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;
 104         unsigned long long k, base;
 105
 106         assert(idx < FIO_IO_U_PLAT_NR);
 107
 108         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 109          * all bits of the sample as index */
 110         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 111                 return idx;
 112
 113         /* Find the group and compute the minimum value of that group */
 114         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 115         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
 116
 117         /* Find its bucket number of the group */
 118         k = idx % FIO_IO_U_PLAT_VAL;
 119
 120         /* Return the mean of the range of the bucket */
 121         return base + ((k + 0.5) * (1 << error_bits));
 122 }
 123
 124 static int double_cmp(const void *a, const void *b)
 125 {
 126         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 127         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 128         int cmp = 0;
 129
 130         if (fa.u.f > fb.u.f)
 131                 cmp = 1;
 132         else if (fa.u.f < fb.u.f)
 133                 cmp = -1;
 134
 135         return cmp;
 136 }
 137
 138 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long long nr,
 139                                    fio_fp64_t *plist, unsigned long long **output,
 140                                    unsigned long long *maxv, unsigned long long *minv)
 141 {
 142         unsigned long long sum = 0;
 143         unsigned int len, i, j = 0;
 144         unsigned int oval_len = 0;
 145         unsigned long long *ovals = NULL;
 146         bool is_last;
 147
 148         *minv = -1ULL;
 149         *maxv = 0;
 150
 151         len = 0;
 152         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 153                 len++;
 154
 155         if (!len)
 156                 return 0;
 157
 158         /*
 159          * Sort the percentile list. Note that it may already be sorted if
 160          * we are using the default values, but since it's a short list this
 161          * isn't a worry. Also note that this does not work for NaN values.
 162          */
 163         if (len > 1)
 164                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 165
 166         /*
 167          * Calculate bucket values, note down max and min values
 168          */
 169         is_last = false;
 170         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 171                 sum += io_u_plat[i];
 172                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 173                         assert(plist[j].u.f <= 100.0);
 174
 175                         if (j == oval_len) {
 176                                 oval_len += 100;
 177                                 ovals = realloc(ovals, oval_len * sizeof(*ovals));
 178                         }
 179
 180                         ovals[j] = plat_idx_to_val(i);
 181                         if (ovals[j] < *minv)
 182                                 *minv = ovals[j];
 183                         if (ovals[j] > *maxv)
 184                                 *maxv = ovals[j];
 185
 186                         is_last = (j == len - 1) != 0;
 187                         if (is_last)
 188                                 break;
 189
 190                         j++;
 191                 }
 192         }
 193
 194         *output = ovals;
 195         return len;
 196 }
 197
 198 /*
 199  * Find and display the p-th percentile of clat
 200  */
 201 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long long nr,
 202                                   fio_fp64_t *plist, unsigned int precision,
 203                                   bool is_clat, struct buf_output *out)
 204 {
 205         unsigned int divisor, len, i, j = 0;
 206         unsigned long long minv, maxv;
 207         unsigned long long *ovals;
 208         int per_line, scale_down, time_width;
 209         const char *pre = is_clat ? "clat" : " lat";
 210         bool is_last;
 211         char fmt[32];
 212
 213         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 214         if (!len)
 215                 goto out;
 216
 217         /*
 218          * We default to nsecs, but if the value range is such that we
 219          * should scale down to usecs or msecs, do that.
 220          */
 221         if (minv > 2000000 && maxv > 99999999ULL) {
 222                 scale_down = 2;
 223                 divisor = 1000000;
 224                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 225         } else if (minv > 2000 && maxv > 99999) {
 226                 scale_down = 1;
 227                 divisor = 1000;
 228                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 229         } else {
 230                 scale_down = 0;
 231                 divisor = 1;
 232                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 233         }
 234
 235
 236         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 237         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 238                         precision, time_width);
 239         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 240         per_line = (80 - 7) / (precision + 10 + time_width);
 241
 242         for (j = 0; j < len; j++) {
 243                 /* for formatting */
 244                 if (j != 0 && (j % per_line) == 0)
 245                         log_buf(out, "     |");
 246
 247                 /* end of the list */
 248                 is_last = (j == len - 1) != 0;
 249
 250                 for (i = 0; i < scale_down; i++)
 251                         ovals[j] = (ovals[j] + 999) / 1000;
 252
 253                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 254
 255                 if (is_last)
 256                         break;
 257
 258                 if ((j % per_line) == per_line - 1)     /* for formatting */
 259                         log_buf(out, "\n");
 260         }
 261
 262 out:
 263         if (ovals)
 264                 free(ovals);
 265 }
 266
 267 bool calc_lat(struct io_stat *is, unsigned long long *min,
 268               unsigned long long *max, double *mean, double *dev)
 269 {
 270         double n = (double) is->samples;
 271
 272         if (n == 0)
 273                 return false;
 274
 275         *min = is->min_val;
 276         *max = is->max_val;
 277         *mean = is->mean.u.f;
 278
 279         if (n > 1.0)
 280                 *dev = sqrt(is->S.u.f / (n - 1.0));
 281         else
 282                 *dev = 0;
 283
 284         return true;
 285 }
 286
 287 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 288 {
 289         char *io, *agg, *min, *max;
 290         char *ioalt, *aggalt, *minalt, *maxalt;
 291         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 292         int i;
 293
 294         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 295
 296         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 297                 const int i2p = is_power_of_2(rs->kb_base);
 298
 299                 if (!rs->max_run[i])
 300                         continue;
 301
 302                 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
 303                 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
 304                 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
 305                 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 306                 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 307                 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 308                 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
 309                 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
 310                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 311                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 312                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 313                                 (unsigned long long) rs->min_run[i],
 314                                 (unsigned long long) rs->max_run[i]);
 315
 316                 free(io);
 317                 free(agg);
 318                 free(min);
 319                 free(max);
 320                 free(ioalt);
 321                 free(aggalt);
 322                 free(minalt);
 323                 free(maxalt);
 324         }
 325 }
 326
 327 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 328 {
 329         int i;
 330
 331         /*
 332          * Do depth distribution calculations
 333          */
 334         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 335                 if (total) {
 336                         io_u_dist[i] = (double) map[i] / (double) total;
 337                         io_u_dist[i] *= 100.0;
 338                         if (io_u_dist[i] < 0.1 && map[i])
 339                                 io_u_dist[i] = 0.1;
 340                 } else
 341                         io_u_dist[i] = 0.0;
 342         }
 343 }
 344
 345 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 346                           unsigned int *src, int nr)
 347 {
 348         unsigned long total = ddir_rw_sum(ts->total_io_u);
 349         int i;
 350
 351         /*
 352          * Do latency distribution calculations
 353          */
 354         for (i = 0; i < nr; i++) {
 355                 if (total) {
 356                         dst[i] = (double) src[i] / (double) total;
 357                         dst[i] *= 100.0;
 358                         if (dst[i] < 0.01 && src[i])
 359                                 dst[i] = 0.01;
 360                 } else
 361                         dst[i] = 0.0;
 362         }
 363 }
 364
 365 /*
 366  * To keep the terse format unaltered, add all of the ns latency
 367  * buckets to the first us latency bucket
 368  */
 369 void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 370 {
 371         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 372         int i;
 373
 374         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 375
 376         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 377                 ntotal += ts->io_u_lat_n[i];
 378
 379         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 380 }
 381
 382 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 383 {
 384         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 385 }
 386
 387 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 388 {
 389         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 390 }
 391
 392 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 393 {
 394         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 395 }
 396
 397 static void display_lat(const char *name, unsigned long long min,
 398                         unsigned long long max, double mean, double dev,
 399                         struct buf_output *out)
 400 {
 401         const char *base = "(nsec)";
 402         char *minp, *maxp;
 403
 404         if (nsec_to_msec(&min, &max, &mean, &dev))
 405                 base = "(msec)";
 406         else if (nsec_to_usec(&min, &max, &mean, &dev))
 407                 base = "(usec)";
 408
 409         minp = num2str(min, 6, 1, 0, N2S_NONE);
 410         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 411
 412         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 413                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 414
 415         free(minp);
 416         free(maxp);
 417 }
 418
 419 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 420                              int ddir, struct buf_output *out)
 421 {
 422         const char *str[] = { " read", "write", " trim" };
 423         unsigned long runt;
 424         unsigned long long min, max, bw, iops;
 425         double mean, dev;
 426         char *io_p, *bw_p, *bw_p_alt, *iops_p;
 427         int i2p;
 428
 429         assert(ddir_rw(ddir));
 430
 431         if (!ts->runtime[ddir])
 432                 return;
 433
 434         i2p = is_power_of_2(rs->kb_base);
 435         runt = ts->runtime[ddir];
 436
 437         bw = (1000 * ts->io_bytes[ddir]) / runt;
 438         io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
 439         bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
 440         bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
 441
 442         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 443         iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
 444
 445         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
 446                         rs->unified_rw_rep ? "mixed" : str[ddir],
 447                         iops_p, bw_p, bw_p_alt, io_p,
 448                         (unsigned long long) ts->runtime[ddir]);
 449
 450         free(io_p);
 451         free(bw_p);
 452         free(bw_p_alt);
 453         free(iops_p);
 454
 455         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 456                 display_lat("slat", min, max, mean, dev, out);
 457         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 458                 display_lat("clat", min, max, mean, dev, out);
 459         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 460                 display_lat(" lat", min, max, mean, dev, out);
 461
 462         if (ts->clat_percentiles || ts->lat_percentiles) {
 463                 uint64_t samples;
 464
 465                 if (ts->clat_percentiles)
 466                         samples = ts->clat_stat[ddir].samples;
 467                 else
 468                         samples = ts->lat_stat[ddir].samples;
 469
 470                 show_clat_percentiles(ts->io_u_plat[ddir],
 471                                         samples,
 472                                         ts->percentile_list,
 473                                         ts->percentile_precision,
 474                                         ts->clat_percentiles, out);
 475         }
 476         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 477                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 478                 const char *bw_str;
 479
 480                 if ((rs->unit_base == 1) && i2p)
 481                         bw_str = "Kibit";
 482                 else if (rs->unit_base == 1)
 483                         bw_str = "kbit";
 484                 else if (i2p)
 485                         bw_str = "KiB";
 486                 else
 487                         bw_str = "kB";
 488
 489                 if (rs->agg[ddir]) {
 490                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 491                         if (p_of_agg > 100.0)
 492                                 p_of_agg = 100.0;
 493                 }
 494
 495                 if (rs->unit_base == 1) {
 496                         min *= 8.0;
 497                         max *= 8.0;
 498                         mean *= 8.0;
 499                         dev *= 8.0;
 500                 }
 501
 502                 if (mean > fkb_base * fkb_base) {
 503                         min /= fkb_base;
 504                         max /= fkb_base;
 505                         mean /= fkb_base;
 506                         dev /= fkb_base;
 507                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 508                 }
 509
 510                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 511                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 512                         bw_str, min, max, p_of_agg, mean, dev,
 513                         (&ts->bw_stat[ddir])->samples);
 514         }
 515         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 516                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 517                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 518                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 519         }
 520 }
 521
 522 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
 523                      const char *msg, struct buf_output *out)
 524 {
 525         bool new_line = true, shown = false;
 526         int i, line = 0;
 527
 528         for (i = 0; i < nr; i++) {
 529                 if (io_u_lat[i] <= 0.0)
 530                         continue;
 531                 shown = true;
 532                 if (new_line) {
 533                         if (line)
 534                                 log_buf(out, "\n");
 535                         log_buf(out, "  lat (%s)   : ", msg);
 536                         new_line = false;
 537                         line = 0;
 538                 }
 539                 if (line)
 540                         log_buf(out, ", ");
 541                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 542                 line++;
 543                 if (line == 5)
 544                         new_line = true;
 545         }
 546
 547         if (shown)
 548                 log_buf(out, "\n");
 549
 550         return true;
 551 }
 552
 553 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 554 {
 555         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 556                                  "250=", "500=", "750=", "1000=", };
 557
 558         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 559 }
 560
 561 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 562 {
 563         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 564                                  "250=", "500=", "750=", "1000=", };
 565
 566         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 567 }
 568
 569 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 570 {
 571         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 572                                  "250=", "500=", "750=", "1000=", "2000=",
 573                                  ">=2000=", };
 574
 575         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 576 }
 577
 578 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 579 {
 580         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 581         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 582         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 583
 584         stat_calc_lat_n(ts, io_u_lat_n);
 585         stat_calc_lat_u(ts, io_u_lat_u);
 586         stat_calc_lat_m(ts, io_u_lat_m);
 587
 588         show_lat_n(io_u_lat_n, out);
 589         show_lat_u(io_u_lat_u, out);
 590         show_lat_m(io_u_lat_m, out);
 591 }
 592
 593 static int block_state_category(int block_state)
 594 {
 595         switch (block_state) {
 596         case BLOCK_STATE_UNINIT:
 597                 return 0;
 598         case BLOCK_STATE_TRIMMED:
 599         case BLOCK_STATE_WRITTEN:
 600                 return 1;
 601         case BLOCK_STATE_WRITE_FAILURE:
 602         case BLOCK_STATE_TRIM_FAILURE:
 603                 return 2;
 604         default:
 605                 /* Silence compile warning on some BSDs and have a return */
 606                 assert(0);
 607                 return -1;
 608         }
 609 }
 610
 611 static int compare_block_infos(const void *bs1, const void *bs2)
 612 {
 613         uint32_t block1 = *(uint32_t *)bs1;
 614         uint32_t block2 = *(uint32_t *)bs2;
 615         int state1 = BLOCK_INFO_STATE(block1);
 616         int state2 = BLOCK_INFO_STATE(block2);
 617         int bscat1 = block_state_category(state1);
 618         int bscat2 = block_state_category(state2);
 619         int cycles1 = BLOCK_INFO_TRIMS(block1);
 620         int cycles2 = BLOCK_INFO_TRIMS(block2);
 621
 622         if (bscat1 < bscat2)
 623                 return -1;
 624         if (bscat1 > bscat2)
 625                 return 1;
 626
 627         if (cycles1 < cycles2)
 628                 return -1;
 629         if (cycles1 > cycles2)
 630                 return 1;
 631
 632         if (state1 < state2)
 633                 return -1;
 634         if (state1 > state2)
 635                 return 1;
 636
 637         assert(block1 == block2);
 638         return 0;
 639 }
 640
 641 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 642                                   fio_fp64_t *plist, unsigned int **percentiles,
 643                                   unsigned int *types)
 644 {
 645         int len = 0;
 646         int i, nr_uninit;
 647
 648         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 649
 650         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 651                 len++;
 652
 653         if (!len)
 654                 return 0;
 655
 656         /*
 657          * Sort the percentile list. Note that it may already be sorted if
 658          * we are using the default values, but since it's a short list this
 659          * isn't a worry. Also note that this does not work for NaN values.
 660          */
 661         if (len > 1)
 662                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 663
 664         nr_uninit = 0;
 665         /* Start only after the uninit entries end */
 666         for (nr_uninit = 0;
 667              nr_uninit < nr_block_infos
 668                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 669              nr_uninit ++)
 670                 ;
 671
 672         if (nr_uninit == nr_block_infos)
 673                 return 0;
 674
 675         *percentiles = calloc(len, sizeof(**percentiles));
 676
 677         for (i = 0; i < len; i++) {
 678                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 679                                 + nr_uninit;
 680                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 681         }
 682
 683         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 684         for (i = 0; i < nr_block_infos; i++)
 685                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 686
 687         return len;
 688 }
 689
 690 static const char *block_state_names[] = {
 691         [BLOCK_STATE_UNINIT] = "unwritten",
 692         [BLOCK_STATE_TRIMMED] = "trimmed",
 693         [BLOCK_STATE_WRITTEN] = "written",
 694         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 695         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 696 };
 697
 698 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 699                              fio_fp64_t *plist, struct buf_output *out)
 700 {
 701         int len, pos, i;
 702         unsigned int *percentiles = NULL;
 703         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 704
 705         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 706                                      &percentiles, block_state_counts);
 707
 708         log_buf(out, "  block lifetime percentiles :\n   |");
 709         pos = 0;
 710         for (i = 0; i < len; i++) {
 711                 uint32_t block_info = percentiles[i];
 712 #define LINE_LENGTH     75
 713                 char str[LINE_LENGTH];
 714                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 715                                      plist[i].u.f, block_info,
 716                                      i == len - 1 ? '\n' : ',');
 717                 assert(strln < LINE_LENGTH);
 718                 if (pos + strln > LINE_LENGTH) {
 719                         pos = 0;
 720                         log_buf(out, "\n   |");
 721                 }
 722                 log_buf(out, "%s", str);
 723                 pos += strln;
 724 #undef LINE_LENGTH
 725         }
 726         if (percentiles)
 727                 free(percentiles);
 728
 729         log_buf(out, "        states               :");
 730         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 731                 log_buf(out, " %s=%u%c",
 732                          block_state_names[i], block_state_counts[i],
 733                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 734 }
 735
 736 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 737 {
 738         char *p1, *p1alt, *p2;
 739         unsigned long long bw_mean, iops_mean;
 740         const int i2p = is_power_of_2(ts->kb_base);
 741
 742         if (!ts->ss_dur)
 743                 return;
 744
 745         bw_mean = steadystate_bw_mean(ts);
 746         iops_mean = steadystate_iops_mean(ts);
 747
 748         p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
 749         p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
 750         p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
 751
 752         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 753                 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
 754                 p1, p1alt, p2,
 755                 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
 756                 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
 757                 ts->ss_criterion.u.f,
 758                 ts->ss_state & FIO_SS_PCT ? "%" : "");
 759
 760         free(p1);
 761         free(p1alt);
 762         free(p2);
 763 }
 764
 765 static void show_thread_status_normal(struct thread_stat *ts,
 766                                       struct group_run_stats *rs,
 767                                       struct buf_output *out)
 768 {
 769         double usr_cpu, sys_cpu;
 770         unsigned long runtime;
 771         double io_u_dist[FIO_IO_U_MAP_NR];
 772         time_t time_p;
 773         char time_buf[32];
 774
 775         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 776                 return;
 777
 778         memset(time_buf, 0, sizeof(time_buf));
 779
 780         time(&time_p);
 781         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 782
 783         if (!ts->error) {
 784                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 785                                         ts->name, ts->groupid, ts->members,
 786                                         ts->error, (int) ts->pid, time_buf);
 787         } else {
 788                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 789                                         ts->name, ts->groupid, ts->members,
 790                                         ts->error, ts->verror, (int) ts->pid,
 791                                         time_buf);
 792         }
 793
 794         if (strlen(ts->description))
 795                 log_buf(out, "  Description  : [%s]\n", ts->description);
 796
 797         if (ts->io_bytes[DDIR_READ])
 798                 show_ddir_status(rs, ts, DDIR_READ, out);
 799         if (ts->io_bytes[DDIR_WRITE])
 800                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 801         if (ts->io_bytes[DDIR_TRIM])
 802                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 803
 804         show_latencies(ts, out);
 805
 806         runtime = ts->total_run_time;
 807         if (runtime) {
 808                 double runt = (double) runtime;
 809
 810                 usr_cpu = (double) ts->usr_time * 100 / runt;
 811                 sys_cpu = (double) ts->sys_time * 100 / runt;
 812         } else {
 813                 usr_cpu = 0;
 814                 sys_cpu = 0;
 815         }
 816
 817         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 818                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 819                         (unsigned long long) ts->ctx,
 820                         (unsigned long long) ts->majf,
 821                         (unsigned long long) ts->minf);
 822
 823         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 824         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 825                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 826                                         io_u_dist[1], io_u_dist[2],
 827                                         io_u_dist[3], io_u_dist[4],
 828                                         io_u_dist[5], io_u_dist[6]);
 829
 830         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 831         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 832                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 833                                         io_u_dist[1], io_u_dist[2],
 834                                         io_u_dist[3], io_u_dist[4],
 835                                         io_u_dist[5], io_u_dist[6]);
 836         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 837         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 838                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 839                                         io_u_dist[1], io_u_dist[2],
 840                                         io_u_dist[3], io_u_dist[4],
 841                                         io_u_dist[5], io_u_dist[6]);
 842         log_buf(out, "     issued rwt: total=%llu,%llu,%llu,"
 843                                  " short=%llu,%llu,%llu,"
 844                                  " dropped=%llu,%llu,%llu\n",
 845                                         (unsigned long long) ts->total_io_u[0],
 846                                         (unsigned long long) ts->total_io_u[1],
 847                                         (unsigned long long) ts->total_io_u[2],
 848                                         (unsigned long long) ts->short_io_u[0],
 849                                         (unsigned long long) ts->short_io_u[1],
 850                                         (unsigned long long) ts->short_io_u[2],
 851                                         (unsigned long long) ts->drop_io_u[0],
 852                                         (unsigned long long) ts->drop_io_u[1],
 853                                         (unsigned long long) ts->drop_io_u[2]);
 854         if (ts->continue_on_error) {
 855                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 856                                         (unsigned long long)ts->total_err_count,
 857                                         ts->first_error,
 858                                         strerror(ts->first_error));
 859         }
 860         if (ts->latency_depth) {
 861                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 862                                         (unsigned long long)ts->latency_target,
 863                                         (unsigned long long)ts->latency_window,
 864                                         ts->latency_percentile.u.f,
 865                                         ts->latency_depth);
 866         }
 867
 868         if (ts->nr_block_infos)
 869                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 870                                   ts->percentile_list, out);
 871
 872         if (ts->ss_dur)
 873                 show_ss_normal(ts, out);
 874 }
 875
 876 static void show_ddir_status_terse(struct thread_stat *ts,
 877                                    struct group_run_stats *rs, int ddir,
 878                                    int ver, struct buf_output *out)
 879 {
 880         unsigned long long min, max, minv, maxv, bw, iops;
 881         unsigned long long *ovals = NULL;
 882         double mean, dev;
 883         unsigned int len;
 884         int i, bw_stat;
 885
 886         assert(ddir_rw(ddir));
 887
 888         iops = bw = 0;
 889         if (ts->runtime[ddir]) {
 890                 uint64_t runt = ts->runtime[ddir];
 891
 892                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 893                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 894         }
 895
 896         log_buf(out, ";%llu;%llu;%llu;%llu",
 897                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 898                                         (unsigned long long) ts->runtime[ddir]);
 899
 900         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 901                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 902         else
 903                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 904
 905         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 906                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 907         else
 908                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 909
 910         if (ts->clat_percentiles || ts->lat_percentiles) {
 911                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 912                                         ts->clat_stat[ddir].samples,
 913                                         ts->percentile_list, &ovals, &maxv,
 914                                         &minv);
 915         } else
 916                 len = 0;
 917
 918         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 919                 if (i >= len) {
 920                         log_buf(out, ";0%%=0");
 921                         continue;
 922                 }
 923                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
 924         }
 925
 926         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 927                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 928         else
 929                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 930
 931         if (ovals)
 932                 free(ovals);
 933
 934         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
 935         if (bw_stat) {
 936                 double p_of_agg = 100.0;
 937
 938                 if (rs->agg[ddir]) {
 939                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 940                         if (p_of_agg > 100.0)
 941                                 p_of_agg = 100.0;
 942                 }
 943
 944                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 945         } else
 946                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
 947
 948         if (ver == 5) {
 949                 if (bw_stat)
 950                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
 951                 else
 952                         log_buf(out, ";%lu", 0UL);
 953
 954                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
 955                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
 956                                 mean, dev, (&ts->iops_stat[ddir])->samples);
 957                 else
 958                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
 959         }
 960 }
 961
 962 static void add_ddir_status_json(struct thread_stat *ts,
 963                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 964 {
 965         unsigned long long min, max, minv, maxv;
 966         unsigned long long bw_bytes, bw;
 967         unsigned long long *ovals = NULL;
 968         double mean, dev, iops;
 969         unsigned int len;
 970         int i;
 971         const char *ddirname[] = {"read", "write", "trim"};
 972         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object = NULL;
 973         char buf[120];
 974         double p_of_agg = 100.0;
 975
 976         assert(ddir_rw(ddir));
 977
 978         if (ts->unified_rw_rep && ddir != DDIR_READ)
 979                 return;
 980
 981         dir_object = json_create_object();
 982         json_object_add_value_object(parent,
 983                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 984
 985         bw_bytes = 0;
 986         bw = 0;
 987         iops = 0.0;
 988         if (ts->runtime[ddir]) {
 989                 uint64_t runt = ts->runtime[ddir];
 990
 991                 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
 992                 bw = bw_bytes / 1024; /* KiB/s */
 993                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 994         }
 995
 996         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
 997         json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
 998         json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
 999         json_object_add_value_int(dir_object, "bw", bw);
1000         json_object_add_value_float(dir_object, "iops", iops);
1001         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1002         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1003         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1004         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1005
1006         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
1007                 min = max = 0;
1008                 mean = dev = 0.0;
1009         }
1010         tmp_object = json_create_object();
1011         json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1012         json_object_add_value_int(tmp_object, "min", min);
1013         json_object_add_value_int(tmp_object, "max", max);
1014         json_object_add_value_float(tmp_object, "mean", mean);
1015         json_object_add_value_float(tmp_object, "stddev", dev);
1016
1017         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1018                 min = max = 0;
1019                 mean = dev = 0.0;
1020         }
1021         tmp_object = json_create_object();
1022         json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1023         json_object_add_value_int(tmp_object, "min", min);
1024         json_object_add_value_int(tmp_object, "max", max);
1025         json_object_add_value_float(tmp_object, "mean", mean);
1026         json_object_add_value_float(tmp_object, "stddev", dev);
1027
1028         if (ts->clat_percentiles || ts->lat_percentiles) {
1029                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1030                                         ts->clat_stat[ddir].samples,
1031                                         ts->percentile_list, &ovals, &maxv,
1032                                         &minv);
1033         } else
1034                 len = 0;
1035
1036         percentile_object = json_create_object();
1037         json_object_add_value_object(tmp_object, "percentile", percentile_object);
1038         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1039                 if (i >= len) {
1040                         json_object_add_value_int(percentile_object, "0.00", 0);
1041                         continue;
1042                 }
1043                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1044                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1045         }
1046
1047         if (output_format & FIO_OUTPUT_JSON_PLUS) {
1048                 clat_bins_object = json_create_object();
1049                 if (ts->clat_percentiles)
1050                         json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1051
1052                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1053                         if (ts->io_u_plat[ddir][i]) {
1054                                 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1055                                 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1056                         }
1057                 }
1058         }
1059
1060         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1061                 min = max = 0;
1062                 mean = dev = 0.0;
1063         }
1064         tmp_object = json_create_object();
1065         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1066         json_object_add_value_int(tmp_object, "min", min);
1067         json_object_add_value_int(tmp_object, "max", max);
1068         json_object_add_value_float(tmp_object, "mean", mean);
1069         json_object_add_value_float(tmp_object, "stddev", dev);
1070         if (output_format & FIO_OUTPUT_JSON_PLUS && ts->lat_percentiles)
1071                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1072
1073         if (ovals)
1074                 free(ovals);
1075
1076         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1077                 if (rs->agg[ddir]) {
1078                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1079                         if (p_of_agg > 100.0)
1080                                 p_of_agg = 100.0;
1081                 }
1082         } else {
1083                 min = max = 0;
1084                 p_of_agg = mean = dev = 0.0;
1085         }
1086         json_object_add_value_int(dir_object, "bw_min", min);
1087         json_object_add_value_int(dir_object, "bw_max", max);
1088         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1089         json_object_add_value_float(dir_object, "bw_mean", mean);
1090         json_object_add_value_float(dir_object, "bw_dev", dev);
1091         json_object_add_value_int(dir_object, "bw_samples",
1092                                 (&ts->bw_stat[ddir])->samples);
1093
1094         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1095                 min = max = 0;
1096                 mean = dev = 0.0;
1097         }
1098         json_object_add_value_int(dir_object, "iops_min", min);
1099         json_object_add_value_int(dir_object, "iops_max", max);
1100         json_object_add_value_float(dir_object, "iops_mean", mean);
1101         json_object_add_value_float(dir_object, "iops_stddev", dev);
1102         json_object_add_value_int(dir_object, "iops_samples",
1103                                 (&ts->iops_stat[ddir])->samples);
1104 }
1105
1106 static void show_thread_status_terse_all(struct thread_stat *ts,
1107                                          struct group_run_stats *rs, int ver,
1108                                          struct buf_output *out)
1109 {
1110         double io_u_dist[FIO_IO_U_MAP_NR];
1111         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1112         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1113         double usr_cpu, sys_cpu;
1114         int i;
1115
1116         /* General Info */
1117         if (ver == 2)
1118                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1119         else
1120                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1121                         ts->name, ts->groupid, ts->error);
1122
1123         /* Log Read Status */
1124         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1125         /* Log Write Status */
1126         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1127         /* Log Trim Status */
1128         if (ver == 2 || ver == 4 || ver == 5)
1129                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1130
1131         /* CPU Usage */
1132         if (ts->total_run_time) {
1133                 double runt = (double) ts->total_run_time;
1134
1135                 usr_cpu = (double) ts->usr_time * 100 / runt;
1136                 sys_cpu = (double) ts->sys_time * 100 / runt;
1137         } else {
1138                 usr_cpu = 0;
1139                 sys_cpu = 0;
1140         }
1141
1142         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1143                                                 (unsigned long long) ts->ctx,
1144                                                 (unsigned long long) ts->majf,
1145                                                 (unsigned long long) ts->minf);
1146
1147         /* Calc % distribution of IO depths, usecond, msecond latency */
1148         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1149         stat_calc_lat_nu(ts, io_u_lat_u);
1150         stat_calc_lat_m(ts, io_u_lat_m);
1151
1152         /* Only show fixed 7 I/O depth levels*/
1153         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1154                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1155                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1156
1157         /* Microsecond latency */
1158         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1159                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1160         /* Millisecond latency */
1161         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1162                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1163
1164         /* disk util stats, if any */
1165         if (ver >= 3)
1166                 show_disk_util(1, NULL, out);
1167
1168         /* Additional output if continue_on_error set - default off*/
1169         if (ts->continue_on_error)
1170                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1171         if (ver == 2)
1172                 log_buf(out, "\n");
1173
1174         /* Additional output if description is set */
1175         if (strlen(ts->description))
1176                 log_buf(out, ";%s", ts->description);
1177
1178         log_buf(out, "\n");
1179 }
1180
1181 static void json_add_job_opts(struct json_object *root, const char *name,
1182                               struct flist_head *opt_list, bool num_jobs)
1183 {
1184         struct json_object *dir_object;
1185         struct flist_head *entry;
1186         struct print_option *p;
1187
1188         if (flist_empty(opt_list))
1189                 return;
1190
1191         dir_object = json_create_object();
1192         json_object_add_value_object(root, name, dir_object);
1193
1194         flist_for_each(entry, opt_list) {
1195                 const char *pos = "";
1196
1197                 p = flist_entry(entry, struct print_option, list);
1198                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1199                         continue;
1200                 if (p->value)
1201                         pos = p->value;
1202                 json_object_add_value_string(dir_object, p->name, pos);
1203         }
1204 }
1205
1206 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1207                                                    struct group_run_stats *rs,
1208                                                    struct flist_head *opt_list)
1209 {
1210         struct json_object *root, *tmp;
1211         struct jobs_eta *je;
1212         double io_u_dist[FIO_IO_U_MAP_NR];
1213         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1214         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1215         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1216         double usr_cpu, sys_cpu;
1217         int i;
1218         size_t size;
1219
1220         root = json_create_object();
1221         json_object_add_value_string(root, "jobname", ts->name);
1222         json_object_add_value_int(root, "groupid", ts->groupid);
1223         json_object_add_value_int(root, "error", ts->error);
1224
1225         /* ETA Info */
1226         je = get_jobs_eta(true, &size);
1227         if (je) {
1228                 json_object_add_value_int(root, "eta", je->eta_sec);
1229                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1230         }
1231
1232         if (opt_list)
1233                 json_add_job_opts(root, "job options", opt_list, true);
1234
1235         add_ddir_status_json(ts, rs, DDIR_READ, root);
1236         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1237         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1238
1239         /* CPU Usage */
1240         if (ts->total_run_time) {
1241                 double runt = (double) ts->total_run_time;
1242
1243                 usr_cpu = (double) ts->usr_time * 100 / runt;
1244                 sys_cpu = (double) ts->sys_time * 100 / runt;
1245         } else {
1246                 usr_cpu = 0;
1247                 sys_cpu = 0;
1248         }
1249         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1250         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1251         json_object_add_value_int(root, "ctx", ts->ctx);
1252         json_object_add_value_int(root, "majf", ts->majf);
1253         json_object_add_value_int(root, "minf", ts->minf);
1254
1255
1256         /* Calc % distribution of IO depths, usecond, msecond latency */
1257         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1258         stat_calc_lat_n(ts, io_u_lat_n);
1259         stat_calc_lat_u(ts, io_u_lat_u);
1260         stat_calc_lat_m(ts, io_u_lat_m);
1261
1262         tmp = json_create_object();
1263         json_object_add_value_object(root, "iodepth_level", tmp);
1264         /* Only show fixed 7 I/O depth levels*/
1265         for (i = 0; i < 7; i++) {
1266                 char name[20];
1267                 if (i < 6)
1268                         snprintf(name, 20, "%d", 1 << i);
1269                 else
1270                         snprintf(name, 20, ">=%d", 1 << i);
1271                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1272         }
1273
1274         /* Nanosecond latency */
1275         tmp = json_create_object();
1276         json_object_add_value_object(root, "latency_ns", tmp);
1277         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1278                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1279                                  "250", "500", "750", "1000", };
1280                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1281         }
1282         /* Microsecond latency */
1283         tmp = json_create_object();
1284         json_object_add_value_object(root, "latency_us", tmp);
1285         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1286                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1287                                  "250", "500", "750", "1000", };
1288                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1289         }
1290         /* Millisecond latency */
1291         tmp = json_create_object();
1292         json_object_add_value_object(root, "latency_ms", tmp);
1293         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1294                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1295                                  "250", "500", "750", "1000", "2000",
1296                                  ">=2000", };
1297                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1298         }
1299
1300         /* Additional output if continue_on_error set - default off*/
1301         if (ts->continue_on_error) {
1302                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1303                 json_object_add_value_int(root, "first_error", ts->first_error);
1304         }
1305
1306         if (ts->latency_depth) {
1307                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1308                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1309                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1310                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1311         }
1312
1313         /* Additional output if description is set */
1314         if (strlen(ts->description))
1315                 json_object_add_value_string(root, "desc", ts->description);
1316
1317         if (ts->nr_block_infos) {
1318                 /* Block error histogram and types */
1319                 int len;
1320                 unsigned int *percentiles = NULL;
1321                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1322
1323                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1324                                              ts->percentile_list,
1325                                              &percentiles, block_state_counts);
1326
1327                 if (len) {
1328                         struct json_object *block, *percentile_object, *states;
1329                         int state;
1330                         block = json_create_object();
1331                         json_object_add_value_object(root, "block", block);
1332
1333                         percentile_object = json_create_object();
1334                         json_object_add_value_object(block, "percentiles",
1335                                                      percentile_object);
1336                         for (i = 0; i < len; i++) {
1337                                 char buf[20];
1338                                 snprintf(buf, sizeof(buf), "%f",
1339                                          ts->percentile_list[i].u.f);
1340                                 json_object_add_value_int(percentile_object,
1341                                                           (const char *)buf,
1342                                                           percentiles[i]);
1343                         }
1344
1345                         states = json_create_object();
1346                         json_object_add_value_object(block, "states", states);
1347                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1348                                 json_object_add_value_int(states,
1349                                         block_state_names[state],
1350                                         block_state_counts[state]);
1351                         }
1352                         free(percentiles);
1353                 }
1354         }
1355
1356         if (ts->ss_dur) {
1357                 struct json_object *data;
1358                 struct json_array *iops, *bw;
1359                 int i, j, k;
1360                 char ss_buf[64];
1361
1362                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1363                         ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1364                         ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1365                         (float) ts->ss_limit.u.f,
1366                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1367
1368                 tmp = json_create_object();
1369                 json_object_add_value_object(root, "steadystate", tmp);
1370                 json_object_add_value_string(tmp, "ss", ss_buf);
1371                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1372                 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1373
1374                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1375                         ts->ss_state & FIO_SS_PCT ? "%" : "");
1376                 json_object_add_value_string(tmp, "criterion", ss_buf);
1377                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1378                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1379
1380                 data = json_create_object();
1381                 json_object_add_value_object(tmp, "data", data);
1382                 bw = json_create_array();
1383                 iops = json_create_array();
1384
1385                 /*
1386                 ** if ss was attained or the buffer is not full,
1387                 ** ss->head points to the first element in the list.
1388                 ** otherwise it actually points to the second element
1389                 ** in the list
1390                 */
1391                 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1392                         j = ts->ss_head;
1393                 else
1394                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1395                 for (i = 0; i < ts->ss_dur; i++) {
1396                         k = (j + i) % ts->ss_dur;
1397                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1398                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1399                 }
1400                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1401                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1402                 json_object_add_value_array(data, "iops", iops);
1403                 json_object_add_value_array(data, "bw", bw);
1404         }
1405
1406         return root;
1407 }
1408
1409 static void show_thread_status_terse(struct thread_stat *ts,
1410                                      struct group_run_stats *rs,
1411                                      struct buf_output *out)
1412 {
1413         if (terse_version >= 2 && terse_version <= 5)
1414                 show_thread_status_terse_all(ts, rs, terse_version, out);
1415         else
1416                 log_err("fio: bad terse version!? %d\n", terse_version);
1417 }
1418
1419 struct json_object *show_thread_status(struct thread_stat *ts,
1420                                        struct group_run_stats *rs,
1421                                        struct flist_head *opt_list,
1422                                        struct buf_output *out)
1423 {
1424         struct json_object *ret = NULL;
1425
1426         if (output_format & FIO_OUTPUT_TERSE)
1427                 show_thread_status_terse(ts, rs,  out);
1428         if (output_format & FIO_OUTPUT_JSON)
1429                 ret = show_thread_status_json(ts, rs, opt_list);
1430         if (output_format & FIO_OUTPUT_NORMAL)
1431                 show_thread_status_normal(ts, rs,  out);
1432
1433         return ret;
1434 }
1435
1436 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1437 {
1438         double mean, S;
1439
1440         if (src->samples == 0)
1441                 return;
1442
1443         dst->min_val = min(dst->min_val, src->min_val);
1444         dst->max_val = max(dst->max_val, src->max_val);
1445
1446         /*
1447          * Compute new mean and S after the merge
1448          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1449          *  #Parallel_algorithm>
1450          */
1451         if (first) {
1452                 mean = src->mean.u.f;
1453                 S = src->S.u.f;
1454         } else {
1455                 double delta = src->mean.u.f - dst->mean.u.f;
1456
1457                 mean = ((src->mean.u.f * src->samples) +
1458                         (dst->mean.u.f * dst->samples)) /
1459                         (dst->samples + src->samples);
1460
1461                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1462                         (dst->samples * src->samples) /
1463                         (dst->samples + src->samples);
1464         }
1465
1466         dst->samples += src->samples;
1467         dst->mean.u.f = mean;
1468         dst->S.u.f = S;
1469 }
1470
1471 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1472 {
1473         int i;
1474
1475         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1476                 if (dst->max_run[i] < src->max_run[i])
1477                         dst->max_run[i] = src->max_run[i];
1478                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1479                         dst->min_run[i] = src->min_run[i];
1480                 if (dst->max_bw[i] < src->max_bw[i])
1481                         dst->max_bw[i] = src->max_bw[i];
1482                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1483                         dst->min_bw[i] = src->min_bw[i];
1484
1485                 dst->iobytes[i] += src->iobytes[i];
1486                 dst->agg[i] += src->agg[i];
1487         }
1488
1489         if (!dst->kb_base)
1490                 dst->kb_base = src->kb_base;
1491         if (!dst->unit_base)
1492                 dst->unit_base = src->unit_base;
1493 }
1494
1495 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1496                       bool first)
1497 {
1498         int l, k;
1499
1500         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1501                 if (!dst->unified_rw_rep) {
1502                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1503                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1504                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1505                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1506                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1507
1508                         dst->io_bytes[l] += src->io_bytes[l];
1509
1510                         if (dst->runtime[l] < src->runtime[l])
1511                                 dst->runtime[l] = src->runtime[l];
1512                 } else {
1513                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1514                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1515                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1516                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1517                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1518
1519                         dst->io_bytes[0] += src->io_bytes[l];
1520
1521                         if (dst->runtime[0] < src->runtime[l])
1522                                 dst->runtime[0] = src->runtime[l];
1523
1524                         /*
1525                          * We're summing to the same destination, so override
1526                          * 'first' after the first iteration of the loop
1527                          */
1528                         first = false;
1529                 }
1530         }
1531
1532         dst->usr_time += src->usr_time;
1533         dst->sys_time += src->sys_time;
1534         dst->ctx += src->ctx;
1535         dst->majf += src->majf;
1536         dst->minf += src->minf;
1537
1538         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1539                 dst->io_u_map[k] += src->io_u_map[k];
1540         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1541                 dst->io_u_submit[k] += src->io_u_submit[k];
1542         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1543                 dst->io_u_complete[k] += src->io_u_complete[k];
1544         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1545                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1546         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1547                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1548         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1549                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1550
1551         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1552                 if (!dst->unified_rw_rep) {
1553                         dst->total_io_u[k] += src->total_io_u[k];
1554                         dst->short_io_u[k] += src->short_io_u[k];
1555                         dst->drop_io_u[k] += src->drop_io_u[k];
1556                 } else {
1557                         dst->total_io_u[0] += src->total_io_u[k];
1558                         dst->short_io_u[0] += src->short_io_u[k];
1559                         dst->drop_io_u[0] += src->drop_io_u[k];
1560                 }
1561         }
1562
1563         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1564                 int m;
1565
1566                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1567                         if (!dst->unified_rw_rep)
1568                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1569                         else
1570                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1571                 }
1572         }
1573
1574         dst->total_run_time += src->total_run_time;
1575         dst->total_submit += src->total_submit;
1576         dst->total_complete += src->total_complete;
1577 }
1578
1579 void init_group_run_stat(struct group_run_stats *gs)
1580 {
1581         int i;
1582         memset(gs, 0, sizeof(*gs));
1583
1584         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1585                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1586 }
1587
1588 void init_thread_stat(struct thread_stat *ts)
1589 {
1590         int j;
1591
1592         memset(ts, 0, sizeof(*ts));
1593
1594         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1595                 ts->lat_stat[j].min_val = -1UL;
1596                 ts->clat_stat[j].min_val = -1UL;
1597                 ts->slat_stat[j].min_val = -1UL;
1598                 ts->bw_stat[j].min_val = -1UL;
1599                 ts->iops_stat[j].min_val = -1UL;
1600         }
1601         ts->groupid = -1;
1602 }
1603
1604 void __show_run_stats(void)
1605 {
1606         struct group_run_stats *runstats, *rs;
1607         struct thread_data *td;
1608         struct thread_stat *threadstats, *ts;
1609         int i, j, k, nr_ts, last_ts, idx;
1610         bool kb_base_warned = false;
1611         bool unit_base_warned = false;
1612         struct json_object *root = NULL;
1613         struct json_array *array = NULL;
1614         struct buf_output output[FIO_OUTPUT_NR];
1615         struct flist_head **opt_lists;
1616
1617         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1618
1619         for (i = 0; i < groupid + 1; i++)
1620                 init_group_run_stat(&runstats[i]);
1621
1622         /*
1623          * find out how many threads stats we need. if group reporting isn't
1624          * enabled, it's one-per-td.
1625          */
1626         nr_ts = 0;
1627         last_ts = -1;
1628         for_each_td(td, i) {
1629                 if (!td->o.group_reporting) {
1630                         nr_ts++;
1631                         continue;
1632                 }
1633                 if (last_ts == td->groupid)
1634                         continue;
1635                 if (!td->o.stats)
1636                         continue;
1637
1638                 last_ts = td->groupid;
1639                 nr_ts++;
1640         }
1641
1642         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1643         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1644
1645         for (i = 0; i < nr_ts; i++) {
1646                 init_thread_stat(&threadstats[i]);
1647                 opt_lists[i] = NULL;
1648         }
1649
1650         j = 0;
1651         last_ts = -1;
1652         idx = 0;
1653         for_each_td(td, i) {
1654                 if (!td->o.stats)
1655                         continue;
1656                 if (idx && (!td->o.group_reporting ||
1657                     (td->o.group_reporting && last_ts != td->groupid))) {
1658                         idx = 0;
1659                         j++;
1660                 }
1661
1662                 last_ts = td->groupid;
1663
1664                 ts = &threadstats[j];
1665
1666                 ts->clat_percentiles = td->o.clat_percentiles;
1667                 ts->lat_percentiles = td->o.lat_percentiles;
1668                 ts->percentile_precision = td->o.percentile_precision;
1669                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1670                 opt_lists[j] = &td->opt_list;
1671
1672                 idx++;
1673                 ts->members++;
1674
1675                 if (ts->groupid == -1) {
1676                         /*
1677                          * These are per-group shared already
1678                          */
1679                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1680                         if (td->o.description)
1681                                 strncpy(ts->description, td->o.description,
1682                                                 FIO_JOBDESC_SIZE - 1);
1683                         else
1684                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1685
1686                         /*
1687                          * If multiple entries in this group, this is
1688                          * the first member.
1689                          */
1690                         ts->thread_number = td->thread_number;
1691                         ts->groupid = td->groupid;
1692
1693                         /*
1694                          * first pid in group, not very useful...
1695                          */
1696                         ts->pid = td->pid;
1697
1698                         ts->kb_base = td->o.kb_base;
1699                         ts->unit_base = td->o.unit_base;
1700                         ts->sig_figs = td->o.sig_figs;
1701                         ts->unified_rw_rep = td->o.unified_rw_rep;
1702                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1703                         log_info("fio: kb_base differs for jobs in group, using"
1704                                  " %u as the base\n", ts->kb_base);
1705                         kb_base_warned = true;
1706                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1707                         log_info("fio: unit_base differs for jobs in group, using"
1708                                  " %u as the base\n", ts->unit_base);
1709                         unit_base_warned = true;
1710                 }
1711
1712                 ts->continue_on_error = td->o.continue_on_error;
1713                 ts->total_err_count += td->total_err_count;
1714                 ts->first_error = td->first_error;
1715                 if (!ts->error) {
1716                         if (!td->error && td->o.continue_on_error &&
1717                             td->first_error) {
1718                                 ts->error = td->first_error;
1719                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1720                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1721                         } else  if (td->error) {
1722                                 ts->error = td->error;
1723                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1724                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1725                         }
1726                 }
1727
1728                 ts->latency_depth = td->latency_qd;
1729                 ts->latency_target = td->o.latency_target;
1730                 ts->latency_percentile = td->o.latency_percentile;
1731                 ts->latency_window = td->o.latency_window;
1732
1733                 ts->nr_block_infos = td->ts.nr_block_infos;
1734                 for (k = 0; k < ts->nr_block_infos; k++)
1735                         ts->block_infos[k] = td->ts.block_infos[k];
1736
1737                 sum_thread_stats(ts, &td->ts, idx == 1);
1738
1739                 if (td->o.ss_dur) {
1740                         ts->ss_state = td->ss.state;
1741                         ts->ss_dur = td->ss.dur;
1742                         ts->ss_head = td->ss.head;
1743                         ts->ss_bw_data = td->ss.bw_data;
1744                         ts->ss_iops_data = td->ss.iops_data;
1745                         ts->ss_limit.u.f = td->ss.limit;
1746                         ts->ss_slope.u.f = td->ss.slope;
1747                         ts->ss_deviation.u.f = td->ss.deviation;
1748                         ts->ss_criterion.u.f = td->ss.criterion;
1749                 }
1750                 else
1751                         ts->ss_dur = ts->ss_state = 0;
1752         }
1753
1754         for (i = 0; i < nr_ts; i++) {
1755                 unsigned long long bw;
1756
1757                 ts = &threadstats[i];
1758                 if (ts->groupid == -1)
1759                         continue;
1760                 rs = &runstats[ts->groupid];
1761                 rs->kb_base = ts->kb_base;
1762                 rs->unit_base = ts->unit_base;
1763                 rs->sig_figs = ts->sig_figs;
1764                 rs->unified_rw_rep += ts->unified_rw_rep;
1765
1766                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1767                         if (!ts->runtime[j])
1768                                 continue;
1769                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1770                                 rs->min_run[j] = ts->runtime[j];
1771                         if (ts->runtime[j] > rs->max_run[j])
1772                                 rs->max_run[j] = ts->runtime[j];
1773
1774                         bw = 0;
1775                         if (ts->runtime[j])
1776                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1777                         if (bw < rs->min_bw[j])
1778                                 rs->min_bw[j] = bw;
1779                         if (bw > rs->max_bw[j])
1780                                 rs->max_bw[j] = bw;
1781
1782                         rs->iobytes[j] += ts->io_bytes[j];
1783                 }
1784         }
1785
1786         for (i = 0; i < groupid + 1; i++) {
1787                 int ddir;
1788
1789                 rs = &runstats[i];
1790
1791                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1792                         if (rs->max_run[ddir])
1793                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1794                                                 rs->max_run[ddir];
1795                 }
1796         }
1797
1798         for (i = 0; i < FIO_OUTPUT_NR; i++)
1799                 buf_output_init(&output[i]);
1800
1801         /*
1802          * don't overwrite last signal output
1803          */
1804         if (output_format & FIO_OUTPUT_NORMAL)
1805                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1806         if (output_format & FIO_OUTPUT_JSON) {
1807                 struct thread_data *global;
1808                 char time_buf[32];
1809                 struct timeval now;
1810                 unsigned long long ms_since_epoch;
1811
1812                 gettimeofday(&now, NULL);
1813                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1814                                  (unsigned long long)(now.tv_usec) / 1000;
1815
1816                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1817                                 sizeof(time_buf));
1818                 if (time_buf[strlen(time_buf) - 1] == '\n')
1819                         time_buf[strlen(time_buf) - 1] = '\0';
1820
1821                 root = json_create_object();
1822                 json_object_add_value_string(root, "fio version", fio_version_string);
1823                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1824                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1825                 json_object_add_value_string(root, "time", time_buf);
1826                 global = get_global_options();
1827                 json_add_job_opts(root, "global options", &global->opt_list, false);
1828                 array = json_create_array();
1829                 json_object_add_value_array(root, "jobs", array);
1830         }
1831
1832         if (is_backend)
1833                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1834
1835         for (i = 0; i < nr_ts; i++) {
1836                 ts = &threadstats[i];
1837                 rs = &runstats[ts->groupid];
1838
1839                 if (is_backend) {
1840                         fio_server_send_job_options(opt_lists[i], i);
1841                         fio_server_send_ts(ts, rs);
1842                 } else {
1843                         if (output_format & FIO_OUTPUT_TERSE)
1844                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1845                         if (output_format & FIO_OUTPUT_JSON) {
1846                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1847                                 json_array_add_value_object(array, tmp);
1848                         }
1849                         if (output_format & FIO_OUTPUT_NORMAL)
1850                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1851                 }
1852         }
1853         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1854                 /* disk util stats, if any */
1855                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1856
1857                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1858
1859                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1860                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1861                 json_free_object(root);
1862         }
1863
1864         for (i = 0; i < groupid + 1; i++) {
1865                 rs = &runstats[i];
1866
1867                 rs->groupid = i;
1868                 if (is_backend)
1869                         fio_server_send_gs(rs);
1870                 else if (output_format & FIO_OUTPUT_NORMAL)
1871                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1872         }
1873
1874         if (is_backend)
1875                 fio_server_send_du();
1876         else if (output_format & FIO_OUTPUT_NORMAL) {
1877                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1878                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1879         }
1880
1881         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1882                 struct buf_output *out = &output[i];
1883
1884                 log_info_buf(out->buf, out->buflen);
1885                 buf_output_free(out);
1886         }
1887
1888         log_info_flush();
1889         free(runstats);
1890         free(threadstats);
1891         free(opt_lists);
1892 }
1893
1894 void show_run_stats(void)
1895 {
1896         fio_mutex_down(stat_mutex);
1897         __show_run_stats();
1898         fio_mutex_up(stat_mutex);
1899 }
1900
1901 void __show_running_run_stats(void)
1902 {
1903         struct thread_data *td;
1904         unsigned long long *rt;
1905         struct timespec ts;
1906         int i;
1907
1908         fio_mutex_down(stat_mutex);
1909
1910         rt = malloc(thread_number * sizeof(unsigned long long));
1911         fio_gettime(&ts, NULL);
1912
1913         for_each_td(td, i) {
1914                 td->update_rusage = 1;
1915                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1916                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1917                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1918                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1919
1920                 rt[i] = mtime_since(&td->start, &ts);
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         for_each_td(td, i) {
1930                 if (td->runstate >= TD_EXITED)
1931                         continue;
1932                 if (td->rusage_sem) {
1933                         td->update_rusage = 1;
1934                         fio_mutex_down(td->rusage_sem);
1935                 }
1936                 td->update_rusage = 0;
1937         }
1938
1939         __show_run_stats();
1940
1941         for_each_td(td, i) {
1942                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1943                         td->ts.runtime[DDIR_READ] -= rt[i];
1944                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1945                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1946                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1947                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1948         }
1949
1950         free(rt);
1951         fio_mutex_up(stat_mutex);
1952 }
1953
1954 static bool status_interval_init;
1955 static struct timespec status_time;
1956 static bool status_file_disabled;
1957
1958 #define FIO_STATUS_FILE         "fio-dump-status"
1959
1960 static int check_status_file(void)
1961 {
1962         struct stat sb;
1963         const char *temp_dir;
1964         char fio_status_file_path[PATH_MAX];
1965
1966         if (status_file_disabled)
1967                 return 0;
1968
1969         temp_dir = getenv("TMPDIR");
1970         if (temp_dir == NULL) {
1971                 temp_dir = getenv("TEMP");
1972                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1973                         temp_dir = NULL;
1974         }
1975         if (temp_dir == NULL)
1976                 temp_dir = "/tmp";
1977
1978         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1979
1980         if (stat(fio_status_file_path, &sb))
1981                 return 0;
1982
1983         if (unlink(fio_status_file_path) < 0) {
1984                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1985                                                         strerror(errno));
1986                 log_err("fio: disabling status file updates\n");
1987                 status_file_disabled = true;
1988         }
1989
1990         return 1;
1991 }
1992
1993 void check_for_running_stats(void)
1994 {
1995         if (status_interval) {
1996                 if (!status_interval_init) {
1997                         fio_gettime(&status_time, NULL);
1998                         status_interval_init = true;
1999                 } else if (mtime_since_now(&status_time) >= status_interval) {
2000                         show_running_run_stats();
2001                         fio_gettime(&status_time, NULL);
2002                         return;
2003                 }
2004         }
2005         if (check_status_file()) {
2006                 show_running_run_stats();
2007                 return;
2008         }
2009 }
2010
2011 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2012 {
2013         double val = data;
2014         double delta;
2015
2016         if (data > is->max_val)
2017                 is->max_val = data;
2018         if (data < is->min_val)
2019                 is->min_val = data;
2020
2021         delta = val - is->mean.u.f;
2022         if (delta) {
2023                 is->mean.u.f += delta / (is->samples + 1.0);
2024                 is->S.u.f += delta * (val - is->mean.u.f);
2025         }
2026
2027         is->samples++;
2028 }
2029
2030 /*
2031  * Return a struct io_logs, which is added to the tail of the log
2032  * list for 'iolog'.
2033  */
2034 static struct io_logs *get_new_log(struct io_log *iolog)
2035 {
2036         size_t new_size, new_samples;
2037         struct io_logs *cur_log;
2038
2039         /*
2040          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2041          * forever
2042          */
2043         if (!iolog->cur_log_max)
2044                 new_samples = DEF_LOG_ENTRIES;
2045         else {
2046                 new_samples = iolog->cur_log_max * 2;
2047                 if (new_samples > MAX_LOG_ENTRIES)
2048                         new_samples = MAX_LOG_ENTRIES;
2049         }
2050
2051         new_size = new_samples * log_entry_sz(iolog);
2052
2053         cur_log = smalloc(sizeof(*cur_log));
2054         if (cur_log) {
2055                 INIT_FLIST_HEAD(&cur_log->list);
2056                 cur_log->log = malloc(new_size);
2057                 if (cur_log->log) {
2058                         cur_log->nr_samples = 0;
2059                         cur_log->max_samples = new_samples;
2060                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2061                         iolog->cur_log_max = new_samples;
2062                         return cur_log;
2063                 }
2064                 sfree(cur_log);
2065         }
2066
2067         return NULL;
2068 }
2069
2070 /*
2071  * Add and return a new log chunk, or return current log if big enough
2072  */
2073 static struct io_logs *regrow_log(struct io_log *iolog)
2074 {
2075         struct io_logs *cur_log;
2076         int i;
2077
2078         if (!iolog || iolog->disabled)
2079                 goto disable;
2080
2081         cur_log = iolog_cur_log(iolog);
2082         if (!cur_log) {
2083                 cur_log = get_new_log(iolog);
2084                 if (!cur_log)
2085                         return NULL;
2086         }
2087
2088         if (cur_log->nr_samples < cur_log->max_samples)
2089                 return cur_log;
2090
2091         /*
2092          * No room for a new sample. If we're compressing on the fly, flush
2093          * out the current chunk
2094          */
2095         if (iolog->log_gz) {
2096                 if (iolog_cur_flush(iolog, cur_log)) {
2097                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2098                         return NULL;
2099                 }
2100         }
2101
2102         /*
2103          * Get a new log array, and add to our list
2104          */
2105         cur_log = get_new_log(iolog);
2106         if (!cur_log) {
2107                 log_err("fio: failed extending iolog! Will stop logging.\n");
2108                 return NULL;
2109         }
2110
2111         if (!iolog->pending || !iolog->pending->nr_samples)
2112                 return cur_log;
2113
2114         /*
2115          * Flush pending items to new log
2116          */
2117         for (i = 0; i < iolog->pending->nr_samples; i++) {
2118                 struct io_sample *src, *dst;
2119
2120                 src = get_sample(iolog, iolog->pending, i);
2121                 dst = get_sample(iolog, cur_log, i);
2122                 memcpy(dst, src, log_entry_sz(iolog));
2123         }
2124         cur_log->nr_samples = iolog->pending->nr_samples;
2125
2126         iolog->pending->nr_samples = 0;
2127         return cur_log;
2128 disable:
2129         if (iolog)
2130                 iolog->disabled = true;
2131         return NULL;
2132 }
2133
2134 void regrow_logs(struct thread_data *td)
2135 {
2136         regrow_log(td->slat_log);
2137         regrow_log(td->clat_log);
2138         regrow_log(td->clat_hist_log);
2139         regrow_log(td->lat_log);
2140         regrow_log(td->bw_log);
2141         regrow_log(td->iops_log);
2142         td->flags &= ~TD_F_REGROW_LOGS;
2143 }
2144
2145 static struct io_logs *get_cur_log(struct io_log *iolog)
2146 {
2147         struct io_logs *cur_log;
2148
2149         cur_log = iolog_cur_log(iolog);
2150         if (!cur_log) {
2151                 cur_log = get_new_log(iolog);
2152                 if (!cur_log)
2153                         return NULL;
2154         }
2155
2156         if (cur_log->nr_samples < cur_log->max_samples)
2157                 return cur_log;
2158
2159         /*
2160          * Out of space. If we're in IO offload mode, or we're not doing
2161          * per unit logging (hence logging happens outside of the IO thread
2162          * as well), add a new log chunk inline. If we're doing inline
2163          * submissions, flag 'td' as needing a log regrow and we'll take
2164          * care of it on the submission side.
2165          */
2166         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2167             !per_unit_log(iolog))
2168                 return regrow_log(iolog);
2169
2170         iolog->td->flags |= TD_F_REGROW_LOGS;
2171         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2172         return iolog->pending;
2173 }
2174
2175 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2176                              enum fio_ddir ddir, unsigned int bs,
2177                              unsigned long t, uint64_t offset)
2178 {
2179         struct io_logs *cur_log;
2180
2181         if (iolog->disabled)
2182                 return;
2183         if (flist_empty(&iolog->io_logs))
2184                 iolog->avg_last[ddir] = t;
2185
2186         cur_log = get_cur_log(iolog);
2187         if (cur_log) {
2188                 struct io_sample *s;
2189
2190                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2191
2192                 s->data = data;
2193                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2194                 io_sample_set_ddir(iolog, s, ddir);
2195                 s->bs = bs;
2196
2197                 if (iolog->log_offset) {
2198                         struct io_sample_offset *so = (void *) s;
2199
2200                         so->offset = offset;
2201                 }
2202
2203                 cur_log->nr_samples++;
2204                 return;
2205         }
2206
2207         iolog->disabled = true;
2208 }
2209
2210 static inline void reset_io_stat(struct io_stat *ios)
2211 {
2212         ios->max_val = ios->min_val = ios->samples = 0;
2213         ios->mean.u.f = ios->S.u.f = 0;
2214 }
2215
2216 void reset_io_stats(struct thread_data *td)
2217 {
2218         struct thread_stat *ts = &td->ts;
2219         int i, j;
2220
2221         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2222                 reset_io_stat(&ts->clat_stat[i]);
2223                 reset_io_stat(&ts->slat_stat[i]);
2224                 reset_io_stat(&ts->lat_stat[i]);
2225                 reset_io_stat(&ts->bw_stat[i]);
2226                 reset_io_stat(&ts->iops_stat[i]);
2227
2228                 ts->io_bytes[i] = 0;
2229                 ts->runtime[i] = 0;
2230                 ts->total_io_u[i] = 0;
2231                 ts->short_io_u[i] = 0;
2232                 ts->drop_io_u[i] = 0;
2233
2234                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2235                         ts->io_u_plat[i][j] = 0;
2236         }
2237
2238         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2239                 ts->io_u_map[i] = 0;
2240                 ts->io_u_submit[i] = 0;
2241                 ts->io_u_complete[i] = 0;
2242         }
2243
2244         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2245                 ts->io_u_lat_n[i] = 0;
2246         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2247                 ts->io_u_lat_u[i] = 0;
2248         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2249                 ts->io_u_lat_m[i] = 0;
2250
2251         ts->total_submit = 0;
2252         ts->total_complete = 0;
2253 }
2254
2255 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2256                               unsigned long elapsed, bool log_max)
2257 {
2258         /*
2259          * Note an entry in the log. Use the mean from the logged samples,
2260          * making sure to properly round up. Only write a log entry if we
2261          * had actual samples done.
2262          */
2263         if (iolog->avg_window[ddir].samples) {
2264                 union io_sample_data data;
2265
2266                 if (log_max)
2267                         data.val = iolog->avg_window[ddir].max_val;
2268                 else
2269                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2270
2271                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2272         }
2273
2274         reset_io_stat(&iolog->avg_window[ddir]);
2275 }
2276
2277 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2278                              bool log_max)
2279 {
2280         int ddir;
2281
2282         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2283                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2284 }
2285
2286 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2287                            union io_sample_data data, enum fio_ddir ddir,
2288                            unsigned int bs, uint64_t offset)
2289 {
2290         unsigned long elapsed, this_window;
2291
2292         if (!ddir_rw(ddir))
2293                 return 0;
2294
2295         elapsed = mtime_since_now(&td->epoch);
2296
2297         /*
2298          * If no time averaging, just add the log sample.
2299          */
2300         if (!iolog->avg_msec) {
2301                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2302                 return 0;
2303         }
2304
2305         /*
2306          * Add the sample. If the time period has passed, then
2307          * add that entry to the log and clear.
2308          */
2309         add_stat_sample(&iolog->avg_window[ddir], data.val);
2310
2311         /*
2312          * If period hasn't passed, adding the above sample is all we
2313          * need to do.
2314          */
2315         this_window = elapsed - iolog->avg_last[ddir];
2316         if (elapsed < iolog->avg_last[ddir])
2317                 return iolog->avg_last[ddir] - elapsed;
2318         else if (this_window < iolog->avg_msec) {
2319                 int diff = iolog->avg_msec - this_window;
2320
2321                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2322                         return diff;
2323         }
2324
2325         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2326
2327         iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2328         return iolog->avg_msec;
2329 }
2330
2331 void finalize_logs(struct thread_data *td, bool unit_logs)
2332 {
2333         unsigned long elapsed;
2334
2335         elapsed = mtime_since_now(&td->epoch);
2336
2337         if (td->clat_log && unit_logs)
2338                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2339         if (td->slat_log && unit_logs)
2340                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2341         if (td->lat_log && unit_logs)
2342                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2343         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2344                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2345         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2346                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2347 }
2348
2349 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2350 {
2351         struct io_log *iolog;
2352
2353         if (!ddir_rw(ddir))
2354                 return;
2355
2356         iolog = agg_io_log[ddir];
2357         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2358 }
2359
2360 static void add_clat_percentile_sample(struct thread_stat *ts,
2361                                 unsigned long long nsec, enum fio_ddir ddir)
2362 {
2363         unsigned int idx = plat_val_to_idx(nsec);
2364         assert(idx < FIO_IO_U_PLAT_NR);
2365
2366         ts->io_u_plat[ddir][idx]++;
2367 }
2368
2369 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2370                      unsigned long long nsec, unsigned int bs, uint64_t offset)
2371 {
2372         unsigned long elapsed, this_window;
2373         struct thread_stat *ts = &td->ts;
2374         struct io_log *iolog = td->clat_hist_log;
2375
2376         td_io_u_lock(td);
2377
2378         add_stat_sample(&ts->clat_stat[ddir], nsec);
2379
2380         if (td->clat_log)
2381                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2382                                offset);
2383
2384         if (ts->clat_percentiles)
2385                 add_clat_percentile_sample(ts, nsec, ddir);
2386
2387         if (iolog && iolog->hist_msec) {
2388                 struct io_hist *hw = &iolog->hist_window[ddir];
2389
2390                 hw->samples++;
2391                 elapsed = mtime_since_now(&td->epoch);
2392                 if (!hw->hist_last)
2393                         hw->hist_last = elapsed;
2394                 this_window = elapsed - hw->hist_last;
2395
2396                 if (this_window >= iolog->hist_msec) {
2397                         unsigned int *io_u_plat;
2398                         struct io_u_plat_entry *dst;
2399
2400                         /*
2401                          * Make a byte-for-byte copy of the latency histogram
2402                          * stored in td->ts.io_u_plat[ddir], recording it in a
2403                          * log sample. Note that the matching call to free() is
2404                          * located in iolog.c after printing this sample to the
2405                          * log file.
2406                          */
2407                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2408                         dst = malloc(sizeof(struct io_u_plat_entry));
2409                         memcpy(&(dst->io_u_plat), io_u_plat,
2410                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2411                         flist_add(&dst->list, &hw->list);
2412                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2413                                                 elapsed, offset);
2414
2415                         /*
2416                          * Update the last time we recorded as being now, minus
2417                          * any drift in time we encountered before actually
2418                          * making the record.
2419                          */
2420                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2421                         hw->samples = 0;
2422                 }
2423         }
2424
2425         td_io_u_unlock(td);
2426 }
2427
2428 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2429                      unsigned long usec, unsigned int bs, uint64_t offset)
2430 {
2431         struct thread_stat *ts = &td->ts;
2432
2433         if (!ddir_rw(ddir))
2434                 return;
2435
2436         td_io_u_lock(td);
2437
2438         add_stat_sample(&ts->slat_stat[ddir], usec);
2439
2440         if (td->slat_log)
2441                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2442
2443         td_io_u_unlock(td);
2444 }
2445
2446 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2447                     unsigned long long nsec, unsigned int bs, uint64_t offset)
2448 {
2449         struct thread_stat *ts = &td->ts;
2450
2451         if (!ddir_rw(ddir))
2452                 return;
2453
2454         td_io_u_lock(td);
2455
2456         add_stat_sample(&ts->lat_stat[ddir], nsec);
2457
2458         if (td->lat_log)
2459                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2460                                offset);
2461
2462         if (ts->lat_percentiles)
2463                 add_clat_percentile_sample(ts, nsec, ddir);
2464
2465         td_io_u_unlock(td);
2466 }
2467
2468 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2469                    unsigned int bytes, unsigned long long spent)
2470 {
2471         struct thread_stat *ts = &td->ts;
2472         unsigned long rate;
2473
2474         if (spent)
2475                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2476         else
2477                 rate = 0;
2478
2479         td_io_u_lock(td);
2480
2481         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2482
2483         if (td->bw_log)
2484                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2485                                bytes, io_u->offset);
2486
2487         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2488         td_io_u_unlock(td);
2489 }
2490
2491 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2492                          struct timespec *t, unsigned int avg_time,
2493                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2494                          struct io_stat *stat, struct io_log *log,
2495                          bool is_kb)
2496 {
2497         unsigned long spent, rate;
2498         enum fio_ddir ddir;
2499         unsigned int next, next_log;
2500
2501         next_log = avg_time;
2502
2503         spent = mtime_since(parent_tv, t);
2504         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2505                 return avg_time - spent;
2506
2507         td_io_u_lock(td);
2508
2509         /*
2510          * Compute both read and write rates for the interval.
2511          */
2512         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2513                 uint64_t delta;
2514
2515                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2516                 if (!delta)
2517                         continue; /* No entries for interval */
2518
2519                 if (spent) {
2520                         if (is_kb)
2521                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2522                         else
2523                                 rate = (delta * 1000) / spent;
2524                 } else
2525                         rate = 0;
2526
2527                 add_stat_sample(&stat[ddir], rate);
2528
2529                 if (log) {
2530                         unsigned int bs = 0;
2531
2532                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2533                                 bs = td->o.min_bs[ddir];
2534
2535                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2536                         next_log = min(next_log, next);
2537                 }
2538
2539                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2540         }
2541
2542         timespec_add_msec(parent_tv, avg_time);
2543
2544         td_io_u_unlock(td);
2545
2546         if (spent <= avg_time)
2547                 next = avg_time;
2548         else
2549                 next = avg_time - (1 + spent - avg_time);
2550
2551         return min(next, next_log);
2552 }
2553
2554 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2555 {
2556         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2557                                 td->this_io_bytes, td->stat_io_bytes,
2558                                 td->ts.bw_stat, td->bw_log, true);
2559 }
2560
2561 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2562                      unsigned int bytes)
2563 {
2564         struct thread_stat *ts = &td->ts;
2565
2566         td_io_u_lock(td);
2567
2568         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2569
2570         if (td->iops_log)
2571                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2572                                bytes, io_u->offset);
2573
2574         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2575         td_io_u_unlock(td);
2576 }
2577
2578 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2579 {
2580         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2581                                 td->this_io_blocks, td->stat_io_blocks,
2582                                 td->ts.iops_stat, td->iops_log, false);
2583 }
2584
2585 /*
2586  * Returns msecs to next event
2587  */
2588 int calc_log_samples(void)
2589 {
2590         struct thread_data *td;
2591         unsigned int next = ~0U, tmp;
2592         struct timespec now;
2593         int i;
2594
2595         fio_gettime(&now, NULL);
2596
2597         for_each_td(td, i) {
2598                 if (!td->o.stats)
2599                         continue;
2600                 if (in_ramp_time(td) ||
2601                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2602                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2603                         continue;
2604                 }
2605                 if (!td->bw_log ||
2606                         (td->bw_log && !per_unit_log(td->bw_log))) {
2607                         tmp = add_bw_samples(td, &now);
2608                         if (tmp < next)
2609                                 next = tmp;
2610                 }
2611                 if (!td->iops_log ||
2612                         (td->iops_log && !per_unit_log(td->iops_log))) {
2613                         tmp = add_iops_samples(td, &now);
2614                         if (tmp < next)
2615                                 next = tmp;
2616                 }
2617         }
2618
2619         return next == ~0U ? 0 : next;
2620 }
2621
2622 void stat_init(void)
2623 {
2624         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2625 }
2626
2627 void stat_exit(void)
2628 {
2629         /*
2630          * When we have the mutex, we know out-of-band access to it
2631          * have ended.
2632          */
2633         fio_mutex_down(stat_mutex);
2634         fio_mutex_remove(stat_mutex);
2635 }
2636
2637 /*
2638  * Called from signal handler. Wake up status thread.
2639  */
2640 void show_running_run_stats(void)
2641 {
2642         helper_do_stat();
2643 }
2644
2645 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2646 {
2647         /* Ignore io_u's which span multiple blocks--they will just get
2648          * inaccurate counts. */
2649         int idx = (io_u->offset - io_u->file->file_offset)
2650                         / td->o.bs[DDIR_TRIM];
2651         uint32_t *info = &td->ts.block_infos[idx];
2652         assert(idx < td->ts.nr_block_infos);
2653         return info;
2654 }