stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/types.h>
   5 #include <sys/stat.h>
   6 #include <dirent.h>
   7 #include <libgen.h>
   8 #include <math.h>
   9
  10 #include "fio.h"
  11 #include "diskutil.h"
  12 #include "lib/ieee754.h"
  13 #include "json.h"
  14 #include "lib/getrusage.h"
  15 #include "idletime.h"
  16 #include "lib/pow2.h"
  17 #include "lib/output_buffer.h"
  18 #include "helper_thread.h"
  19 #include "smalloc.h"
  20
  21 #define LOG_MSEC_SLACK  10
  22
  23 struct fio_mutex *stat_mutex;
  24
  25 void clear_rusage_stat(struct thread_data *td)
  26 {
  27         struct thread_stat *ts = &td->ts;
  28
  29         fio_getrusage(&td->ru_start);
  30         ts->usr_time = ts->sys_time = 0;
  31         ts->ctx = 0;
  32         ts->minf = ts->majf = 0;
  33 }
  34
  35 void update_rusage_stat(struct thread_data *td)
  36 {
  37         struct thread_stat *ts = &td->ts;
  38
  39         fio_getrusage(&td->ru_end);
  40         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
  41                                         &td->ru_end.ru_utime);
  42         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
  43                                         &td->ru_end.ru_stime);
  44         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  45                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  46         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  47         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  48
  49         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  50 }
  51
  52 /*
  53  * Given a latency, return the index of the corresponding bucket in
  54  * the structure tracking percentiles.
  55  *
  56  * (1) find the group (and error bits) that the value (latency)
  57  * belongs to by looking at its MSB. (2) find the bucket number in the
  58  * group by looking at the index bits.
  59  *
  60  */
  61 static unsigned int plat_val_to_idx(unsigned long long val)
  62 {
  63         unsigned int msb, error_bits, base, offset, idx;
  64
  65         /* Find MSB starting from bit 0 */
  66         if (val == 0)
  67                 msb = 0;
  68         else
  69                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
  70
  71         /*
  72          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  73          * all bits of the sample as index
  74          */
  75         if (msb <= FIO_IO_U_PLAT_BITS)
  76                 return val;
  77
  78         /* Compute the number of error bits to discard*/
  79         error_bits = msb - FIO_IO_U_PLAT_BITS;
  80
  81         /* Compute the number of buckets before the group */
  82         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  83
  84         /*
  85          * Discard the error bits and apply the mask to find the
  86          * index for the buckets in the group
  87          */
  88         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  89
  90         /* Make sure the index does not exceed (array size - 1) */
  91         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  92                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  93
  94         return idx;
  95 }
  96
  97 /*
  98  * Convert the given index of the bucket array to the value
  99  * represented by the bucket
 100  */
 101 static unsigned long long plat_idx_to_val(unsigned int idx)
 102 {
 103         unsigned int error_bits, k, base;
 104
 105         assert(idx < FIO_IO_U_PLAT_NR);
 106
 107         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 108          * all bits of the sample as index */
 109         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 110                 return idx;
 111
 112         /* Find the group and compute the minimum value of that group */
 113         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 114         base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
 115
 116         /* Find its bucket number of the group */
 117         k = idx % FIO_IO_U_PLAT_VAL;
 118
 119         /* Return the mean of the range of the bucket */
 120         return base + ((k + 0.5) * (1 << error_bits));
 121 }
 122
 123 static int double_cmp(const void *a, const void *b)
 124 {
 125         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 126         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 127         int cmp = 0;
 128
 129         if (fa.u.f > fb.u.f)
 130                 cmp = 1;
 131         else if (fa.u.f < fb.u.f)
 132                 cmp = -1;
 133
 134         return cmp;
 135 }
 136
 137 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 138                                    fio_fp64_t *plist, unsigned long long **output,
 139                                    unsigned long long *maxv, unsigned long long *minv)
 140 {
 141         unsigned long sum = 0;
 142         unsigned int len, i, j = 0;
 143         unsigned int oval_len = 0;
 144         unsigned long long *ovals = NULL;
 145         int is_last;
 146
 147         *minv = -1ULL;
 148         *maxv = 0;
 149
 150         len = 0;
 151         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 152                 len++;
 153
 154         if (!len)
 155                 return 0;
 156
 157         /*
 158          * Sort the percentile list. Note that it may already be sorted if
 159          * we are using the default values, but since it's a short list this
 160          * isn't a worry. Also note that this does not work for NaN values.
 161          */
 162         if (len > 1)
 163                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 164
 165         /*
 166          * Calculate bucket values, note down max and min values
 167          */
 168         is_last = 0;
 169         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 170                 sum += io_u_plat[i];
 171                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 172                         assert(plist[j].u.f <= 100.0);
 173
 174                         if (j == oval_len) {
 175                                 oval_len += 100;
 176                                 ovals = realloc(ovals, oval_len * sizeof(*ovals));
 177                         }
 178
 179                         ovals[j] = plat_idx_to_val(i);
 180                         if (ovals[j] < *minv)
 181                                 *minv = ovals[j];
 182                         if (ovals[j] > *maxv)
 183                                 *maxv = ovals[j];
 184
 185                         is_last = (j == len - 1);
 186                         if (is_last)
 187                                 break;
 188
 189                         j++;
 190                 }
 191         }
 192
 193         *output = ovals;
 194         return len;
 195 }
 196
 197 /*
 198  * Find and display the p-th percentile of clat
 199  */
 200 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 201                                   fio_fp64_t *plist, unsigned int precision,
 202                                   struct buf_output *out)
 203 {
 204         unsigned int divisor, len, i, j = 0;
 205         unsigned long long minv, maxv;
 206         unsigned long long *ovals;
 207         int is_last, per_line, scale_down, time_width;
 208         char fmt[32];
 209
 210         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 211         if (!len)
 212                 goto out;
 213
 214         /*
 215          * We default to nsecs, but if the value range is such that we
 216          * should scale down to usecs or msecs, do that.
 217          */
 218         if (minv > 2000000 && maxv > 99999999ULL) {
 219                 scale_down = 2;
 220                 divisor = 1000000;
 221                 log_buf(out, "    clat percentiles (msec):\n     |");
 222         } else if (minv > 2000 && maxv > 99999) {
 223                 scale_down = 1;
 224                 divisor = 1000;
 225                 log_buf(out, "    clat percentiles (usec):\n     |");
 226         } else {
 227                 scale_down = 0;
 228                 divisor = 1;
 229                 log_buf(out, "    clat percentiles (nsec):\n     |");
 230         }
 231
 232
 233         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 234         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 235                         precision, time_width);
 236         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 237         per_line = (80 - 7) / (precision + 10 + time_width);
 238
 239         for (j = 0; j < len; j++) {
 240                 /* for formatting */
 241                 if (j != 0 && (j % per_line) == 0)
 242                         log_buf(out, "     |");
 243
 244                 /* end of the list */
 245                 is_last = (j == len - 1);
 246
 247                 for (i = 0; i < scale_down; i++)
 248                         ovals[j] = (ovals[j] + 999) / 1000;
 249
 250                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 251
 252                 if (is_last)
 253                         break;
 254
 255                 if ((j % per_line) == per_line - 1)     /* for formatting */
 256                         log_buf(out, "\n");
 257         }
 258
 259 out:
 260         if (ovals)
 261                 free(ovals);
 262 }
 263
 264 bool calc_lat(struct io_stat *is, unsigned long long *min,
 265               unsigned long long *max, double *mean, double *dev)
 266 {
 267         double n = (double) is->samples;
 268
 269         if (n == 0)
 270                 return false;
 271
 272         *min = is->min_val;
 273         *max = is->max_val;
 274         *mean = is->mean.u.f;
 275
 276         if (n > 1.0)
 277                 *dev = sqrt(is->S.u.f / (n - 1.0));
 278         else
 279                 *dev = 0;
 280
 281         return true;
 282 }
 283
 284 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 285 {
 286         char *io, *agg, *min, *max;
 287         char *ioalt, *aggalt, *minalt, *maxalt;
 288         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 289         int i;
 290
 291         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 292
 293         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 294                 const int i2p = is_power_of_2(rs->kb_base);
 295
 296                 if (!rs->max_run[i])
 297                         continue;
 298
 299                 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
 300                 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
 301                 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
 302                 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
 303                 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
 304                 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
 305                 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
 306                 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
 307                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 308                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 309                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 310                                 (unsigned long long) rs->min_run[i],
 311                                 (unsigned long long) rs->max_run[i]);
 312
 313                 free(io);
 314                 free(agg);
 315                 free(min);
 316                 free(max);
 317                 free(ioalt);
 318                 free(aggalt);
 319                 free(minalt);
 320                 free(maxalt);
 321         }
 322 }
 323
 324 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 325 {
 326         int i;
 327
 328         /*
 329          * Do depth distribution calculations
 330          */
 331         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 332                 if (total) {
 333                         io_u_dist[i] = (double) map[i] / (double) total;
 334                         io_u_dist[i] *= 100.0;
 335                         if (io_u_dist[i] < 0.1 && map[i])
 336                                 io_u_dist[i] = 0.1;
 337                 } else
 338                         io_u_dist[i] = 0.0;
 339         }
 340 }
 341
 342 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 343                           unsigned int *src, int nr)
 344 {
 345         unsigned long total = ddir_rw_sum(ts->total_io_u);
 346         int i;
 347
 348         /*
 349          * Do latency distribution calculations
 350          */
 351         for (i = 0; i < nr; i++) {
 352                 if (total) {
 353                         dst[i] = (double) src[i] / (double) total;
 354                         dst[i] *= 100.0;
 355                         if (dst[i] < 0.01 && src[i])
 356                                 dst[i] = 0.01;
 357                 } else
 358                         dst[i] = 0.0;
 359         }
 360 }
 361
 362 /*
 363  * To keep the terse format unaltered, add all of the ns latency
 364  * buckets to the first us latency bucket
 365  */
 366 void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 367 {
 368         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 369         int i;
 370
 371         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 372
 373         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 374                 ntotal += ts->io_u_lat_n[i];
 375
 376         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 377 }
 378
 379 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 380 {
 381         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 382 }
 383
 384 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 385 {
 386         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 387 }
 388
 389 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 390 {
 391         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 392 }
 393
 394 static void display_lat(const char *name, unsigned long long min,
 395                         unsigned long long max, double mean, double dev,
 396                         struct buf_output *out)
 397 {
 398         const char *base = "(nsec)";
 399         char *minp, *maxp;
 400
 401         if (nsec_to_msec(&min, &max, &mean, &dev))
 402                 base = "(msec)";
 403         else if (nsec_to_usec(&min, &max, &mean, &dev))
 404                 base = "(usec)";
 405
 406         minp = num2str(min, 6, 1, 0, N2S_NONE);
 407         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 408
 409         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 410                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 411
 412         free(minp);
 413         free(maxp);
 414 }
 415
 416 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 417                              int ddir, struct buf_output *out)
 418 {
 419         const char *str[] = { " read", "write", " trim" };
 420         unsigned long runt;
 421         unsigned long long min, max, bw, iops;
 422         double mean, dev;
 423         char *io_p, *bw_p, *bw_p_alt, *iops_p;
 424         int i2p;
 425
 426         assert(ddir_rw(ddir));
 427
 428         if (!ts->runtime[ddir])
 429                 return;
 430
 431         i2p = is_power_of_2(rs->kb_base);
 432         runt = ts->runtime[ddir];
 433
 434         bw = (1000 * ts->io_bytes[ddir]) / runt;
 435         io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
 436         bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
 437         bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
 438
 439         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 440         iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
 441
 442         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
 443                         rs->unified_rw_rep ? "mixed" : str[ddir],
 444                         iops_p, bw_p, bw_p_alt, io_p,
 445                         (unsigned long long) ts->runtime[ddir]);
 446
 447         free(io_p);
 448         free(bw_p);
 449         free(bw_p_alt);
 450         free(iops_p);
 451
 452         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 453                 display_lat("slat", min, max, mean, dev, out);
 454         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 455                 display_lat("clat", min, max, mean, dev, out);
 456         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 457                 display_lat(" lat", min, max, mean, dev, out);
 458
 459         if (ts->clat_percentiles) {
 460                 show_clat_percentiles(ts->io_u_plat[ddir],
 461                                         ts->clat_stat[ddir].samples,
 462                                         ts->percentile_list,
 463                                         ts->percentile_precision, out);
 464         }
 465         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 466                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 467                 const char *bw_str;
 468
 469                 if ((rs->unit_base == 1) && i2p)
 470                         bw_str = "Kibit";
 471                 else if (rs->unit_base == 1)
 472                         bw_str = "kbit";
 473                 else if (i2p)
 474                         bw_str = "KiB";
 475                 else
 476                         bw_str = "kB";
 477
 478                 if (rs->unit_base == 1) {
 479                         min *= 8.0;
 480                         max *= 8.0;
 481                         mean *= 8.0;
 482                         dev *= 8.0;
 483                 }
 484
 485                 if (rs->agg[ddir]) {
 486                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 487                         if (p_of_agg > 100.0)
 488                                 p_of_agg = 100.0;
 489                 }
 490
 491                 if (mean > fkb_base * fkb_base) {
 492                         min /= fkb_base;
 493                         max /= fkb_base;
 494                         mean /= fkb_base;
 495                         dev /= fkb_base;
 496                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 497                 }
 498
 499                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 500                         "avg=%5.02f, stdev=%5.02f, samples=%5lu\n",
 501                         bw_str, min, max, p_of_agg, mean, dev,
 502                         (&ts->bw_stat[ddir])->samples);
 503         }
 504         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 505                 log_buf(out, "   iops : min=%5llu, max=%5llu, avg=%5.02f, "
 506                         "stdev=%5.02f, samples=%5lu\n",
 507                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 508         }
 509 }
 510
 511 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 512                     const char *msg, struct buf_output *out)
 513 {
 514         int new_line = 1, i, line = 0, shown = 0;
 515
 516         for (i = 0; i < nr; i++) {
 517                 if (io_u_lat[i] <= 0.0)
 518                         continue;
 519                 shown = 1;
 520                 if (new_line) {
 521                         if (line)
 522                                 log_buf(out, "\n");
 523                         log_buf(out, "    lat (%s) : ", msg);
 524                         new_line = 0;
 525                         line = 0;
 526                 }
 527                 if (line)
 528                         log_buf(out, ", ");
 529                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 530                 line++;
 531                 if (line == 5)
 532                         new_line = 1;
 533         }
 534
 535         if (shown)
 536                 log_buf(out, "\n");
 537
 538         return shown;
 539 }
 540
 541 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 542 {
 543         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 544                                  "250=", "500=", "750=", "1000=", };
 545
 546         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 547 }
 548
 549 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 550 {
 551         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 552                                  "250=", "500=", "750=", "1000=", };
 553
 554         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 555 }
 556
 557 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 558 {
 559         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 560                                  "250=", "500=", "750=", "1000=", "2000=",
 561                                  ">=2000=", };
 562
 563         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 564 }
 565
 566 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 567 {
 568         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 569         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 570         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 571
 572         stat_calc_lat_n(ts, io_u_lat_n);
 573         stat_calc_lat_u(ts, io_u_lat_u);
 574         stat_calc_lat_m(ts, io_u_lat_m);
 575
 576         show_lat_n(io_u_lat_n, out);
 577         show_lat_u(io_u_lat_u, out);
 578         show_lat_m(io_u_lat_m, out);
 579 }
 580
 581 static int block_state_category(int block_state)
 582 {
 583         switch (block_state) {
 584         case BLOCK_STATE_UNINIT:
 585                 return 0;
 586         case BLOCK_STATE_TRIMMED:
 587         case BLOCK_STATE_WRITTEN:
 588                 return 1;
 589         case BLOCK_STATE_WRITE_FAILURE:
 590         case BLOCK_STATE_TRIM_FAILURE:
 591                 return 2;
 592         default:
 593                 /* Silence compile warning on some BSDs and have a return */
 594                 assert(0);
 595                 return -1;
 596         }
 597 }
 598
 599 static int compare_block_infos(const void *bs1, const void *bs2)
 600 {
 601         uint32_t block1 = *(uint32_t *)bs1;
 602         uint32_t block2 = *(uint32_t *)bs2;
 603         int state1 = BLOCK_INFO_STATE(block1);
 604         int state2 = BLOCK_INFO_STATE(block2);
 605         int bscat1 = block_state_category(state1);
 606         int bscat2 = block_state_category(state2);
 607         int cycles1 = BLOCK_INFO_TRIMS(block1);
 608         int cycles2 = BLOCK_INFO_TRIMS(block2);
 609
 610         if (bscat1 < bscat2)
 611                 return -1;
 612         if (bscat1 > bscat2)
 613                 return 1;
 614
 615         if (cycles1 < cycles2)
 616                 return -1;
 617         if (cycles1 > cycles2)
 618                 return 1;
 619
 620         if (state1 < state2)
 621                 return -1;
 622         if (state1 > state2)
 623                 return 1;
 624
 625         assert(block1 == block2);
 626         return 0;
 627 }
 628
 629 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 630                                   fio_fp64_t *plist, unsigned int **percentiles,
 631                                   unsigned int *types)
 632 {
 633         int len = 0;
 634         int i, nr_uninit;
 635
 636         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 637
 638         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 639                 len++;
 640
 641         if (!len)
 642                 return 0;
 643
 644         /*
 645          * Sort the percentile list. Note that it may already be sorted if
 646          * we are using the default values, but since it's a short list this
 647          * isn't a worry. Also note that this does not work for NaN values.
 648          */
 649         if (len > 1)
 650                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 651
 652         nr_uninit = 0;
 653         /* Start only after the uninit entries end */
 654         for (nr_uninit = 0;
 655              nr_uninit < nr_block_infos
 656                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 657              nr_uninit ++)
 658                 ;
 659
 660         if (nr_uninit == nr_block_infos)
 661                 return 0;
 662
 663         *percentiles = calloc(len, sizeof(**percentiles));
 664
 665         for (i = 0; i < len; i++) {
 666                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 667                                 + nr_uninit;
 668                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 669         }
 670
 671         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 672         for (i = 0; i < nr_block_infos; i++)
 673                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 674
 675         return len;
 676 }
 677
 678 static const char *block_state_names[] = {
 679         [BLOCK_STATE_UNINIT] = "unwritten",
 680         [BLOCK_STATE_TRIMMED] = "trimmed",
 681         [BLOCK_STATE_WRITTEN] = "written",
 682         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 683         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 684 };
 685
 686 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 687                              fio_fp64_t *plist, struct buf_output *out)
 688 {
 689         int len, pos, i;
 690         unsigned int *percentiles = NULL;
 691         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 692
 693         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 694                                      &percentiles, block_state_counts);
 695
 696         log_buf(out, "  block lifetime percentiles :\n   |");
 697         pos = 0;
 698         for (i = 0; i < len; i++) {
 699                 uint32_t block_info = percentiles[i];
 700 #define LINE_LENGTH     75
 701                 char str[LINE_LENGTH];
 702                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 703                                      plist[i].u.f, block_info,
 704                                      i == len - 1 ? '\n' : ',');
 705                 assert(strln < LINE_LENGTH);
 706                 if (pos + strln > LINE_LENGTH) {
 707                         pos = 0;
 708                         log_buf(out, "\n   |");
 709                 }
 710                 log_buf(out, "%s", str);
 711                 pos += strln;
 712 #undef LINE_LENGTH
 713         }
 714         if (percentiles)
 715                 free(percentiles);
 716
 717         log_buf(out, "        states               :");
 718         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 719                 log_buf(out, " %s=%u%c",
 720                          block_state_names[i], block_state_counts[i],
 721                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 722 }
 723
 724 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 725 {
 726         char *p1, *p1alt, *p2;
 727         unsigned long long bw_mean, iops_mean;
 728         const int i2p = is_power_of_2(ts->kb_base);
 729
 730         if (!ts->ss_dur)
 731                 return;
 732
 733         bw_mean = steadystate_bw_mean(ts);
 734         iops_mean = steadystate_iops_mean(ts);
 735
 736         p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
 737         p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
 738         p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
 739
 740         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 741                 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
 742                 p1, p1alt, p2,
 743                 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
 744                 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
 745                 ts->ss_criterion.u.f,
 746                 ts->ss_state & __FIO_SS_PCT ? "%" : "");
 747
 748         free(p1);
 749         free(p1alt);
 750         free(p2);
 751 }
 752
 753 static void show_thread_status_normal(struct thread_stat *ts,
 754                                       struct group_run_stats *rs,
 755                                       struct buf_output *out)
 756 {
 757         double usr_cpu, sys_cpu;
 758         unsigned long runtime;
 759         double io_u_dist[FIO_IO_U_MAP_NR];
 760         time_t time_p;
 761         char time_buf[32];
 762
 763         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 764                 return;
 765
 766         memset(time_buf, 0, sizeof(time_buf));
 767
 768         time(&time_p);
 769         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 770
 771         if (!ts->error) {
 772                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 773                                         ts->name, ts->groupid, ts->members,
 774                                         ts->error, (int) ts->pid, time_buf);
 775         } else {
 776                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 777                                         ts->name, ts->groupid, ts->members,
 778                                         ts->error, ts->verror, (int) ts->pid,
 779                                         time_buf);
 780         }
 781
 782         if (strlen(ts->description))
 783                 log_buf(out, "  Description  : [%s]\n", ts->description);
 784
 785         if (ts->io_bytes[DDIR_READ])
 786                 show_ddir_status(rs, ts, DDIR_READ, out);
 787         if (ts->io_bytes[DDIR_WRITE])
 788                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 789         if (ts->io_bytes[DDIR_TRIM])
 790                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 791
 792         show_latencies(ts, out);
 793
 794         runtime = ts->total_run_time;
 795         if (runtime) {
 796                 double runt = (double) runtime;
 797
 798                 usr_cpu = (double) ts->usr_time * 100 / runt;
 799                 sys_cpu = (double) ts->sys_time * 100 / runt;
 800         } else {
 801                 usr_cpu = 0;
 802                 sys_cpu = 0;
 803         }
 804
 805         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 806                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 807                         (unsigned long long) ts->ctx,
 808                         (unsigned long long) ts->majf,
 809                         (unsigned long long) ts->minf);
 810
 811         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 812         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 813                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 814                                         io_u_dist[1], io_u_dist[2],
 815                                         io_u_dist[3], io_u_dist[4],
 816                                         io_u_dist[5], io_u_dist[6]);
 817
 818         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 819         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 820                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 821                                         io_u_dist[1], io_u_dist[2],
 822                                         io_u_dist[3], io_u_dist[4],
 823                                         io_u_dist[5], io_u_dist[6]);
 824         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 825         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 826                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 827                                         io_u_dist[1], io_u_dist[2],
 828                                         io_u_dist[3], io_u_dist[4],
 829                                         io_u_dist[5], io_u_dist[6]);
 830         log_buf(out, "     issued rwt: total=%llu,%llu,%llu,"
 831                                  " short=%llu,%llu,%llu,"
 832                                  " dropped=%llu,%llu,%llu\n",
 833                                         (unsigned long long) ts->total_io_u[0],
 834                                         (unsigned long long) ts->total_io_u[1],
 835                                         (unsigned long long) ts->total_io_u[2],
 836                                         (unsigned long long) ts->short_io_u[0],
 837                                         (unsigned long long) ts->short_io_u[1],
 838                                         (unsigned long long) ts->short_io_u[2],
 839                                         (unsigned long long) ts->drop_io_u[0],
 840                                         (unsigned long long) ts->drop_io_u[1],
 841                                         (unsigned long long) ts->drop_io_u[2]);
 842         if (ts->continue_on_error) {
 843                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 844                                         (unsigned long long)ts->total_err_count,
 845                                         ts->first_error,
 846                                         strerror(ts->first_error));
 847         }
 848         if (ts->latency_depth) {
 849                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 850                                         (unsigned long long)ts->latency_target,
 851                                         (unsigned long long)ts->latency_window,
 852                                         ts->latency_percentile.u.f,
 853                                         ts->latency_depth);
 854         }
 855
 856         if (ts->nr_block_infos)
 857                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 858                                   ts->percentile_list, out);
 859
 860         if (ts->ss_dur)
 861                 show_ss_normal(ts, out);
 862 }
 863
 864 static void show_ddir_status_terse(struct thread_stat *ts,
 865                                    struct group_run_stats *rs, int ddir,
 866                                    struct buf_output *out)
 867 {
 868         unsigned long long min, max, minv, maxv, bw, iops;
 869         unsigned long long *ovals = NULL;
 870         double mean, dev;
 871         unsigned int len;
 872         int i;
 873
 874         assert(ddir_rw(ddir));
 875
 876         iops = bw = 0;
 877         if (ts->runtime[ddir]) {
 878                 uint64_t runt = ts->runtime[ddir];
 879
 880                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 881                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 882         }
 883
 884         log_buf(out, ";%llu;%llu;%llu;%llu",
 885                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 886                                         (unsigned long long) ts->runtime[ddir]);
 887
 888         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 889                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 890         else
 891                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 892
 893         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 894                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 895         else
 896                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 897
 898         if (ts->clat_percentiles) {
 899                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 900                                         ts->clat_stat[ddir].samples,
 901                                         ts->percentile_list, &ovals, &maxv,
 902                                         &minv);
 903         } else
 904                 len = 0;
 905
 906         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 907                 if (i >= len) {
 908                         log_buf(out, ";0%%=0");
 909                         continue;
 910                 }
 911                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
 912         }
 913
 914         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 915                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 916         else
 917                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 918
 919         if (ovals)
 920                 free(ovals);
 921
 922         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 923                 double p_of_agg = 100.0;
 924
 925                 if (rs->agg[ddir]) {
 926                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 927                         if (p_of_agg > 100.0)
 928                                 p_of_agg = 100.0;
 929                 }
 930
 931                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 932         } else
 933                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
 934 }
 935
 936 static void add_ddir_status_json(struct thread_stat *ts,
 937                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 938 {
 939         unsigned long long min, max, minv, maxv;
 940         unsigned long long bw;
 941         unsigned long long *ovals = NULL;
 942         double mean, dev, iops;
 943         unsigned int len;
 944         int i;
 945         const char *ddirname[] = {"read", "write", "trim"};
 946         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 947         char buf[120];
 948         double p_of_agg = 100.0;
 949
 950         assert(ddir_rw(ddir));
 951
 952         if (ts->unified_rw_rep && ddir != DDIR_READ)
 953                 return;
 954
 955         dir_object = json_create_object();
 956         json_object_add_value_object(parent,
 957                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 958
 959         bw = 0;
 960         iops = 0.0;
 961         if (ts->runtime[ddir]) {
 962                 uint64_t runt = ts->runtime[ddir];
 963
 964                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 965                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 966         }
 967
 968         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
 969         json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
 970         json_object_add_value_int(dir_object, "bw", bw);
 971         json_object_add_value_float(dir_object, "iops", iops);
 972         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 973         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 974         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 975         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 976
 977         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 978                 min = max = 0;
 979                 mean = dev = 0.0;
 980         }
 981         tmp_object = json_create_object();
 982         json_object_add_value_object(dir_object, "slat_ns", tmp_object);
 983         json_object_add_value_int(tmp_object, "min", min);
 984         json_object_add_value_int(tmp_object, "max", max);
 985         json_object_add_value_float(tmp_object, "mean", mean);
 986         json_object_add_value_float(tmp_object, "stddev", dev);
 987
 988         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 989                 min = max = 0;
 990                 mean = dev = 0.0;
 991         }
 992         tmp_object = json_create_object();
 993         json_object_add_value_object(dir_object, "clat_ns", tmp_object);
 994         json_object_add_value_int(tmp_object, "min", min);
 995         json_object_add_value_int(tmp_object, "max", max);
 996         json_object_add_value_float(tmp_object, "mean", mean);
 997         json_object_add_value_float(tmp_object, "stddev", dev);
 998
 999         if (ts->clat_percentiles) {
1000                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1001                                         ts->clat_stat[ddir].samples,
1002                                         ts->percentile_list, &ovals, &maxv,
1003                                         &minv);
1004         } else
1005                 len = 0;
1006
1007         percentile_object = json_create_object();
1008         json_object_add_value_object(tmp_object, "percentile", percentile_object);
1009         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1010                 if (i >= len) {
1011                         json_object_add_value_int(percentile_object, "0.00", 0);
1012                         continue;
1013                 }
1014                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1015                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1016         }
1017
1018         if (output_format & FIO_OUTPUT_JSON_PLUS) {
1019                 clat_bins_object = json_create_object();
1020                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1021                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1022                         if (ts->io_u_plat[ddir][i]) {
1023                                 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1024                                 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1025                         }
1026                 }
1027         }
1028
1029         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1030                 min = max = 0;
1031                 mean = dev = 0.0;
1032         }
1033         tmp_object = json_create_object();
1034         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1035         json_object_add_value_int(tmp_object, "min", min);
1036         json_object_add_value_int(tmp_object, "max", max);
1037         json_object_add_value_float(tmp_object, "mean", mean);
1038         json_object_add_value_float(tmp_object, "stddev", dev);
1039         if (ovals)
1040                 free(ovals);
1041
1042         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1043                 if (rs->agg[ddir]) {
1044                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
1045                         if (p_of_agg > 100.0)
1046                                 p_of_agg = 100.0;
1047                 }
1048         } else {
1049                 min = max = 0;
1050                 p_of_agg = mean = dev = 0.0;
1051         }
1052         json_object_add_value_int(dir_object, "bw_min", min);
1053         json_object_add_value_int(dir_object, "bw_max", max);
1054         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1055         json_object_add_value_float(dir_object, "bw_mean", mean);
1056         json_object_add_value_float(dir_object, "bw_dev", dev);
1057         json_object_add_value_int(dir_object, "bw_samples",
1058                                 (&ts->bw_stat[ddir])->samples);
1059
1060         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1061                 min = max = 0;
1062                 mean = dev = 0.0;
1063         }
1064         json_object_add_value_int(dir_object, "iops_min", min);
1065         json_object_add_value_int(dir_object, "iops_max", max);
1066         json_object_add_value_float(dir_object, "iops_mean", mean);
1067         json_object_add_value_float(dir_object, "iops_stddev", dev);
1068         json_object_add_value_int(dir_object, "iops_samples",
1069                                 (&ts->iops_stat[ddir])->samples);
1070 }
1071
1072 static void show_thread_status_terse_all(struct thread_stat *ts,
1073                                          struct group_run_stats *rs, int ver,
1074                                          struct buf_output *out)
1075 {
1076         double io_u_dist[FIO_IO_U_MAP_NR];
1077         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1078         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1079         double usr_cpu, sys_cpu;
1080         int i;
1081
1082         /* General Info */
1083         if (ver == 2)
1084                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1085         else
1086                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1087                         ts->name, ts->groupid, ts->error);
1088
1089         /* Log Read Status */
1090         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1091         /* Log Write Status */
1092         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1093         /* Log Trim Status */
1094         if (ver == 2 || ver == 4)
1095                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1096
1097         /* CPU Usage */
1098         if (ts->total_run_time) {
1099                 double runt = (double) ts->total_run_time;
1100
1101                 usr_cpu = (double) ts->usr_time * 100 / runt;
1102                 sys_cpu = (double) ts->sys_time * 100 / runt;
1103         } else {
1104                 usr_cpu = 0;
1105                 sys_cpu = 0;
1106         }
1107
1108         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1109                                                 (unsigned long long) ts->ctx,
1110                                                 (unsigned long long) ts->majf,
1111                                                 (unsigned long long) ts->minf);
1112
1113         /* Calc % distribution of IO depths, usecond, msecond latency */
1114         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1115         stat_calc_lat_nu(ts, io_u_lat_u);
1116         stat_calc_lat_m(ts, io_u_lat_m);
1117
1118         /* Only show fixed 7 I/O depth levels*/
1119         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1120                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1121                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1122
1123         /* Microsecond latency */
1124         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1125                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1126         /* Millisecond latency */
1127         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1128                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1129
1130         /* disk util stats, if any */
1131         if (ver >= 3)
1132                 show_disk_util(1, NULL, out);
1133
1134         /* Additional output if continue_on_error set - default off*/
1135         if (ts->continue_on_error)
1136                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1137         if (ver == 2)
1138                 log_buf(out, "\n");
1139
1140         /* Additional output if description is set */
1141         if (strlen(ts->description))
1142                 log_buf(out, ";%s", ts->description);
1143
1144         log_buf(out, "\n");
1145 }
1146
1147 static void json_add_job_opts(struct json_object *root, const char *name,
1148                               struct flist_head *opt_list, bool num_jobs)
1149 {
1150         struct json_object *dir_object;
1151         struct flist_head *entry;
1152         struct print_option *p;
1153
1154         if (flist_empty(opt_list))
1155                 return;
1156
1157         dir_object = json_create_object();
1158         json_object_add_value_object(root, name, dir_object);
1159
1160         flist_for_each(entry, opt_list) {
1161                 const char *pos = "";
1162
1163                 p = flist_entry(entry, struct print_option, list);
1164                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1165                         continue;
1166                 if (p->value)
1167                         pos = p->value;
1168                 json_object_add_value_string(dir_object, p->name, pos);
1169         }
1170 }
1171
1172 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1173                                                    struct group_run_stats *rs,
1174                                                    struct flist_head *opt_list)
1175 {
1176         struct json_object *root, *tmp;
1177         struct jobs_eta *je;
1178         double io_u_dist[FIO_IO_U_MAP_NR];
1179         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1180         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1181         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1182         double usr_cpu, sys_cpu;
1183         int i;
1184         size_t size;
1185
1186         root = json_create_object();
1187         json_object_add_value_string(root, "jobname", ts->name);
1188         json_object_add_value_int(root, "groupid", ts->groupid);
1189         json_object_add_value_int(root, "error", ts->error);
1190
1191         /* ETA Info */
1192         je = get_jobs_eta(true, &size);
1193         if (je) {
1194                 json_object_add_value_int(root, "eta", je->eta_sec);
1195                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1196         }
1197
1198         if (opt_list)
1199                 json_add_job_opts(root, "job options", opt_list, true);
1200
1201         add_ddir_status_json(ts, rs, DDIR_READ, root);
1202         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1203         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1204
1205         /* CPU Usage */
1206         if (ts->total_run_time) {
1207                 double runt = (double) ts->total_run_time;
1208
1209                 usr_cpu = (double) ts->usr_time * 100 / runt;
1210                 sys_cpu = (double) ts->sys_time * 100 / runt;
1211         } else {
1212                 usr_cpu = 0;
1213                 sys_cpu = 0;
1214         }
1215         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1216         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1217         json_object_add_value_int(root, "ctx", ts->ctx);
1218         json_object_add_value_int(root, "majf", ts->majf);
1219         json_object_add_value_int(root, "minf", ts->minf);
1220
1221
1222         /* Calc % distribution of IO depths, usecond, msecond latency */
1223         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1224         stat_calc_lat_n(ts, io_u_lat_n);
1225         stat_calc_lat_u(ts, io_u_lat_u);
1226         stat_calc_lat_m(ts, io_u_lat_m);
1227
1228         tmp = json_create_object();
1229         json_object_add_value_object(root, "iodepth_level", tmp);
1230         /* Only show fixed 7 I/O depth levels*/
1231         for (i = 0; i < 7; i++) {
1232                 char name[20];
1233                 if (i < 6)
1234                         snprintf(name, 20, "%d", 1 << i);
1235                 else
1236                         snprintf(name, 20, ">=%d", 1 << i);
1237                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1238         }
1239
1240         /* Nanosecond latency */
1241         tmp = json_create_object();
1242         json_object_add_value_object(root, "latency_ns", tmp);
1243         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1244                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1245                                  "250", "500", "750", "1000", };
1246                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1247         }
1248         /* Microsecond latency */
1249         tmp = json_create_object();
1250         json_object_add_value_object(root, "latency_us", tmp);
1251         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1252                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1253                                  "250", "500", "750", "1000", };
1254                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1255         }
1256         /* Millisecond latency */
1257         tmp = json_create_object();
1258         json_object_add_value_object(root, "latency_ms", tmp);
1259         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1260                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1261                                  "250", "500", "750", "1000", "2000",
1262                                  ">=2000", };
1263                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1264         }
1265
1266         /* Additional output if continue_on_error set - default off*/
1267         if (ts->continue_on_error) {
1268                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1269                 json_object_add_value_int(root, "first_error", ts->first_error);
1270         }
1271
1272         if (ts->latency_depth) {
1273                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1274                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1275                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1276                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1277         }
1278
1279         /* Additional output if description is set */
1280         if (strlen(ts->description))
1281                 json_object_add_value_string(root, "desc", ts->description);
1282
1283         if (ts->nr_block_infos) {
1284                 /* Block error histogram and types */
1285                 int len;
1286                 unsigned int *percentiles = NULL;
1287                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1288
1289                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1290                                              ts->percentile_list,
1291                                              &percentiles, block_state_counts);
1292
1293                 if (len) {
1294                         struct json_object *block, *percentile_object, *states;
1295                         int state;
1296                         block = json_create_object();
1297                         json_object_add_value_object(root, "block", block);
1298
1299                         percentile_object = json_create_object();
1300                         json_object_add_value_object(block, "percentiles",
1301                                                      percentile_object);
1302                         for (i = 0; i < len; i++) {
1303                                 char buf[20];
1304                                 snprintf(buf, sizeof(buf), "%f",
1305                                          ts->percentile_list[i].u.f);
1306                                 json_object_add_value_int(percentile_object,
1307                                                           (const char *)buf,
1308                                                           percentiles[i]);
1309                         }
1310
1311                         states = json_create_object();
1312                         json_object_add_value_object(block, "states", states);
1313                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1314                                 json_object_add_value_int(states,
1315                                         block_state_names[state],
1316                                         block_state_counts[state]);
1317                         }
1318                         free(percentiles);
1319                 }
1320         }
1321
1322         if (ts->ss_dur) {
1323                 struct json_object *data;
1324                 struct json_array *iops, *bw;
1325                 int i, j, k;
1326                 char ss_buf[64];
1327
1328                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1329                         ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1330                         ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1331                         (float) ts->ss_limit.u.f,
1332                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1333
1334                 tmp = json_create_object();
1335                 json_object_add_value_object(root, "steadystate", tmp);
1336                 json_object_add_value_string(tmp, "ss", ss_buf);
1337                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1338                 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1339
1340                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1341                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1342                 json_object_add_value_string(tmp, "criterion", ss_buf);
1343                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1344                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1345
1346                 data = json_create_object();
1347                 json_object_add_value_object(tmp, "data", data);
1348                 bw = json_create_array();
1349                 iops = json_create_array();
1350
1351                 /*
1352                 ** if ss was attained or the buffer is not full,
1353                 ** ss->head points to the first element in the list.
1354                 ** otherwise it actually points to the second element
1355                 ** in the list
1356                 */
1357                 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1358                         j = ts->ss_head;
1359                 else
1360                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1361                 for (i = 0; i < ts->ss_dur; i++) {
1362                         k = (j + i) % ts->ss_dur;
1363                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1364                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1365                 }
1366                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1367                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1368                 json_object_add_value_array(data, "iops", iops);
1369                 json_object_add_value_array(data, "bw", bw);
1370         }
1371
1372         return root;
1373 }
1374
1375 static void show_thread_status_terse(struct thread_stat *ts,
1376                                      struct group_run_stats *rs,
1377                                      struct buf_output *out)
1378 {
1379         if (terse_version >= 2 && terse_version <= 4)
1380                 show_thread_status_terse_all(ts, rs, terse_version, out);
1381         else
1382                 log_err("fio: bad terse version!? %d\n", terse_version);
1383 }
1384
1385 struct json_object *show_thread_status(struct thread_stat *ts,
1386                                        struct group_run_stats *rs,
1387                                        struct flist_head *opt_list,
1388                                        struct buf_output *out)
1389 {
1390         struct json_object *ret = NULL;
1391
1392         if (output_format & FIO_OUTPUT_TERSE)
1393                 show_thread_status_terse(ts, rs,  out);
1394         if (output_format & FIO_OUTPUT_JSON)
1395                 ret = show_thread_status_json(ts, rs, opt_list);
1396         if (output_format & FIO_OUTPUT_NORMAL)
1397                 show_thread_status_normal(ts, rs,  out);
1398
1399         return ret;
1400 }
1401
1402 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1403 {
1404         double mean, S;
1405
1406         if (src->samples == 0)
1407                 return;
1408
1409         dst->min_val = min(dst->min_val, src->min_val);
1410         dst->max_val = max(dst->max_val, src->max_val);
1411
1412         /*
1413          * Compute new mean and S after the merge
1414          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1415          *  #Parallel_algorithm>
1416          */
1417         if (first) {
1418                 mean = src->mean.u.f;
1419                 S = src->S.u.f;
1420         } else {
1421                 double delta = src->mean.u.f - dst->mean.u.f;
1422
1423                 mean = ((src->mean.u.f * src->samples) +
1424                         (dst->mean.u.f * dst->samples)) /
1425                         (dst->samples + src->samples);
1426
1427                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1428                         (dst->samples * src->samples) /
1429                         (dst->samples + src->samples);
1430         }
1431
1432         dst->samples += src->samples;
1433         dst->mean.u.f = mean;
1434         dst->S.u.f = S;
1435 }
1436
1437 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1438 {
1439         int i;
1440
1441         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1442                 if (dst->max_run[i] < src->max_run[i])
1443                         dst->max_run[i] = src->max_run[i];
1444                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1445                         dst->min_run[i] = src->min_run[i];
1446                 if (dst->max_bw[i] < src->max_bw[i])
1447                         dst->max_bw[i] = src->max_bw[i];
1448                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1449                         dst->min_bw[i] = src->min_bw[i];
1450
1451                 dst->iobytes[i] += src->iobytes[i];
1452                 dst->agg[i] += src->agg[i];
1453         }
1454
1455         if (!dst->kb_base)
1456                 dst->kb_base = src->kb_base;
1457         if (!dst->unit_base)
1458                 dst->unit_base = src->unit_base;
1459 }
1460
1461 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1462                       bool first)
1463 {
1464         int l, k;
1465
1466         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1467                 if (!dst->unified_rw_rep) {
1468                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1469                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1470                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1471                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1472                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1473
1474                         dst->io_bytes[l] += src->io_bytes[l];
1475
1476                         if (dst->runtime[l] < src->runtime[l])
1477                                 dst->runtime[l] = src->runtime[l];
1478                 } else {
1479                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1480                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1481                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1482                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1483                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1484
1485                         dst->io_bytes[0] += src->io_bytes[l];
1486
1487                         if (dst->runtime[0] < src->runtime[l])
1488                                 dst->runtime[0] = src->runtime[l];
1489
1490                         /*
1491                          * We're summing to the same destination, so override
1492                          * 'first' after the first iteration of the loop
1493                          */
1494                         first = false;
1495                 }
1496         }
1497
1498         dst->usr_time += src->usr_time;
1499         dst->sys_time += src->sys_time;
1500         dst->ctx += src->ctx;
1501         dst->majf += src->majf;
1502         dst->minf += src->minf;
1503
1504         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1505                 dst->io_u_map[k] += src->io_u_map[k];
1506         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1507                 dst->io_u_submit[k] += src->io_u_submit[k];
1508         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1509                 dst->io_u_complete[k] += src->io_u_complete[k];
1510         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1511                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1512         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1513                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1514         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1515                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1516
1517         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1518                 if (!dst->unified_rw_rep) {
1519                         dst->total_io_u[k] += src->total_io_u[k];
1520                         dst->short_io_u[k] += src->short_io_u[k];
1521                         dst->drop_io_u[k] += src->drop_io_u[k];
1522                 } else {
1523                         dst->total_io_u[0] += src->total_io_u[k];
1524                         dst->short_io_u[0] += src->short_io_u[k];
1525                         dst->drop_io_u[0] += src->drop_io_u[k];
1526                 }
1527         }
1528
1529         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1530                 int m;
1531
1532                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1533                         if (!dst->unified_rw_rep)
1534                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1535                         else
1536                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1537                 }
1538         }
1539
1540         dst->total_run_time += src->total_run_time;
1541         dst->total_submit += src->total_submit;
1542         dst->total_complete += src->total_complete;
1543 }
1544
1545 void init_group_run_stat(struct group_run_stats *gs)
1546 {
1547         int i;
1548         memset(gs, 0, sizeof(*gs));
1549
1550         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1551                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1552 }
1553
1554 void init_thread_stat(struct thread_stat *ts)
1555 {
1556         int j;
1557
1558         memset(ts, 0, sizeof(*ts));
1559
1560         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1561                 ts->lat_stat[j].min_val = -1UL;
1562                 ts->clat_stat[j].min_val = -1UL;
1563                 ts->slat_stat[j].min_val = -1UL;
1564                 ts->bw_stat[j].min_val = -1UL;
1565                 ts->iops_stat[j].min_val = -1UL;
1566         }
1567         ts->groupid = -1;
1568 }
1569
1570 void __show_run_stats(void)
1571 {
1572         struct group_run_stats *runstats, *rs;
1573         struct thread_data *td;
1574         struct thread_stat *threadstats, *ts;
1575         int i, j, k, nr_ts, last_ts, idx;
1576         int kb_base_warned = 0;
1577         int unit_base_warned = 0;
1578         struct json_object *root = NULL;
1579         struct json_array *array = NULL;
1580         struct buf_output output[FIO_OUTPUT_NR];
1581         struct flist_head **opt_lists;
1582
1583         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1584
1585         for (i = 0; i < groupid + 1; i++)
1586                 init_group_run_stat(&runstats[i]);
1587
1588         /*
1589          * find out how many threads stats we need. if group reporting isn't
1590          * enabled, it's one-per-td.
1591          */
1592         nr_ts = 0;
1593         last_ts = -1;
1594         for_each_td(td, i) {
1595                 if (!td->o.group_reporting) {
1596                         nr_ts++;
1597                         continue;
1598                 }
1599                 if (last_ts == td->groupid)
1600                         continue;
1601                 if (!td->o.stats)
1602                         continue;
1603
1604                 last_ts = td->groupid;
1605                 nr_ts++;
1606         }
1607
1608         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1609         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1610
1611         for (i = 0; i < nr_ts; i++) {
1612                 init_thread_stat(&threadstats[i]);
1613                 opt_lists[i] = NULL;
1614         }
1615
1616         j = 0;
1617         last_ts = -1;
1618         idx = 0;
1619         for_each_td(td, i) {
1620                 if (!td->o.stats)
1621                         continue;
1622                 if (idx && (!td->o.group_reporting ||
1623                     (td->o.group_reporting && last_ts != td->groupid))) {
1624                         idx = 0;
1625                         j++;
1626                 }
1627
1628                 last_ts = td->groupid;
1629
1630                 ts = &threadstats[j];
1631
1632                 ts->clat_percentiles = td->o.clat_percentiles;
1633                 ts->percentile_precision = td->o.percentile_precision;
1634                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1635                 opt_lists[j] = &td->opt_list;
1636
1637                 idx++;
1638                 ts->members++;
1639
1640                 if (ts->groupid == -1) {
1641                         /*
1642                          * These are per-group shared already
1643                          */
1644                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1645                         if (td->o.description)
1646                                 strncpy(ts->description, td->o.description,
1647                                                 FIO_JOBDESC_SIZE - 1);
1648                         else
1649                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1650
1651                         /*
1652                          * If multiple entries in this group, this is
1653                          * the first member.
1654                          */
1655                         ts->thread_number = td->thread_number;
1656                         ts->groupid = td->groupid;
1657
1658                         /*
1659                          * first pid in group, not very useful...
1660                          */
1661                         ts->pid = td->pid;
1662
1663                         ts->kb_base = td->o.kb_base;
1664                         ts->unit_base = td->o.unit_base;
1665                         ts->unified_rw_rep = td->o.unified_rw_rep;
1666                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1667                         log_info("fio: kb_base differs for jobs in group, using"
1668                                  " %u as the base\n", ts->kb_base);
1669                         kb_base_warned = 1;
1670                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1671                         log_info("fio: unit_base differs for jobs in group, using"
1672                                  " %u as the base\n", ts->unit_base);
1673                         unit_base_warned = 1;
1674                 }
1675
1676                 ts->continue_on_error = td->o.continue_on_error;
1677                 ts->total_err_count += td->total_err_count;
1678                 ts->first_error = td->first_error;
1679                 if (!ts->error) {
1680                         if (!td->error && td->o.continue_on_error &&
1681                             td->first_error) {
1682                                 ts->error = td->first_error;
1683                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1684                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1685                         } else  if (td->error) {
1686                                 ts->error = td->error;
1687                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1688                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1689                         }
1690                 }
1691
1692                 ts->latency_depth = td->latency_qd;
1693                 ts->latency_target = td->o.latency_target;
1694                 ts->latency_percentile = td->o.latency_percentile;
1695                 ts->latency_window = td->o.latency_window;
1696
1697                 ts->nr_block_infos = td->ts.nr_block_infos;
1698                 for (k = 0; k < ts->nr_block_infos; k++)
1699                         ts->block_infos[k] = td->ts.block_infos[k];
1700
1701                 sum_thread_stats(ts, &td->ts, idx == 1);
1702
1703                 if (td->o.ss_dur) {
1704                         ts->ss_state = td->ss.state;
1705                         ts->ss_dur = td->ss.dur;
1706                         ts->ss_head = td->ss.head;
1707                         ts->ss_bw_data = td->ss.bw_data;
1708                         ts->ss_iops_data = td->ss.iops_data;
1709                         ts->ss_limit.u.f = td->ss.limit;
1710                         ts->ss_slope.u.f = td->ss.slope;
1711                         ts->ss_deviation.u.f = td->ss.deviation;
1712                         ts->ss_criterion.u.f = td->ss.criterion;
1713                 }
1714                 else
1715                         ts->ss_dur = ts->ss_state = 0;
1716         }
1717
1718         for (i = 0; i < nr_ts; i++) {
1719                 unsigned long long bw;
1720
1721                 ts = &threadstats[i];
1722                 if (ts->groupid == -1)
1723                         continue;
1724                 rs = &runstats[ts->groupid];
1725                 rs->kb_base = ts->kb_base;
1726                 rs->unit_base = ts->unit_base;
1727                 rs->unified_rw_rep += ts->unified_rw_rep;
1728
1729                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1730                         if (!ts->runtime[j])
1731                                 continue;
1732                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1733                                 rs->min_run[j] = ts->runtime[j];
1734                         if (ts->runtime[j] > rs->max_run[j])
1735                                 rs->max_run[j] = ts->runtime[j];
1736
1737                         bw = 0;
1738                         if (ts->runtime[j])
1739                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1740                         if (bw < rs->min_bw[j])
1741                                 rs->min_bw[j] = bw;
1742                         if (bw > rs->max_bw[j])
1743                                 rs->max_bw[j] = bw;
1744
1745                         rs->iobytes[j] += ts->io_bytes[j];
1746                 }
1747         }
1748
1749         for (i = 0; i < groupid + 1; i++) {
1750                 int ddir;
1751
1752                 rs = &runstats[i];
1753
1754                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1755                         if (rs->max_run[ddir])
1756                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1757                                                 rs->max_run[ddir];
1758                 }
1759         }
1760
1761         for (i = 0; i < FIO_OUTPUT_NR; i++)
1762                 buf_output_init(&output[i]);
1763
1764         /*
1765          * don't overwrite last signal output
1766          */
1767         if (output_format & FIO_OUTPUT_NORMAL)
1768                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1769         if (output_format & FIO_OUTPUT_JSON) {
1770                 struct thread_data *global;
1771                 char time_buf[32];
1772                 struct timeval now;
1773                 unsigned long long ms_since_epoch;
1774
1775                 gettimeofday(&now, NULL);
1776                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1777                                  (unsigned long long)(now.tv_usec) / 1000;
1778
1779                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1780                                 sizeof(time_buf));
1781                 if (time_buf[strlen(time_buf) - 1] == '\n')
1782                         time_buf[strlen(time_buf) - 1] = '\0';
1783
1784                 root = json_create_object();
1785                 json_object_add_value_string(root, "fio version", fio_version_string);
1786                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1787                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1788                 json_object_add_value_string(root, "time", time_buf);
1789                 global = get_global_options();
1790                 json_add_job_opts(root, "global options", &global->opt_list, false);
1791                 array = json_create_array();
1792                 json_object_add_value_array(root, "jobs", array);
1793         }
1794
1795         if (is_backend)
1796                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1797
1798         for (i = 0; i < nr_ts; i++) {
1799                 ts = &threadstats[i];
1800                 rs = &runstats[ts->groupid];
1801
1802                 if (is_backend) {
1803                         fio_server_send_job_options(opt_lists[i], i);
1804                         fio_server_send_ts(ts, rs);
1805                 } else {
1806                         if (output_format & FIO_OUTPUT_TERSE)
1807                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1808                         if (output_format & FIO_OUTPUT_JSON) {
1809                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1810                                 json_array_add_value_object(array, tmp);
1811                         }
1812                         if (output_format & FIO_OUTPUT_NORMAL)
1813                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1814                 }
1815         }
1816         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1817                 /* disk util stats, if any */
1818                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1819
1820                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1821
1822                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1823                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1824                 json_free_object(root);
1825         }
1826
1827         for (i = 0; i < groupid + 1; i++) {
1828                 rs = &runstats[i];
1829
1830                 rs->groupid = i;
1831                 if (is_backend)
1832                         fio_server_send_gs(rs);
1833                 else if (output_format & FIO_OUTPUT_NORMAL)
1834                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1835         }
1836
1837         if (is_backend)
1838                 fio_server_send_du();
1839         else if (output_format & FIO_OUTPUT_NORMAL) {
1840                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1841                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1842         }
1843
1844         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1845                 struct buf_output *out = &output[i];
1846
1847                 log_info_buf(out->buf, out->buflen);
1848                 buf_output_free(out);
1849         }
1850
1851         log_info_flush();
1852         free(runstats);
1853         free(threadstats);
1854         free(opt_lists);
1855 }
1856
1857 void show_run_stats(void)
1858 {
1859         fio_mutex_down(stat_mutex);
1860         __show_run_stats();
1861         fio_mutex_up(stat_mutex);
1862 }
1863
1864 void __show_running_run_stats(void)
1865 {
1866         struct thread_data *td;
1867         unsigned long long *rt;
1868         struct timespec ts;
1869         int i;
1870
1871         fio_mutex_down(stat_mutex);
1872
1873         rt = malloc(thread_number * sizeof(unsigned long long));
1874         fio_gettime(&ts, NULL);
1875
1876         for_each_td(td, i) {
1877                 td->update_rusage = 1;
1878                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1879                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1880                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1881                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1882
1883                 rt[i] = mtime_since(&td->start, &ts);
1884                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1885                         td->ts.runtime[DDIR_READ] += rt[i];
1886                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1887                         td->ts.runtime[DDIR_WRITE] += rt[i];
1888                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1889                         td->ts.runtime[DDIR_TRIM] += rt[i];
1890         }
1891
1892         for_each_td(td, i) {
1893                 if (td->runstate >= TD_EXITED)
1894                         continue;
1895                 if (td->rusage_sem) {
1896                         td->update_rusage = 1;
1897                         fio_mutex_down(td->rusage_sem);
1898                 }
1899                 td->update_rusage = 0;
1900         }
1901
1902         __show_run_stats();
1903
1904         for_each_td(td, i) {
1905                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1906                         td->ts.runtime[DDIR_READ] -= rt[i];
1907                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1908                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1909                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1910                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1911         }
1912
1913         free(rt);
1914         fio_mutex_up(stat_mutex);
1915 }
1916
1917 static int status_interval_init;
1918 static struct timespec status_time;
1919 static int status_file_disabled;
1920
1921 #define FIO_STATUS_FILE         "fio-dump-status"
1922
1923 static int check_status_file(void)
1924 {
1925         struct stat sb;
1926         const char *temp_dir;
1927         char fio_status_file_path[PATH_MAX];
1928
1929         if (status_file_disabled)
1930                 return 0;
1931
1932         temp_dir = getenv("TMPDIR");
1933         if (temp_dir == NULL) {
1934                 temp_dir = getenv("TEMP");
1935                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1936                         temp_dir = NULL;
1937         }
1938         if (temp_dir == NULL)
1939                 temp_dir = "/tmp";
1940
1941         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1942
1943         if (stat(fio_status_file_path, &sb))
1944                 return 0;
1945
1946         if (unlink(fio_status_file_path) < 0) {
1947                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1948                                                         strerror(errno));
1949                 log_err("fio: disabling status file updates\n");
1950                 status_file_disabled = 1;
1951         }
1952
1953         return 1;
1954 }
1955
1956 void check_for_running_stats(void)
1957 {
1958         if (status_interval) {
1959                 if (!status_interval_init) {
1960                         fio_gettime(&status_time, NULL);
1961                         status_interval_init = 1;
1962                 } else if (mtime_since_now(&status_time) >= status_interval) {
1963                         show_running_run_stats();
1964                         fio_gettime(&status_time, NULL);
1965                         return;
1966                 }
1967         }
1968         if (check_status_file()) {
1969                 show_running_run_stats();
1970                 return;
1971         }
1972 }
1973
1974 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
1975 {
1976         double val = data;
1977         double delta;
1978
1979         if (data > is->max_val)
1980                 is->max_val = data;
1981         if (data < is->min_val)
1982                 is->min_val = data;
1983
1984         delta = val - is->mean.u.f;
1985         if (delta) {
1986                 is->mean.u.f += delta / (is->samples + 1.0);
1987                 is->S.u.f += delta * (val - is->mean.u.f);
1988         }
1989
1990         is->samples++;
1991 }
1992
1993 /*
1994  * Return a struct io_logs, which is added to the tail of the log
1995  * list for 'iolog'.
1996  */
1997 static struct io_logs *get_new_log(struct io_log *iolog)
1998 {
1999         size_t new_size, new_samples;
2000         struct io_logs *cur_log;
2001
2002         /*
2003          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2004          * forever
2005          */
2006         if (!iolog->cur_log_max)
2007                 new_samples = DEF_LOG_ENTRIES;
2008         else {
2009                 new_samples = iolog->cur_log_max * 2;
2010                 if (new_samples > MAX_LOG_ENTRIES)
2011                         new_samples = MAX_LOG_ENTRIES;
2012         }
2013
2014         new_size = new_samples * log_entry_sz(iolog);
2015
2016         cur_log = smalloc(sizeof(*cur_log));
2017         if (cur_log) {
2018                 INIT_FLIST_HEAD(&cur_log->list);
2019                 cur_log->log = malloc(new_size);
2020                 if (cur_log->log) {
2021                         cur_log->nr_samples = 0;
2022                         cur_log->max_samples = new_samples;
2023                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2024                         iolog->cur_log_max = new_samples;
2025                         return cur_log;
2026                 }
2027                 sfree(cur_log);
2028         }
2029
2030         return NULL;
2031 }
2032
2033 /*
2034  * Add and return a new log chunk, or return current log if big enough
2035  */
2036 static struct io_logs *regrow_log(struct io_log *iolog)
2037 {
2038         struct io_logs *cur_log;
2039         int i;
2040
2041         if (!iolog || iolog->disabled)
2042                 goto disable;
2043
2044         cur_log = iolog_cur_log(iolog);
2045         if (!cur_log) {
2046                 cur_log = get_new_log(iolog);
2047                 if (!cur_log)
2048                         return NULL;
2049         }
2050
2051         if (cur_log->nr_samples < cur_log->max_samples)
2052                 return cur_log;
2053
2054         /*
2055          * No room for a new sample. If we're compressing on the fly, flush
2056          * out the current chunk
2057          */
2058         if (iolog->log_gz) {
2059                 if (iolog_cur_flush(iolog, cur_log)) {
2060                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2061                         return NULL;
2062                 }
2063         }
2064
2065         /*
2066          * Get a new log array, and add to our list
2067          */
2068         cur_log = get_new_log(iolog);
2069         if (!cur_log) {
2070                 log_err("fio: failed extending iolog! Will stop logging.\n");
2071                 return NULL;
2072         }
2073
2074         if (!iolog->pending || !iolog->pending->nr_samples)
2075                 return cur_log;
2076
2077         /*
2078          * Flush pending items to new log
2079          */
2080         for (i = 0; i < iolog->pending->nr_samples; i++) {
2081                 struct io_sample *src, *dst;
2082
2083                 src = get_sample(iolog, iolog->pending, i);
2084                 dst = get_sample(iolog, cur_log, i);
2085                 memcpy(dst, src, log_entry_sz(iolog));
2086         }
2087         cur_log->nr_samples = iolog->pending->nr_samples;
2088
2089         iolog->pending->nr_samples = 0;
2090         return cur_log;
2091 disable:
2092         if (iolog)
2093                 iolog->disabled = true;
2094         return NULL;
2095 }
2096
2097 void regrow_logs(struct thread_data *td)
2098 {
2099         regrow_log(td->slat_log);
2100         regrow_log(td->clat_log);
2101         regrow_log(td->clat_hist_log);
2102         regrow_log(td->lat_log);
2103         regrow_log(td->bw_log);
2104         regrow_log(td->iops_log);
2105         td->flags &= ~TD_F_REGROW_LOGS;
2106 }
2107
2108 static struct io_logs *get_cur_log(struct io_log *iolog)
2109 {
2110         struct io_logs *cur_log;
2111
2112         cur_log = iolog_cur_log(iolog);
2113         if (!cur_log) {
2114                 cur_log = get_new_log(iolog);
2115                 if (!cur_log)
2116                         return NULL;
2117         }
2118
2119         if (cur_log->nr_samples < cur_log->max_samples)
2120                 return cur_log;
2121
2122         /*
2123          * Out of space. If we're in IO offload mode, or we're not doing
2124          * per unit logging (hence logging happens outside of the IO thread
2125          * as well), add a new log chunk inline. If we're doing inline
2126          * submissions, flag 'td' as needing a log regrow and we'll take
2127          * care of it on the submission side.
2128          */
2129         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2130             !per_unit_log(iolog))
2131                 return regrow_log(iolog);
2132
2133         iolog->td->flags |= TD_F_REGROW_LOGS;
2134         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2135         return iolog->pending;
2136 }
2137
2138 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2139                              enum fio_ddir ddir, unsigned int bs,
2140                              unsigned long t, uint64_t offset)
2141 {
2142         struct io_logs *cur_log;
2143
2144         if (iolog->disabled)
2145                 return;
2146         if (flist_empty(&iolog->io_logs))
2147                 iolog->avg_last = t;
2148
2149         cur_log = get_cur_log(iolog);
2150         if (cur_log) {
2151                 struct io_sample *s;
2152
2153                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2154
2155                 s->data = data;
2156                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2157                 io_sample_set_ddir(iolog, s, ddir);
2158                 s->bs = bs;
2159
2160                 if (iolog->log_offset) {
2161                         struct io_sample_offset *so = (void *) s;
2162
2163                         so->offset = offset;
2164                 }
2165
2166                 cur_log->nr_samples++;
2167                 return;
2168         }
2169
2170         iolog->disabled = true;
2171 }
2172
2173 static inline void reset_io_stat(struct io_stat *ios)
2174 {
2175         ios->max_val = ios->min_val = ios->samples = 0;
2176         ios->mean.u.f = ios->S.u.f = 0;
2177 }
2178
2179 void reset_io_stats(struct thread_data *td)
2180 {
2181         struct thread_stat *ts = &td->ts;
2182         int i, j;
2183
2184         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2185                 reset_io_stat(&ts->clat_stat[i]);
2186                 reset_io_stat(&ts->slat_stat[i]);
2187                 reset_io_stat(&ts->lat_stat[i]);
2188                 reset_io_stat(&ts->bw_stat[i]);
2189                 reset_io_stat(&ts->iops_stat[i]);
2190
2191                 ts->io_bytes[i] = 0;
2192                 ts->runtime[i] = 0;
2193                 ts->total_io_u[i] = 0;
2194                 ts->short_io_u[i] = 0;
2195                 ts->drop_io_u[i] = 0;
2196
2197                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2198                         ts->io_u_plat[i][j] = 0;
2199         }
2200
2201         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2202                 ts->io_u_map[i] = 0;
2203                 ts->io_u_submit[i] = 0;
2204                 ts->io_u_complete[i] = 0;
2205         }
2206
2207         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2208                 ts->io_u_lat_n[i] = 0;
2209         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2210                 ts->io_u_lat_u[i] = 0;
2211         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2212                 ts->io_u_lat_m[i] = 0;
2213
2214         ts->total_submit = 0;
2215         ts->total_complete = 0;
2216 }
2217
2218 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2219                               unsigned long elapsed, bool log_max)
2220 {
2221         /*
2222          * Note an entry in the log. Use the mean from the logged samples,
2223          * making sure to properly round up. Only write a log entry if we
2224          * had actual samples done.
2225          */
2226         if (iolog->avg_window[ddir].samples) {
2227                 union io_sample_data data;
2228
2229                 if (log_max)
2230                         data.val = iolog->avg_window[ddir].max_val;
2231                 else
2232                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2233
2234                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2235         }
2236
2237         reset_io_stat(&iolog->avg_window[ddir]);
2238 }
2239
2240 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2241                              bool log_max)
2242 {
2243         int ddir;
2244
2245         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2246                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2247 }
2248
2249 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2250                            union io_sample_data data, enum fio_ddir ddir,
2251                            unsigned int bs, uint64_t offset)
2252 {
2253         unsigned long elapsed, this_window;
2254
2255         if (!ddir_rw(ddir))
2256                 return 0;
2257
2258         elapsed = mtime_since_now(&td->epoch);
2259
2260         /*
2261          * If no time averaging, just add the log sample.
2262          */
2263         if (!iolog->avg_msec) {
2264                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2265                 return 0;
2266         }
2267
2268         /*
2269          * Add the sample. If the time period has passed, then
2270          * add that entry to the log and clear.
2271          */
2272         add_stat_sample(&iolog->avg_window[ddir], data.val);
2273
2274         /*
2275          * If period hasn't passed, adding the above sample is all we
2276          * need to do.
2277          */
2278         this_window = elapsed - iolog->avg_last;
2279         if (elapsed < iolog->avg_last)
2280                 return iolog->avg_last - elapsed;
2281         else if (this_window < iolog->avg_msec) {
2282                 int diff = iolog->avg_msec - this_window;
2283
2284                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2285                         return diff;
2286         }
2287
2288         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2289
2290         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2291         return iolog->avg_msec;
2292 }
2293
2294 void finalize_logs(struct thread_data *td, bool unit_logs)
2295 {
2296         unsigned long elapsed;
2297
2298         elapsed = mtime_since_now(&td->epoch);
2299
2300         if (td->clat_log && unit_logs)
2301                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2302         if (td->slat_log && unit_logs)
2303                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2304         if (td->lat_log && unit_logs)
2305                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2306         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2307                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2308         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2309                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2310 }
2311
2312 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2313 {
2314         struct io_log *iolog;
2315
2316         if (!ddir_rw(ddir))
2317                 return;
2318
2319         iolog = agg_io_log[ddir];
2320         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2321 }
2322
2323 static void add_clat_percentile_sample(struct thread_stat *ts,
2324                                 unsigned long long nsec, enum fio_ddir ddir)
2325 {
2326         unsigned int idx = plat_val_to_idx(nsec);
2327         assert(idx < FIO_IO_U_PLAT_NR);
2328
2329         ts->io_u_plat[ddir][idx]++;
2330 }
2331
2332 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2333                      unsigned long long nsec, unsigned int bs, uint64_t offset)
2334 {
2335         unsigned long elapsed, this_window;
2336         struct thread_stat *ts = &td->ts;
2337         struct io_log *iolog = td->clat_hist_log;
2338
2339         td_io_u_lock(td);
2340
2341         add_stat_sample(&ts->clat_stat[ddir], nsec);
2342
2343         if (td->clat_log)
2344                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2345                                offset);
2346
2347         if (ts->clat_percentiles)
2348                 add_clat_percentile_sample(ts, nsec, ddir);
2349
2350         if (iolog && iolog->hist_msec) {
2351                 struct io_hist *hw = &iolog->hist_window[ddir];
2352
2353                 hw->samples++;
2354                 elapsed = mtime_since_now(&td->epoch);
2355                 if (!hw->hist_last)
2356                         hw->hist_last = elapsed;
2357                 this_window = elapsed - hw->hist_last;
2358
2359                 if (this_window >= iolog->hist_msec) {
2360                         unsigned int *io_u_plat;
2361                         struct io_u_plat_entry *dst;
2362
2363                         /*
2364                          * Make a byte-for-byte copy of the latency histogram
2365                          * stored in td->ts.io_u_plat[ddir], recording it in a
2366                          * log sample. Note that the matching call to free() is
2367                          * located in iolog.c after printing this sample to the
2368                          * log file.
2369                          */
2370                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2371                         dst = malloc(sizeof(struct io_u_plat_entry));
2372                         memcpy(&(dst->io_u_plat), io_u_plat,
2373                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2374                         flist_add(&dst->list, &hw->list);
2375                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2376                                                 elapsed, offset);
2377
2378                         /*
2379                          * Update the last time we recorded as being now, minus
2380                          * any drift in time we encountered before actually
2381                          * making the record.
2382                          */
2383                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2384                         hw->samples = 0;
2385                 }
2386         }
2387
2388         td_io_u_unlock(td);
2389 }
2390
2391 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2392                      unsigned long usec, unsigned int bs, uint64_t offset)
2393 {
2394         struct thread_stat *ts = &td->ts;
2395
2396         if (!ddir_rw(ddir))
2397                 return;
2398
2399         td_io_u_lock(td);
2400
2401         add_stat_sample(&ts->slat_stat[ddir], usec);
2402
2403         if (td->slat_log)
2404                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2405
2406         td_io_u_unlock(td);
2407 }
2408
2409 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2410                     unsigned long long nsec, unsigned int bs, uint64_t offset)
2411 {
2412         struct thread_stat *ts = &td->ts;
2413
2414         if (!ddir_rw(ddir))
2415                 return;
2416
2417         td_io_u_lock(td);
2418
2419         add_stat_sample(&ts->lat_stat[ddir], nsec);
2420
2421         if (td->lat_log)
2422                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2423                                offset);
2424
2425         td_io_u_unlock(td);
2426 }
2427
2428 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2429                    unsigned int bytes, unsigned long long spent)
2430 {
2431         struct thread_stat *ts = &td->ts;
2432         unsigned long rate;
2433
2434         if (spent)
2435                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2436         else
2437                 rate = 0;
2438
2439         td_io_u_lock(td);
2440
2441         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2442
2443         if (td->bw_log)
2444                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2445                                bytes, io_u->offset);
2446
2447         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2448         td_io_u_unlock(td);
2449 }
2450
2451 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2452                          struct timespec *t, unsigned int avg_time,
2453                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2454                          struct io_stat *stat, struct io_log *log,
2455                          bool is_kb)
2456 {
2457         unsigned long spent, rate;
2458         enum fio_ddir ddir;
2459         unsigned int next, next_log;
2460
2461         next_log = avg_time;
2462
2463         spent = mtime_since(parent_tv, t);
2464         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2465                 return avg_time - spent;
2466
2467         td_io_u_lock(td);
2468
2469         /*
2470          * Compute both read and write rates for the interval.
2471          */
2472         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2473                 uint64_t delta;
2474
2475                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2476                 if (!delta)
2477                         continue; /* No entries for interval */
2478
2479                 if (spent) {
2480                         if (is_kb)
2481                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2482                         else
2483                                 rate = (delta * 1000) / spent;
2484                 } else
2485                         rate = 0;
2486
2487                 add_stat_sample(&stat[ddir], rate);
2488
2489                 if (log) {
2490                         unsigned int bs = 0;
2491
2492                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2493                                 bs = td->o.min_bs[ddir];
2494
2495                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2496                         next_log = min(next_log, next);
2497                 }
2498
2499                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2500         }
2501
2502         timespec_add_msec(parent_tv, avg_time);
2503
2504         td_io_u_unlock(td);
2505
2506         if (spent <= avg_time)
2507                 next = avg_time;
2508         else
2509                 next = avg_time - (1 + spent - avg_time);
2510
2511         return min(next, next_log);
2512 }
2513
2514 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2515 {
2516         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2517                                 td->this_io_bytes, td->stat_io_bytes,
2518                                 td->ts.bw_stat, td->bw_log, true);
2519 }
2520
2521 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2522                      unsigned int bytes)
2523 {
2524         struct thread_stat *ts = &td->ts;
2525
2526         td_io_u_lock(td);
2527
2528         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2529
2530         if (td->iops_log)
2531                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2532                                bytes, io_u->offset);
2533
2534         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2535         td_io_u_unlock(td);
2536 }
2537
2538 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2539 {
2540         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2541                                 td->this_io_blocks, td->stat_io_blocks,
2542                                 td->ts.iops_stat, td->iops_log, false);
2543 }
2544
2545 /*
2546  * Returns msecs to next event
2547  */
2548 int calc_log_samples(void)
2549 {
2550         struct thread_data *td;
2551         unsigned int next = ~0U, tmp;
2552         struct timespec now;
2553         int i;
2554
2555         fio_gettime(&now, NULL);
2556
2557         for_each_td(td, i) {
2558                 if (!td->o.stats)
2559                         continue;
2560                 if (in_ramp_time(td) ||
2561                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2562                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2563                         continue;
2564                 }
2565                 if (!td->bw_log ||
2566                         (td->bw_log && !per_unit_log(td->bw_log))) {
2567                         tmp = add_bw_samples(td, &now);
2568                         if (tmp < next)
2569                                 next = tmp;
2570                 }
2571                 if (!td->iops_log ||
2572                         (td->iops_log && !per_unit_log(td->iops_log))) {
2573                         tmp = add_iops_samples(td, &now);
2574                         if (tmp < next)
2575                                 next = tmp;
2576                 }
2577         }
2578
2579         return next == ~0U ? 0 : next;
2580 }
2581
2582 void stat_init(void)
2583 {
2584         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2585 }
2586
2587 void stat_exit(void)
2588 {
2589         /*
2590          * When we have the mutex, we know out-of-band access to it
2591          * have ended.
2592          */
2593         fio_mutex_down(stat_mutex);
2594         fio_mutex_remove(stat_mutex);
2595 }
2596
2597 /*
2598  * Called from signal handler. Wake up status thread.
2599  */
2600 void show_running_run_stats(void)
2601 {
2602         helper_do_stat();
2603 }
2604
2605 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2606 {
2607         /* Ignore io_u's which span multiple blocks--they will just get
2608          * inaccurate counts. */
2609         int idx = (io_u->offset - io_u->file->file_offset)
2610                         / td->o.bs[DDIR_TRIM];
2611         uint32_t *info = &td->ts.block_infos[idx];
2612         assert(idx < td->ts.nr_block_infos);
2613         return info;
2614 }