stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/types.h>
   5 #include <sys/stat.h>
   6 #include <dirent.h>
   7 #include <libgen.h>
   8 #include <math.h>
   9
  10 #include "fio.h"
  11 #include "diskutil.h"
  12 #include "lib/ieee754.h"
  13 #include "json.h"
  14 #include "lib/getrusage.h"
  15 #include "idletime.h"
  16 #include "lib/pow2.h"
  17 #include "lib/output_buffer.h"
  18 #include "helper_thread.h"
  19 #include "smalloc.h"
  20
  21 #define LOG_MSEC_SLACK  10
  22
  23 struct fio_mutex *stat_mutex;
  24
  25 void clear_rusage_stat(struct thread_data *td)
  26 {
  27         struct thread_stat *ts = &td->ts;
  28
  29         fio_getrusage(&td->ru_start);
  30         ts->usr_time = ts->sys_time = 0;
  31         ts->ctx = 0;
  32         ts->minf = ts->majf = 0;
  33 }
  34
  35 void update_rusage_stat(struct thread_data *td)
  36 {
  37         struct thread_stat *ts = &td->ts;
  38
  39         fio_getrusage(&td->ru_end);
  40         ts->usr_time += mtime_since(&td->ru_start.ru_utime,
  41                                         &td->ru_end.ru_utime);
  42         ts->sys_time += mtime_since(&td->ru_start.ru_stime,
  43                                         &td->ru_end.ru_stime);
  44         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  45                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  46         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  47         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  48
  49         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  50 }
  51
  52 /*
  53  * Given a latency, return the index of the corresponding bucket in
  54  * the structure tracking percentiles.
  55  *
  56  * (1) find the group (and error bits) that the value (latency)
  57  * belongs to by looking at its MSB. (2) find the bucket number in the
  58  * group by looking at the index bits.
  59  *
  60  */
  61 static unsigned int plat_val_to_idx(unsigned int val)
  62 {
  63         unsigned int msb, error_bits, base, offset, idx;
  64
  65         /* Find MSB starting from bit 0 */
  66         if (val == 0)
  67                 msb = 0;
  68         else
  69                 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
  70
  71         /*
  72          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  73          * all bits of the sample as index
  74          */
  75         if (msb <= FIO_IO_U_PLAT_BITS)
  76                 return val;
  77
  78         /* Compute the number of error bits to discard*/
  79         error_bits = msb - FIO_IO_U_PLAT_BITS;
  80
  81         /* Compute the number of buckets before the group */
  82         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  83
  84         /*
  85          * Discard the error bits and apply the mask to find the
  86          * index for the buckets in the group
  87          */
  88         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  89
  90         /* Make sure the index does not exceed (array size - 1) */
  91         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  92                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  93
  94         return idx;
  95 }
  96
  97 /*
  98  * Convert the given index of the bucket array to the value
  99  * represented by the bucket
 100  */
 101 static unsigned int plat_idx_to_val(unsigned int idx)
 102 {
 103         unsigned int error_bits, k, base;
 104
 105         assert(idx < FIO_IO_U_PLAT_NR);
 106
 107         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 108          * all bits of the sample as index */
 109         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 110                 return idx;
 111
 112         /* Find the group and compute the minimum value of that group */
 113         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 114         base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
 115
 116         /* Find its bucket number of the group */
 117         k = idx % FIO_IO_U_PLAT_VAL;
 118
 119         /* Return the mean of the range of the bucket */
 120         return base + ((k + 0.5) * (1 << error_bits));
 121 }
 122
 123 static int double_cmp(const void *a, const void *b)
 124 {
 125         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 126         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 127         int cmp = 0;
 128
 129         if (fa.u.f > fb.u.f)
 130                 cmp = 1;
 131         else if (fa.u.f < fb.u.f)
 132                 cmp = -1;
 133
 134         return cmp;
 135 }
 136
 137 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 138                                    fio_fp64_t *plist, unsigned int **output,
 139                                    unsigned int *maxv, unsigned int *minv)
 140 {
 141         unsigned long sum = 0;
 142         unsigned int len, i, j = 0;
 143         unsigned int oval_len = 0;
 144         unsigned int *ovals = NULL;
 145         int is_last;
 146
 147         *minv = -1U;
 148         *maxv = 0;
 149
 150         len = 0;
 151         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 152                 len++;
 153
 154         if (!len)
 155                 return 0;
 156
 157         /*
 158          * Sort the percentile list. Note that it may already be sorted if
 159          * we are using the default values, but since it's a short list this
 160          * isn't a worry. Also note that this does not work for NaN values.
 161          */
 162         if (len > 1)
 163                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 164
 165         /*
 166          * Calculate bucket values, note down max and min values
 167          */
 168         is_last = 0;
 169         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 170                 sum += io_u_plat[i];
 171                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 172                         assert(plist[j].u.f <= 100.0);
 173
 174                         if (j == oval_len) {
 175                                 oval_len += 100;
 176                                 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
 177                         }
 178
 179                         ovals[j] = plat_idx_to_val(i);
 180                         if (ovals[j] < *minv)
 181                                 *minv = ovals[j];
 182                         if (ovals[j] > *maxv)
 183                                 *maxv = ovals[j];
 184
 185                         is_last = (j == len - 1);
 186                         if (is_last)
 187                                 break;
 188
 189                         j++;
 190                 }
 191         }
 192
 193         *output = ovals;
 194         return len;
 195 }
 196
 197 /*
 198  * Find and display the p-th percentile of clat
 199  */
 200 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 201                                   fio_fp64_t *plist, unsigned int precision,
 202                                   struct buf_output *out)
 203 {
 204         unsigned int len, j = 0, minv, maxv;
 205         unsigned int *ovals;
 206         int is_last, per_line, scale_down;
 207         char fmt[32];
 208
 209         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 210         if (!len)
 211                 goto out;
 212
 213         /*
 214          * We default to usecs, but if the value range is such that we
 215          * should scale down to msecs, do that.
 216          */
 217         if (minv > 2000 && maxv > 99999) {
 218                 scale_down = 1;
 219                 log_buf(out, "    clat percentiles (msec):\n     |");
 220         } else {
 221                 scale_down = 0;
 222                 log_buf(out, "    clat percentiles (usec):\n     |");
 223         }
 224
 225         snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
 226         per_line = (80 - 7) / (precision + 14);
 227
 228         for (j = 0; j < len; j++) {
 229                 char fbuf[16], *ptr = fbuf;
 230
 231                 /* for formatting */
 232                 if (j != 0 && (j % per_line) == 0)
 233                         log_buf(out, "     |");
 234
 235                 /* end of the list */
 236                 is_last = (j == len - 1);
 237
 238                 if (plist[j].u.f < 10.0)
 239                         ptr += sprintf(fbuf, " ");
 240
 241                 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
 242
 243                 if (scale_down)
 244                         ovals[j] = (ovals[j] + 999) / 1000;
 245
 246                 log_buf(out, " %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
 247
 248                 if (is_last)
 249                         break;
 250
 251                 if ((j % per_line) == per_line - 1)     /* for formatting */
 252                         log_buf(out, "\n");
 253         }
 254
 255 out:
 256         if (ovals)
 257                 free(ovals);
 258 }
 259
 260 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 261              double *mean, double *dev)
 262 {
 263         double n = (double) is->samples;
 264
 265         if (n == 0)
 266                 return 0;
 267
 268         *min = is->min_val;
 269         *max = is->max_val;
 270         *mean = is->mean.u.f;
 271
 272         if (n > 1.0)
 273                 *dev = sqrt(is->S.u.f / (n - 1.0));
 274         else
 275                 *dev = 0;
 276
 277         return 1;
 278 }
 279
 280 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 281 {
 282         char *p1, *p2, *p3, *p4;
 283         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 284         int i;
 285
 286         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 287
 288         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 289                 const int i2p = is_power_of_2(rs->kb_base);
 290
 291                 if (!rs->max_run[i])
 292                         continue;
 293
 294                 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
 295                 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
 296                 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 297                 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 298
 299                 log_buf(out, "%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
 300                          " mint=%llumsec, maxt=%llumsec\n",
 301                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 302                                 p1, p2, p3, p4,
 303                                 (unsigned long long) rs->min_run[i],
 304                                 (unsigned long long) rs->max_run[i]);
 305
 306                 free(p1);
 307                 free(p2);
 308                 free(p3);
 309                 free(p4);
 310         }
 311 }
 312
 313 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 314 {
 315         int i;
 316
 317         /*
 318          * Do depth distribution calculations
 319          */
 320         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 321                 if (total) {
 322                         io_u_dist[i] = (double) map[i] / (double) total;
 323                         io_u_dist[i] *= 100.0;
 324                         if (io_u_dist[i] < 0.1 && map[i])
 325                                 io_u_dist[i] = 0.1;
 326                 } else
 327                         io_u_dist[i] = 0.0;
 328         }
 329 }
 330
 331 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 332                           unsigned int *src, int nr)
 333 {
 334         unsigned long total = ddir_rw_sum(ts->total_io_u);
 335         int i;
 336
 337         /*
 338          * Do latency distribution calculations
 339          */
 340         for (i = 0; i < nr; i++) {
 341                 if (total) {
 342                         dst[i] = (double) src[i] / (double) total;
 343                         dst[i] *= 100.0;
 344                         if (dst[i] < 0.01 && src[i])
 345                                 dst[i] = 0.01;
 346                 } else
 347                         dst[i] = 0.0;
 348         }
 349 }
 350
 351 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 352 {
 353         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 354 }
 355
 356 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 357 {
 358         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 359 }
 360
 361 static void display_lat(const char *name, unsigned long min, unsigned long max,
 362                         double mean, double dev, struct buf_output *out)
 363 {
 364         const char *base = "(usec)";
 365         char *minp, *maxp;
 366
 367         if (!usec_to_msec(&min, &max, &mean, &dev))
 368                 base = "(msec)";
 369
 370         minp = num2str(min, 6, 1, 0, 0);
 371         maxp = num2str(max, 6, 1, 0, 0);
 372
 373         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 374                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 375
 376         free(minp);
 377         free(maxp);
 378 }
 379
 380 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 381                              int ddir, struct buf_output *out)
 382 {
 383         const char *str[] = { "read ", "write", "trim" };
 384         unsigned long min, max, runt;
 385         unsigned long long bw, iops;
 386         double mean, dev;
 387         char *io_p, *bw_p, *iops_p;
 388         int i2p;
 389
 390         assert(ddir_rw(ddir));
 391
 392         if (!ts->runtime[ddir])
 393                 return;
 394
 395         i2p = is_power_of_2(rs->kb_base);
 396         runt = ts->runtime[ddir];
 397
 398         bw = (1000 * ts->io_bytes[ddir]) / runt;
 399         io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
 400         bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
 401
 402         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 403         iops_p = num2str(iops, 6, 1, 0, 0);
 404
 405         log_buf(out, "  %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
 406                                 rs->unified_rw_rep ? "mixed" : str[ddir],
 407                                 io_p, bw_p, iops_p,
 408                                 (unsigned long long) ts->runtime[ddir]);
 409
 410         free(io_p);
 411         free(bw_p);
 412         free(iops_p);
 413
 414         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 415                 display_lat("slat", min, max, mean, dev, out);
 416         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 417                 display_lat("clat", min, max, mean, dev, out);
 418         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 419                 display_lat(" lat", min, max, mean, dev, out);
 420
 421         if (ts->clat_percentiles) {
 422                 show_clat_percentiles(ts->io_u_plat[ddir],
 423                                         ts->clat_stat[ddir].samples,
 424                                         ts->percentile_list,
 425                                         ts->percentile_precision, out);
 426         }
 427         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 428                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 429                 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
 430
 431                 if (rs->unit_base == 1) {
 432                         min *= 8.0;
 433                         max *= 8.0;
 434                         mean *= 8.0;
 435                         dev *= 8.0;
 436                 }
 437
 438                 if (rs->agg[ddir]) {
 439                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 440                         if (p_of_agg > 100.0)
 441                                 p_of_agg = 100.0;
 442                 }
 443
 444                 if (mean > fkb_base * fkb_base) {
 445                         min /= fkb_base;
 446                         max /= fkb_base;
 447                         mean /= fkb_base;
 448                         dev /= fkb_base;
 449                         bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
 450                 }
 451
 452                 log_buf(out, "    bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
 453                          " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
 454                                                         p_of_agg, mean, dev);
 455         }
 456 }
 457
 458 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 459                     const char *msg, struct buf_output *out)
 460 {
 461         int new_line = 1, i, line = 0, shown = 0;
 462
 463         for (i = 0; i < nr; i++) {
 464                 if (io_u_lat[i] <= 0.0)
 465                         continue;
 466                 shown = 1;
 467                 if (new_line) {
 468                         if (line)
 469                                 log_buf(out, "\n");
 470                         log_buf(out, "    lat (%s) : ", msg);
 471                         new_line = 0;
 472                         line = 0;
 473                 }
 474                 if (line)
 475                         log_buf(out, ", ");
 476                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 477                 line++;
 478                 if (line == 5)
 479                         new_line = 1;
 480         }
 481
 482         if (shown)
 483                 log_buf(out, "\n");
 484
 485         return shown;
 486 }
 487
 488 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 489 {
 490         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 491                                  "250=", "500=", "750=", "1000=", };
 492
 493         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 494 }
 495
 496 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 497 {
 498         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 499                                  "250=", "500=", "750=", "1000=", "2000=",
 500                                  ">=2000=", };
 501
 502         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 503 }
 504
 505 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 506 {
 507         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 508         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 509
 510         stat_calc_lat_u(ts, io_u_lat_u);
 511         stat_calc_lat_m(ts, io_u_lat_m);
 512
 513         show_lat_u(io_u_lat_u, out);
 514         show_lat_m(io_u_lat_m, out);
 515 }
 516
 517 static int block_state_category(int block_state)
 518 {
 519         switch (block_state) {
 520         case BLOCK_STATE_UNINIT:
 521                 return 0;
 522         case BLOCK_STATE_TRIMMED:
 523         case BLOCK_STATE_WRITTEN:
 524                 return 1;
 525         case BLOCK_STATE_WRITE_FAILURE:
 526         case BLOCK_STATE_TRIM_FAILURE:
 527                 return 2;
 528         default:
 529                 /* Silence compile warning on some BSDs and have a return */
 530                 assert(0);
 531                 return -1;
 532         }
 533 }
 534
 535 static int compare_block_infos(const void *bs1, const void *bs2)
 536 {
 537         uint32_t block1 = *(uint32_t *)bs1;
 538         uint32_t block2 = *(uint32_t *)bs2;
 539         int state1 = BLOCK_INFO_STATE(block1);
 540         int state2 = BLOCK_INFO_STATE(block2);
 541         int bscat1 = block_state_category(state1);
 542         int bscat2 = block_state_category(state2);
 543         int cycles1 = BLOCK_INFO_TRIMS(block1);
 544         int cycles2 = BLOCK_INFO_TRIMS(block2);
 545
 546         if (bscat1 < bscat2)
 547                 return -1;
 548         if (bscat1 > bscat2)
 549                 return 1;
 550
 551         if (cycles1 < cycles2)
 552                 return -1;
 553         if (cycles1 > cycles2)
 554                 return 1;
 555
 556         if (state1 < state2)
 557                 return -1;
 558         if (state1 > state2)
 559                 return 1;
 560
 561         assert(block1 == block2);
 562         return 0;
 563 }
 564
 565 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 566                                   fio_fp64_t *plist, unsigned int **percentiles,
 567                                   unsigned int *types)
 568 {
 569         int len = 0;
 570         int i, nr_uninit;
 571
 572         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 573
 574         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 575                 len++;
 576
 577         if (!len)
 578                 return 0;
 579
 580         /*
 581          * Sort the percentile list. Note that it may already be sorted if
 582          * we are using the default values, but since it's a short list this
 583          * isn't a worry. Also note that this does not work for NaN values.
 584          */
 585         if (len > 1)
 586                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 587
 588         nr_uninit = 0;
 589         /* Start only after the uninit entries end */
 590         for (nr_uninit = 0;
 591              nr_uninit < nr_block_infos
 592                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 593              nr_uninit ++)
 594                 ;
 595
 596         if (nr_uninit == nr_block_infos)
 597                 return 0;
 598
 599         *percentiles = calloc(len, sizeof(**percentiles));
 600
 601         for (i = 0; i < len; i++) {
 602                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 603                                 + nr_uninit;
 604                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 605         }
 606
 607         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 608         for (i = 0; i < nr_block_infos; i++)
 609                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 610
 611         return len;
 612 }
 613
 614 static const char *block_state_names[] = {
 615         [BLOCK_STATE_UNINIT] = "unwritten",
 616         [BLOCK_STATE_TRIMMED] = "trimmed",
 617         [BLOCK_STATE_WRITTEN] = "written",
 618         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 619         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 620 };
 621
 622 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 623                              fio_fp64_t *plist, struct buf_output *out)
 624 {
 625         int len, pos, i;
 626         unsigned int *percentiles = NULL;
 627         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 628
 629         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 630                                      &percentiles, block_state_counts);
 631
 632         log_buf(out, "  block lifetime percentiles :\n   |");
 633         pos = 0;
 634         for (i = 0; i < len; i++) {
 635                 uint32_t block_info = percentiles[i];
 636 #define LINE_LENGTH     75
 637                 char str[LINE_LENGTH];
 638                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 639                                      plist[i].u.f, block_info,
 640                                      i == len - 1 ? '\n' : ',');
 641                 assert(strln < LINE_LENGTH);
 642                 if (pos + strln > LINE_LENGTH) {
 643                         pos = 0;
 644                         log_buf(out, "\n   |");
 645                 }
 646                 log_buf(out, "%s", str);
 647                 pos += strln;
 648 #undef LINE_LENGTH
 649         }
 650         if (percentiles)
 651                 free(percentiles);
 652
 653         log_buf(out, "        states               :");
 654         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 655                 log_buf(out, " %s=%u%c",
 656                          block_state_names[i], block_state_counts[i],
 657                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 658 }
 659
 660 static void show_thread_status_normal(struct thread_stat *ts,
 661                                       struct group_run_stats *rs,
 662                                       struct buf_output *out)
 663 {
 664         double usr_cpu, sys_cpu;
 665         unsigned long runtime;
 666         double io_u_dist[FIO_IO_U_MAP_NR];
 667         time_t time_p;
 668         char time_buf[32];
 669
 670         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 671                 return;
 672
 673         time(&time_p);
 674         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 675
 676         if (!ts->error) {
 677                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 678                                         ts->name, ts->groupid, ts->members,
 679                                         ts->error, (int) ts->pid, time_buf);
 680         } else {
 681                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 682                                         ts->name, ts->groupid, ts->members,
 683                                         ts->error, ts->verror, (int) ts->pid,
 684                                         time_buf);
 685         }
 686
 687         if (strlen(ts->description))
 688                 log_buf(out, "  Description  : [%s]\n", ts->description);
 689
 690         if (ts->io_bytes[DDIR_READ])
 691                 show_ddir_status(rs, ts, DDIR_READ, out);
 692         if (ts->io_bytes[DDIR_WRITE])
 693                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 694         if (ts->io_bytes[DDIR_TRIM])
 695                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 696
 697         show_latencies(ts, out);
 698
 699         runtime = ts->total_run_time;
 700         if (runtime) {
 701                 double runt = (double) runtime;
 702
 703                 usr_cpu = (double) ts->usr_time * 100 / runt;
 704                 sys_cpu = (double) ts->sys_time * 100 / runt;
 705         } else {
 706                 usr_cpu = 0;
 707                 sys_cpu = 0;
 708         }
 709
 710         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 711                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 712                         (unsigned long long) ts->ctx,
 713                         (unsigned long long) ts->majf,
 714                         (unsigned long long) ts->minf);
 715
 716         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 717         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 718                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 719                                         io_u_dist[1], io_u_dist[2],
 720                                         io_u_dist[3], io_u_dist[4],
 721                                         io_u_dist[5], io_u_dist[6]);
 722
 723         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 724         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 725                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 726                                         io_u_dist[1], io_u_dist[2],
 727                                         io_u_dist[3], io_u_dist[4],
 728                                         io_u_dist[5], io_u_dist[6]);
 729         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 730         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 731                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 732                                         io_u_dist[1], io_u_dist[2],
 733                                         io_u_dist[3], io_u_dist[4],
 734                                         io_u_dist[5], io_u_dist[6]);
 735         log_buf(out, "     issued    : total=r=%llu/w=%llu/d=%llu,"
 736                                  " short=r=%llu/w=%llu/d=%llu,"
 737                                  " drop=r=%llu/w=%llu/d=%llu\n",
 738                                         (unsigned long long) ts->total_io_u[0],
 739                                         (unsigned long long) ts->total_io_u[1],
 740                                         (unsigned long long) ts->total_io_u[2],
 741                                         (unsigned long long) ts->short_io_u[0],
 742                                         (unsigned long long) ts->short_io_u[1],
 743                                         (unsigned long long) ts->short_io_u[2],
 744                                         (unsigned long long) ts->drop_io_u[0],
 745                                         (unsigned long long) ts->drop_io_u[1],
 746                                         (unsigned long long) ts->drop_io_u[2]);
 747         if (ts->continue_on_error) {
 748                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 749                                         (unsigned long long)ts->total_err_count,
 750                                         ts->first_error,
 751                                         strerror(ts->first_error));
 752         }
 753         if (ts->latency_depth) {
 754                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 755                                         (unsigned long long)ts->latency_target,
 756                                         (unsigned long long)ts->latency_window,
 757                                         ts->latency_percentile.u.f,
 758                                         ts->latency_depth);
 759         }
 760
 761         if (ts->nr_block_infos)
 762                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 763                                   ts->percentile_list, out);
 764 }
 765
 766 static void show_ddir_status_terse(struct thread_stat *ts,
 767                                    struct group_run_stats *rs, int ddir,
 768                                    struct buf_output *out)
 769 {
 770         unsigned long min, max;
 771         unsigned long long bw, iops;
 772         unsigned int *ovals = NULL;
 773         double mean, dev;
 774         unsigned int len, minv, maxv;
 775         int i;
 776
 777         assert(ddir_rw(ddir));
 778
 779         iops = bw = 0;
 780         if (ts->runtime[ddir]) {
 781                 uint64_t runt = ts->runtime[ddir];
 782
 783                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 784                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 785         }
 786
 787         log_buf(out, ";%llu;%llu;%llu;%llu",
 788                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 789                                         (unsigned long long) ts->runtime[ddir]);
 790
 791         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 792                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 793         else
 794                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 795
 796         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 797                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 798         else
 799                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 800
 801         if (ts->clat_percentiles) {
 802                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 803                                         ts->clat_stat[ddir].samples,
 804                                         ts->percentile_list, &ovals, &maxv,
 805                                         &minv);
 806         } else
 807                 len = 0;
 808
 809         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 810                 if (i >= len) {
 811                         log_buf(out, ";0%%=0");
 812                         continue;
 813                 }
 814                 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
 815         }
 816
 817         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 818                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 819         else
 820                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 821
 822         if (ovals)
 823                 free(ovals);
 824
 825         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 826                 double p_of_agg = 100.0;
 827
 828                 if (rs->agg[ddir]) {
 829                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 830                         if (p_of_agg > 100.0)
 831                                 p_of_agg = 100.0;
 832                 }
 833
 834                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 835         } else
 836                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 837 }
 838
 839 static void add_ddir_status_json(struct thread_stat *ts,
 840                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 841 {
 842         unsigned long min, max;
 843         unsigned long long bw;
 844         unsigned int *ovals = NULL;
 845         double mean, dev, iops;
 846         unsigned int len, minv, maxv;
 847         int i;
 848         const char *ddirname[] = {"read", "write", "trim"};
 849         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 850         char buf[120];
 851         double p_of_agg = 100.0;
 852
 853         assert(ddir_rw(ddir));
 854
 855         if (ts->unified_rw_rep && ddir != DDIR_READ)
 856                 return;
 857
 858         dir_object = json_create_object();
 859         json_object_add_value_object(parent,
 860                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 861
 862         bw = 0;
 863         iops = 0.0;
 864         if (ts->runtime[ddir]) {
 865                 uint64_t runt = ts->runtime[ddir];
 866
 867                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 868                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 869         }
 870
 871         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
 872         json_object_add_value_int(dir_object, "bw", bw);
 873         json_object_add_value_float(dir_object, "iops", iops);
 874         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 875         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 876         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 877         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 878
 879         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 880                 min = max = 0;
 881                 mean = dev = 0.0;
 882         }
 883         tmp_object = json_create_object();
 884         json_object_add_value_object(dir_object, "slat", tmp_object);
 885         json_object_add_value_int(tmp_object, "min", min);
 886         json_object_add_value_int(tmp_object, "max", max);
 887         json_object_add_value_float(tmp_object, "mean", mean);
 888         json_object_add_value_float(tmp_object, "stddev", dev);
 889
 890         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 891                 min = max = 0;
 892                 mean = dev = 0.0;
 893         }
 894         tmp_object = json_create_object();
 895         json_object_add_value_object(dir_object, "clat", tmp_object);
 896         json_object_add_value_int(tmp_object, "min", min);
 897         json_object_add_value_int(tmp_object, "max", max);
 898         json_object_add_value_float(tmp_object, "mean", mean);
 899         json_object_add_value_float(tmp_object, "stddev", dev);
 900
 901         if (ts->clat_percentiles) {
 902                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 903                                         ts->clat_stat[ddir].samples,
 904                                         ts->percentile_list, &ovals, &maxv,
 905                                         &minv);
 906         } else
 907                 len = 0;
 908
 909         percentile_object = json_create_object();
 910         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 911         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 912                 if (i >= len) {
 913                         json_object_add_value_int(percentile_object, "0.00", 0);
 914                         continue;
 915                 }
 916                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 917                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 918         }
 919
 920         if (output_format & FIO_OUTPUT_JSON_PLUS) {
 921                 clat_bins_object = json_create_object();
 922                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
 923                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
 924                         snprintf(buf, sizeof(buf), "%d", i);
 925                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
 926                 }
 927                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS);
 928                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL);
 929                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR);
 930         }
 931
 932         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
 933                 min = max = 0;
 934                 mean = dev = 0.0;
 935         }
 936         tmp_object = json_create_object();
 937         json_object_add_value_object(dir_object, "lat", tmp_object);
 938         json_object_add_value_int(tmp_object, "min", min);
 939         json_object_add_value_int(tmp_object, "max", max);
 940         json_object_add_value_float(tmp_object, "mean", mean);
 941         json_object_add_value_float(tmp_object, "stddev", dev);
 942         if (ovals)
 943                 free(ovals);
 944
 945         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 946                 if (rs->agg[ddir]) {
 947                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 948                         if (p_of_agg > 100.0)
 949                                 p_of_agg = 100.0;
 950                 }
 951         } else {
 952                 min = max = 0;
 953                 p_of_agg = mean = dev = 0.0;
 954         }
 955         json_object_add_value_int(dir_object, "bw_min", min);
 956         json_object_add_value_int(dir_object, "bw_max", max);
 957         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
 958         json_object_add_value_float(dir_object, "bw_mean", mean);
 959         json_object_add_value_float(dir_object, "bw_dev", dev);
 960 }
 961
 962 static void show_thread_status_terse_v2(struct thread_stat *ts,
 963                                         struct group_run_stats *rs,
 964                                         struct buf_output *out)
 965 {
 966         double io_u_dist[FIO_IO_U_MAP_NR];
 967         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 968         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 969         double usr_cpu, sys_cpu;
 970         int i;
 971
 972         /* General Info */
 973         log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
 974         /* Log Read Status */
 975         show_ddir_status_terse(ts, rs, DDIR_READ, out);
 976         /* Log Write Status */
 977         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
 978         /* Log Trim Status */
 979         show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
 980
 981         /* CPU Usage */
 982         if (ts->total_run_time) {
 983                 double runt = (double) ts->total_run_time;
 984
 985                 usr_cpu = (double) ts->usr_time * 100 / runt;
 986                 sys_cpu = (double) ts->sys_time * 100 / runt;
 987         } else {
 988                 usr_cpu = 0;
 989                 sys_cpu = 0;
 990         }
 991
 992         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
 993                                                 (unsigned long long) ts->ctx,
 994                                                 (unsigned long long) ts->majf,
 995                                                 (unsigned long long) ts->minf);
 996
 997         /* Calc % distribution of IO depths, usecond, msecond latency */
 998         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 999         stat_calc_lat_u(ts, io_u_lat_u);
1000         stat_calc_lat_m(ts, io_u_lat_m);
1001
1002         /* Only show fixed 7 I/O depth levels*/
1003         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1004                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1005                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1006
1007         /* Microsecond latency */
1008         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1009                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1010         /* Millisecond latency */
1011         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1012                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1013         /* Additional output if continue_on_error set - default off*/
1014         if (ts->continue_on_error)
1015                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1016         log_buf(out, "\n");
1017
1018         /* Additional output if description is set */
1019         if (strlen(ts->description))
1020                 log_buf(out, ";%s", ts->description);
1021
1022         log_buf(out, "\n");
1023 }
1024
1025 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1026                                            struct group_run_stats *rs, int ver,
1027                                            struct buf_output *out)
1028 {
1029         double io_u_dist[FIO_IO_U_MAP_NR];
1030         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1031         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1032         double usr_cpu, sys_cpu;
1033         int i;
1034
1035         /* General Info */
1036         log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1037                                         ts->name, ts->groupid, ts->error);
1038         /* Log Read Status */
1039         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1040         /* Log Write Status */
1041         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1042         /* Log Trim Status */
1043         if (ver == 4)
1044                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1045
1046         /* CPU Usage */
1047         if (ts->total_run_time) {
1048                 double runt = (double) ts->total_run_time;
1049
1050                 usr_cpu = (double) ts->usr_time * 100 / runt;
1051                 sys_cpu = (double) ts->sys_time * 100 / runt;
1052         } else {
1053                 usr_cpu = 0;
1054                 sys_cpu = 0;
1055         }
1056
1057         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1058                                                 (unsigned long long) ts->ctx,
1059                                                 (unsigned long long) ts->majf,
1060                                                 (unsigned long long) ts->minf);
1061
1062         /* Calc % distribution of IO depths, usecond, msecond latency */
1063         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1064         stat_calc_lat_u(ts, io_u_lat_u);
1065         stat_calc_lat_m(ts, io_u_lat_m);
1066
1067         /* Only show fixed 7 I/O depth levels*/
1068         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1069                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1070                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1071
1072         /* Microsecond latency */
1073         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1074                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1075         /* Millisecond latency */
1076         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1077                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1078
1079         /* disk util stats, if any */
1080         show_disk_util(1, NULL, out);
1081
1082         /* Additional output if continue_on_error set - default off*/
1083         if (ts->continue_on_error)
1084                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1085
1086         /* Additional output if description is set */
1087         if (strlen(ts->description))
1088                 log_buf(out, ";%s", ts->description);
1089
1090         log_buf(out, "\n");
1091 }
1092
1093 void json_add_job_opts(struct json_object *root, const char *name,
1094                        struct flist_head *opt_list, bool num_jobs)
1095 {
1096         struct json_object *dir_object;
1097         struct flist_head *entry;
1098         struct print_option *p;
1099
1100         if (flist_empty(opt_list))
1101                 return;
1102
1103         dir_object = json_create_object();
1104         json_object_add_value_object(root, name, dir_object);
1105
1106         flist_for_each(entry, opt_list) {
1107                 const char *pos = "";
1108
1109                 p = flist_entry(entry, struct print_option, list);
1110                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1111                         continue;
1112                 if (p->value)
1113                         pos = p->value;
1114                 json_object_add_value_string(dir_object, p->name, pos);
1115         }
1116 }
1117
1118 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1119                                                    struct group_run_stats *rs,
1120                                                    struct flist_head *opt_list)
1121 {
1122         struct json_object *root, *tmp;
1123         struct jobs_eta *je;
1124         double io_u_dist[FIO_IO_U_MAP_NR];
1125         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1126         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1127         double usr_cpu, sys_cpu;
1128         int i;
1129         size_t size;
1130
1131         root = json_create_object();
1132         json_object_add_value_string(root, "jobname", ts->name);
1133         json_object_add_value_int(root, "groupid", ts->groupid);
1134         json_object_add_value_int(root, "error", ts->error);
1135
1136         /* ETA Info */
1137         je = get_jobs_eta(true, &size);
1138         if (je) {
1139                 json_object_add_value_int(root, "eta", je->eta_sec);
1140                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1141         }
1142
1143         if (opt_list)
1144                 json_add_job_opts(root, "job options", opt_list, true);
1145
1146         add_ddir_status_json(ts, rs, DDIR_READ, root);
1147         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1148         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1149
1150         /* CPU Usage */
1151         if (ts->total_run_time) {
1152                 double runt = (double) ts->total_run_time;
1153
1154                 usr_cpu = (double) ts->usr_time * 100 / runt;
1155                 sys_cpu = (double) ts->sys_time * 100 / runt;
1156         } else {
1157                 usr_cpu = 0;
1158                 sys_cpu = 0;
1159         }
1160         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1161         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1162         json_object_add_value_int(root, "ctx", ts->ctx);
1163         json_object_add_value_int(root, "majf", ts->majf);
1164         json_object_add_value_int(root, "minf", ts->minf);
1165
1166
1167         /* Calc % distribution of IO depths, usecond, msecond latency */
1168         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1169         stat_calc_lat_u(ts, io_u_lat_u);
1170         stat_calc_lat_m(ts, io_u_lat_m);
1171
1172         tmp = json_create_object();
1173         json_object_add_value_object(root, "iodepth_level", tmp);
1174         /* Only show fixed 7 I/O depth levels*/
1175         for (i = 0; i < 7; i++) {
1176                 char name[20];
1177                 if (i < 6)
1178                         snprintf(name, 20, "%d", 1 << i);
1179                 else
1180                         snprintf(name, 20, ">=%d", 1 << i);
1181                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1182         }
1183
1184         tmp = json_create_object();
1185         json_object_add_value_object(root, "latency_us", tmp);
1186         /* Microsecond latency */
1187         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1188                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1189                                  "250", "500", "750", "1000", };
1190                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1191         }
1192         /* Millisecond latency */
1193         tmp = json_create_object();
1194         json_object_add_value_object(root, "latency_ms", tmp);
1195         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1196                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1197                                  "250", "500", "750", "1000", "2000",
1198                                  ">=2000", };
1199                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1200         }
1201
1202         /* Additional output if continue_on_error set - default off*/
1203         if (ts->continue_on_error) {
1204                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1205                 json_object_add_value_int(root, "first_error", ts->first_error);
1206         }
1207
1208         if (ts->latency_depth) {
1209                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1210                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1211                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1212                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1213         }
1214
1215         /* Additional output if description is set */
1216         if (strlen(ts->description))
1217                 json_object_add_value_string(root, "desc", ts->description);
1218
1219         if (ts->nr_block_infos) {
1220                 /* Block error histogram and types */
1221                 int len;
1222                 unsigned int *percentiles = NULL;
1223                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1224
1225                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1226                                              ts->percentile_list,
1227                                              &percentiles, block_state_counts);
1228
1229                 if (len) {
1230                         struct json_object *block, *percentile_object, *states;
1231                         int state;
1232                         block = json_create_object();
1233                         json_object_add_value_object(root, "block", block);
1234
1235                         percentile_object = json_create_object();
1236                         json_object_add_value_object(block, "percentiles",
1237                                                      percentile_object);
1238                         for (i = 0; i < len; i++) {
1239                                 char buf[20];
1240                                 snprintf(buf, sizeof(buf), "%f",
1241                                          ts->percentile_list[i].u.f);
1242                                 json_object_add_value_int(percentile_object,
1243                                                           (const char *)buf,
1244                                                           percentiles[i]);
1245                         }
1246
1247                         states = json_create_object();
1248                         json_object_add_value_object(block, "states", states);
1249                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1250                                 json_object_add_value_int(states,
1251                                         block_state_names[state],
1252                                         block_state_counts[state]);
1253                         }
1254                         free(percentiles);
1255                 }
1256         }
1257
1258         /* steady state detection; move this behind json+? */
1259         if (ts->ss) {
1260                 struct json_object *data;
1261                 struct json_array *iops, *bw;
1262                 struct steadystate_data *ss = ts->ss;
1263                 unsigned long long sum_iops, sum_bw;
1264                 double mean_iops, mean_bw;
1265                 int i, j, k;
1266                 char ss_buf[64];
1267
1268                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1269                         ss->check_iops ? "iops" : "bw",
1270                         ss->check_slope ? "_slope" : "",
1271                         (float) ss->limit,
1272                         ss->pct ? "%" : "");
1273
1274                 tmp = json_create_object();
1275                 json_object_add_value_object(root, "steadystate", tmp);
1276                 json_object_add_value_string(tmp, "ss", ss_buf);
1277                 json_object_add_value_int(tmp, "duration", (int)ss->dur);
1278                 json_object_add_value_int(tmp, "steadystate_ramptime", ss->ramp_time / 1000000L);
1279                 json_object_add_value_int(tmp, "attained", ss->attained);
1280
1281                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ss->criterion, ss->pct ? "%" : "");
1282                 json_object_add_value_string(tmp, "criterion", ss_buf);
1283                 json_object_add_value_float(tmp, "max_deviation", ss->deviation);
1284                 json_object_add_value_float(tmp, "slope", ss->slope);
1285
1286                 data = json_create_object();
1287                 json_object_add_value_object(tmp, "data", data);
1288                 bw = json_create_array();
1289                 iops = json_create_array();
1290
1291                 /*
1292                 ** if ss was attained or the buffer is not full,
1293                 ** ss->head points to the first element in the list.
1294                 ** otherwise it actually points to the second element
1295                 ** in the list
1296                 */
1297                 if (ss->attained || ss->sum_y == 0)
1298                         j = ss->head;
1299                 else
1300                         j = ss->head == 0 ? ss->dur - 1 : ss->head - 1;
1301                 for (i = 0, sum_iops = 0, sum_bw = 0; i < ss->dur; i++) {
1302                         k = (j + i) % ss->dur;
1303                         sum_bw += ss->bw_data[k];
1304                         sum_iops += ss->iops_data[k];
1305                         json_array_add_value_int(bw, ss->bw_data[k]);
1306                         json_array_add_value_int(iops, ss->iops_data[k]);
1307                 }
1308                 mean_bw = (double) sum_bw / ss->dur;
1309                 mean_iops = (double) sum_iops / ss->dur;
1310                 json_object_add_value_float(data, "bw_mean", mean_bw);
1311                 json_object_add_value_float(data, "iops_mean", mean_iops);
1312                 json_object_add_value_array(data, "iops", iops);
1313                 json_object_add_value_array(data, "bw", bw);
1314         }
1315
1316         return root;
1317 }
1318
1319 static void show_thread_status_terse(struct thread_stat *ts,
1320                                      struct group_run_stats *rs,
1321                                      struct buf_output *out)
1322 {
1323         if (terse_version == 2)
1324                 show_thread_status_terse_v2(ts, rs, out);
1325         else if (terse_version == 3 || terse_version == 4)
1326                 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1327         else
1328                 log_err("fio: bad terse version!? %d\n", terse_version);
1329 }
1330
1331 struct json_object *show_thread_status(struct thread_stat *ts,
1332                                        struct group_run_stats *rs,
1333                                        struct flist_head *opt_list,
1334                                        struct buf_output *out)
1335 {
1336         struct json_object *ret = NULL;
1337
1338         if (output_format & FIO_OUTPUT_TERSE)
1339                 show_thread_status_terse(ts, rs,  out);
1340         if (output_format & FIO_OUTPUT_JSON)
1341                 ret = show_thread_status_json(ts, rs, opt_list);
1342         if (output_format & FIO_OUTPUT_NORMAL)
1343                 show_thread_status_normal(ts, rs,  out);
1344
1345         return ret;
1346 }
1347
1348 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1349 {
1350         double mean, S;
1351
1352         if (src->samples == 0)
1353                 return;
1354
1355         dst->min_val = min(dst->min_val, src->min_val);
1356         dst->max_val = max(dst->max_val, src->max_val);
1357
1358         /*
1359          * Compute new mean and S after the merge
1360          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1361          *  #Parallel_algorithm>
1362          */
1363         if (first) {
1364                 mean = src->mean.u.f;
1365                 S = src->S.u.f;
1366         } else {
1367                 double delta = src->mean.u.f - dst->mean.u.f;
1368
1369                 mean = ((src->mean.u.f * src->samples) +
1370                         (dst->mean.u.f * dst->samples)) /
1371                         (dst->samples + src->samples);
1372
1373                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1374                         (dst->samples * src->samples) /
1375                         (dst->samples + src->samples);
1376         }
1377
1378         dst->samples += src->samples;
1379         dst->mean.u.f = mean;
1380         dst->S.u.f = S;
1381 }
1382
1383 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1384 {
1385         int i;
1386
1387         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1388                 if (dst->max_run[i] < src->max_run[i])
1389                         dst->max_run[i] = src->max_run[i];
1390                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1391                         dst->min_run[i] = src->min_run[i];
1392                 if (dst->max_bw[i] < src->max_bw[i])
1393                         dst->max_bw[i] = src->max_bw[i];
1394                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1395                         dst->min_bw[i] = src->min_bw[i];
1396
1397                 dst->io_kb[i] += src->io_kb[i];
1398                 dst->agg[i] += src->agg[i];
1399         }
1400
1401         if (!dst->kb_base)
1402                 dst->kb_base = src->kb_base;
1403         if (!dst->unit_base)
1404                 dst->unit_base = src->unit_base;
1405 }
1406
1407 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1408                       bool first)
1409 {
1410         int l, k;
1411
1412         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1413                 if (!dst->unified_rw_rep) {
1414                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1415                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1416                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1417                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1418
1419                         dst->io_bytes[l] += src->io_bytes[l];
1420
1421                         if (dst->runtime[l] < src->runtime[l])
1422                                 dst->runtime[l] = src->runtime[l];
1423                 } else {
1424                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1425                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1426                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1427                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1428
1429                         dst->io_bytes[0] += src->io_bytes[l];
1430
1431                         if (dst->runtime[0] < src->runtime[l])
1432                                 dst->runtime[0] = src->runtime[l];
1433
1434                         /*
1435                          * We're summing to the same destination, so override
1436                          * 'first' after the first iteration of the loop
1437                          */
1438                         first = false;
1439                 }
1440         }
1441
1442         dst->usr_time += src->usr_time;
1443         dst->sys_time += src->sys_time;
1444         dst->ctx += src->ctx;
1445         dst->majf += src->majf;
1446         dst->minf += src->minf;
1447
1448         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1449                 dst->io_u_map[k] += src->io_u_map[k];
1450         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1451                 dst->io_u_submit[k] += src->io_u_submit[k];
1452         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1453                 dst->io_u_complete[k] += src->io_u_complete[k];
1454         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1455                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1456         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1457                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1458
1459         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1460                 if (!dst->unified_rw_rep) {
1461                         dst->total_io_u[k] += src->total_io_u[k];
1462                         dst->short_io_u[k] += src->short_io_u[k];
1463                         dst->drop_io_u[k] += src->drop_io_u[k];
1464                 } else {
1465                         dst->total_io_u[0] += src->total_io_u[k];
1466                         dst->short_io_u[0] += src->short_io_u[k];
1467                         dst->drop_io_u[0] += src->drop_io_u[k];
1468                 }
1469         }
1470
1471         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1472                 int m;
1473
1474                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1475                         if (!dst->unified_rw_rep)
1476                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1477                         else
1478                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1479                 }
1480         }
1481
1482         dst->total_run_time += src->total_run_time;
1483         dst->total_submit += src->total_submit;
1484         dst->total_complete += src->total_complete;
1485 }
1486
1487 void init_group_run_stat(struct group_run_stats *gs)
1488 {
1489         int i;
1490         memset(gs, 0, sizeof(*gs));
1491
1492         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1493                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1494 }
1495
1496 void init_thread_stat(struct thread_stat *ts)
1497 {
1498         int j;
1499
1500         memset(ts, 0, sizeof(*ts));
1501
1502         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1503                 ts->lat_stat[j].min_val = -1UL;
1504                 ts->clat_stat[j].min_val = -1UL;
1505                 ts->slat_stat[j].min_val = -1UL;
1506                 ts->bw_stat[j].min_val = -1UL;
1507         }
1508         ts->groupid = -1;
1509 }
1510
1511 void __show_run_stats(void)
1512 {
1513         struct group_run_stats *runstats, *rs;
1514         struct thread_data *td;
1515         struct thread_stat *threadstats, *ts;
1516         int i, j, k, nr_ts, last_ts, idx;
1517         int kb_base_warned = 0;
1518         int unit_base_warned = 0;
1519         struct json_object *root = NULL;
1520         struct json_array *array = NULL;
1521         struct buf_output output[FIO_OUTPUT_NR];
1522         struct flist_head **opt_lists;
1523
1524         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1525
1526         for (i = 0; i < groupid + 1; i++)
1527                 init_group_run_stat(&runstats[i]);
1528
1529         /*
1530          * find out how many threads stats we need. if group reporting isn't
1531          * enabled, it's one-per-td.
1532          */
1533         nr_ts = 0;
1534         last_ts = -1;
1535         for_each_td(td, i) {
1536                 if (!td->o.group_reporting) {
1537                         nr_ts++;
1538                         continue;
1539                 }
1540                 if (last_ts == td->groupid)
1541                         continue;
1542
1543                 last_ts = td->groupid;
1544                 nr_ts++;
1545         }
1546
1547         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1548         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1549
1550         for (i = 0; i < nr_ts; i++) {
1551                 init_thread_stat(&threadstats[i]);
1552                 opt_lists[i] = NULL;
1553         }
1554
1555         j = 0;
1556         last_ts = -1;
1557         idx = 0;
1558         for_each_td(td, i) {
1559                 if (idx && (!td->o.group_reporting ||
1560                     (td->o.group_reporting && last_ts != td->groupid))) {
1561                         idx = 0;
1562                         j++;
1563                 }
1564
1565                 last_ts = td->groupid;
1566
1567                 ts = &threadstats[j];
1568
1569                 ts->clat_percentiles = td->o.clat_percentiles;
1570                 ts->percentile_precision = td->o.percentile_precision;
1571                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1572                 opt_lists[j] = &td->opt_list;
1573
1574                 idx++;
1575                 ts->members++;
1576
1577                 if (ts->groupid == -1) {
1578                         /*
1579                          * These are per-group shared already
1580                          */
1581                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1582                         if (td->o.description)
1583                                 strncpy(ts->description, td->o.description,
1584                                                 FIO_JOBDESC_SIZE - 1);
1585                         else
1586                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1587
1588                         /*
1589                          * If multiple entries in this group, this is
1590                          * the first member.
1591                          */
1592                         ts->thread_number = td->thread_number;
1593                         ts->groupid = td->groupid;
1594
1595                         /*
1596                          * first pid in group, not very useful...
1597                          */
1598                         ts->pid = td->pid;
1599
1600                         ts->kb_base = td->o.kb_base;
1601                         ts->unit_base = td->o.unit_base;
1602                         ts->unified_rw_rep = td->o.unified_rw_rep;
1603                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1604                         log_info("fio: kb_base differs for jobs in group, using"
1605                                  " %u as the base\n", ts->kb_base);
1606                         kb_base_warned = 1;
1607                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1608                         log_info("fio: unit_base differs for jobs in group, using"
1609                                  " %u as the base\n", ts->unit_base);
1610                         unit_base_warned = 1;
1611                 }
1612
1613                 ts->continue_on_error = td->o.continue_on_error;
1614                 ts->total_err_count += td->total_err_count;
1615                 ts->first_error = td->first_error;
1616                 if (!ts->error) {
1617                         if (!td->error && td->o.continue_on_error &&
1618                             td->first_error) {
1619                                 ts->error = td->first_error;
1620                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1621                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1622                         } else  if (td->error) {
1623                                 ts->error = td->error;
1624                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1625                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1626                         }
1627                 }
1628
1629                 ts->latency_depth = td->latency_qd;
1630                 ts->latency_target = td->o.latency_target;
1631                 ts->latency_percentile = td->o.latency_percentile;
1632                 ts->latency_window = td->o.latency_window;
1633
1634                 ts->nr_block_infos = td->ts.nr_block_infos;
1635                 for (k = 0; k < ts->nr_block_infos; k++)
1636                         ts->block_infos[k] = td->ts.block_infos[k];
1637
1638                 sum_thread_stats(ts, &td->ts, idx == 1);
1639
1640                 if (td->o.ss_dur)
1641                         ts->ss = &td->ss;
1642                 else
1643                         ts->ss = NULL;
1644         }
1645
1646         for (i = 0; i < nr_ts; i++) {
1647                 unsigned long long bw;
1648
1649                 ts = &threadstats[i];
1650                 if (ts->groupid == -1)
1651                         continue;
1652                 rs = &runstats[ts->groupid];
1653                 rs->kb_base = ts->kb_base;
1654                 rs->unit_base = ts->unit_base;
1655                 rs->unified_rw_rep += ts->unified_rw_rep;
1656
1657                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1658                         if (!ts->runtime[j])
1659                                 continue;
1660                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1661                                 rs->min_run[j] = ts->runtime[j];
1662                         if (ts->runtime[j] > rs->max_run[j])
1663                                 rs->max_run[j] = ts->runtime[j];
1664
1665                         bw = 0;
1666                         if (ts->runtime[j]) {
1667                                 unsigned long runt = ts->runtime[j];
1668                                 unsigned long long kb;
1669
1670                                 kb = ts->io_bytes[j] / rs->kb_base;
1671                                 bw = kb * 1000 / runt;
1672                         }
1673                         if (bw < rs->min_bw[j])
1674                                 rs->min_bw[j] = bw;
1675                         if (bw > rs->max_bw[j])
1676                                 rs->max_bw[j] = bw;
1677
1678                         rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1679                 }
1680         }
1681
1682         for (i = 0; i < groupid + 1; i++) {
1683                 int ddir;
1684
1685                 rs = &runstats[i];
1686
1687                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1688                         if (rs->max_run[ddir])
1689                                 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1690                                                 rs->max_run[ddir];
1691                 }
1692         }
1693
1694         for (i = 0; i < FIO_OUTPUT_NR; i++)
1695                 buf_output_init(&output[i]);
1696
1697         /*
1698          * don't overwrite last signal output
1699          */
1700         if (output_format & FIO_OUTPUT_NORMAL)
1701                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1702         if (output_format & FIO_OUTPUT_JSON) {
1703                 struct thread_data *global;
1704                 char time_buf[32];
1705                 struct timeval now;
1706                 unsigned long long ms_since_epoch;
1707
1708                 gettimeofday(&now, NULL);
1709                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1710                                  (unsigned long long)(now.tv_usec) / 1000;
1711
1712                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1713                                 sizeof(time_buf));
1714                 time_buf[strlen(time_buf) - 1] = '\0';
1715
1716                 root = json_create_object();
1717                 json_object_add_value_string(root, "fio version", fio_version_string);
1718                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1719                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1720                 json_object_add_value_string(root, "time", time_buf);
1721                 global = get_global_options();
1722                 json_add_job_opts(root, "global options", &global->opt_list, false);
1723                 array = json_create_array();
1724                 json_object_add_value_array(root, "jobs", array);
1725         }
1726
1727         if (is_backend)
1728                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1729
1730         for (i = 0; i < nr_ts; i++) {
1731                 ts = &threadstats[i];
1732                 rs = &runstats[ts->groupid];
1733
1734                 if (is_backend) {
1735                         fio_server_send_job_options(opt_lists[i], i);
1736                         fio_server_send_ts(ts, rs);
1737                 } else {
1738                         if (output_format & FIO_OUTPUT_TERSE)
1739                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1740                         if (output_format & FIO_OUTPUT_JSON) {
1741                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1742                                 json_array_add_value_object(array, tmp);
1743                         }
1744                         if (output_format & FIO_OUTPUT_NORMAL)
1745                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1746                 }
1747         }
1748         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1749                 /* disk util stats, if any */
1750                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1751
1752                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1753
1754                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1755                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1756                 json_free_object(root);
1757         }
1758
1759         for (i = 0; i < groupid + 1; i++) {
1760                 rs = &runstats[i];
1761
1762                 rs->groupid = i;
1763                 if (is_backend)
1764                         fio_server_send_gs(rs);
1765                 else if (output_format & FIO_OUTPUT_NORMAL)
1766                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1767         }
1768
1769         if (is_backend)
1770                 fio_server_send_du();
1771         else if (output_format & FIO_OUTPUT_NORMAL) {
1772                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1773                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1774         }
1775
1776         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1777                 buf_output_flush(&output[i]);
1778                 buf_output_free(&output[i]);
1779         }
1780
1781         log_info_flush();
1782         free(runstats);
1783         free(threadstats);
1784         free(opt_lists);
1785 }
1786
1787 void show_run_stats(void)
1788 {
1789         fio_mutex_down(stat_mutex);
1790         __show_run_stats();
1791         fio_mutex_up(stat_mutex);
1792 }
1793
1794 void __show_running_run_stats(void)
1795 {
1796         struct thread_data *td;
1797         unsigned long long *rt;
1798         struct timeval tv;
1799         int i;
1800
1801         fio_mutex_down(stat_mutex);
1802
1803         rt = malloc(thread_number * sizeof(unsigned long long));
1804         fio_gettime(&tv, NULL);
1805
1806         for_each_td(td, i) {
1807                 td->update_rusage = 1;
1808                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1809                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1810                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1811                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1812
1813                 rt[i] = mtime_since(&td->start, &tv);
1814                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1815                         td->ts.runtime[DDIR_READ] += rt[i];
1816                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1817                         td->ts.runtime[DDIR_WRITE] += rt[i];
1818                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1819                         td->ts.runtime[DDIR_TRIM] += rt[i];
1820         }
1821
1822         for_each_td(td, i) {
1823                 if (td->runstate >= TD_EXITED)
1824                         continue;
1825                 if (td->rusage_sem) {
1826                         td->update_rusage = 1;
1827                         fio_mutex_down(td->rusage_sem);
1828                 }
1829                 td->update_rusage = 0;
1830         }
1831
1832         __show_run_stats();
1833
1834         for_each_td(td, i) {
1835                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1836                         td->ts.runtime[DDIR_READ] -= rt[i];
1837                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1838                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1839                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1840                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1841         }
1842
1843         free(rt);
1844         fio_mutex_up(stat_mutex);
1845 }
1846
1847 static int status_interval_init;
1848 static struct timeval status_time;
1849 static int status_file_disabled;
1850
1851 #define FIO_STATUS_FILE         "fio-dump-status"
1852
1853 static int check_status_file(void)
1854 {
1855         struct stat sb;
1856         const char *temp_dir;
1857         char fio_status_file_path[PATH_MAX];
1858
1859         if (status_file_disabled)
1860                 return 0;
1861
1862         temp_dir = getenv("TMPDIR");
1863         if (temp_dir == NULL) {
1864                 temp_dir = getenv("TEMP");
1865                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1866                         temp_dir = NULL;
1867         }
1868         if (temp_dir == NULL)
1869                 temp_dir = "/tmp";
1870
1871         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1872
1873         if (stat(fio_status_file_path, &sb))
1874                 return 0;
1875
1876         if (unlink(fio_status_file_path) < 0) {
1877                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1878                                                         strerror(errno));
1879                 log_err("fio: disabling status file updates\n");
1880                 status_file_disabled = 1;
1881         }
1882
1883         return 1;
1884 }
1885
1886 void check_for_running_stats(void)
1887 {
1888         if (status_interval) {
1889                 if (!status_interval_init) {
1890                         fio_gettime(&status_time, NULL);
1891                         status_interval_init = 1;
1892                 } else if (mtime_since_now(&status_time) >= status_interval) {
1893                         show_running_run_stats();
1894                         fio_gettime(&status_time, NULL);
1895                         return;
1896                 }
1897         }
1898         if (check_status_file()) {
1899                 show_running_run_stats();
1900                 return;
1901         }
1902 }
1903
1904 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1905 {
1906         double val = data;
1907         double delta;
1908
1909         if (data > is->max_val)
1910                 is->max_val = data;
1911         if (data < is->min_val)
1912                 is->min_val = data;
1913
1914         delta = val - is->mean.u.f;
1915         if (delta) {
1916                 is->mean.u.f += delta / (is->samples + 1.0);
1917                 is->S.u.f += delta * (val - is->mean.u.f);
1918         }
1919
1920         is->samples++;
1921 }
1922
1923 /*
1924  * Return a struct io_logs, which is added to the tail of the log
1925  * list for 'iolog'.
1926  */
1927 static struct io_logs *get_new_log(struct io_log *iolog)
1928 {
1929         size_t new_size, new_samples;
1930         struct io_logs *cur_log;
1931
1932         /*
1933          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1934          * forever
1935          */
1936         if (!iolog->cur_log_max)
1937                 new_samples = DEF_LOG_ENTRIES;
1938         else {
1939                 new_samples = iolog->cur_log_max * 2;
1940                 if (new_samples > MAX_LOG_ENTRIES)
1941                         new_samples = MAX_LOG_ENTRIES;
1942         }
1943
1944         new_size = new_samples * log_entry_sz(iolog);
1945
1946         cur_log = smalloc(sizeof(*cur_log));
1947         if (cur_log) {
1948                 INIT_FLIST_HEAD(&cur_log->list);
1949                 cur_log->log = malloc(new_size);
1950                 if (cur_log->log) {
1951                         cur_log->nr_samples = 0;
1952                         cur_log->max_samples = new_samples;
1953                         flist_add_tail(&cur_log->list, &iolog->io_logs);
1954                         iolog->cur_log_max = new_samples;
1955                         return cur_log;
1956                 }
1957                 sfree(cur_log);
1958         }
1959
1960         return NULL;
1961 }
1962
1963 /*
1964  * Add and return a new log chunk, or return current log if big enough
1965  */
1966 static struct io_logs *regrow_log(struct io_log *iolog)
1967 {
1968         struct io_logs *cur_log;
1969         int i;
1970
1971         if (!iolog || iolog->disabled)
1972                 goto disable;
1973
1974         cur_log = iolog_cur_log(iolog);
1975         if (!cur_log) {
1976                 cur_log = get_new_log(iolog);
1977                 if (!cur_log)
1978                         return NULL;
1979         }
1980
1981         if (cur_log->nr_samples < cur_log->max_samples)
1982                 return cur_log;
1983
1984         /*
1985          * No room for a new sample. If we're compressing on the fly, flush
1986          * out the current chunk
1987          */
1988         if (iolog->log_gz) {
1989                 if (iolog_cur_flush(iolog, cur_log)) {
1990                         log_err("fio: failed flushing iolog! Will stop logging.\n");
1991                         return NULL;
1992                 }
1993         }
1994
1995         /*
1996          * Get a new log array, and add to our list
1997          */
1998         cur_log = get_new_log(iolog);
1999         if (!cur_log) {
2000                 log_err("fio: failed extending iolog! Will stop logging.\n");
2001                 return NULL;
2002         }
2003
2004         if (!iolog->pending || !iolog->pending->nr_samples)
2005                 return cur_log;
2006
2007         /*
2008          * Flush pending items to new log
2009          */
2010         for (i = 0; i < iolog->pending->nr_samples; i++) {
2011                 struct io_sample *src, *dst;
2012
2013                 src = get_sample(iolog, iolog->pending, i);
2014                 dst = get_sample(iolog, cur_log, i);
2015                 memcpy(dst, src, log_entry_sz(iolog));
2016         }
2017         cur_log->nr_samples = iolog->pending->nr_samples;
2018
2019         iolog->pending->nr_samples = 0;
2020         return cur_log;
2021 disable:
2022         if (iolog)
2023                 iolog->disabled = true;
2024         return NULL;
2025 }
2026
2027 void regrow_logs(struct thread_data *td)
2028 {
2029         regrow_log(td->slat_log);
2030         regrow_log(td->clat_log);
2031         regrow_log(td->clat_hist_log);
2032         regrow_log(td->lat_log);
2033         regrow_log(td->bw_log);
2034         regrow_log(td->iops_log);
2035         td->flags &= ~TD_F_REGROW_LOGS;
2036 }
2037
2038 static struct io_logs *get_cur_log(struct io_log *iolog)
2039 {
2040         struct io_logs *cur_log;
2041
2042         cur_log = iolog_cur_log(iolog);
2043         if (!cur_log) {
2044                 cur_log = get_new_log(iolog);
2045                 if (!cur_log)
2046                         return NULL;
2047         }
2048
2049         if (cur_log->nr_samples < cur_log->max_samples)
2050                 return cur_log;
2051
2052         /*
2053          * Out of space. If we're in IO offload mode, or we're not doing
2054          * per unit logging (hence logging happens outside of the IO thread
2055          * as well), add a new log chunk inline. If we're doing inline
2056          * submissions, flag 'td' as needing a log regrow and we'll take
2057          * care of it on the submission side.
2058          */
2059         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2060             !per_unit_log(iolog))
2061                 return regrow_log(iolog);
2062
2063         iolog->td->flags |= TD_F_REGROW_LOGS;
2064         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2065         return iolog->pending;
2066 }
2067
2068 static void __add_log_sample(struct io_log *iolog, unsigned long val,
2069                              enum fio_ddir ddir, unsigned int bs,
2070                              unsigned long t, uint64_t offset)
2071 {
2072         struct io_logs *cur_log;
2073
2074         if (iolog->disabled)
2075                 return;
2076         if (flist_empty(&iolog->io_logs))
2077                 iolog->avg_last = t;
2078
2079         cur_log = get_cur_log(iolog);
2080         if (cur_log) {
2081                 struct io_sample *s;
2082
2083                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2084
2085                 s->val = val;
2086                 s->time = t;
2087                 io_sample_set_ddir(iolog, s, ddir);
2088                 s->bs = bs;
2089
2090                 if (iolog->log_offset) {
2091                         struct io_sample_offset *so = (void *) s;
2092
2093                         so->offset = offset;
2094                 }
2095
2096                 cur_log->nr_samples++;
2097                 return;
2098         }
2099
2100         iolog->disabled = true;
2101 }
2102
2103 static inline void reset_io_stat(struct io_stat *ios)
2104 {
2105         ios->max_val = ios->min_val = ios->samples = 0;
2106         ios->mean.u.f = ios->S.u.f = 0;
2107 }
2108
2109 void reset_io_stats(struct thread_data *td)
2110 {
2111         struct thread_stat *ts = &td->ts;
2112         int i, j;
2113
2114         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2115                 reset_io_stat(&ts->clat_stat[i]);
2116                 reset_io_stat(&ts->slat_stat[i]);
2117                 reset_io_stat(&ts->lat_stat[i]);
2118                 reset_io_stat(&ts->bw_stat[i]);
2119                 reset_io_stat(&ts->iops_stat[i]);
2120
2121                 ts->io_bytes[i] = 0;
2122                 ts->runtime[i] = 0;
2123
2124                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2125                         ts->io_u_plat[i][j] = 0;
2126         }
2127
2128         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2129                 ts->io_u_map[i] = 0;
2130                 ts->io_u_submit[i] = 0;
2131                 ts->io_u_complete[i] = 0;
2132                 ts->io_u_lat_u[i] = 0;
2133                 ts->io_u_lat_m[i] = 0;
2134                 ts->total_submit = 0;
2135                 ts->total_complete = 0;
2136         }
2137
2138         for (i = 0; i < 3; i++) {
2139                 ts->total_io_u[i] = 0;
2140                 ts->short_io_u[i] = 0;
2141                 ts->drop_io_u[i] = 0;
2142         }
2143 }
2144
2145 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2146                               unsigned long elapsed, bool log_max)
2147 {
2148         /*
2149          * Note an entry in the log. Use the mean from the logged samples,
2150          * making sure to properly round up. Only write a log entry if we
2151          * had actual samples done.
2152          */
2153         if (iolog->avg_window[ddir].samples) {
2154                 unsigned long val;
2155
2156                 if (log_max)
2157                         val = iolog->avg_window[ddir].max_val;
2158                 else
2159                         val = iolog->avg_window[ddir].mean.u.f + 0.50;
2160
2161                 __add_log_sample(iolog, val, ddir, 0, elapsed, 0);
2162         }
2163
2164         reset_io_stat(&iolog->avg_window[ddir]);
2165 }
2166
2167 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2168                              bool log_max)
2169 {
2170         int ddir;
2171
2172         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2173                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2174 }
2175
2176 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2177                            unsigned long val, enum fio_ddir ddir,
2178                            unsigned int bs, uint64_t offset)
2179 {
2180         unsigned long elapsed, this_window;
2181
2182         if (!ddir_rw(ddir))
2183                 return 0;
2184
2185         elapsed = mtime_since_now(&td->epoch);
2186
2187         /*
2188          * If no time averaging, just add the log sample.
2189          */
2190         if (!iolog->avg_msec) {
2191                 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
2192                 return 0;
2193         }
2194
2195         /*
2196          * Add the sample. If the time period has passed, then
2197          * add that entry to the log and clear.
2198          */
2199         add_stat_sample(&iolog->avg_window[ddir], val);
2200
2201         /*
2202          * If period hasn't passed, adding the above sample is all we
2203          * need to do.
2204          */
2205         this_window = elapsed - iolog->avg_last;
2206         if (elapsed < iolog->avg_last)
2207                 return iolog->avg_last - elapsed;
2208         else if (this_window < iolog->avg_msec) {
2209                 int diff = iolog->avg_msec - this_window;
2210
2211                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2212                         return diff;
2213         }
2214
2215         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2216
2217         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2218         return iolog->avg_msec;
2219 }
2220
2221 void finalize_logs(struct thread_data *td, bool unit_logs)
2222 {
2223         unsigned long elapsed;
2224
2225         elapsed = mtime_since_now(&td->epoch);
2226
2227         if (td->clat_log && unit_logs)
2228                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2229         if (td->slat_log && unit_logs)
2230                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2231         if (td->lat_log && unit_logs)
2232                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2233         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2234                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2235         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2236                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2237 }
2238
2239 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
2240 {
2241         struct io_log *iolog;
2242
2243         if (!ddir_rw(ddir))
2244                 return;
2245
2246         iolog = agg_io_log[ddir];
2247         __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
2248 }
2249
2250 static void add_clat_percentile_sample(struct thread_stat *ts,
2251                                 unsigned long usec, enum fio_ddir ddir)
2252 {
2253         unsigned int idx = plat_val_to_idx(usec);
2254         assert(idx < FIO_IO_U_PLAT_NR);
2255
2256         ts->io_u_plat[ddir][idx]++;
2257 }
2258
2259 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2260                      unsigned long usec, unsigned int bs, uint64_t offset)
2261 {
2262         unsigned long elapsed, this_window;
2263         struct thread_stat *ts = &td->ts;
2264         struct io_log *iolog = td->clat_hist_log;
2265
2266         td_io_u_lock(td);
2267
2268         add_stat_sample(&ts->clat_stat[ddir], usec);
2269
2270         if (td->clat_log)
2271                 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
2272
2273         if (ts->clat_percentiles)
2274                 add_clat_percentile_sample(ts, usec, ddir);
2275
2276         if (iolog && iolog->hist_msec) {
2277                 struct io_hist *hw = &iolog->hist_window[ddir];
2278
2279                 hw->samples++;
2280                 elapsed = mtime_since_now(&td->epoch);
2281                 if (!hw->hist_last)
2282                         hw->hist_last = elapsed;
2283                 this_window = elapsed - hw->hist_last;
2284
2285                 if (this_window >= iolog->hist_msec) {
2286                         unsigned int *io_u_plat;
2287                         unsigned int *dst;
2288
2289                         /*
2290                          * Make a byte-for-byte copy of the latency histogram
2291                          * stored in td->ts.io_u_plat[ddir], recording it in a
2292                          * log sample. Note that the matching call to free() is
2293                          * located in iolog.c after printing this sample to the
2294                          * log file.
2295                          */
2296                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2297                         dst = malloc(FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2298                         memcpy(dst, io_u_plat,
2299                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2300                         __add_log_sample(iolog, (unsigned long )dst, ddir, bs,
2301                                                 elapsed, offset);
2302
2303                         /*
2304                          * Update the last time we recorded as being now, minus
2305                          * any drift in time we encountered before actually
2306                          * making the record.
2307                          */
2308                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2309                         hw->samples = 0;
2310                 }
2311         }
2312
2313         td_io_u_unlock(td);
2314 }
2315
2316 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2317                      unsigned long usec, unsigned int bs, uint64_t offset)
2318 {
2319         struct thread_stat *ts = &td->ts;
2320
2321         if (!ddir_rw(ddir))
2322                 return;
2323
2324         td_io_u_lock(td);
2325
2326         add_stat_sample(&ts->slat_stat[ddir], usec);
2327
2328         if (td->slat_log)
2329                 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
2330
2331         td_io_u_unlock(td);
2332 }
2333
2334 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2335                     unsigned long usec, unsigned int bs, uint64_t offset)
2336 {
2337         struct thread_stat *ts = &td->ts;
2338
2339         if (!ddir_rw(ddir))
2340                 return;
2341
2342         td_io_u_lock(td);
2343
2344         add_stat_sample(&ts->lat_stat[ddir], usec);
2345
2346         if (td->lat_log)
2347                 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2348
2349         td_io_u_unlock(td);
2350 }
2351
2352 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2353                    unsigned int bytes, unsigned long spent)
2354 {
2355         struct thread_stat *ts = &td->ts;
2356         unsigned long rate;
2357
2358         if (spent)
2359                 rate = bytes * 1000 / spent;
2360         else
2361                 rate = 0;
2362
2363         td_io_u_lock(td);
2364
2365         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2366
2367         if (td->bw_log)
2368                 add_log_sample(td, td->bw_log, rate, io_u->ddir, bytes, io_u->offset);
2369
2370         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2371         td_io_u_unlock(td);
2372 }
2373
2374 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2375 {
2376         struct thread_stat *ts = &td->ts;
2377         unsigned long spent, rate;
2378         enum fio_ddir ddir;
2379         unsigned int next, next_log;
2380
2381         next_log = td->o.bw_avg_time;
2382
2383         spent = mtime_since(&td->bw_sample_time, t);
2384         if (spent < td->o.bw_avg_time &&
2385             td->o.bw_avg_time - spent >= LOG_MSEC_SLACK)
2386                 return td->o.bw_avg_time - spent;
2387
2388         td_io_u_lock(td);
2389
2390         /*
2391          * Compute both read and write rates for the interval.
2392          */
2393         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2394                 uint64_t delta;
2395
2396                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2397                 if (!delta)
2398                         continue; /* No entries for interval */
2399
2400                 if (spent)
2401                         rate = delta * 1000 / spent / 1024;
2402                 else
2403                         rate = 0;
2404
2405                 add_stat_sample(&ts->bw_stat[ddir], rate);
2406
2407                 if (td->bw_log) {
2408                         unsigned int bs = 0;
2409
2410                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2411                                 bs = td->o.min_bs[ddir];
2412
2413                         next = add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2414                         next_log = min(next_log, next);
2415                 }
2416
2417                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2418         }
2419
2420         timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time);
2421
2422         td_io_u_unlock(td);
2423
2424         if (spent <= td->o.bw_avg_time)
2425                 return min(next_log, td->o.bw_avg_time);
2426
2427         next = td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time);
2428         return min(next, next_log);
2429 }
2430
2431 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2432                      unsigned int bytes)
2433 {
2434         struct thread_stat *ts = &td->ts;
2435
2436         td_io_u_lock(td);
2437
2438         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2439
2440         if (td->iops_log)
2441                 add_log_sample(td, td->iops_log, 1, io_u->ddir, bytes, io_u->offset);
2442
2443         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2444         td_io_u_unlock(td);
2445 }
2446
2447 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2448 {
2449         struct thread_stat *ts = &td->ts;
2450         unsigned long spent, iops;
2451         enum fio_ddir ddir;
2452         unsigned int next, next_log;
2453
2454         next_log = td->o.iops_avg_time;
2455
2456         spent = mtime_since(&td->iops_sample_time, t);
2457         if (spent < td->o.iops_avg_time &&
2458             td->o.iops_avg_time - spent >= LOG_MSEC_SLACK)
2459                 return td->o.iops_avg_time - spent;
2460
2461         td_io_u_lock(td);
2462
2463         /*
2464          * Compute both read and write rates for the interval.
2465          */
2466         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2467                 uint64_t delta;
2468
2469                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2470                 if (!delta)
2471                         continue; /* No entries for interval */
2472
2473                 if (spent)
2474                         iops = (delta * 1000) / spent;
2475                 else
2476                         iops = 0;
2477
2478                 add_stat_sample(&ts->iops_stat[ddir], iops);
2479
2480                 if (td->iops_log) {
2481                         unsigned int bs = 0;
2482
2483                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2484                                 bs = td->o.min_bs[ddir];
2485
2486                         next = add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2487                         next_log = min(next_log, next);
2488                 }
2489
2490                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2491         }
2492
2493         timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time);
2494
2495         td_io_u_unlock(td);
2496
2497         if (spent <= td->o.iops_avg_time)
2498                 return min(next_log, td->o.iops_avg_time);
2499
2500         next = td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
2501         return min(next, next_log);
2502 }
2503
2504 /*
2505  * Returns msecs to next event
2506  */
2507 int calc_log_samples(void)
2508 {
2509         struct thread_data *td;
2510         unsigned int next = ~0U, tmp;
2511         struct timeval now;
2512         int i;
2513
2514         fio_gettime(&now, NULL);
2515
2516         for_each_td(td, i) {
2517                 if (in_ramp_time(td) ||
2518                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2519                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2520                         continue;
2521                 }
2522                 if (!per_unit_log(td->bw_log)) {
2523                         tmp = add_bw_samples(td, &now);
2524                         if (tmp < next)
2525                                 next = tmp;
2526                 }
2527                 if (!per_unit_log(td->iops_log)) {
2528                         tmp = add_iops_samples(td, &now);
2529                         if (tmp < next)
2530                                 next = tmp;
2531                 }
2532         }
2533
2534         return next == ~0U ? 0 : next;
2535 }
2536
2537 void stat_init(void)
2538 {
2539         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2540 }
2541
2542 void stat_exit(void)
2543 {
2544         /*
2545          * When we have the mutex, we know out-of-band access to it
2546          * have ended.
2547          */
2548         fio_mutex_down(stat_mutex);
2549         fio_mutex_remove(stat_mutex);
2550 }
2551
2552 /*
2553  * Called from signal handler. Wake up status thread.
2554  */
2555 void show_running_run_stats(void)
2556 {
2557         helper_do_stat();
2558 }
2559
2560 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2561 {
2562         /* Ignore io_u's which span multiple blocks--they will just get
2563          * inaccurate counts. */
2564         int idx = (io_u->offset - io_u->file->file_offset)
2565                         / td->o.bs[DDIR_TRIM];
2566         uint32_t *info = &td->ts.block_infos[idx];
2567         assert(idx < td->ts.nr_block_infos);
2568         return info;
2569 }