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