }
 780         if (ts->latency_depth) {
 781                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 782                                         (unsigned long long)ts->latency_target,
 783                                         (unsigned long long)ts->latency_window,
 784                                         ts->latency_percentile.u.f,
 785                                         ts->latency_depth);
 786         }
 787
 788         if (ts->nr_block_infos)
 789                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 790                                   ts->percentile_list, out);
 791
 792         if (ts->ss_dur)
 793                 show_ss_normal(ts, out);
 794 }
 795
 796 static void show_ddir_status_terse(struct thread_stat *ts,
 797                                    struct group_run_stats *rs, int ddir,
 798                                    struct buf_output *out)
 799 {
 800         unsigned long min, max;
 801         unsigned long long bw, iops;
 802         unsigned int *ovals = NULL;
 803         double mean, dev;
 804         unsigned int len, minv, maxv;
 805         int i;
 806
 807         assert(ddir_rw(ddir));
 808
 809         iops = bw = 0;
 810         if (ts->runtime[ddir]) {
 811                 uint64_t runt = ts->runtime[ddir];
 812
 813                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 814                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 815         }
 816
 817         log_buf(out, ";%llu;%llu;%llu;%llu",
 818                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 819                                         (unsigned long long) ts->runtime[ddir]);
 820
 821         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 822                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 823         else
 824                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 825
 826         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 827                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 828         else
 829                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 830
 831         if (ts->clat_percentiles) {
 832                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 833                                         ts->clat_stat[ddir].samples,
 834                                         ts->percentile_list, &ovals, &maxv,
 835                                         &minv);
 836         } else
 837                 len = 0;
 838
 839         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 840                 if (i >= len) {
 841                         log_buf(out, ";0%%=0");
 842                         continue;
 843                 }
 844                 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
 845         }
 846
 847         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 848                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 849         else
 850                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 851
 852         if (ovals)
 853                 free(ovals);
 854
 855         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 856                 double p_of_agg = 100.0;
 857
 858                 if (rs->agg[ddir]) {
 859                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 860                         if (p_of_agg > 100.0)
 861                                 p_of_agg = 100.0;
 862                 }
 863
 864                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 865         } else
 866                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 867 }
 868
 869 static void add_ddir_status_json(struct thread_stat *ts,
 870                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 871 {
 872         unsigned long min, max;
 873         unsigned long long bw;
 874         unsigned int *ovals = NULL;
 875         double mean, dev, iops;
 876         unsigned int len, minv, maxv;
 877         int i;
 878         const char *ddirname[] = {"read", "write", "trim"};
 879         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 880         char buf[120];
 881         double p_of_agg = 100.0;
 882
 883         assert(ddir_rw(ddir));
 884
 885         if (ts->unified_rw_rep && ddir != DDIR_READ)
 886                 return;
 887
 888         dir_object = json_create_object();
 889         json_object_add_value_object(parent,
 890                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 891
 892         bw = 0;
 893         iops = 0.0;
 894         if (ts->runtime[ddir]) {
 895                 uint64_t runt = ts->runtime[ddir];
 896
 897                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 898                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 899         }
 900
 901         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
 902         json_object_add_value_int(dir_object, "bw", bw);
 903         json_object_add_value_float(dir_object, "iops", iops);
 904         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 905         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 906         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 907         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 908
 909         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 910                 min = max = 0;
 911                 mean = dev = 0.0;
 912         }
 913         tmp_object = json_create_object();
 914         json_object_add_value_object(dir_object, "slat", tmp_object);
 915         json_object_add_value_int(tmp_object, "min", min);
 916         json_object_add_value_int(tmp_object, "max", max);
 917         json_object_add_value_float(tmp_object, "mean", mean);
 918         json_object_add_value_float(tmp_object, "stddev", dev);
 919
 920         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 921                 min = max = 0;
 922                 mean = dev = 0.0;
 923         }
 924         tmp_object = json_create_object();
 925         json_object_add_value_object(dir_object, "clat", tmp_object);
 926         json_object_add_value_int(tmp_object, "min", min);
 927         json_object_add_value_int(tmp_object, "max", max);
 928         json_object_add_value_float(tmp_object, "mean", mean);
 929         json_object_add_value_float(tmp_object, "stddev", dev);
 930
 931         if (ts->clat_percentiles) {
 932                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 933                                         ts->clat_stat[ddir].samples,
 934                                         ts->percentile_list, &ovals, &maxv,
 935                                         &minv);
 936         } else
 937                 len = 0;
 938
 939         percentile_object = json_create_object();
 940         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 941         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 942                 if (i >= len) {
 943                         json_object_add_value_int(percentile_object, "0.00", 0);
 944                         continue;
 945                 }
 946                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 947                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 948         }
 949
 950         if (output_format & FIO_OUTPUT_JSON_PLUS) {
 951                 clat_bins_object = json_create_object();
 952                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
 953                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
 954                         snprintf(buf, sizeof(buf), "%d", i);
 955                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
 956                 }
 957                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS);
 958                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL);
 959                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR);
 960         }
 961
 962         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
 963                 min = max = 0;
 964                 mean = dev = 0.0;
 965         }
 966         tmp_object = json_create_object();
 967         json_object_add_value_object(dir_object, "lat", tmp_object);
 968         json_object_add_value_int(tmp_object, "min", min);
 969         json_object_add_value_int(tmp_object, "max", max);
 970         json_object_add_value_float(tmp_object, "mean", mean);
 971         json_object_add_value_float(tmp_object, "stddev", dev);
 972         if (ovals)
 973                 free(ovals);
 974
 975         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 976                 if (rs->agg[ddir]) {
 977                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 978                         if (p_of_agg > 100.0)
 979                                 p_of_agg = 100.0;
 980                 }
 981         } else {
 982                 min = max = 0;
 983                 p_of_agg = mean = dev = 0.0;
 984         }
 985         json_object_add_value_int(dir_object, "bw_min", min);
 986         json_object_add_value_int(dir_object, "bw_max", max);
 987         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
 988         json_object_add_value_float(dir_object, "bw_mean", mean);
 989         json_object_add_value_float(dir_object, "bw_dev", dev);
 990 }
 991
 992 static void show_thread_status_terse_v2(struct thread_stat *ts,
 993                                         struct group_run_stats *rs,
 994                                         struct buf_output *out)
 995 {
 996         double io_u_dist[FIO_IO_U_MAP_NR];
 997         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 998         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 999         double usr_cpu, sys_cpu;
1000         int i;
1001
1002         /* General Info */
1003         log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1004         /* Log Read Status */
1005         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1006         /* Log Write Status */
1007         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1008         /* Log Trim Status */
1009         show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1010
1011         /* CPU Usage */
1012         if (ts->total_run_time) {
1013                 double runt = (double) ts->total_run_time;
1014
1015                 usr_cpu = (double) ts->usr_time * 100 / runt;
1016                 sys_cpu = (double) ts->sys_time * 100 / runt;
1017         } else {
1018                 usr_cpu = 0;
1019                 sys_cpu = 0;
1020         }
1021
1022         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1023                                                 (unsigned long long) ts->ctx,
1024                                                 (unsigned long long) ts->majf,
1025                                                 (unsigned long long) ts->minf);
1026
1027         /* Calc % distribution of IO depths, usecond, msecond latency */
1028         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1029         stat_calc_lat_u(ts, io_u_lat_u);
1030         stat_calc_lat_m(ts, io_u_lat_m);
1031
1032         /* Only show fixed 7 I/O depth levels*/
1033         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1034                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1035                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1036
1037         /* Microsecond latency */
1038         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1039                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1040         /* Millisecond latency */
1041         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1042                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1043         /* Additional output if continue_on_error set - default off*/
1044         if (ts->continue_on_error)
1045                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1046         log_buf(out, "\n");
1047
1048         /* Additional output if description is set */
1049         if (strlen(ts->description))
1050                 log_buf(out, ";%s", ts->description);
1051
1052         log_buf(out, "\n");
1053 }
1054
1055 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1056                                            struct group_run_stats *rs, int ver,
1057                                            struct buf_output *out)
1058 {
1059         double io_u_dist[FIO_IO_U_MAP_NR];
1060         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1061         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1062         double usr_cpu, sys_cpu;
1063         int i;
1064
1065         /* General Info */
1066         log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1067                                         ts->name, ts->groupid, ts->error);
1068         /* Log Read Status */
1069         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1070         /* Log Write Status */
1071         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1072         /* Log Trim Status */
1073         if (ver == 4)
1074                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1075
1076         /* CPU Usage */
1077         if (ts->total_run_time) {
1078                 double runt = (double) ts->total_run_time;
1079
1080                 usr_cpu = (double) ts->usr_time * 100 / runt;
1081                 sys_cpu = (double) ts->sys_time * 100 / runt;
1082         } else {
1083                 usr_cpu = 0;
1084                 sys_cpu = 0;
1085         }
1086
1087         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1088                                                 (unsigned long long) ts->ctx,
1089                                                 (unsigned long long) ts->majf,
1090                                                 (unsigned long long) ts->minf);
1091
1092         /* Calc % distribution of IO depths, usecond, msecond latency */
1093         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1094         stat_calc_lat_u(ts, io_u_lat_u);
1095         stat_calc_lat_m(ts, io_u_lat_m);
1096
1097         /* Only show fixed 7 I/O depth levels*/
1098         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1099                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1100                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1101
1102         /* Microsecond latency */
1103         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1104                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1105         /* Millisecond latency */
1106         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1107                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1108
1109         /* disk util stats, if any */
1110         show_disk_util(1, NULL, out);
1111
1112         /* Additional output if continue_on_error set - default off*/
1113         if (ts->continue_on_error)
1114                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1115
1116         /* Additional output if description is set */
1117         if (strlen(ts->description))
1118                 log_buf(out, ";%s", ts->description);
1119
1120         log_buf(out, "\n");
1121 }
1122
1123 void json_add_job_opts(struct json_object *root, const char *name,
1124                        struct flist_head *opt_list, bool num_jobs)
1125 {
1126         struct json_object *dir_object;
1127         struct flist_head *entry;
1128         struct print_option *p;
1129
1130         if (flist_empty(opt_list))
1131                 return;
1132
1133         dir_object = json_create_object();
1134         json_object_add_value_object(root, name, dir_object);
1135
1136         flist_for_each(entry, opt_list) {
1137                 const char *pos = "";
1138
1139                 p = flist_entry(entry, struct print_option, list);
1140                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1141                         continue;
1142                 if (p->value)
1143                         pos = p->value;
1144                 json_object_add_value_string(dir_object, p->name, pos);
1145         }
1146 }
1147
1148 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1149                                                    struct group_run_stats *rs,
1150                                                    struct flist_head *opt_list)
1151 {
1152         struct json_object *root, *tmp;
1153         struct jobs_eta *je;
1154         double io_u_dist[FIO_IO_U_MAP_NR];
1155         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1156         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1157         double usr_cpu, sys_cpu;
1158         int i;
1159         size_t size;
1160
1161         root = json_create_object();
1162         json_object_add_value_string(root, "jobname", ts->name);
1163         json_object_add_value_int(root, "groupid", ts->groupid);
1164         json_object_add_value_int(root, "error", ts->error);
1165
1166         /* ETA Info */
1167         je = get_jobs_eta(true, &size);
1168         if (je) {
1169                 json_object_add_value_int(root, "eta", je->eta_sec);
1170                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1171         }
1172
1173         if (opt_list)
1174                 json_add_job_opts(root, "job options", opt_list, true);
1175
1176         add_ddir_status_json(ts, rs, DDIR_READ, root);
1177         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1178         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1179
1180         /* CPU Usage */
1181         if (ts->total_run_time) {
1182                 double runt = (double) ts->total_run_time;
1183
1184                 usr_cpu = (double) ts->usr_time * 100 / runt;
1185                 sys_cpu = (double) ts->sys_time * 100 / runt;
1186         } else {
1187                 usr_cpu = 0;
1188                 sys_cpu = 0;
1189         }
1190         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1191         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1192         json_object_add_value_int(root, "ctx", ts->ctx);
1193         json_object_add_value_int(root, "majf", ts->majf);
1194         json_object_add_value_int(root, "minf", ts->minf);
1195
1196
1197         /* Calc % distribution of IO depths, usecond, msecond latency */
1198         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1199         stat_calc_lat_u(ts, io_u_lat_u);
1200         stat_calc_lat_m(ts, io_u_lat_m);
1201
1202         tmp = json_create_object();
1203         json_object_add_value_object(root, "iodepth_level", tmp);
1204         /* Only show fixed 7 I/O depth levels*/
1205         for (i = 0; i < 7; i++) {
1206                 char name[20];
1207                 if (i < 6)
1208                         snprintf(name, 20, "%d", 1 << i);
1209                 else
1210                         snprintf(name, 20, ">=%d", 1 << i);
1211                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1212         }
1213
1214         tmp = json_create_object();
1215         json_object_add_value_object(root, "latency_us", tmp);
1216         /* Microsecond latency */
1217         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1218                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1219                                  "250", "500", "750", "1000", };
1220                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1221         }
1222         /* Millisecond latency */
1223         tmp = json_create_object();
1224         json_object_add_value_object(root, "latency_ms", tmp);
1225         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1226                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1227                                  "250", "500", "750", "1000", "2000",
1228                                  ">=2000", };
1229                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1230         }
1231
1232         /* Additional output if continue_on_error set - default off*/
1233         if (ts->continue_on_error) {
1234                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1235                 json_object_add_value_int(root, "first_error", ts->first_error);
1236         }
1237
1238         if (ts->latency_depth) {
1239                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1240                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1241                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1242                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1243         }
1244
1245         /* Additional output if description is set */
1246         if (strlen(ts->description))
1247                 json_object_add_value_string(root, "desc", ts->description);
1248
1249         if (ts->nr_block_infos) {
1250                 /* Block error histogram and types */
1251                 int len;
1252                 unsigned int *percentiles = NULL;
1253                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1254
1255                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1256                                              ts->percentile_list,
1257                                              &percentiles, block_state_counts);
1258
1259                 if (len) {
1260                         struct json_object *block, *percentile_object, *states;
1261                         int state;
1262                         block = json_create_object();
1263                         json_object_add_value_object(root, "block", block);
1264
1265                         percentile_object = json_create_object();
1266                         json_object_add_value_object(block, "percentiles",
1267                                                      percentile_object);
1268                         for (i = 0; i < len; i++) {
1269                                 char buf[20];
1270                                 snprintf(buf, sizeof(buf), "%f",
1271                                          ts->percentile_list[i].u.f);
1272                                 json_object_add_value_int(percentile_object,
1273                                                           (const char *)buf,
1274                                                           percentiles[i]);
1275                         }
1276
1277                         states = json_create_object();
1278                         json_object_add_value_object(block, "states", states);
1279                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1280                                 json_object_add_value_int(states,
1281                                         block_state_names[state],
1282                                         block_state_counts[state]);
1283                         }
1284                         free(percentiles);
1285                 }
1286         }
1287
1288         if (ts->ss_dur) {
1289                 struct json_object *data;
1290                 struct json_array *iops, *bw;
1291                 int i, j, k;
1292                 char ss_buf[64];
1293
1294                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1295                         ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1296                         ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1297                         (float) ts->ss_limit.u.f,
1298                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1299
1300                 tmp = json_create_object();
1301                 json_object_add_value_object(root, "steadystate", tmp);
1302                 json_object_add_value_string(tmp, "ss", ss_buf);
1303                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1304                 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1305
1306                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1307                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1308                 json_object_add_value_string(tmp, "criterion", ss_buf);
1309                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1310                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1311
1312                 data = json_create_object();
1313                 json_object_add_value_object(tmp, "data", data);
1314                 bw = json_create_array();
1315                 iops = json_create_array();
1316
1317                 /*
1318                 ** if ss was attained or the buffer is not full,
1319                 ** ss->head points to the first element in the list.
1320                 ** otherwise it actually points to the second element
1321                 ** in the list
1322                 */
1323                 if ((ts->ss_state & __FIO_SS_ATTAINED) || ts->ss_sum_y == 0)
1324                         j = ts->ss_head;
1325                 else
1326                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1327                 for (i = 0; i < ts->ss_dur; i++) {
1328                         k = (j + i) % ts->ss_dur;
1329                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1330                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1331                 }
1332                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1333                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1334                 json_object_add_value_array(data, "iops", iops);
1335                 json_object_add_value_array(data, "bw", bw);
1336         }
1337
1338         return root;
1339 }
1340
1341 static void show_thread_status_terse(struct thread_stat *ts,
1342                                      struct group_run_stats *rs,
1343                                      struct buf_output *out)
1344 {
1345         if (terse_version == 2)
1346                 show_thread_status_terse_v2(ts, rs, out);
1347         else if (terse_version == 3 || terse_version == 4)
1348                 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1349         else
1350                 log_err("fio: bad terse version!? %d\n", terse_version);
1351 }
1352
1353 struct json_object *show_thread_status(struct thread_stat *ts,
1354                                        struct group_run_stats *rs,
1355                                        struct flist_head *opt_list,
1356                                        struct buf_output *out)
1357 {
1358         struct json_object *ret = NULL;
1359
1360         if (output_format & FIO_OUTPUT_TERSE)
1361                 show_thread_status_terse(ts, rs,  out);
1362         if (output_format & FIO_OUTPUT_JSON)
1363                 ret = show_thread_status_json(ts, rs, opt_list);
1364         if (output_format & FIO_OUTPUT_NORMAL)
1365                 show_thread_status_normal(ts, rs,  out);
1366
1367         return ret;
1368 }
1369
1370 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1371 {
1372         double mean, S;
1373
1374         if (src->samples == 0)
1375                 return;
1376
1377         dst->min_val = min(dst->min_val, src->min_val);
1378         dst->max_val = max(dst->max_val, src->max_val);
1379
1380         /*
1381          * Compute new mean and S after the merge
1382          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1383          *  #Parallel_algorithm>
1384          */
1385         if (first) {
1386                 mean = src->mean.u.f;
1387                 S = src->S.u.f;
1388         } else {
1389                 double delta = src->mean.u.f - dst->mean.u.f;
1390
1391                 mean = ((src->mean.u.f * src->samples) +
1392                         (dst->mean.u.f * dst->samples)) /
1393                         (dst->samples + src->samples);
1394
1395                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1396                         (dst->samples * src->samples) /
1397                         (dst->samples + src->samples);
1398         }
1399
1400         dst->samples += src->samples;
1401         dst->mean.u.f = mean;
1402         dst->S.u.f = S;
1403 }
1404
1405 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1406 {
1407         int i;
1408
1409         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1410                 if (dst->max_run[i] < src->max_run[i])
1411                         dst->max_run[i] = src->max_run[i];
1412                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1413                         dst->min_run[i] = src->min_run[i];
1414                 if (dst->max_bw[i] < src->max_bw[i])
1415                         dst->max_bw[i] = src->max_bw[i];
1416                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1417                         dst->min_bw[i] = src->min_bw[i];
1418
1419                 dst->io_kb[i] += src->io_kb[i];
1420                 dst->agg[i] += src->agg[i];
1421         }
1422
1423         if (!dst->kb_base)
1424                 dst->kb_base = src->kb_base;
1425         if (!dst->unit_base)
1426                 dst->unit_base = src->unit_base;
1427 }
1428
1429 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1430                       bool first)
1431 {
1432         int l, k;
1433
1434         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1435                 if (!dst->unified_rw_rep) {
1436                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1437                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1438                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1439                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1440
1441                         dst->io_bytes[l] += src->io_bytes[l];
1442
1443                         if (dst->runtime[l] < src->runtime[l])
1444                                 dst->runtime[l] = src->runtime[l];
1445                 } else {
1446                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1447                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1448                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1449                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1450
1451                         dst->io_bytes[0] += src->io_bytes[l];
1452
1453                         if (dst->runtime[0] < src->runtime[l])
1454                                 dst->runtime[0] = src->runtime[l];
1455
1456                         /*
1457                          * We're summing to the same destination, so override
1458                          * 'first' after the first iteration of the loop
1459                          */
1460                         first = false;
1461                 }
1462         }
1463
1464         dst->usr_time += src->usr_time;
1465         dst->sys_time += src->sys_time;
1466         dst->ctx += src->ctx;
1467         dst->majf += src->majf;
1468         dst->minf += src->minf;
1469
1470         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1471                 dst->io_u_map[k] += src->io_u_map[k];
1472         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1473                 dst->io_u_submit[k] += src->io_u_submit[k];
1474         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1475                 dst->io_u_complete[k] += src->io_u_complete[k];
1476         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1477                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1478         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1479                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1480
1481         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1482                 if (!dst->unified_rw_rep) {
1483                         dst->total_io_u[k] += src->total_io_u[k];
1484                         dst->short_io_u[k] += src->short_io_u[k];
1485                         dst->drop_io_u[k] += src->drop_io_u[k];
1486                 } else {
1487                         dst->total_io_u[0] += src->total_io_u[k];
1488                         dst->short_io_u[0] += src->short_io_u[k];
1489                         dst->drop_io_u[0] += src->drop_io_u[k];
1490                 }
1491         }
1492
1493         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1494                 int m;
1495
1496                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1497                         if (!dst->unified_rw_rep)
1498                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1499                         else
1500                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1501                 }
1502         }
1503
1504         dst->total_run_time += src->total_run_time;
1505         dst->total_submit += src->total_submit;
1506         dst->total_complete += src->total_complete;
1507 }
1508
1509 void init_group_run_stat(struct group_run_stats *gs)
1510 {
1511         int i;
1512         memset(gs, 0, sizeof(*gs));
1513
1514         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1515                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1516 }
1517
1518 void init_thread_stat(struct thread_stat *ts)
1519 {
1520         int j;
1521
1522         memset(ts, 0, sizeof(*ts));
1523
1524         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1525                 ts->lat_stat[j].min_val = -1UL;
1526                 ts->clat_stat[j].min_val = -1UL;
1527                 ts->slat_stat[j].min_val = -1UL;
1528                 ts->bw_stat[j].min_val = -1UL;
1529         }
1530         ts->groupid = -1;
1531 }
1532
1533 void __show_run_stats(void)
1534 {
1535         struct group_run_stats *runstats, *rs;
1536         struct thread_data *td;
1537         struct thread_stat *threadstats, *ts;
1538         int i, j, k, nr_ts, last_ts, idx;
1539         int kb_base_warned = 0;
1540         int unit_base_warned = 0;
1541         struct json_object *root = NULL;
1542         struct json_array *array = NULL;
1543         struct buf_output output[FIO_OUTPUT_NR];
1544         struct flist_head **opt_lists;
1545
1546         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1547
1548         for (i = 0; i < groupid + 1; i++)
1549                 init_group_run_stat(&runstats[i]);
1550
1551         /*
1552          * find out how many threads stats we need. if group reporting isn't
1553          * enabled, it's one-per-td.
1554          */
1555         nr_ts = 0;
1556         last_ts = -1;
1557         for_each_td(td, i) {
1558                 if (!td->o.group_reporting) {
1559                         nr_ts++;
1560                         continue;
1561                 }
1562                 if (last_ts == td->groupid)
1563                         continue;
1564
1565                 last_ts = td->groupid;
1566                 nr_ts++;
1567         }
1568
1569         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1570         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1571
1572         for (i = 0; i < nr_ts; i++) {
1573                 init_thread_stat(&threadstats[i]);
1574                 opt_lists[i] = NULL;
1575         }
1576
1577         j = 0;
1578         last_ts = -1;
1579         idx = 0;
1580         for_each_td(td, i) {
1581                 if (idx && (!td->o.group_reporting ||
1582                     (td->o.group_reporting && last_ts != td->groupid))) {
1583                         idx = 0;
1584                         j++;
1585                 }
1586
1587                 last_ts = td->groupid;
1588
1589                 ts = &threadstats[j];
1590
1591                 ts->clat_percentiles = td->o.clat_percentiles;
1592                 ts->percentile_precision = td->o.percentile_precision;
1593                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1594                 opt_lists[j] = &td->opt_list;
1595
1596                 idx++;
1597                 ts->members++;
1598
1599                 if (ts->groupid == -1) {
1600                         /*
1601                          * These are per-group shared already
1602                          */
1603                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1604                         if (td->o.description)
1605                                 strncpy(ts->description, td->o.description,
1606                                                 FIO_JOBDESC_SIZE - 1);
1607                         else
1608                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1609
1610                         /*
1611                          * If multiple entries in this group, this is
1612                          * the first member.
1613                          */
1614                         ts->thread_number = td->thread_number;
1615                         ts->groupid = td->groupid;
1616
1617                         /*
1618                          * first pid in group, not very useful...
1619                          */
1620                         ts->pid = td->pid;
1621
1622                         ts->kb_base = td->o.kb_base;
1623                         ts->unit_base = td->o.unit_base;
1624                         ts->unified_rw_rep = td->o.unified_rw_rep;
1625                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1626                         log_info("fio: kb_base differs for jobs in group, using"
1627                                  " %u as the base\n", ts->kb_base);
1628                         kb_base_warned = 1;
1629                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1630                         log_info("fio: unit_base differs for jobs in group, using"
1631                                  " %u as the base\n", ts->unit_base);
1632                         unit_base_warned = 1;
1633                 }
1634
1635                 ts->continue_on_error = td->o.continue_on_error;
1636                 ts->total_err_count += td->total_err_count;
1637                 ts->first_error = td->first_error;
1638                 if (!ts->error) {
1639                         if (!td->error && td->o.continue_on_error &&
1640                             td->first_error) {
1641                                 ts->error = td->first_error;
1642                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1643                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1644                         } else  if (td->error) {
1645                                 ts->error = td->error;
1646                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1647                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1648                         }
1649                 }
1650
1651                 ts->latency_depth = td->latency_qd;
1652                 ts->latency_target = td->o.latency_target;
1653                 ts->latency_percentile = td->o.latency_percentile;
1654                 ts->latency_window = td->o.latency_window;
1655
1656                 ts->nr_block_infos = td->ts.nr_block_infos;
1657                 for (k = 0; k < ts->nr_block_infos; k++)
1658                         ts->block_infos[k] = td->ts.block_infos[k];
1659
1660                 sum_thread_stats(ts, &td->ts, idx == 1);
1661
1662                 if (td->o.ss_dur) {
1663                         ts->ss_state = td->ss.state;
1664                         ts->ss_dur = td->ss.dur;
1665                         ts->ss_head = td->ss.head;
1666                         ts->ss_sum_y = td->ss.sum_y;
1667                         ts->ss_bw_data = td->ss.bw_data;
1668                         ts->ss_iops_data = td->ss.iops_data;
1669                         ts->ss_limit.u.f = td->ss.limit;
1670                         ts->ss_slope.u.f = td->ss.slope;
1671                         ts->ss_deviation.u.f = td->ss.deviation;
1672                         ts->ss_criterion.u.f = td->ss.criterion;
1673                 }
1674                 else
1675                         ts->ss_dur = ts->ss_state = 0;
1676         }
1677
1678         for (i = 0; i < nr_ts; i++) {
1679                 unsigned long long bw;
1680
1681                 ts = &threadstats[i];
1682                 if (ts->groupid == -1)
1683                         continue;
1684                 rs = &runstats[ts->groupid];
1685                 rs->kb_base = ts->kb_base;
1686                 rs->unit_base = ts->unit_base;
1687                 rs->unified_rw_rep += ts->unified_rw_rep;
1688
1689                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1690                         if (!ts->runtime[j])
1691                                 continue;
1692                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1693                                 rs->min_run[j] = ts->runtime[j];
1694                         if (ts->runtime[j] > rs->max_run[j])
1695                                 rs->max_run[j] = ts->runtime[j];
1696
1697                         bw = 0;
1698                         if (ts->runtime[j]) {
1699                                 unsigned long runt = ts->runtime[j];
1700                                 unsigned long long kb;
1701
1702                                 kb = ts->io_bytes[j] / rs->kb_base;
1703                                 bw = kb * 1000 / runt;
1704                         }
1705                         if (bw < rs->min_bw[j])
1706                                 rs->min_bw[j] = bw;
1707                         if (bw > rs->max_bw[j])
1708                                 rs->max_bw[j] = bw;
1709
1710                         rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1711                 }
1712         }
1713
1714         for (i = 0; i < groupid + 1; i++) {
1715                 int ddir;
1716
1717                 rs = &runstats[i];
1718
1719                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1720                         if (rs->max_run[ddir])
1721                                 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1722                                                 rs->max_run[ddir];
1723                 }
1724         }
1725
1726         for (i = 0; i < FIO_OUTPUT_NR; i++)
1727                 buf_output_init(&output[i]);
1728
1729         /*
1730          * don't overwrite last signal output
1731          */
1732         if (output_format & FIO_OUTPUT_NORMAL)
1733                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1734         if (output_format & FIO_OUTPUT_JSON) {
1735                 struct thread_data *global;
1736                 char time_buf[32];
1737                 struct timeval now;
1738                 unsigned long long ms_since_epoch;
1739
1740                 gettimeofday(&now, NULL);
1741                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1742                                  (unsigned long long)(now.tv_usec) / 1000;
1743
1744                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1745                                 sizeof(time_buf));
1746                 time_buf[strlen(time_buf) - 1] = '\0';
1747
1748                 root = json_create_object();
1749                 json_object_add_value_string(root, "fio version", fio_version_string);
1750                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1751                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1752                 json_object_add_value_string(root, "time", time_buf);
1753                 global = get_global_options();
1754                 json_add_job_opts(root, "global options", &global->opt_list, false);
1755                 array = json_create_array();
1756                 json_object_add_value_array(root, "jobs", array);
1757         }
1758
1759         if (is_backend)
1760                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1761
1762         for (i = 0; i < nr_ts; i++) {
1763                 ts = &threadstats[i];
1764                 rs = &runstats[ts->groupid];
1765
1766                 if (is_backend) {
1767                         fio_server_send_job_options(opt_lists[i], i);
1768                         fio_server_send_ts(ts, rs);
1769                 } else {
1770                         if (output_format & FIO_OUTPUT_TERSE)
1771                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1772                         if (output_format & FIO_OUTPUT_JSON) {
1773                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1774                                 json_array_add_value_object(array, tmp);
1775                         }
1776                         if (output_format & FIO_OUTPUT_NORMAL)
1777                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1778                 }
1779         }
1780         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1781                 /* disk util stats, if any */
1782                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1783
1784                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1785
1786                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1787                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1788                 json_free_object(root);
1789         }
1790
1791         for (i = 0; i < groupid + 1; i++) {
1792                 rs = &runstats[i];
1793
1794                 rs->groupid = i;
1795                 if (is_backend)
1796                         fio_server_send_gs(rs);
1797                 else if (output_format & FIO_OUTPUT_NORMAL)
1798                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1799         }
1800
1801         if (is_backend)
1802                 fio_server_send_du();
1803         else if (output_format & FIO_OUTPUT_NORMAL) {
1804                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1805                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1806         }
1807
1808         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1809                 buf_output_flush(&output[i]);
1810                 buf_output_free(&output[i]);
1811         }
1812
1813         log_info_flush();
1814         free(runstats);
1815         free(threadstats);
1816         free(opt_lists);
1817 }
1818
1819 void show_run_stats(void)
1820 {
1821         fio_mutex_down(stat_mutex);
1822         __show_run_stats();
1823         fio_mutex_up(stat_mutex);
1824 }
1825
1826 void __show_running_run_stats(void)
1827 {
1828         struct thread_data *td;
1829         unsigned long long *rt;
1830         struct timeval tv;
1831         int i;
1832
1833         fio_mutex_down(stat_mutex);
1834
1835         rt = malloc(thread_number * sizeof(unsigned long long));
1836         fio_gettime(&tv, NULL);
1837
1838         for_each_td(td, i) {
1839                 td->update_rusage = 1;
1840                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1841                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1842                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1843                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1844
1845                 rt[i] = mtime_since(&td->start, &tv);
1846                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1847                         td->ts.runtime[DDIR_READ] += rt[i];
1848                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1849                         td->ts.runtime[DDIR_WRITE] += rt[i];
1850                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1851                         td->ts.runtime[DDIR_TRIM] += rt[i];
1852         }
1853
1854         for_each_td(td, i) {
1855                 if (td->runstate >= TD_EXITED)
1856                         continue;
1857                 if (td->rusage_sem) {
1858                         td->update_rusage = 1;
1859                         fio_mutex_down(td->rusage_sem);
1860                 }
1861                 td->update_rusage = 0;
1862         }
1863
1864         __show_run_stats();
1865
1866         for_each_td(td, i) {
1867                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1868                         td->ts.runtime[DDIR_READ] -= rt[i];
1869                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1870                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1871                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1872                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1873         }
1874
1875         free(rt);
1876         fio_mutex_up(stat_mutex);
1877 }
1878
1879 static int status_interval_init;
1880 static struct timeval status_time;
1881 static int status_file_disabled;
1882
1883 #define FIO_STATUS_FILE         "fio-dump-status"
1884
1885 static int check_status_file(void)
1886 {
1887         struct stat sb;
1888         const char *temp_dir;
1889         char fio_status_file_path[PATH_MAX];
1890
1891         if (status_file_disabled)
1892                 return 0;
1893
1894         temp_dir = getenv("TMPDIR");
1895         if (temp_dir == NULL) {
1896                 temp_dir = getenv("TEMP");
1897                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1898                         temp_dir = NULL;
1899         }
1900         if (temp_dir == NULL)
1901                 temp_dir = "/tmp";
1902
1903         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1904
1905         if (stat(fio_status_file_path, &sb))
1906                 return 0;
1907
1908         if (unlink(fio_status_file_path) < 0) {
1909                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1910                                                         strerror(errno));
1911                 log_err("fio: disabling status file updates\n");
1912                 status_file_disabled = 1;
1913         }
1914
1915         return 1;
1916 }
1917
1918 void check_for_running_stats(void)
1919 {
1920         if (status_interval) {
1921                 if (!status_interval_init) {
1922                         fio_gettime(&status_time, NULL);
1923                         status_interval_init = 1;
1924                 } else if (mtime_since_now(&status_time) >= status_interval) {
1925                         show_running_run_stats();
1926                         fio_gettime(&status_time, NULL);
1927                         return;
1928                 }
1929         }
1930         if (check_status_file()) {
1931                 show_running_run_stats();
1932                 return;
1933         }
1934 }
1935
1936 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1937 {
1938         double val = data;
1939         double delta;
1940
1941         if (data > is->max_val)
1942                 is->max_val = data;
1943         if (data < is->min_val)
1944                 is->min_val = data;
1945
1946         delta = val - is->mean.u.f;
1947         if (delta) {
1948                 is->mean.u.f += delta / (is->samples + 1.0);
1949                 is->S.u.f += delta * (val - is->mean.u.f);
1950         }
1951
1952         is->samples++;
1953 }
1954
1955 /*
1956  * Return a struct io_logs, which is added to the tail of the log
1957  * list for 'iolog'.
1958  */
1959 static struct io_logs *get_new_log(struct io_log *iolog)
1960 {
1961         size_t new_size, new_samples;
1962         struct io_logs *cur_log;
1963
1964         /*
1965          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1966          * forever
1967          */
1968         if (!iolog->cur_log_max)
1969                 new_samples = DEF_LOG_ENTRIES;
1970         else {
1971                 new_samples = iolog->cur_log_max * 2;
1972                 if (new_samples > MAX_LOG_ENTRIES)
1973                         new_samples = MAX_LOG_ENTRIES;
1974         }
1975
1976         new_size = new_samples * log_entry_sz(iolog);
1977
1978         cur_log = smalloc(sizeof(*cur_log));
1979         if (cur_log) {
1980                 INIT_FLIST_HEAD(&cur_log->list);
1981                 cur_log->log = malloc(new_size);
1982                 if (cur_log->log) {
1983                         cur_log->nr_samples = 0;
1984                         cur_log->max_samples = new_samples;
1985                         flist_add_tail(&cur_log->list, &iolog->io_logs);
1986                         iolog->cur_log_max = new_samples;
1987                         return cur_log;
1988                 }
1989                 sfree(cur_log);
1990         }
1991
1992         return NULL;
1993 }
1994
1995 /*
1996  * Add and return a new log chunk, or return current log if big enough
1997  */
1998 static struct io_logs *regrow_log(struct io_log *iolog)
1999 {
2000         struct io_logs *cur_log;
2001         int i;
2002
2003         if (!iolog || iolog->disabled)
2004                 goto disable;
2005
2006         cur_log = iolog_cur_log(iolog);
2007         if (!cur_log) {
2008                 cur_log = get_new_log(iolog);
2009                 if (!cur_log)
2010                         return NULL;
2011         }
2012
2013         if (cur_log->nr_samples < cur_log->max_samples)
2014                 return cur_log;
2015
2016         /*
2017          * No room for a new sample. If we're compressing on the fly, flush
2018          * out the current chunk
2019          */
2020         if (iolog->log_gz) {
2021                 if (iolog_cur_flush(iolog, cur_log)) {
2022                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2023                         return NULL;
2024                 }
2025         }
2026
2027         /*
2028          * Get a new log array, and add to our list
2029          */
2030         cur_log = get_new_log(iolog);
2031         if (!cur_log) {
2032                 log_err("fio: failed extending iolog! Will stop logging.\n");
2033                 return NULL;
2034         }
2035
2036         if (!iolog->pending || !iolog->pending->nr_samples)
2037                 return cur_log;
2038
2039         /*
2040          * Flush pending items to new log
2041          */
2042         for (i = 0; i < iolog->pending->nr_samples; i++) {
2043                 struct io_sample *src, *dst;
2044
2045                 src = get_sample(iolog, iolog->pending, i);
2046                 dst = get_sample(iolog, cur_log, i);
2047                 memcpy(dst, src, log_entry_sz(iolog));
2048         }
2049         cur_log->nr_samples = iolog->pending->nr_samples;
2050
2051         iolog->pending->nr_samples = 0;
2052         return cur_log;
2053 disable:
2054         if (iolog)
2055                 iolog->disabled = true;
2056         return NULL;
2057 }
2058
2059 void regrow_logs(struct thread_data *td)
2060 {
2061         regrow_log(td->slat_log);
2062         regrow_log(td->clat_log);
2063         regrow_log(td->clat_hist_log);
2064         regrow_log(td->lat_log);
2065         regrow_log(td->bw_log);
2066         regrow_log(td->iops_log);
2067         td->flags &= ~TD_F_REGROW_LOGS;
2068 }
2069
2070 static struct io_logs *get_cur_log(struct io_log *iolog)
2071 {
2072         struct io_logs *cur_log;
2073
2074         cur_log = iolog_cur_log(iolog);
2075         if (!cur_log) {
2076                 cur_log = get_new_log(iolog);
2077                 if (!cur_log)
2078                         return NULL;
2079         }
2080
2081         if (cur_log->nr_samples < cur_log->max_samples)
2082                 return cur_log;
2083
2084         /*
2085          * Out of space. If we're in IO offload mode, or we're not doing
2086          * per unit logging (hence logging happens outside of the IO thread
2087          * as well), add a new log chunk inline. If we're doing inline
2088          * submissions, flag 'td' as needing a log regrow and we'll take
2089          * care of it on the submission side.
2090          */
2091         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2092             !per_unit_log(iolog))
2093                 return regrow_log(iolog);
2094
2095         iolog->td->flags |= TD_F_REGROW_LOGS;
2096         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2097         return iolog->pending;
2098 }
2099
2100 static void __add_log_sample(struct io_log *iolog, unsigned long val,
2101                              enum fio_ddir ddir, unsigned int bs,
2102                              unsigned long t, uint64_t offset)
2103 {
2104         struct io_logs *cur_log;
2105
2106         if (iolog->disabled)
2107                 return;
2108         if (flist_empty(&iolog->io_logs))
2109                 iolog->avg_last = t;
2110
2111         cur_log = get_cur_log(iolog);
2112         if (cur_log) {
2113                 struct io_sample *s;
2114
2115                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2116
2117                 s->val = val;
2118                 s->time = t;
2119                 io_sample_set_ddir(iolog, s, ddir);
2120                 s->bs = bs;
2121
2122                 if (iolog->log_offset) {
2123                         struct io_sample_offset *so = (void *) s;
2124
2125                         so->offset = offset;
2126                 }
2127
2128                 cur_log->nr_samples++;
2129                 return;
2130         }
2131
2132         iolog->disabled = true;
2133 }
2134
2135 static inline void reset_io_stat(struct io_stat *ios)
2136 {
2137         ios->max_val = ios->min_val = ios->samples = 0;
2138         ios->mean.u.f = ios->S.u.f = 0;
2139 }
2140
2141 void reset_io_stats(struct thread_data *td)
2142 {
2143         struct thread_stat *ts = &td->ts;
2144         int i, j;
2145
2146         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2147                 reset_io_stat(&ts->clat_stat[i]);
2148                 reset_io_stat(&ts->slat_stat[i]);
2149                 reset_io_stat(&ts->lat_stat[i]);
2150                 reset_io_stat(&ts->bw_stat[i]);
2151                 reset_io_stat(&ts->iops_stat[i]);
2152
2153                 ts->io_bytes[i] = 0;
2154                 ts->runtime[i] = 0;
2155
2156                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2157                         ts->io_u_plat[i][j] = 0;
2158         }
2159
2160         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2161                 ts->io_u_map[i] = 0;
2162                 ts->io_u_submit[i] = 0;
2163                 ts->io_u_complete[i] = 0;
2164                 ts->io_u_lat_u[i] = 0;
2165                 ts->io_u_lat_m[i] = 0;
2166                 ts->total_submit = 0;
2167                 ts->total_complete = 0;
2168         }
2169
2170         for (i = 0; i < 3; i++) {
2171                 ts->total_io_u[i] = 0;
2172                 ts->short_io_u[i] = 0;
2173                 ts->drop_io_u[i] = 0;
2174         }
2175 }
2176
2177 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2178                               unsigned long elapsed, bool log_max)
2179 {
2180         /*
2181          * Note an entry in the log. Use the mean from the logged samples,
2182          * making sure to properly round up. Only write a log entry if we
2183          * had actual samples done.
2184          */
2185         if (iolog->avg_window[ddir].samples) {
2186                 unsigned long val;
2187
2188                 if (log_max)
2189                         val = iolog->avg_window[ddir].max_val;
2190                 else
2191                         val = iolog->avg_window[ddir].mean.u.f + 0.50;
2192
2193                 __add_log_sample(iolog, val, ddir, 0, elapsed, 0);
2194         }
2195
2196         reset_io_stat(&iolog->avg_window[ddir]);
2197 }
2198
2199 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2200                              bool log_max)
2201 {
2202         int ddir;
2203
2204         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2205                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2206 }
2207
2208 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2209                            unsigned long val, enum fio_ddir ddir,
2210                            unsigned int bs, uint64_t offset)
2211 {
2212         unsigned long elapsed, this_window;
2213
2214         if (!ddir_rw(ddir))
2215                 return 0;
2216
2217         elapsed = mtime_since_now(&td->epoch);
2218
2219         /*
2220          * If no time averaging, just add the log sample.
2221          */
2222         if (!iolog->avg_msec) {
2223                 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
2224                 return 0;
2225         }
2226
2227         /*
2228          * Add the sample. If the time period has passed, then
2229          * add that entry to the log and clear.
2230          */
2231         add_stat_sample(&iolog->avg_window[ddir], val);
2232
2233         /*
2234          * If period hasn't passed, adding the above sample is all we
2235          * need to do.
2236          */
2237         this_window = elapsed - iolog->avg_last;
2238         if (elapsed < iolog->avg_last)
2239                 return iolog->avg_last - elapsed;
2240         else if (this_window < iolog->avg_msec) {
2241                 int diff = iolog->avg_msec - this_window;
2242
2243                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2244                         return diff;
2245         }
2246
2247         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2248
2249         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2250         return iolog->avg_msec;
2251 }
2252
2253 void finalize_logs(struct thread_data *td, bool unit_logs)
2254 {
2255         unsigned long elapsed;
2256
2257         elapsed = mtime_since_now(&td->epoch);
2258
2259         if (td->clat_log && unit_logs)
2260                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2261         if (td->slat_log && unit_logs)
2262                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2263         if (td->lat_log && unit_logs)
2264                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2265         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2266                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2267         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2268                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2269 }
2270
2271 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
2272 {
2273         struct io_log *iolog;
2274
2275         if (!ddir_rw(ddir))
2276                 return;
2277
2278         iolog = agg_io_log[ddir];
2279         __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
2280 }
2281
2282 static void add_clat_percentile_sample(struct thread_stat *ts,
2283                                 unsigned long usec, enum fio_ddir ddir)
2284 {
2285         unsigned int idx = plat_val_to_idx(usec);
2286         assert(idx < FIO_IO_U_PLAT_NR);
2287
2288         ts->io_u_plat[ddir][idx]++;
2289 }
2290
2291 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2292                      unsigned long usec, unsigned int bs, uint64_t offset)
2293 {
2294         unsigned long elapsed, this_window;
2295         struct thread_stat *ts = &td->ts;
2296         struct io_log *iolog = td->clat_hist_log;
2297
2298         td_io_u_lock(td);
2299
2300         add_stat_sample(&ts->clat_stat[ddir], usec);
2301
2302         if (td->clat_log)
2303                 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
2304
2305         if (ts->clat_percentiles)
2306                 add_clat_percentile_sample(ts, usec, ddir);
2307
2308         if (iolog && iolog->hist_msec) {
2309                 struct io_hist *hw = &iolog->hist_window[ddir];
2310
2311                 hw->samples++;
2312                 elapsed = mtime_since_now(&td->epoch);
2313                 if (!hw->hist_last)
2314                         hw->hist_last = elapsed;
2315                 this_window = elapsed - hw->hist_last;
2316
2317                 if (this_window >= iolog->hist_msec) {
2318                         unsigned int *io_u_plat;
2319                         unsigned int *dst;
2320
2321                         /*
2322                          * Make a byte-for-byte copy of the latency histogram
2323                          * stored in td->ts.io_u_plat[ddir], recording it in a
2324                          * log sample. Note that the matching call to free() is
2325                          * located in iolog.c after printing this sample to the
2326                          * log file.
2327                          */
2328                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2329                         dst = malloc(FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2330                         memcpy(dst, io_u_plat,
2331                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2332                         __add_log_sample(iolog, (unsigned long )dst, ddir, bs,
2333                                                 elapsed, offset);
2334
2335                         /*
2336                          * Update the last time we recorded as being now, minus
2337                          * any drift in time we encountered before actually
2338                          * making the record.
2339                          */
2340                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2341                         hw->samples = 0;
2342                 }
2343         }
2344
2345         td_io_u_unlock(td);
2346 }
2347
2348 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2349                      unsigned long usec, unsigned int bs, uint64_t offset)
2350 {
2351         struct thread_stat *ts = &td->ts;
2352
2353         if (!ddir_rw(ddir))
2354                 return;
2355
2356         td_io_u_lock(td);
2357
2358         add_stat_sample(&ts->slat_stat[ddir], usec);
2359
2360         if (td->slat_log)
2361                 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
2362
2363         td_io_u_unlock(td);
2364 }
2365
2366 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2367                     unsigned long usec, unsigned int bs, uint64_t offset)
2368 {
2369         struct thread_stat *ts = &td->ts;
2370
2371         if (!ddir_rw(ddir))
2372                 return;
2373
2374         td_io_u_lock(td);
2375
2376         add_stat_sample(&ts->lat_stat[ddir], usec);
2377
2378         if (td->lat_log)
2379                 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2380
2381         td_io_u_unlock(td);
2382 }
2383
2384 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2385                    unsigned int bytes, unsigned long spent)
2386 {
2387         struct thread_stat *ts = &td->ts;
2388         unsigned long rate;
2389
2390         if (spent)
2391                 rate = bytes * 1000 / spent;
2392         else
2393                 rate = 0;
2394
2395         td_io_u_lock(td);
2396
2397         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2398
2399         if (td->bw_log)
2400                 add_log_sample(td, td->bw_log, rate, io_u->ddir, bytes, io_u->offset);
2401
2402         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2403         td_io_u_unlock(td);
2404 }
2405
2406 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2407 {
2408         struct thread_stat *ts = &td->ts;
2409         unsigned long spent, rate;
2410         enum fio_ddir ddir;
2411         unsigned int next, next_log;
2412
2413         next_log = td->o.bw_avg_time;
2414
2415         spent = mtime_since(&td->bw_sample_time, t);
2416         if (spent < td->o.bw_avg_time &&
2417             td->o.bw_avg_time - spent >= LOG_MSEC_SLACK)
2418                 return td->o.bw_avg_time - spent;
2419
2420         td_io_u_lock(td);
2421
2422         /*
2423          * Compute both read and write rates for the interval.
2424          */
2425         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2426                 uint64_t delta;
2427
2428                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2429                 if (!delta)
2430                         continue; /* No entries for interval */
2431
2432                 if (spent)
2433                         rate = delta * 1000 / spent / 1024;
2434                 else
2435                         rate = 0;
2436
2437                 add_stat_sample(&ts->bw_stat[ddir], rate);
2438
2439                 if (td->bw_log) {
2440                         unsigned int bs = 0;
2441
2442                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2443                                 bs = td->o.min_bs[ddir];
2444
2445                         next = add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2446                         next_log = min(next_log, next);
2447                 }
2448
2449                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2450         }
2451
2452         timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time);
2453
2454         td_io_u_unlock(td);
2455
2456         if (spent <= td->o.bw_avg_time)
2457                 return min(next_log, td->o.bw_avg_time);
2458
2459         next = td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time);
2460         return min(next, next_log);
2461 }
2462
2463 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2464                      unsigned int bytes)
2465 {
2466         struct thread_stat *ts = &td->ts;
2467
2468         td_io_u_lock(td);
2469
2470         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2471
2472         if (td->iops_log)
2473                 add_log_sample(td, td->iops_log, 1, io_u->ddir, bytes, io_u->offset);
2474
2475         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2476         td_io_u_unlock(td);
2477 }
2478
2479 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2480 {
2481         struct thread_stat *ts = &td->ts;
2482         unsigned long spent, iops;
2483         enum fio_ddir ddir;
2484         unsigned int next, next_log;
2485
2486         next_log = td->o.iops_avg_time;
2487
2488         spent = mtime_since(&td->iops_sample_time, t);
2489         if (spent < td->o.iops_avg_time &&
2490             td->o.iops_avg_time - spent >= LOG_MSEC_SLACK)
2491                 return td->o.iops_avg_time - spent;
2492
2493         td_io_u_lock(td);
2494
2495         /*
2496          * Compute both read and write rates for the interval.
2497          */
2498         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2499                 uint64_t delta;
2500
2501                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2502                 if (!delta)
2503                         continue; /* No entries for interval */
2504
2505                 if (spent)
2506                         iops = (delta * 1000) / spent;
2507                 else
2508                         iops = 0;
2509
2510                 add_stat_sample(&ts->iops_stat[ddir], iops);
2511
2512                 if (td->iops_log) {
2513                         unsigned int bs = 0;
2514
2515                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2516                                 bs = td->o.min_bs[ddir];
2517
2518                         next = add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2519                         next_log = min(next_log, next);
2520                 }
2521
2522                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2523         }
2524
2525         timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time);
2526
2527         td_io_u_unlock(td);
2528
2529         if (spent <= td->o.iops_avg_time)
2530                 return min(next_log, td->o.iops_avg_time);
2531
2532         next = td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
2533         return min(next, next_log);
2534 }
2535
2536 /*
2537  * Returns msecs to next event
2538  */
2539 int calc_log_samples(void)
2540 {
2541         struct thread_data *td;
2542         unsigned int next = ~0U, tmp;
2543         struct timeval now;
2544         int i;
2545
2546         fio_gettime(&now, NULL);
2547
2548         for_each_td(td, i) {
2549                 if (in_ramp_time(td) ||
2550                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2551                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2552                         continue;
2553                 }
2554                 if (!per_unit_log(td->bw_log)) {
2555                         tmp = add_bw_samples(td, &now);
2556                         if (tmp < next)
2557                                 next = tmp;
2558                 }
2559                 if (!per_unit_log(td->iops_log)) {
2560                         tmp = add_iops_samples(td, &now);
2561                         if (tmp < next)
2562                                 next = tmp;
2563                 }
2564         }
2565
2566         return next == ~0U ? 0 : next;
2567 }
2568
2569 void stat_init(void)
2570 {
2571         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2572 }
2573
2574 void stat_exit(void)
2575 {
2576         /*
2577          * When we have the mutex, we know out-of-band access to it
2578          * have ended.
2579          */
2580         fio_mutex_down(stat_mutex);
2581         fio_mutex_remove(stat_mutex);
2582 }
2583
2584 /*
2585  * Called from signal handler. Wake up status thread.
2586  */
2587 void show_running_run_stats(void)
2588 {
2589         helper_do_stat();
2590 }
2591
2592 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2593 {
2594         /* Ignore io_u's which span multiple blocks--they will just get
2595          * inaccurate counts. */
2596         int idx = (io_u->offset - io_u->file->file_offset)
2597                         / td->o.bs[DDIR_TRIM];
2598         uint32_t *info = &td->ts.block_infos[idx];
2599         assert(idx < td->ts.nr_block_infos);
2600         return info;
2601 }