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_tv(&td->ru_start.ru_utime,
  41                                         &td->ru_end.ru_utime);
  42         ts->sys_time += mtime_since_tv(&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 long long 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_clzll(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 long long 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 long long **output,
 139                                    unsigned long long *maxv, unsigned long long *minv)
 140 {
 141         unsigned long sum = 0;
 142         unsigned int len, i, j = 0;
 143         unsigned int oval_len = 0;
 144         unsigned long long *ovals = NULL;
 145         int is_last;
 146
 147         *minv = -1ULL;
 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(*ovals));
 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 divisor, len, i, j = 0;
 205         unsigned long long minv, maxv;
 206         unsigned long long *ovals;
 207         int is_last, per_line, scale_down, time_width;
 208         char fmt[32];
 209
 210         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 211         if (!len)
 212                 goto out;
 213
 214         /*
 215          * We default to nsecs, but if the value range is such that we
 216          * should scale down to usecs or msecs, do that.
 217          */
 218         if (minv > 2000000 && maxv > 99999999ULL) {
 219                 scale_down = 2;
 220                 divisor = 1000000;
 221                 log_buf(out, "    clat percentiles (msec):\n     |");
 222         } else if (minv > 2000 && maxv > 99999) {
 223                 scale_down = 1;
 224                 divisor = 1000;
 225                 log_buf(out, "    clat percentiles (usec):\n     |");
 226         } else {
 227                 scale_down = 0;
 228                 divisor = 1;
 229                 log_buf(out, "    clat percentiles (nsec):\n     |");
 230         }
 231
 232
 233         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 234         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision+3, precision, time_width);
 235         // fmt will be something like " %5.2fth=[%4llu]%c"
 236         per_line = (80 - 7) / (precision + 10 + time_width);
 237
 238         for (j = 0; j < len; j++) {
 239                 /* for formatting */
 240                 if (j != 0 && (j % per_line) == 0)
 241                         log_buf(out, "     |");
 242
 243                 /* end of the list */
 244                 is_last = (j == len - 1);
 245
 246                 for (i = 0; i < scale_down; i++)
 247                         ovals[j] = (ovals[j] + 999) / 1000;
 248
 249                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 250
 251                 if (is_last)
 252                         break;
 253
 254                 if ((j % per_line) == per_line - 1)     /* for formatting */
 255                         log_buf(out, "\n");
 256         }
 257
 258 out:
 259         if (ovals)
 260                 free(ovals);
 261 }
 262
 263 bool calc_lat(struct io_stat *is, unsigned long long *min, unsigned long long *max,
 264               double *mean, double *dev)
 265 {
 266         double n = (double) is->samples;
 267
 268         if (n == 0)
 269                 return false;
 270
 271         *min = is->min_val;
 272         *max = is->max_val;
 273         *mean = is->mean.u.f;
 274
 275         if (n > 1.0)
 276                 *dev = sqrt(is->S.u.f / (n - 1.0));
 277         else
 278                 *dev = 0;
 279
 280         return true;
 281 }
 282
 283 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 284 {
 285         char *io, *agg, *min, *max;
 286         char *ioalt, *aggalt, *minalt, *maxalt;
 287         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 288         int i;
 289
 290         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 291
 292         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 293                 const int i2p = is_power_of_2(rs->kb_base);
 294
 295                 if (!rs->max_run[i])
 296                         continue;
 297
 298                 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
 299                 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
 300                 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
 301                 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
 302                 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
 303                 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
 304                 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
 305                 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
 306                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 307                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 308                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 309                                 (unsigned long long) rs->min_run[i],
 310                                 (unsigned long long) rs->max_run[i]);
 311
 312                 free(io);
 313                 free(agg);
 314                 free(min);
 315                 free(max);
 316                 free(ioalt);
 317                 free(aggalt);
 318                 free(minalt);
 319                 free(maxalt);
 320         }
 321 }
 322
 323 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 324 {
 325         int i;
 326
 327         /*
 328          * Do depth distribution calculations
 329          */
 330         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 331                 if (total) {
 332                         io_u_dist[i] = (double) map[i] / (double) total;
 333                         io_u_dist[i] *= 100.0;
 334                         if (io_u_dist[i] < 0.1 && map[i])
 335                                 io_u_dist[i] = 0.1;
 336                 } else
 337                         io_u_dist[i] = 0.0;
 338         }
 339 }
 340
 341 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 342                           unsigned int *src, int nr)
 343 {
 344         unsigned long total = ddir_rw_sum(ts->total_io_u);
 345         int i;
 346
 347         /*
 348          * Do latency distribution calculations
 349          */
 350         for (i = 0; i < nr; i++) {
 351                 if (total) {
 352                         dst[i] = (double) src[i] / (double) total;
 353                         dst[i] *= 100.0;
 354                         if (dst[i] < 0.01 && src[i])
 355                                 dst[i] = 0.01;
 356                 } else
 357                         dst[i] = 0.0;
 358         }
 359 }
 360
 361 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 362 {
 363         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 364 }
 365
 366 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 367 {
 368         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 369 }
 370
 371 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 372 {
 373         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 374 }
 375
 376 static void display_lat(const char *name, unsigned long long min, unsigned long long max,
 377                         double mean, double dev, struct buf_output *out)
 378 {
 379         const char *base = "(nsec)";
 380         char *minp, *maxp;
 381
 382         if (nsec_to_msec(&min, &max, &mean, &dev))
 383                 base = "(msec)";
 384         else if (nsec_to_usec(&min, &max, &mean, &dev))
 385                 base = "(usec)";
 386
 387         minp = num2str(min, 6, 1, 0, N2S_NONE);
 388         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 389
 390         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 391                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 392
 393         free(minp);
 394         free(maxp);
 395 }
 396
 397 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 398                              int ddir, struct buf_output *out)
 399 {
 400         const char *str[] = { " read", "write", " trim" };
 401         unsigned long runt;
 402         unsigned long long min, max, bw, iops;
 403         double mean, dev;
 404         char *io_p, *bw_p, *bw_p_alt, *iops_p;
 405         int i2p;
 406
 407         assert(ddir_rw(ddir));
 408
 409         if (!ts->runtime[ddir])
 410                 return;
 411
 412         i2p = is_power_of_2(rs->kb_base);
 413         runt = ts->runtime[ddir];
 414
 415         bw = (1000 * ts->io_bytes[ddir]) / runt;
 416         io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
 417         bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
 418         bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
 419
 420         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 421         iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
 422
 423         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
 424                         rs->unified_rw_rep ? "mixed" : str[ddir],
 425                         iops_p, bw_p, bw_p_alt, io_p,
 426                         (unsigned long long) ts->runtime[ddir]);
 427
 428         free(io_p);
 429         free(bw_p);
 430         free(bw_p_alt);
 431         free(iops_p);
 432
 433         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 434                 display_lat("slat", min, max, mean, dev, out);
 435         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 436                 display_lat("clat", min, max, mean, dev, out);
 437         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 438                 display_lat(" lat", min, max, mean, dev, out);
 439
 440         if (ts->clat_percentiles) {
 441                 show_clat_percentiles(ts->io_u_plat[ddir],
 442                                         ts->clat_stat[ddir].samples,
 443                                         ts->percentile_list,
 444                                         ts->percentile_precision, out);
 445         }
 446         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 447                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 448                 const char *bw_str;
 449
 450                 if ((rs->unit_base == 1) && i2p)
 451                         bw_str = "Kibit";
 452                 else if (rs->unit_base == 1)
 453                         bw_str = "kbit";
 454                 else if (i2p)
 455                         bw_str = "KiB";
 456                 else
 457                         bw_str = "kB";
 458
 459                 if (rs->unit_base == 1) {
 460                         min *= 8.0;
 461                         max *= 8.0;
 462                         mean *= 8.0;
 463                         dev *= 8.0;
 464                 }
 465
 466                 if (rs->agg[ddir]) {
 467                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 468                         if (p_of_agg > 100.0)
 469                                 p_of_agg = 100.0;
 470                 }
 471
 472                 if (mean > fkb_base * fkb_base) {
 473                         min /= fkb_base;
 474                         max /= fkb_base;
 475                         mean /= fkb_base;
 476                         dev /= fkb_base;
 477                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 478                 }
 479
 480                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, avg=%5.02f, stdev=%5.02f\n",
 481                         bw_str, min, max, p_of_agg, mean, dev);
 482         }
 483 }
 484
 485 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 486                     const char *msg, struct buf_output *out)
 487 {
 488         int new_line = 1, i, line = 0, shown = 0;
 489
 490         for (i = 0; i < nr; i++) {
 491                 if (io_u_lat[i] <= 0.0)
 492                         continue;
 493                 shown = 1;
 494                 if (new_line) {
 495                         if (line)
 496                                 log_buf(out, "\n");
 497                         log_buf(out, "    lat (%s) : ", msg);
 498                         new_line = 0;
 499                         line = 0;
 500                 }
 501                 if (line)
 502                         log_buf(out, ", ");
 503                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 504                 line++;
 505                 if (line == 5)
 506                         new_line = 1;
 507         }
 508
 509         if (shown)
 510                 log_buf(out, "\n");
 511
 512         return shown;
 513 }
 514
 515 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 516 {
 517         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 518                                  "250=", "500=", "750=", "1000=", };
 519
 520         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 521 }
 522
 523 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 524 {
 525         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 526                                  "250=", "500=", "750=", "1000=", };
 527
 528         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 529 }
 530
 531 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 532 {
 533         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 534                                  "250=", "500=", "750=", "1000=", "2000=",
 535                                  ">=2000=", };
 536
 537         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 538 }
 539
 540 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 541 {
 542         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 543         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 544         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 545
 546         stat_calc_lat_n(ts, io_u_lat_n);
 547         stat_calc_lat_u(ts, io_u_lat_u);
 548         stat_calc_lat_m(ts, io_u_lat_m);
 549
 550         show_lat_n(io_u_lat_n, out);
 551         show_lat_u(io_u_lat_u, out);
 552         show_lat_m(io_u_lat_m, out);
 553 }
 554
 555 static int block_state_category(int block_state)
 556 {
 557         switch (block_state) {
 558         case BLOCK_STATE_UNINIT:
 559                 return 0;
 560         case BLOCK_STATE_TRIMMED:
 561         case BLOCK_STATE_WRITTEN:
 562                 return 1;
 563         case BLOCK_STATE_WRITE_FAILURE:
 564         case BLOCK_STATE_TRIM_FAILURE:
 565                 return 2;
 566         default:
 567                 /* Silence compile warning on some BSDs and have a return */
 568                 assert(0);
 569                 return -1;
 570         }
 571 }
 572
 573 static int compare_block_infos(const void *bs1, const void *bs2)
 574 {
 575         uint32_t block1 = *(uint32_t *)bs1;
 576         uint32_t block2 = *(uint32_t *)bs2;
 577         int state1 = BLOCK_INFO_STATE(block1);
 578         int state2 = BLOCK_INFO_STATE(block2);
 579         int bscat1 = block_state_category(state1);
 580         int bscat2 = block_state_category(state2);
 581         int cycles1 = BLOCK_INFO_TRIMS(block1);
 582         int cycles2 = BLOCK_INFO_TRIMS(block2);
 583
 584         if (bscat1 < bscat2)
 585                 return -1;
 586         if (bscat1 > bscat2)
 587                 return 1;
 588
 589         if (cycles1 < cycles2)
 590                 return -1;
 591         if (cycles1 > cycles2)
 592                 return 1;
 593
 594         if (state1 < state2)
 595                 return -1;
 596         if (state1 > state2)
 597                 return 1;
 598
 599         assert(block1 == block2);
 600         return 0;
 601 }
 602
 603 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 604                                   fio_fp64_t *plist, unsigned int **percentiles,
 605                                   unsigned int *types)
 606 {
 607         int len = 0;
 608         int i, nr_uninit;
 609
 610         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 611
 612         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 613                 len++;
 614
 615         if (!len)
 616                 return 0;
 617
 618         /*
 619          * Sort the percentile list. Note that it may already be sorted if
 620          * we are using the default values, but since it's a short list this
 621          * isn't a worry. Also note that this does not work for NaN values.
 622          */
 623         if (len > 1)
 624                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 625
 626         nr_uninit = 0;
 627         /* Start only after the uninit entries end */
 628         for (nr_uninit = 0;
 629              nr_uninit < nr_block_infos
 630                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 631              nr_uninit ++)
 632                 ;
 633
 634         if (nr_uninit == nr_block_infos)
 635                 return 0;
 636
 637         *percentiles = calloc(len, sizeof(**percentiles));
 638
 639         for (i = 0; i < len; i++) {
 640                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 641                                 + nr_uninit;
 642                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 643         }
 644
 645         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 646         for (i = 0; i < nr_block_infos; i++)
 647                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 648
 649         return len;
 650 }
 651
 652 static const char *block_state_names[] = {
 653         [BLOCK_STATE_UNINIT] = "unwritten",
 654         [BLOCK_STATE_TRIMMED] = "trimmed",
 655         [BLOCK_STATE_WRITTEN] = "written",
 656         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 657         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 658 };
 659
 660 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 661                              fio_fp64_t *plist, struct buf_output *out)
 662 {
 663         int len, pos, i;
 664         unsigned int *percentiles = NULL;
 665         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 666
 667         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 668                                      &percentiles, block_state_counts);
 669
 670         log_buf(out, "  block lifetime percentiles :\n   |");
 671         pos = 0;
 672         for (i = 0; i < len; i++) {
 673                 uint32_t block_info = percentiles[i];
 674 #define LINE_LENGTH     75
 675                 char str[LINE_LENGTH];
 676                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 677                                      plist[i].u.f, block_info,
 678                                      i == len - 1 ? '\n' : ',');
 679                 assert(strln < LINE_LENGTH);
 680                 if (pos + strln > LINE_LENGTH) {
 681                         pos = 0;
 682                         log_buf(out, "\n   |");
 683                 }
 684                 log_buf(out, "%s", str);
 685                 pos += strln;
 686 #undef LINE_LENGTH
 687         }
 688         if (percentiles)
 689                 free(percentiles);
 690
 691         log_buf(out, "        states               :");
 692         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 693                 log_buf(out, " %s=%u%c",
 694                          block_state_names[i], block_state_counts[i],
 695                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 696 }
 697
 698 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 699 {
 700         char *p1, *p1alt, *p2;
 701         unsigned long long bw_mean, iops_mean;
 702         const int i2p = is_power_of_2(ts->kb_base);
 703
 704         if (!ts->ss_dur)
 705                 return;
 706
 707         bw_mean = steadystate_bw_mean(ts);
 708         iops_mean = steadystate_iops_mean(ts);
 709
 710         p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
 711         p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
 712         p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
 713
 714         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 715                 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
 716                 p1, p1alt, p2,
 717                 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
 718                 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
 719                 ts->ss_criterion.u.f,
 720                 ts->ss_state & __FIO_SS_PCT ? "%" : "");
 721
 722         free(p1);
 723         free(p1alt);
 724         free(p2);
 725 }
 726
 727 static void show_thread_status_normal(struct thread_stat *ts,
 728                                       struct group_run_stats *rs,
 729                                       struct buf_output *out)
 730 {
 731         double usr_cpu, sys_cpu;
 732         unsigned long runtime;
 733         double io_u_dist[FIO_IO_U_MAP_NR];
 734         time_t time_p;
 735         char time_buf[32];
 736
 737         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 738                 return;
 739
 740         memset(time_buf, 0, sizeof(time_buf));
 741
 742         time(&time_p);
 743         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 744
 745         if (!ts->error) {
 746                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 747                                         ts->name, ts->groupid, ts->members,
 748                                         ts->error, (int) ts->pid, time_buf);
 749         } else {
 750                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 751                                         ts->name, ts->groupid, ts->members,
 752                                         ts->error, ts->verror, (int) ts->pid,
 753                                         time_buf);
 754         }
 755
 756         if (strlen(ts->description))
 757                 log_buf(out, "  Description  : [%s]\n", ts->description);
 758
 759         if (ts->io_bytes[DDIR_READ])
 760                 show_ddir_status(rs, ts, DDIR_READ, out);
 761         if (ts->io_bytes[DDIR_WRITE])
 762                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 763         if (ts->io_bytes[DDIR_TRIM])
 764                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 765
 766         show_latencies(ts, out);
 767
 768         runtime = ts->total_run_time;
 769         if (runtime) {
 770                 double runt = (double) runtime;
 771
 772                 usr_cpu = (double) ts->usr_time * 100 / runt;
 773                 sys_cpu = (double) ts->sys_time * 100 / runt;
 774         } else {
 775                 usr_cpu = 0;
 776                 sys_cpu = 0;
 777         }
 778
 779         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 780                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 781                         (unsigned long long) ts->ctx,
 782                         (unsigned long long) ts->majf,
 783                         (unsigned long long) ts->minf);
 784
 785         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 786         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 787                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 788                                         io_u_dist[1], io_u_dist[2],
 789                                         io_u_dist[3], io_u_dist[4],
 790                                         io_u_dist[5], io_u_dist[6]);
 791
 792         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 793         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 794                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 795                                         io_u_dist[1], io_u_dist[2],
 796                                         io_u_dist[3], io_u_dist[4],
 797                                         io_u_dist[5], io_u_dist[6]);
 798         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 799         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 800                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 801                                         io_u_dist[1], io_u_dist[2],
 802                                         io_u_dist[3], io_u_dist[4],
 803                                         io_u_dist[5], io_u_dist[6]);
 804         log_buf(out, "     issued rwt: total=%llu,%llu,%llu,"
 805                                  " short=%llu,%llu,%llu,"
 806                                  " dropped=%llu,%llu,%llu\n",
 807                                         (unsigned long long) ts->total_io_u[0],
 808                                         (unsigned long long) ts->total_io_u[1],
 809                                         (unsigned long long) ts->total_io_u[2],
 810                                         (unsigned long long) ts->short_io_u[0],
 811                                         (unsigned long long) ts->short_io_u[1],
 812                                         (unsigned long long) ts->short_io_u[2],
 813                                         (unsigned long long) ts->drop_io_u[0],
 814                                         (unsigned long long) ts->drop_io_u[1],
 815                                         (unsigned long long) ts->drop_io_u[2]);
 816         if (ts->continue_on_error) {
 817                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 818                                         (unsigned long long)ts->total_err_count,
 819                                         ts->first_error,
 820                                         strerror(ts->first_error));
 821         }
 822         if (ts->latency_depth) {
 823                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 824                                         (unsigned long long)ts->latency_target,
 825                                         (unsigned long long)ts->latency_window,
 826                                         ts->latency_percentile.u.f,
 827                                         ts->latency_depth);
 828         }
 829
 830         if (ts->nr_block_infos)
 831                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 832                                   ts->percentile_list, out);
 833
 834         if (ts->ss_dur)
 835                 show_ss_normal(ts, out);
 836 }
 837
 838 static void show_ddir_status_terse(struct thread_stat *ts,
 839                                    struct group_run_stats *rs, int ddir,
 840                                    struct buf_output *out)
 841 {
 842         unsigned long long min, max, minv, maxv, bw, iops;
 843         unsigned long long *ovals = NULL;
 844         double mean, dev;
 845         unsigned int len;
 846         int i;
 847
 848         assert(ddir_rw(ddir));
 849
 850         iops = bw = 0;
 851         if (ts->runtime[ddir]) {
 852                 uint64_t runt = ts->runtime[ddir];
 853
 854                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 855                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 856         }
 857
 858         log_buf(out, ";%llu;%llu;%llu;%llu",
 859                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 860                                         (unsigned long long) ts->runtime[ddir]);
 861
 862         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 863                 log_buf(out, ";%llu;%llu;%f;%f", min, max, mean, dev);
 864         else
 865                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 866
 867         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 868                 log_buf(out, ";%llu;%llu;%f;%f", min, max, mean, dev);
 869         else
 870                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 871
 872         if (ts->clat_percentiles) {
 873                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 874                                         ts->clat_stat[ddir].samples,
 875                                         ts->percentile_list, &ovals, &maxv,
 876                                         &minv);
 877         } else
 878                 len = 0;
 879
 880         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 881                 if (i >= len) {
 882                         log_buf(out, ";0%%=0");
 883                         continue;
 884                 }
 885                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]);
 886         }
 887
 888         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 889                 log_buf(out, ";%llu;%llu;%f;%f", min, max, mean, dev);
 890         else
 891                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 892
 893         if (ovals)
 894                 free(ovals);
 895
 896         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 897                 double p_of_agg = 100.0;
 898
 899                 if (rs->agg[ddir]) {
 900                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 901                         if (p_of_agg > 100.0)
 902                                 p_of_agg = 100.0;
 903                 }
 904
 905                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 906         } else
 907                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
 908 }
 909
 910 static void add_ddir_status_json(struct thread_stat *ts,
 911                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 912 {
 913         unsigned long long min, max, minv, maxv;
 914         unsigned long long bw;
 915         unsigned long long *ovals = NULL;
 916         double mean, dev, iops;
 917         unsigned int len;
 918         int i;
 919         const char *ddirname[] = {"read", "write", "trim"};
 920         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 921         char buf[120];
 922         double p_of_agg = 100.0;
 923
 924         assert(ddir_rw(ddir));
 925
 926         if (ts->unified_rw_rep && ddir != DDIR_READ)
 927                 return;
 928
 929         dir_object = json_create_object();
 930         json_object_add_value_object(parent,
 931                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 932
 933         bw = 0;
 934         iops = 0.0;
 935         if (ts->runtime[ddir]) {
 936                 uint64_t runt = ts->runtime[ddir];
 937
 938                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 939                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 940         }
 941
 942         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
 943         json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
 944         json_object_add_value_int(dir_object, "bw", bw);
 945         json_object_add_value_float(dir_object, "iops", iops);
 946         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 947         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 948         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 949         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 950
 951         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 952                 min = max = 0;
 953                 mean = dev = 0.0;
 954         }
 955         tmp_object = json_create_object();
 956         json_object_add_value_object(dir_object, "slat_ns", tmp_object);
 957         json_object_add_value_int(tmp_object, "min", min);
 958         json_object_add_value_int(tmp_object, "max", max);
 959         json_object_add_value_float(tmp_object, "mean", mean);
 960         json_object_add_value_float(tmp_object, "stddev", dev);
 961
 962         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 963                 min = max = 0;
 964                 mean = dev = 0.0;
 965         }
 966         tmp_object = json_create_object();
 967         json_object_add_value_object(dir_object, "clat_ns", tmp_object);
 968         json_object_add_value_int(tmp_object, "min", min);
 969         json_object_add_value_int(tmp_object, "max", max);
 970         json_object_add_value_float(tmp_object, "mean", mean);
 971         json_object_add_value_float(tmp_object, "stddev", dev);
 972
 973         if (ts->clat_percentiles) {
 974                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 975                                         ts->clat_stat[ddir].samples,
 976                                         ts->percentile_list, &ovals, &maxv,
 977                                         &minv);
 978         } else
 979                 len = 0;
 980
 981         percentile_object = json_create_object();
 982         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 983         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 984                 if (i >= len) {
 985                         json_object_add_value_int(percentile_object, "0.00", 0);
 986                         continue;
 987                 }
 988                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 989                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 990         }
 991
 992         if (output_format & FIO_OUTPUT_JSON_PLUS) {
 993                 clat_bins_object = json_create_object();
 994                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
 995                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
 996                         if (ts->io_u_plat[ddir][i]) {
 997                                 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
 998                                 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
 999                         }
1000                 }
1001         }
1002
1003         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1004                 min = max = 0;
1005                 mean = dev = 0.0;
1006         }
1007         tmp_object = json_create_object();
1008         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1009         json_object_add_value_int(tmp_object, "min", min);
1010         json_object_add_value_int(tmp_object, "max", max);
1011         json_object_add_value_float(tmp_object, "mean", mean);
1012         json_object_add_value_float(tmp_object, "stddev", dev);
1013         if (ovals)
1014                 free(ovals);
1015
1016         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1017                 if (rs->agg[ddir]) {
1018                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
1019                         if (p_of_agg > 100.0)
1020                                 p_of_agg = 100.0;
1021                 }
1022         } else {
1023                 min = max = 0;
1024                 p_of_agg = mean = dev = 0.0;
1025         }
1026         json_object_add_value_int(dir_object, "bw_min", min);
1027         json_object_add_value_int(dir_object, "bw_max", max);
1028         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1029         json_object_add_value_float(dir_object, "bw_mean", mean);
1030         json_object_add_value_float(dir_object, "bw_dev", dev);
1031 }
1032
1033 static void show_thread_status_terse_v2(struct thread_stat *ts,
1034                                         struct group_run_stats *rs,
1035                                         struct buf_output *out)
1036 {
1037         double io_u_dist[FIO_IO_U_MAP_NR];
1038         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1039         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1040         double usr_cpu, sys_cpu;
1041         int i;
1042
1043         /* General Info */
1044         log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1045         /* Log Read Status */
1046         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1047         /* Log Write Status */
1048         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1049         /* Log Trim Status */
1050         show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1051
1052         /* CPU Usage */
1053         if (ts->total_run_time) {
1054                 double runt = (double) ts->total_run_time;
1055
1056                 usr_cpu = (double) ts->usr_time * 100 / runt;
1057                 sys_cpu = (double) ts->sys_time * 100 / runt;
1058         } else {
1059                 usr_cpu = 0;
1060                 sys_cpu = 0;
1061         }
1062
1063         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1064                                                 (unsigned long long) ts->ctx,
1065                                                 (unsigned long long) ts->majf,
1066                                                 (unsigned long long) ts->minf);
1067
1068         /* Calc % distribution of IO depths, usecond, msecond latency */
1069         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1070         stat_calc_lat_u(ts, io_u_lat_u);
1071         stat_calc_lat_m(ts, io_u_lat_m);
1072
1073         /* Only show fixed 7 I/O depth levels*/
1074         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1075                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1076                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1077
1078         /* Microsecond latency */
1079         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1080                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1081         /* Millisecond latency */
1082         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1083                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1084         /* Additional output if continue_on_error set - default off*/
1085         if (ts->continue_on_error)
1086                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1087         log_buf(out, "\n");
1088
1089         /* Additional output if description is set */
1090         if (strlen(ts->description))
1091                 log_buf(out, ";%s", ts->description);
1092
1093         log_buf(out, "\n");
1094 }
1095
1096 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1097                                            struct group_run_stats *rs, int ver,
1098                                            struct buf_output *out)
1099 {
1100         double io_u_dist[FIO_IO_U_MAP_NR];
1101         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1102         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1103         double usr_cpu, sys_cpu;
1104         int i;
1105
1106         /* General Info */
1107         log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1108                                         ts->name, ts->groupid, ts->error);
1109         /* Log Read Status */
1110         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1111         /* Log Write Status */
1112         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1113         /* Log Trim Status */
1114         if (ver == 4)
1115                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
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
1128         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1129                                                 (unsigned long long) ts->ctx,
1130                                                 (unsigned long long) ts->majf,
1131                                                 (unsigned long long) ts->minf);
1132
1133         /* Calc % distribution of IO depths, usecond, msecond latency */
1134         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1135         stat_calc_lat_u(ts, io_u_lat_u);
1136         stat_calc_lat_m(ts, io_u_lat_m);
1137
1138         /* Only show fixed 7 I/O depth levels*/
1139         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1140                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1141                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1142
1143         /* Microsecond latency */
1144         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1145                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1146         /* Millisecond latency */
1147         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1148                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1149
1150         /* disk util stats, if any */
1151         show_disk_util(1, NULL, out);
1152
1153         /* Additional output if continue_on_error set - default off*/
1154         if (ts->continue_on_error)
1155                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1156
1157         /* Additional output if description is set */
1158         if (strlen(ts->description))
1159                 log_buf(out, ";%s", ts->description);
1160
1161         log_buf(out, "\n");
1162 }
1163
1164 static void json_add_job_opts(struct json_object *root, const char *name,
1165                               struct flist_head *opt_list, bool num_jobs)
1166 {
1167         struct json_object *dir_object;
1168         struct flist_head *entry;
1169         struct print_option *p;
1170
1171         if (flist_empty(opt_list))
1172                 return;
1173
1174         dir_object = json_create_object();
1175         json_object_add_value_object(root, name, dir_object);
1176
1177         flist_for_each(entry, opt_list) {
1178                 const char *pos = "";
1179
1180                 p = flist_entry(entry, struct print_option, list);
1181                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1182                         continue;
1183                 if (p->value)
1184                         pos = p->value;
1185                 json_object_add_value_string(dir_object, p->name, pos);
1186         }
1187 }
1188
1189 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1190                                                    struct group_run_stats *rs,
1191                                                    struct flist_head *opt_list)
1192 {
1193         struct json_object *root, *tmp;
1194         struct jobs_eta *je;
1195         double io_u_dist[FIO_IO_U_MAP_NR];
1196         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1197         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1198         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1199         double usr_cpu, sys_cpu;
1200         int i;
1201         size_t size;
1202
1203         root = json_create_object();
1204         json_object_add_value_string(root, "jobname", ts->name);
1205         json_object_add_value_int(root, "groupid", ts->groupid);
1206         json_object_add_value_int(root, "error", ts->error);
1207
1208         /* ETA Info */
1209         je = get_jobs_eta(true, &size);
1210         if (je) {
1211                 json_object_add_value_int(root, "eta", je->eta_sec);
1212                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1213         }
1214
1215         if (opt_list)
1216                 json_add_job_opts(root, "job options", opt_list, true);
1217
1218         add_ddir_status_json(ts, rs, DDIR_READ, root);
1219         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1220         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1221
1222         /* CPU Usage */
1223         if (ts->total_run_time) {
1224                 double runt = (double) ts->total_run_time;
1225
1226                 usr_cpu = (double) ts->usr_time * 100 / runt;
1227                 sys_cpu = (double) ts->sys_time * 100 / runt;
1228         } else {
1229                 usr_cpu = 0;
1230                 sys_cpu = 0;
1231         }
1232         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1233         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1234         json_object_add_value_int(root, "ctx", ts->ctx);
1235         json_object_add_value_int(root, "majf", ts->majf);
1236         json_object_add_value_int(root, "minf", ts->minf);
1237
1238
1239         /* Calc % distribution of IO depths, usecond, msecond latency */
1240         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1241         stat_calc_lat_n(ts, io_u_lat_n);
1242         stat_calc_lat_u(ts, io_u_lat_u);
1243         stat_calc_lat_m(ts, io_u_lat_m);
1244
1245         tmp = json_create_object();
1246         json_object_add_value_object(root, "iodepth_level", tmp);
1247         /* Only show fixed 7 I/O depth levels*/
1248         for (i = 0; i < 7; i++) {
1249                 char name[20];
1250                 if (i < 6)
1251                         snprintf(name, 20, "%d", 1 << i);
1252                 else
1253                         snprintf(name, 20, ">=%d", 1 << i);
1254                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1255         }
1256
1257         /* Nanosecond latency */
1258         tmp = json_create_object();
1259         json_object_add_value_object(root, "latency_ns", tmp);
1260         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1261                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1262                                  "250", "500", "750", "1000", };
1263                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1264         }
1265         /* Microsecond latency */
1266         tmp = json_create_object();
1267         json_object_add_value_object(root, "latency_us", tmp);
1268         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1269                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1270                                  "250", "500", "750", "1000", };
1271                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1272         }
1273         /* Millisecond latency */
1274         tmp = json_create_object();
1275         json_object_add_value_object(root, "latency_ms", tmp);
1276         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1277                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1278                                  "250", "500", "750", "1000", "2000",
1279                                  ">=2000", };
1280                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1281         }
1282
1283         /* Additional output if continue_on_error set - default off*/
1284         if (ts->continue_on_error) {
1285                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1286                 json_object_add_value_int(root, "first_error", ts->first_error);
1287         }
1288
1289         if (ts->latency_depth) {
1290                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1291                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1292                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1293                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1294         }
1295
1296         /* Additional output if description is set */
1297         if (strlen(ts->description))
1298                 json_object_add_value_string(root, "desc", ts->description);
1299
1300         if (ts->nr_block_infos) {
1301                 /* Block error histogram and types */
1302                 int len;
1303                 unsigned int *percentiles = NULL;
1304                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1305
1306                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1307                                              ts->percentile_list,
1308                                              &percentiles, block_state_counts);
1309
1310                 if (len) {
1311                         struct json_object *block, *percentile_object, *states;
1312                         int state;
1313                         block = json_create_object();
1314                         json_object_add_value_object(root, "block", block);
1315
1316                         percentile_object = json_create_object();
1317                         json_object_add_value_object(block, "percentiles",
1318                                                      percentile_object);
1319                         for (i = 0; i < len; i++) {
1320                                 char buf[20];
1321                                 snprintf(buf, sizeof(buf), "%f",
1322                                          ts->percentile_list[i].u.f);
1323                                 json_object_add_value_int(percentile_object,
1324                                                           (const char *)buf,
1325                                                           percentiles[i]);
1326                         }
1327
1328                         states = json_create_object();
1329                         json_object_add_value_object(block, "states", states);
1330                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1331                                 json_object_add_value_int(states,
1332                                         block_state_names[state],
1333                                         block_state_counts[state]);
1334                         }
1335                         free(percentiles);
1336                 }
1337         }
1338
1339         if (ts->ss_dur) {
1340                 struct json_object *data;
1341                 struct json_array *iops, *bw;
1342                 int i, j, k;
1343                 char ss_buf[64];
1344
1345                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1346                         ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1347                         ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1348                         (float) ts->ss_limit.u.f,
1349                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1350
1351                 tmp = json_create_object();
1352                 json_object_add_value_object(root, "steadystate", tmp);
1353                 json_object_add_value_string(tmp, "ss", ss_buf);
1354                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1355                 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1356
1357                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1358                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1359                 json_object_add_value_string(tmp, "criterion", ss_buf);
1360                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1361                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1362
1363                 data = json_create_object();
1364                 json_object_add_value_object(tmp, "data", data);
1365                 bw = json_create_array();
1366                 iops = json_create_array();
1367
1368                 /*
1369                 ** if ss was attained or the buffer is not full,
1370                 ** ss->head points to the first element in the list.
1371                 ** otherwise it actually points to the second element
1372                 ** in the list
1373                 */
1374                 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1375                         j = ts->ss_head;
1376                 else
1377                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1378                 for (i = 0; i < ts->ss_dur; i++) {
1379                         k = (j + i) % ts->ss_dur;
1380                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1381                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1382                 }
1383                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1384                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1385                 json_object_add_value_array(data, "iops", iops);
1386                 json_object_add_value_array(data, "bw", bw);
1387         }
1388
1389         return root;
1390 }
1391
1392 static void show_thread_status_terse(struct thread_stat *ts,
1393                                      struct group_run_stats *rs,
1394                                      struct buf_output *out)
1395 {
1396         if (terse_version == 2)
1397                 show_thread_status_terse_v2(ts, rs, out);
1398         else if (terse_version == 3 || terse_version == 4)
1399                 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1400         else
1401                 log_err("fio: bad terse version!? %d\n", terse_version);
1402 }
1403
1404 struct json_object *show_thread_status(struct thread_stat *ts,
1405                                        struct group_run_stats *rs,
1406                                        struct flist_head *opt_list,
1407                                        struct buf_output *out)
1408 {
1409         struct json_object *ret = NULL;
1410
1411         if (output_format & FIO_OUTPUT_TERSE)
1412                 show_thread_status_terse(ts, rs,  out);
1413         if (output_format & FIO_OUTPUT_JSON)
1414                 ret = show_thread_status_json(ts, rs, opt_list);
1415         if (output_format & FIO_OUTPUT_NORMAL)
1416                 show_thread_status_normal(ts, rs,  out);
1417
1418         return ret;
1419 }
1420
1421 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1422 {
1423         double mean, S;
1424
1425         if (src->samples == 0)
1426                 return;
1427
1428         dst->min_val = min(dst->min_val, src->min_val);
1429         dst->max_val = max(dst->max_val, src->max_val);
1430
1431         /*
1432          * Compute new mean and S after the merge
1433          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1434          *  #Parallel_algorithm>
1435          */
1436         if (first) {
1437                 mean = src->mean.u.f;
1438                 S = src->S.u.f;
1439         } else {
1440                 double delta = src->mean.u.f - dst->mean.u.f;
1441
1442                 mean = ((src->mean.u.f * src->samples) +
1443                         (dst->mean.u.f * dst->samples)) /
1444                         (dst->samples + src->samples);
1445
1446                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1447                         (dst->samples * src->samples) /
1448                         (dst->samples + src->samples);
1449         }
1450
1451         dst->samples += src->samples;
1452         dst->mean.u.f = mean;
1453         dst->S.u.f = S;
1454 }
1455
1456 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1457 {
1458         int i;
1459
1460         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1461                 if (dst->max_run[i] < src->max_run[i])
1462                         dst->max_run[i] = src->max_run[i];
1463                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1464                         dst->min_run[i] = src->min_run[i];
1465                 if (dst->max_bw[i] < src->max_bw[i])
1466                         dst->max_bw[i] = src->max_bw[i];
1467                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1468                         dst->min_bw[i] = src->min_bw[i];
1469
1470                 dst->iobytes[i] += src->iobytes[i];
1471                 dst->agg[i] += src->agg[i];
1472         }
1473
1474         if (!dst->kb_base)
1475                 dst->kb_base = src->kb_base;
1476         if (!dst->unit_base)
1477                 dst->unit_base = src->unit_base;
1478 }
1479
1480 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1481                       bool first)
1482 {
1483         int l, k;
1484
1485         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1486                 if (!dst->unified_rw_rep) {
1487                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1488                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1489                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1490                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1491
1492                         dst->io_bytes[l] += src->io_bytes[l];
1493
1494                         if (dst->runtime[l] < src->runtime[l])
1495                                 dst->runtime[l] = src->runtime[l];
1496                 } else {
1497                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1498                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1499                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1500                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1501
1502                         dst->io_bytes[0] += src->io_bytes[l];
1503
1504                         if (dst->runtime[0] < src->runtime[l])
1505                                 dst->runtime[0] = src->runtime[l];
1506
1507                         /*
1508                          * We're summing to the same destination, so override
1509                          * 'first' after the first iteration of the loop
1510                          */
1511                         first = false;
1512                 }
1513         }
1514
1515         dst->usr_time += src->usr_time;
1516         dst->sys_time += src->sys_time;
1517         dst->ctx += src->ctx;
1518         dst->majf += src->majf;
1519         dst->minf += src->minf;
1520
1521         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1522                 dst->io_u_map[k] += src->io_u_map[k];
1523         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1524                 dst->io_u_submit[k] += src->io_u_submit[k];
1525         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1526                 dst->io_u_complete[k] += src->io_u_complete[k];
1527         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1528                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1529         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1530                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1531         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1532                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1533
1534         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1535                 if (!dst->unified_rw_rep) {
1536                         dst->total_io_u[k] += src->total_io_u[k];
1537                         dst->short_io_u[k] += src->short_io_u[k];
1538                         dst->drop_io_u[k] += src->drop_io_u[k];
1539                 } else {
1540                         dst->total_io_u[0] += src->total_io_u[k];
1541                         dst->short_io_u[0] += src->short_io_u[k];
1542                         dst->drop_io_u[0] += src->drop_io_u[k];
1543                 }
1544         }
1545
1546         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1547                 int m;
1548
1549                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1550                         if (!dst->unified_rw_rep)
1551                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1552                         else
1553                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1554                 }
1555         }
1556
1557         dst->total_run_time += src->total_run_time;
1558         dst->total_submit += src->total_submit;
1559         dst->total_complete += src->total_complete;
1560 }
1561
1562 void init_group_run_stat(struct group_run_stats *gs)
1563 {
1564         int i;
1565         memset(gs, 0, sizeof(*gs));
1566
1567         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1568                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1569 }
1570
1571 void init_thread_stat(struct thread_stat *ts)
1572 {
1573         int j;
1574
1575         memset(ts, 0, sizeof(*ts));
1576
1577         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1578                 ts->lat_stat[j].min_val = -1UL;
1579                 ts->clat_stat[j].min_val = -1UL;
1580                 ts->slat_stat[j].min_val = -1UL;
1581                 ts->bw_stat[j].min_val = -1UL;
1582         }
1583         ts->groupid = -1;
1584 }
1585
1586 void __show_run_stats(void)
1587 {
1588         struct group_run_stats *runstats, *rs;
1589         struct thread_data *td;
1590         struct thread_stat *threadstats, *ts;
1591         int i, j, k, nr_ts, last_ts, idx;
1592         int kb_base_warned = 0;
1593         int unit_base_warned = 0;
1594         struct json_object *root = NULL;
1595         struct json_array *array = NULL;
1596         struct buf_output output[FIO_OUTPUT_NR];
1597         struct flist_head **opt_lists;
1598
1599         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1600
1601         for (i = 0; i < groupid + 1; i++)
1602                 init_group_run_stat(&runstats[i]);
1603
1604         /*
1605          * find out how many threads stats we need. if group reporting isn't
1606          * enabled, it's one-per-td.
1607          */
1608         nr_ts = 0;
1609         last_ts = -1;
1610         for_each_td(td, i) {
1611                 if (!td->o.group_reporting) {
1612                         nr_ts++;
1613                         continue;
1614                 }
1615                 if (last_ts == td->groupid)
1616                         continue;
1617                 if (!td->o.stats)
1618                         continue;
1619
1620                 last_ts = td->groupid;
1621                 nr_ts++;
1622         }
1623
1624         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1625         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1626
1627         for (i = 0; i < nr_ts; i++) {
1628                 init_thread_stat(&threadstats[i]);
1629                 opt_lists[i] = NULL;
1630         }
1631
1632         j = 0;
1633         last_ts = -1;
1634         idx = 0;
1635         for_each_td(td, i) {
1636                 if (!td->o.stats)
1637                         continue;
1638                 if (idx && (!td->o.group_reporting ||
1639                     (td->o.group_reporting && last_ts != td->groupid))) {
1640                         idx = 0;
1641                         j++;
1642                 }
1643
1644                 last_ts = td->groupid;
1645
1646                 ts = &threadstats[j];
1647
1648                 ts->clat_percentiles = td->o.clat_percentiles;
1649                 ts->percentile_precision = td->o.percentile_precision;
1650                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1651                 opt_lists[j] = &td->opt_list;
1652
1653                 idx++;
1654                 ts->members++;
1655
1656                 if (ts->groupid == -1) {
1657                         /*
1658                          * These are per-group shared already
1659                          */
1660                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1661                         if (td->o.description)
1662                                 strncpy(ts->description, td->o.description,
1663                                                 FIO_JOBDESC_SIZE - 1);
1664                         else
1665                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1666
1667                         /*
1668                          * If multiple entries in this group, this is
1669                          * the first member.
1670                          */
1671                         ts->thread_number = td->thread_number;
1672                         ts->groupid = td->groupid;
1673
1674                         /*
1675                          * first pid in group, not very useful...
1676                          */
1677                         ts->pid = td->pid;
1678
1679                         ts->kb_base = td->o.kb_base;
1680                         ts->unit_base = td->o.unit_base;
1681                         ts->unified_rw_rep = td->o.unified_rw_rep;
1682                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1683                         log_info("fio: kb_base differs for jobs in group, using"
1684                                  " %u as the base\n", ts->kb_base);
1685                         kb_base_warned = 1;
1686                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1687                         log_info("fio: unit_base differs for jobs in group, using"
1688                                  " %u as the base\n", ts->unit_base);
1689                         unit_base_warned = 1;
1690                 }
1691
1692                 ts->continue_on_error = td->o.continue_on_error;
1693                 ts->total_err_count += td->total_err_count;
1694                 ts->first_error = td->first_error;
1695                 if (!ts->error) {
1696                         if (!td->error && td->o.continue_on_error &&
1697                             td->first_error) {
1698                                 ts->error = td->first_error;
1699                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1700                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1701                         } else  if (td->error) {
1702                                 ts->error = td->error;
1703                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1704                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1705                         }
1706                 }
1707
1708                 ts->latency_depth = td->latency_qd;
1709                 ts->latency_target = td->o.latency_target;
1710                 ts->latency_percentile = td->o.latency_percentile;
1711                 ts->latency_window = td->o.latency_window;
1712
1713                 ts->nr_block_infos = td->ts.nr_block_infos;
1714                 for (k = 0; k < ts->nr_block_infos; k++)
1715                         ts->block_infos[k] = td->ts.block_infos[k];
1716
1717                 sum_thread_stats(ts, &td->ts, idx == 1);
1718
1719                 if (td->o.ss_dur) {
1720                         ts->ss_state = td->ss.state;
1721                         ts->ss_dur = td->ss.dur;
1722                         ts->ss_head = td->ss.head;
1723                         ts->ss_bw_data = td->ss.bw_data;
1724                         ts->ss_iops_data = td->ss.iops_data;
1725                         ts->ss_limit.u.f = td->ss.limit;
1726                         ts->ss_slope.u.f = td->ss.slope;
1727                         ts->ss_deviation.u.f = td->ss.deviation;
1728                         ts->ss_criterion.u.f = td->ss.criterion;
1729                 }
1730                 else
1731                         ts->ss_dur = ts->ss_state = 0;
1732         }
1733
1734         for (i = 0; i < nr_ts; i++) {
1735                 unsigned long long bw;
1736
1737                 ts = &threadstats[i];
1738                 if (ts->groupid == -1)
1739                         continue;
1740                 rs = &runstats[ts->groupid];
1741                 rs->kb_base = ts->kb_base;
1742                 rs->unit_base = ts->unit_base;
1743                 rs->unified_rw_rep += ts->unified_rw_rep;
1744
1745                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1746                         if (!ts->runtime[j])
1747                                 continue;
1748                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1749                                 rs->min_run[j] = ts->runtime[j];
1750                         if (ts->runtime[j] > rs->max_run[j])
1751                                 rs->max_run[j] = ts->runtime[j];
1752
1753                         bw = 0;
1754                         if (ts->runtime[j])
1755                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1756                         if (bw < rs->min_bw[j])
1757                                 rs->min_bw[j] = bw;
1758                         if (bw > rs->max_bw[j])
1759                                 rs->max_bw[j] = bw;
1760
1761                         rs->iobytes[j] += ts->io_bytes[j];
1762                 }
1763         }
1764
1765         for (i = 0; i < groupid + 1; i++) {
1766                 int ddir;
1767
1768                 rs = &runstats[i];
1769
1770                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1771                         if (rs->max_run[ddir])
1772                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1773                                                 rs->max_run[ddir];
1774                 }
1775         }
1776
1777         for (i = 0; i < FIO_OUTPUT_NR; i++)
1778                 buf_output_init(&output[i]);
1779
1780         /*
1781          * don't overwrite last signal output
1782          */
1783         if (output_format & FIO_OUTPUT_NORMAL)
1784                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1785         if (output_format & FIO_OUTPUT_JSON) {
1786                 struct thread_data *global;
1787                 char time_buf[32];
1788                 struct timeval now;
1789                 unsigned long long ms_since_epoch;
1790
1791                 gettimeofday(&now, NULL);
1792                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1793                                  (unsigned long long)(now.tv_usec) / 1000;
1794
1795                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1796                                 sizeof(time_buf));
1797                 if (time_buf[strlen(time_buf) - 1] == '\n')
1798                         time_buf[strlen(time_buf) - 1] = '\0';
1799
1800                 root = json_create_object();
1801                 json_object_add_value_string(root, "fio version", fio_version_string);
1802                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1803                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1804                 json_object_add_value_string(root, "time", time_buf);
1805                 global = get_global_options();
1806                 json_add_job_opts(root, "global options", &global->opt_list, false);
1807                 array = json_create_array();
1808                 json_object_add_value_array(root, "jobs", array);
1809         }
1810
1811         if (is_backend)
1812                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1813
1814         for (i = 0; i < nr_ts; i++) {
1815                 ts = &threadstats[i];
1816                 rs = &runstats[ts->groupid];
1817
1818                 if (is_backend) {
1819                         fio_server_send_job_options(opt_lists[i], i);
1820                         fio_server_send_ts(ts, rs);
1821                 } else {
1822                         if (output_format & FIO_OUTPUT_TERSE)
1823                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1824                         if (output_format & FIO_OUTPUT_JSON) {
1825                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1826                                 json_array_add_value_object(array, tmp);
1827                         }
1828                         if (output_format & FIO_OUTPUT_NORMAL)
1829                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1830                 }
1831         }
1832         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1833                 /* disk util stats, if any */
1834                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1835
1836                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1837
1838                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1839                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1840                 json_free_object(root);
1841         }
1842
1843         for (i = 0; i < groupid + 1; i++) {
1844                 rs = &runstats[i];
1845
1846                 rs->groupid = i;
1847                 if (is_backend)
1848                         fio_server_send_gs(rs);
1849                 else if (output_format & FIO_OUTPUT_NORMAL)
1850                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1851         }
1852
1853         if (is_backend)
1854                 fio_server_send_du();
1855         else if (output_format & FIO_OUTPUT_NORMAL) {
1856                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1857                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1858         }
1859
1860         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1861                 struct buf_output *out = &output[i];
1862
1863                 log_info_buf(out->buf, out->buflen);
1864                 buf_output_free(out);
1865         }
1866
1867         log_info_flush();
1868         free(runstats);
1869         free(threadstats);
1870         free(opt_lists);
1871 }
1872
1873 void show_run_stats(void)
1874 {
1875         fio_mutex_down(stat_mutex);
1876         __show_run_stats();
1877         fio_mutex_up(stat_mutex);
1878 }
1879
1880 void __show_running_run_stats(void)
1881 {
1882         struct thread_data *td;
1883         unsigned long long *rt;
1884         struct timespec ts;
1885         int i;
1886
1887         fio_mutex_down(stat_mutex);
1888
1889         rt = malloc(thread_number * sizeof(unsigned long long));
1890         fio_gettime(&ts, NULL);
1891
1892         for_each_td(td, i) {
1893                 td->update_rusage = 1;
1894                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1895                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1896                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1897                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1898
1899                 rt[i] = mtime_since(&td->start, &ts);
1900                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1901                         td->ts.runtime[DDIR_READ] += rt[i];
1902                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1903                         td->ts.runtime[DDIR_WRITE] += rt[i];
1904                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1905                         td->ts.runtime[DDIR_TRIM] += rt[i];
1906         }
1907
1908         for_each_td(td, i) {
1909                 if (td->runstate >= TD_EXITED)
1910                         continue;
1911                 if (td->rusage_sem) {
1912                         td->update_rusage = 1;
1913                         fio_mutex_down(td->rusage_sem);
1914                 }
1915                 td->update_rusage = 0;
1916         }
1917
1918         __show_run_stats();
1919
1920         for_each_td(td, i) {
1921                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1922                         td->ts.runtime[DDIR_READ] -= rt[i];
1923                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1924                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1925                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1926                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1927         }
1928
1929         free(rt);
1930         fio_mutex_up(stat_mutex);
1931 }
1932
1933 static int status_interval_init;
1934 static struct timespec status_time;
1935 static int status_file_disabled;
1936
1937 #define FIO_STATUS_FILE         "fio-dump-status"
1938
1939 static int check_status_file(void)
1940 {
1941         struct stat sb;
1942         const char *temp_dir;
1943         char fio_status_file_path[PATH_MAX];
1944
1945         if (status_file_disabled)
1946                 return 0;
1947
1948         temp_dir = getenv("TMPDIR");
1949         if (temp_dir == NULL) {
1950                 temp_dir = getenv("TEMP");
1951                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1952                         temp_dir = NULL;
1953         }
1954         if (temp_dir == NULL)
1955                 temp_dir = "/tmp";
1956
1957         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1958
1959         if (stat(fio_status_file_path, &sb))
1960                 return 0;
1961
1962         if (unlink(fio_status_file_path) < 0) {
1963                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1964                                                         strerror(errno));
1965                 log_err("fio: disabling status file updates\n");
1966                 status_file_disabled = 1;
1967         }
1968
1969         return 1;
1970 }
1971
1972 void check_for_running_stats(void)
1973 {
1974         if (status_interval) {
1975                 if (!status_interval_init) {
1976                         fio_gettime(&status_time, NULL);
1977                         status_interval_init = 1;
1978                 } else if (mtime_since_now(&status_time) >= status_interval) {
1979                         show_running_run_stats();
1980                         fio_gettime(&status_time, NULL);
1981                         return;
1982                 }
1983         }
1984         if (check_status_file()) {
1985                 show_running_run_stats();
1986                 return;
1987         }
1988 }
1989
1990 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
1991 {
1992         double val = data;
1993         double delta;
1994
1995         if (data > is->max_val)
1996                 is->max_val = data;
1997         if (data < is->min_val)
1998                 is->min_val = data;
1999
2000         delta = val - is->mean.u.f;
2001         if (delta) {
2002                 is->mean.u.f += delta / (is->samples + 1.0);
2003                 is->S.u.f += delta * (val - is->mean.u.f);
2004         }
2005
2006         is->samples++;
2007 }
2008
2009 /*
2010  * Return a struct io_logs, which is added to the tail of the log
2011  * list for 'iolog'.
2012  */
2013 static struct io_logs *get_new_log(struct io_log *iolog)
2014 {
2015         size_t new_size, new_samples;
2016         struct io_logs *cur_log;
2017
2018         /*
2019          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2020          * forever
2021          */
2022         if (!iolog->cur_log_max)
2023                 new_samples = DEF_LOG_ENTRIES;
2024         else {
2025                 new_samples = iolog->cur_log_max * 2;
2026                 if (new_samples > MAX_LOG_ENTRIES)
2027                         new_samples = MAX_LOG_ENTRIES;
2028         }
2029
2030         new_size = new_samples * log_entry_sz(iolog);
2031
2032         cur_log = smalloc(sizeof(*cur_log));
2033         if (cur_log) {
2034                 INIT_FLIST_HEAD(&cur_log->list);
2035                 cur_log->log = malloc(new_size);
2036                 if (cur_log->log) {
2037                         cur_log->nr_samples = 0;
2038                         cur_log->max_samples = new_samples;
2039                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2040                         iolog->cur_log_max = new_samples;
2041                         return cur_log;
2042                 }
2043                 sfree(cur_log);
2044         }
2045
2046         return NULL;
2047 }
2048
2049 /*
2050  * Add and return a new log chunk, or return current log if big enough
2051  */
2052 static struct io_logs *regrow_log(struct io_log *iolog)
2053 {
2054         struct io_logs *cur_log;
2055         int i;
2056
2057         if (!iolog || iolog->disabled)
2058                 goto disable;
2059
2060         cur_log = iolog_cur_log(iolog);
2061         if (!cur_log) {
2062                 cur_log = get_new_log(iolog);
2063                 if (!cur_log)
2064                         return NULL;
2065         }
2066
2067         if (cur_log->nr_samples < cur_log->max_samples)
2068                 return cur_log;
2069
2070         /*
2071          * No room for a new sample. If we're compressing on the fly, flush
2072          * out the current chunk
2073          */
2074         if (iolog->log_gz) {
2075                 if (iolog_cur_flush(iolog, cur_log)) {
2076                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2077                         return NULL;
2078                 }
2079         }
2080
2081         /*
2082          * Get a new log array, and add to our list
2083          */
2084         cur_log = get_new_log(iolog);
2085         if (!cur_log) {
2086                 log_err("fio: failed extending iolog! Will stop logging.\n");
2087                 return NULL;
2088         }
2089
2090         if (!iolog->pending || !iolog->pending->nr_samples)
2091                 return cur_log;
2092
2093         /*
2094          * Flush pending items to new log
2095          */
2096         for (i = 0; i < iolog->pending->nr_samples; i++) {
2097                 struct io_sample *src, *dst;
2098
2099                 src = get_sample(iolog, iolog->pending, i);
2100                 dst = get_sample(iolog, cur_log, i);
2101                 memcpy(dst, src, log_entry_sz(iolog));
2102         }
2103         cur_log->nr_samples = iolog->pending->nr_samples;
2104
2105         iolog->pending->nr_samples = 0;
2106         return cur_log;
2107 disable:
2108         if (iolog)
2109                 iolog->disabled = true;
2110         return NULL;
2111 }
2112
2113 void regrow_logs(struct thread_data *td)
2114 {
2115         regrow_log(td->slat_log);
2116         regrow_log(td->clat_log);
2117         regrow_log(td->clat_hist_log);
2118         regrow_log(td->lat_log);
2119         regrow_log(td->bw_log);
2120         regrow_log(td->iops_log);
2121         td->flags &= ~TD_F_REGROW_LOGS;
2122 }
2123
2124 static struct io_logs *get_cur_log(struct io_log *iolog)
2125 {
2126         struct io_logs *cur_log;
2127
2128         cur_log = iolog_cur_log(iolog);
2129         if (!cur_log) {
2130                 cur_log = get_new_log(iolog);
2131                 if (!cur_log)
2132                         return NULL;
2133         }
2134
2135         if (cur_log->nr_samples < cur_log->max_samples)
2136                 return cur_log;
2137
2138         /*
2139          * Out of space. If we're in IO offload mode, or we're not doing
2140          * per unit logging (hence logging happens outside of the IO thread
2141          * as well), add a new log chunk inline. If we're doing inline
2142          * submissions, flag 'td' as needing a log regrow and we'll take
2143          * care of it on the submission side.
2144          */
2145         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2146             !per_unit_log(iolog))
2147                 return regrow_log(iolog);
2148
2149         iolog->td->flags |= TD_F_REGROW_LOGS;
2150         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2151         return iolog->pending;
2152 }
2153
2154 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2155                              enum fio_ddir ddir, unsigned int bs,
2156                              unsigned long t, uint64_t offset)
2157 {
2158         struct io_logs *cur_log;
2159
2160         if (iolog->disabled)
2161                 return;
2162         if (flist_empty(&iolog->io_logs))
2163                 iolog->avg_last = t;
2164
2165         cur_log = get_cur_log(iolog);
2166         if (cur_log) {
2167                 struct io_sample *s;
2168
2169                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2170
2171                 s->data = data;
2172                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2173                 io_sample_set_ddir(iolog, s, ddir);
2174                 s->bs = bs;
2175
2176                 if (iolog->log_offset) {
2177                         struct io_sample_offset *so = (void *) s;
2178
2179                         so->offset = offset;
2180                 }
2181
2182                 cur_log->nr_samples++;
2183                 return;
2184         }
2185
2186         iolog->disabled = true;
2187 }
2188
2189 static inline void reset_io_stat(struct io_stat *ios)
2190 {
2191         ios->max_val = ios->min_val = ios->samples = 0;
2192         ios->mean.u.f = ios->S.u.f = 0;
2193 }
2194
2195 void reset_io_stats(struct thread_data *td)
2196 {
2197         struct thread_stat *ts = &td->ts;
2198         int i, j;
2199
2200         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2201                 reset_io_stat(&ts->clat_stat[i]);
2202                 reset_io_stat(&ts->slat_stat[i]);
2203                 reset_io_stat(&ts->lat_stat[i]);
2204                 reset_io_stat(&ts->bw_stat[i]);
2205                 reset_io_stat(&ts->iops_stat[i]);
2206
2207                 ts->io_bytes[i] = 0;
2208                 ts->runtime[i] = 0;
2209                 ts->total_io_u[i] = 0;
2210                 ts->short_io_u[i] = 0;
2211                 ts->drop_io_u[i] = 0;
2212
2213                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2214                         ts->io_u_plat[i][j] = 0;
2215         }
2216
2217         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2218                 ts->io_u_map[i] = 0;
2219                 ts->io_u_submit[i] = 0;
2220                 ts->io_u_complete[i] = 0;
2221         }
2222
2223         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2224                 ts->io_u_lat_n[i] = 0;
2225         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2226                 ts->io_u_lat_u[i] = 0;
2227         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2228                 ts->io_u_lat_m[i] = 0;
2229
2230         ts->total_submit = 0;
2231         ts->total_complete = 0;
2232 }
2233
2234 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2235                               unsigned long elapsed, bool log_max)
2236 {
2237         /*
2238          * Note an entry in the log. Use the mean from the logged samples,
2239          * making sure to properly round up. Only write a log entry if we
2240          * had actual samples done.
2241          */
2242         if (iolog->avg_window[ddir].samples) {
2243                 union io_sample_data data;
2244
2245                 if (log_max)
2246                         data.val = iolog->avg_window[ddir].max_val;
2247                 else
2248                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2249
2250                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2251         }
2252
2253         reset_io_stat(&iolog->avg_window[ddir]);
2254 }
2255
2256 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2257                              bool log_max)
2258 {
2259         int ddir;
2260
2261         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2262                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2263 }
2264
2265 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2266                            union io_sample_data data, enum fio_ddir ddir,
2267                            unsigned int bs, uint64_t offset)
2268 {
2269         unsigned long elapsed, this_window;
2270
2271         if (!ddir_rw(ddir))
2272                 return 0;
2273
2274         elapsed = mtime_since_now(&td->epoch);
2275
2276         /*
2277          * If no time averaging, just add the log sample.
2278          */
2279         if (!iolog->avg_msec) {
2280                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2281                 return 0;
2282         }
2283
2284         /*
2285          * Add the sample. If the time period has passed, then
2286          * add that entry to the log and clear.
2287          */
2288         add_stat_sample(&iolog->avg_window[ddir], data.val);
2289
2290         /*
2291          * If period hasn't passed, adding the above sample is all we
2292          * need to do.
2293          */
2294         this_window = elapsed - iolog->avg_last;
2295         if (elapsed < iolog->avg_last)
2296                 return iolog->avg_last - elapsed;
2297         else if (this_window < iolog->avg_msec) {
2298                 int diff = iolog->avg_msec - this_window;
2299
2300                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2301                         return diff;
2302         }
2303
2304         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2305
2306         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2307         return iolog->avg_msec;
2308 }
2309
2310 void finalize_logs(struct thread_data *td, bool unit_logs)
2311 {
2312         unsigned long elapsed;
2313
2314         elapsed = mtime_since_now(&td->epoch);
2315
2316         if (td->clat_log && unit_logs)
2317                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2318         if (td->slat_log && unit_logs)
2319                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2320         if (td->lat_log && unit_logs)
2321                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2322         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2323                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2324         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2325                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2326 }
2327
2328 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2329 {
2330         struct io_log *iolog;
2331
2332         if (!ddir_rw(ddir))
2333                 return;
2334
2335         iolog = agg_io_log[ddir];
2336         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2337 }
2338
2339 static void add_clat_percentile_sample(struct thread_stat *ts,
2340                                 unsigned long long nsec, enum fio_ddir ddir)
2341 {
2342         unsigned int idx = plat_val_to_idx(nsec);
2343         assert(idx < FIO_IO_U_PLAT_NR);
2344
2345         ts->io_u_plat[ddir][idx]++;
2346 }
2347
2348 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2349                      unsigned long long nsec, unsigned int bs, uint64_t offset)
2350 {
2351         unsigned long elapsed, this_window;
2352         struct thread_stat *ts = &td->ts;
2353         struct io_log *iolog = td->clat_hist_log;
2354
2355         td_io_u_lock(td);
2356
2357         add_stat_sample(&ts->clat_stat[ddir], nsec);
2358
2359         if (td->clat_log)
2360                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2361                                offset);
2362
2363         if (ts->clat_percentiles)
2364                 add_clat_percentile_sample(ts, nsec, ddir);
2365
2366         if (iolog && iolog->hist_msec) {
2367                 struct io_hist *hw = &iolog->hist_window[ddir];
2368
2369                 hw->samples++;
2370                 elapsed = mtime_since_now(&td->epoch);
2371                 if (!hw->hist_last)
2372                         hw->hist_last = elapsed;
2373                 this_window = elapsed - hw->hist_last;
2374
2375                 if (this_window >= iolog->hist_msec) {
2376                         unsigned int *io_u_plat;
2377                         struct io_u_plat_entry *dst;
2378
2379                         /*
2380                          * Make a byte-for-byte copy of the latency histogram
2381                          * stored in td->ts.io_u_plat[ddir], recording it in a
2382                          * log sample. Note that the matching call to free() is
2383                          * located in iolog.c after printing this sample to the
2384                          * log file.
2385                          */
2386                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2387                         dst = malloc(sizeof(struct io_u_plat_entry));
2388                         memcpy(&(dst->io_u_plat), io_u_plat,
2389                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2390                         flist_add(&dst->list, &hw->list);
2391                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2392                                                 elapsed, offset);
2393
2394                         /*
2395                          * Update the last time we recorded as being now, minus
2396                          * any drift in time we encountered before actually
2397                          * making the record.
2398                          */
2399                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2400                         hw->samples = 0;
2401                 }
2402         }
2403
2404         td_io_u_unlock(td);
2405 }
2406
2407 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2408                      unsigned long usec, unsigned int bs, uint64_t offset)
2409 {
2410         struct thread_stat *ts = &td->ts;
2411
2412         if (!ddir_rw(ddir))
2413                 return;
2414
2415         td_io_u_lock(td);
2416
2417         add_stat_sample(&ts->slat_stat[ddir], usec);
2418
2419         if (td->slat_log)
2420                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2421
2422         td_io_u_unlock(td);
2423 }
2424
2425 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2426                     unsigned long long nsec, unsigned int bs, uint64_t offset)
2427 {
2428         struct thread_stat *ts = &td->ts;
2429
2430         if (!ddir_rw(ddir))
2431                 return;
2432
2433         td_io_u_lock(td);
2434
2435         add_stat_sample(&ts->lat_stat[ddir], nsec);
2436
2437         if (td->lat_log)
2438                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2439                                offset);
2440
2441         td_io_u_unlock(td);
2442 }
2443
2444 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2445                    unsigned int bytes, unsigned long long spent)
2446 {
2447         struct thread_stat *ts = &td->ts;
2448         unsigned long rate;
2449
2450         if (spent)
2451                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2452         else
2453                 rate = 0;
2454
2455         td_io_u_lock(td);
2456
2457         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2458
2459         if (td->bw_log)
2460                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2461                                bytes, io_u->offset);
2462
2463         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2464         td_io_u_unlock(td);
2465 }
2466
2467 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2468                          struct timespec *t, unsigned int avg_time,
2469                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2470                          struct io_stat *stat, struct io_log *log,
2471                          bool is_kb)
2472 {
2473         unsigned long spent, rate;
2474         enum fio_ddir ddir;
2475         unsigned int next, next_log;
2476
2477         next_log = avg_time;
2478
2479         spent = mtime_since(parent_tv, t);
2480         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2481                 return avg_time - spent;
2482
2483         td_io_u_lock(td);
2484
2485         /*
2486          * Compute both read and write rates for the interval.
2487          */
2488         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2489                 uint64_t delta;
2490
2491                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2492                 if (!delta)
2493                         continue; /* No entries for interval */
2494
2495                 if (spent) {
2496                         if (is_kb)
2497                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2498                         else
2499                                 rate = (delta * 1000) / spent;
2500                 } else
2501                         rate = 0;
2502
2503                 add_stat_sample(&stat[ddir], rate);
2504
2505                 if (log) {
2506                         unsigned int bs = 0;
2507
2508                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2509                                 bs = td->o.min_bs[ddir];
2510
2511                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2512                         next_log = min(next_log, next);
2513                 }
2514
2515                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2516         }
2517
2518         timespec_add_msec(parent_tv, avg_time);
2519
2520         td_io_u_unlock(td);
2521
2522         if (spent <= avg_time)
2523                 next = avg_time;
2524         else
2525                 next = avg_time - (1 + spent - avg_time);
2526
2527         return min(next, next_log);
2528 }
2529
2530 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2531 {
2532         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2533                                 td->this_io_bytes, td->stat_io_bytes,
2534                                 td->ts.bw_stat, td->bw_log, true);
2535 }
2536
2537 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2538                      unsigned int bytes)
2539 {
2540         struct thread_stat *ts = &td->ts;
2541
2542         td_io_u_lock(td);
2543
2544         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2545
2546         if (td->iops_log)
2547                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2548                                bytes, io_u->offset);
2549
2550         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2551         td_io_u_unlock(td);
2552 }
2553
2554 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2555 {
2556         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2557                                 td->this_io_blocks, td->stat_io_blocks,
2558                                 td->ts.iops_stat, td->iops_log, false);
2559 }
2560
2561 /*
2562  * Returns msecs to next event
2563  */
2564 int calc_log_samples(void)
2565 {
2566         struct thread_data *td;
2567         unsigned int next = ~0U, tmp;
2568         struct timespec now;
2569         int i;
2570
2571         fio_gettime(&now, NULL);
2572
2573         for_each_td(td, i) {
2574                 if (!td->o.stats)
2575                         continue;
2576                 if (in_ramp_time(td) ||
2577                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2578                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2579                         continue;
2580                 }
2581                 if (!td->bw_log ||
2582                         (td->bw_log && !per_unit_log(td->bw_log))) {
2583                         tmp = add_bw_samples(td, &now);
2584                         if (tmp < next)
2585                                 next = tmp;
2586                 }
2587                 if (!td->iops_log ||
2588                         (td->iops_log && !per_unit_log(td->iops_log))) {
2589                         tmp = add_iops_samples(td, &now);
2590                         if (tmp < next)
2591                                 next = tmp;
2592                 }
2593         }
2594
2595         return next == ~0U ? 0 : next;
2596 }
2597
2598 void stat_init(void)
2599 {
2600         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2601 }
2602
2603 void stat_exit(void)
2604 {
2605         /*
2606          * When we have the mutex, we know out-of-band access to it
2607          * have ended.
2608          */
2609         fio_mutex_down(stat_mutex);
2610         fio_mutex_remove(stat_mutex);
2611 }
2612
2613 /*
2614  * Called from signal handler. Wake up status thread.
2615  */
2616 void show_running_run_stats(void)
2617 {
2618         helper_do_stat();
2619 }
2620
2621 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2622 {
2623         /* Ignore io_u's which span multiple blocks--they will just get
2624          * inaccurate counts. */
2625         int idx = (io_u->offset - io_u->file->file_offset)
2626                         / td->o.bs[DDIR_TRIM];
2627         uint32_t *info = &td->ts.block_infos[idx];
2628         assert(idx < td->ts.nr_block_infos);
2629         return info;
2630 }