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%%, avg=%5.02f, stdev=%5.02f\n",
 500                         bw_str, min, max, p_of_agg, mean, dev);
 501         }
 502         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 503                 log_buf(out, "   iops : min=%5llu, max=%5llu, avg=%5.02f, "
 504                         "stdev=%5.02f\n", min, max, mean, dev);
 505         }
 506 }
 507
 508 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 509                     const char *msg, struct buf_output *out)
 510 {
 511         int new_line = 1, i, line = 0, shown = 0;
 512
 513         for (i = 0; i < nr; i++) {
 514                 if (io_u_lat[i] <= 0.0)
 515                         continue;
 516                 shown = 1;
 517                 if (new_line) {
 518                         if (line)
 519                                 log_buf(out, "\n");
 520                         log_buf(out, "    lat (%s) : ", msg);
 521                         new_line = 0;
 522                         line = 0;
 523                 }
 524                 if (line)
 525                         log_buf(out, ", ");
 526                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 527                 line++;
 528                 if (line == 5)
 529                         new_line = 1;
 530         }
 531
 532         if (shown)
 533                 log_buf(out, "\n");
 534
 535         return shown;
 536 }
 537
 538 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 539 {
 540         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 541                                  "250=", "500=", "750=", "1000=", };
 542
 543         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 544 }
 545
 546 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 547 {
 548         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 549                                  "250=", "500=", "750=", "1000=", };
 550
 551         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 552 }
 553
 554 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 555 {
 556         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 557                                  "250=", "500=", "750=", "1000=", "2000=",
 558                                  ">=2000=", };
 559
 560         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 561 }
 562
 563 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 564 {
 565         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 566         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 567         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 568
 569         stat_calc_lat_n(ts, io_u_lat_n);
 570         stat_calc_lat_u(ts, io_u_lat_u);
 571         stat_calc_lat_m(ts, io_u_lat_m);
 572
 573         show_lat_n(io_u_lat_n, out);
 574         show_lat_u(io_u_lat_u, out);
 575         show_lat_m(io_u_lat_m, out);
 576 }
 577
 578 static int block_state_category(int block_state)
 579 {
 580         switch (block_state) {
 581         case BLOCK_STATE_UNINIT:
 582                 return 0;
 583         case BLOCK_STATE_TRIMMED:
 584         case BLOCK_STATE_WRITTEN:
 585                 return 1;
 586         case BLOCK_STATE_WRITE_FAILURE:
 587         case BLOCK_STATE_TRIM_FAILURE:
 588                 return 2;
 589         default:
 590                 /* Silence compile warning on some BSDs and have a return */
 591                 assert(0);
 592                 return -1;
 593         }
 594 }
 595
 596 static int compare_block_infos(const void *bs1, const void *bs2)
 597 {
 598         uint32_t block1 = *(uint32_t *)bs1;
 599         uint32_t block2 = *(uint32_t *)bs2;
 600         int state1 = BLOCK_INFO_STATE(block1);
 601         int state2 = BLOCK_INFO_STATE(block2);
 602         int bscat1 = block_state_category(state1);
 603         int bscat2 = block_state_category(state2);
 604         int cycles1 = BLOCK_INFO_TRIMS(block1);
 605         int cycles2 = BLOCK_INFO_TRIMS(block2);
 606
 607         if (bscat1 < bscat2)
 608                 return -1;
 609         if (bscat1 > bscat2)
 610                 return 1;
 611
 612         if (cycles1 < cycles2)
 613                 return -1;
 614         if (cycles1 > cycles2)
 615                 return 1;
 616
 617         if (state1 < state2)
 618                 return -1;
 619         if (state1 > state2)
 620                 return 1;
 621
 622         assert(block1 == block2);
 623         return 0;
 624 }
 625
 626 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 627                                   fio_fp64_t *plist, unsigned int **percentiles,
 628                                   unsigned int *types)
 629 {
 630         int len = 0;
 631         int i, nr_uninit;
 632
 633         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 634
 635         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 636                 len++;
 637
 638         if (!len)
 639                 return 0;
 640
 641         /*
 642          * Sort the percentile list. Note that it may already be sorted if
 643          * we are using the default values, but since it's a short list this
 644          * isn't a worry. Also note that this does not work for NaN values.
 645          */
 646         if (len > 1)
 647                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 648
 649         nr_uninit = 0;
 650         /* Start only after the uninit entries end */
 651         for (nr_uninit = 0;
 652              nr_uninit < nr_block_infos
 653                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 654              nr_uninit ++)
 655                 ;
 656
 657         if (nr_uninit == nr_block_infos)
 658                 return 0;
 659
 660         *percentiles = calloc(len, sizeof(**percentiles));
 661
 662         for (i = 0; i < len; i++) {
 663                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 664                                 + nr_uninit;
 665                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 666         }
 667
 668         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 669         for (i = 0; i < nr_block_infos; i++)
 670                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 671
 672         return len;
 673 }
 674
 675 static const char *block_state_names[] = {
 676         [BLOCK_STATE_UNINIT] = "unwritten",
 677         [BLOCK_STATE_TRIMMED] = "trimmed",
 678         [BLOCK_STATE_WRITTEN] = "written",
 679         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 680         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 681 };
 682
 683 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 684                              fio_fp64_t *plist, struct buf_output *out)
 685 {
 686         int len, pos, i;
 687         unsigned int *percentiles = NULL;
 688         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 689
 690         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 691                                      &percentiles, block_state_counts);
 692
 693         log_buf(out, "  block lifetime percentiles :\n   |");
 694         pos = 0;
 695         for (i = 0; i < len; i++) {
 696                 uint32_t block_info = percentiles[i];
 697 #define LINE_LENGTH     75
 698                 char str[LINE_LENGTH];
 699                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 700                                      plist[i].u.f, block_info,
 701                                      i == len - 1 ? '\n' : ',');
 702                 assert(strln < LINE_LENGTH);
 703                 if (pos + strln > LINE_LENGTH) {
 704                         pos = 0;
 705                         log_buf(out, "\n   |");
 706                 }
 707                 log_buf(out, "%s", str);
 708                 pos += strln;
 709 #undef LINE_LENGTH
 710         }
 711         if (percentiles)
 712                 free(percentiles);
 713
 714         log_buf(out, "        states               :");
 715         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 716                 log_buf(out, " %s=%u%c",
 717                          block_state_names[i], block_state_counts[i],
 718                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 719 }
 720
 721 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 722 {
 723         char *p1, *p1alt, *p2;
 724         unsigned long long bw_mean, iops_mean;
 725         const int i2p = is_power_of_2(ts->kb_base);
 726
 727         if (!ts->ss_dur)
 728                 return;
 729
 730         bw_mean = steadystate_bw_mean(ts);
 731         iops_mean = steadystate_iops_mean(ts);
 732
 733         p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
 734         p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
 735         p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
 736
 737         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 738                 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
 739                 p1, p1alt, p2,
 740                 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
 741                 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
 742                 ts->ss_criterion.u.f,
 743                 ts->ss_state & __FIO_SS_PCT ? "%" : "");
 744
 745         free(p1);
 746         free(p1alt);
 747         free(p2);
 748 }
 749
 750 static void show_thread_status_normal(struct thread_stat *ts,
 751                                       struct group_run_stats *rs,
 752                                       struct buf_output *out)
 753 {
 754         double usr_cpu, sys_cpu;
 755         unsigned long runtime;
 756         double io_u_dist[FIO_IO_U_MAP_NR];
 757         time_t time_p;
 758         char time_buf[32];
 759
 760         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 761                 return;
 762
 763         memset(time_buf, 0, sizeof(time_buf));
 764
 765         time(&time_p);
 766         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 767
 768         if (!ts->error) {
 769                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 770                                         ts->name, ts->groupid, ts->members,
 771                                         ts->error, (int) ts->pid, time_buf);
 772         } else {
 773                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 774                                         ts->name, ts->groupid, ts->members,
 775                                         ts->error, ts->verror, (int) ts->pid,
 776                                         time_buf);
 777         }
 778
 779         if (strlen(ts->description))
 780                 log_buf(out, "  Description  : [%s]\n", ts->description);
 781
 782         if (ts->io_bytes[DDIR_READ])
 783                 show_ddir_status(rs, ts, DDIR_READ, out);
 784         if (ts->io_bytes[DDIR_WRITE])
 785                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 786         if (ts->io_bytes[DDIR_TRIM])
 787                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 788
 789         show_latencies(ts, out);
 790
 791         runtime = ts->total_run_time;
 792         if (runtime) {
 793                 double runt = (double) runtime;
 794
 795                 usr_cpu = (double) ts->usr_time * 100 / runt;
 796                 sys_cpu = (double) ts->sys_time * 100 / runt;
 797         } else {
 798                 usr_cpu = 0;
 799                 sys_cpu = 0;
 800         }
 801
 802         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 803                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 804                         (unsigned long long) ts->ctx,
 805                         (unsigned long long) ts->majf,
 806                         (unsigned long long) ts->minf);
 807
 808         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 809         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 810                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 811                                         io_u_dist[1], io_u_dist[2],
 812                                         io_u_dist[3], io_u_dist[4],
 813                                         io_u_dist[5], io_u_dist[6]);
 814
 815         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 816         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 817                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 818                                         io_u_dist[1], io_u_dist[2],
 819                                         io_u_dist[3], io_u_dist[4],
 820                                         io_u_dist[5], io_u_dist[6]);
 821         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 822         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 823                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 824                                         io_u_dist[1], io_u_dist[2],
 825                                         io_u_dist[3], io_u_dist[4],
 826                                         io_u_dist[5], io_u_dist[6]);
 827         log_buf(out, "     issued rwt: total=%llu,%llu,%llu,"
 828                                  " short=%llu,%llu,%llu,"
 829                                  " dropped=%llu,%llu,%llu\n",
 830                                         (unsigned long long) ts->total_io_u[0],
 831                                         (unsigned long long) ts->total_io_u[1],
 832                                         (unsigned long long) ts->total_io_u[2],
 833                                         (unsigned long long) ts->short_io_u[0],
 834                                         (unsigned long long) ts->short_io_u[1],
 835                                         (unsigned long long) ts->short_io_u[2],
 836                                         (unsigned long long) ts->drop_io_u[0],
 837                                         (unsigned long long) ts->drop_io_u[1],
 838                                         (unsigned long long) ts->drop_io_u[2]);
 839         if (ts->continue_on_error) {
 840                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 841                                         (unsigned long long)ts->total_err_count,
 842                                         ts->first_error,
 843                                         strerror(ts->first_error));
 844         }
 845         if (ts->latency_depth) {
 846                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 847                                         (unsigned long long)ts->latency_target,
 848                                         (unsigned long long)ts->latency_window,
 849                                         ts->latency_percentile.u.f,
 850                                         ts->latency_depth);
 851         }
 852
 853         if (ts->nr_block_infos)
 854                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 855                                   ts->percentile_list, out);
 856
 857         if (ts->ss_dur)
 858                 show_ss_normal(ts, out);
 859 }
 860
 861 static void show_ddir_status_terse(struct thread_stat *ts,
 862                                    struct group_run_stats *rs, int ddir,
 863                                    struct buf_output *out)
 864 {
 865         unsigned long long min, max, minv, maxv, bw, iops;
 866         unsigned long long *ovals = NULL;
 867         double mean, dev;
 868         unsigned int len;
 869         int i;
 870
 871         assert(ddir_rw(ddir));
 872
 873         iops = bw = 0;
 874         if (ts->runtime[ddir]) {
 875                 uint64_t runt = ts->runtime[ddir];
 876
 877                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 878                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 879         }
 880
 881         log_buf(out, ";%llu;%llu;%llu;%llu",
 882                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 883                                         (unsigned long long) ts->runtime[ddir]);
 884
 885         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 886                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 887         else
 888                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 889
 890         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 891                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 892         else
 893                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 894
 895         if (ts->clat_percentiles) {
 896                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 897                                         ts->clat_stat[ddir].samples,
 898                                         ts->percentile_list, &ovals, &maxv,
 899                                         &minv);
 900         } else
 901                 len = 0;
 902
 903         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 904                 if (i >= len) {
 905                         log_buf(out, ";0%%=0");
 906                         continue;
 907                 }
 908                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
 909         }
 910
 911         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 912                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 913         else
 914                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 915
 916         if (ovals)
 917                 free(ovals);
 918
 919         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 920                 double p_of_agg = 100.0;
 921
 922                 if (rs->agg[ddir]) {
 923                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 924                         if (p_of_agg > 100.0)
 925                                 p_of_agg = 100.0;
 926                 }
 927
 928                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 929         } else
 930                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
 931 }
 932
 933 static void add_ddir_status_json(struct thread_stat *ts,
 934                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 935 {
 936         unsigned long long min, max, minv, maxv;
 937         unsigned long long bw;
 938         unsigned long long *ovals = NULL;
 939         double mean, dev, iops;
 940         unsigned int len;
 941         int i;
 942         const char *ddirname[] = {"read", "write", "trim"};
 943         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 944         char buf[120];
 945         double p_of_agg = 100.0;
 946
 947         assert(ddir_rw(ddir));
 948
 949         if (ts->unified_rw_rep && ddir != DDIR_READ)
 950                 return;
 951
 952         dir_object = json_create_object();
 953         json_object_add_value_object(parent,
 954                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 955
 956         bw = 0;
 957         iops = 0.0;
 958         if (ts->runtime[ddir]) {
 959                 uint64_t runt = ts->runtime[ddir];
 960
 961                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 962                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 963         }
 964
 965         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
 966         json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
 967         json_object_add_value_int(dir_object, "bw", bw);
 968         json_object_add_value_float(dir_object, "iops", iops);
 969         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 970         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 971         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 972         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 973
 974         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 975                 min = max = 0;
 976                 mean = dev = 0.0;
 977         }
 978         tmp_object = json_create_object();
 979         json_object_add_value_object(dir_object, "slat_ns", tmp_object);
 980         json_object_add_value_int(tmp_object, "min", min);
 981         json_object_add_value_int(tmp_object, "max", max);
 982         json_object_add_value_float(tmp_object, "mean", mean);
 983         json_object_add_value_float(tmp_object, "stddev", dev);
 984
 985         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 986                 min = max = 0;
 987                 mean = dev = 0.0;
 988         }
 989         tmp_object = json_create_object();
 990         json_object_add_value_object(dir_object, "clat_ns", tmp_object);
 991         json_object_add_value_int(tmp_object, "min", min);
 992         json_object_add_value_int(tmp_object, "max", max);
 993         json_object_add_value_float(tmp_object, "mean", mean);
 994         json_object_add_value_float(tmp_object, "stddev", dev);
 995
 996         if (ts->clat_percentiles) {
 997                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 998                                         ts->clat_stat[ddir].samples,
 999                                         ts->percentile_list, &ovals, &maxv,
1000                                         &minv);
1001         } else
1002                 len = 0;
1003
1004         percentile_object = json_create_object();
1005         json_object_add_value_object(tmp_object, "percentile", percentile_object);
1006         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1007                 if (i >= len) {
1008                         json_object_add_value_int(percentile_object, "0.00", 0);
1009                         continue;
1010                 }
1011                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1012                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1013         }
1014
1015         if (output_format & FIO_OUTPUT_JSON_PLUS) {
1016                 clat_bins_object = json_create_object();
1017                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1018                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1019                         if (ts->io_u_plat[ddir][i]) {
1020                                 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1021                                 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1022                         }
1023                 }
1024         }
1025
1026         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1027                 min = max = 0;
1028                 mean = dev = 0.0;
1029         }
1030         tmp_object = json_create_object();
1031         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1032         json_object_add_value_int(tmp_object, "min", min);
1033         json_object_add_value_int(tmp_object, "max", max);
1034         json_object_add_value_float(tmp_object, "mean", mean);
1035         json_object_add_value_float(tmp_object, "stddev", dev);
1036         if (ovals)
1037                 free(ovals);
1038
1039         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1040                 if (rs->agg[ddir]) {
1041                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
1042                         if (p_of_agg > 100.0)
1043                                 p_of_agg = 100.0;
1044                 }
1045         } else {
1046                 min = max = 0;
1047                 p_of_agg = mean = dev = 0.0;
1048         }
1049         json_object_add_value_int(dir_object, "bw_min", min);
1050         json_object_add_value_int(dir_object, "bw_max", max);
1051         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1052         json_object_add_value_float(dir_object, "bw_mean", mean);
1053         json_object_add_value_float(dir_object, "bw_dev", dev);
1054
1055         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1056                 min = max = 0;
1057                 mean = dev = 0.0;
1058         }
1059         json_object_add_value_int(dir_object, "iops_min", min);
1060         json_object_add_value_int(dir_object, "iops_max", max);
1061         json_object_add_value_float(dir_object, "iops_mean", mean);
1062         json_object_add_value_float(dir_object, "iops_stddev", dev);
1063 }
1064
1065 static void show_thread_status_terse_v2(struct thread_stat *ts,
1066                                         struct group_run_stats *rs,
1067                                         struct buf_output *out)
1068 {
1069         double io_u_dist[FIO_IO_U_MAP_NR];
1070         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1071         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1072         double usr_cpu, sys_cpu;
1073         int i;
1074
1075         /* General Info */
1076         log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1077         /* Log Read Status */
1078         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1079         /* Log Write Status */
1080         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1081         /* Log Trim Status */
1082         show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1083
1084         /* CPU Usage */
1085         if (ts->total_run_time) {
1086                 double runt = (double) ts->total_run_time;
1087
1088                 usr_cpu = (double) ts->usr_time * 100 / runt;
1089                 sys_cpu = (double) ts->sys_time * 100 / runt;
1090         } else {
1091                 usr_cpu = 0;
1092                 sys_cpu = 0;
1093         }
1094
1095         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1096                                                 (unsigned long long) ts->ctx,
1097                                                 (unsigned long long) ts->majf,
1098                                                 (unsigned long long) ts->minf);
1099
1100         /* Calc % distribution of IO depths, usecond, msecond latency */
1101         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1102         stat_calc_lat_nu(ts, io_u_lat_u);
1103         stat_calc_lat_m(ts, io_u_lat_m);
1104
1105         /* Only show fixed 7 I/O depth levels*/
1106         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1107                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1108                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1109
1110         /* Microsecond latency */
1111         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1112                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1113         /* Millisecond latency */
1114         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1115                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1116         /* Additional output if continue_on_error set - default off*/
1117         if (ts->continue_on_error)
1118                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1119         log_buf(out, "\n");
1120
1121         /* Additional output if description is set */
1122         if (strlen(ts->description))
1123                 log_buf(out, ";%s", ts->description);
1124
1125         log_buf(out, "\n");
1126 }
1127
1128 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1129                                            struct group_run_stats *rs, int ver,
1130                                            struct buf_output *out)
1131 {
1132         double io_u_dist[FIO_IO_U_MAP_NR];
1133         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1134         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1135         double usr_cpu, sys_cpu;
1136         int i;
1137
1138         /* General Info */
1139         log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1140                                         ts->name, ts->groupid, ts->error);
1141         /* Log Read Status */
1142         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1143         /* Log Write Status */
1144         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1145         /* Log Trim Status */
1146         if (ver == 4)
1147                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1148
1149         /* CPU Usage */
1150         if (ts->total_run_time) {
1151                 double runt = (double) ts->total_run_time;
1152
1153                 usr_cpu = (double) ts->usr_time * 100 / runt;
1154                 sys_cpu = (double) ts->sys_time * 100 / runt;
1155         } else {
1156                 usr_cpu = 0;
1157                 sys_cpu = 0;
1158         }
1159
1160         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1161                                                 (unsigned long long) ts->ctx,
1162                                                 (unsigned long long) ts->majf,
1163                                                 (unsigned long long) ts->minf);
1164
1165         /* Calc % distribution of IO depths, usecond, msecond latency */
1166         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1167         stat_calc_lat_nu(ts, io_u_lat_u);
1168         stat_calc_lat_m(ts, io_u_lat_m);
1169
1170         /* Only show fixed 7 I/O depth levels*/
1171         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1172                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1173                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1174
1175         /* Microsecond latency */
1176         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1177                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1178         /* Millisecond latency */
1179         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1180                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1181
1182         /* disk util stats, if any */
1183         show_disk_util(1, NULL, out);
1184
1185         /* Additional output if continue_on_error set - default off*/
1186         if (ts->continue_on_error)
1187                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1188
1189         /* Additional output if description is set */
1190         if (strlen(ts->description))
1191                 log_buf(out, ";%s", ts->description);
1192
1193         log_buf(out, "\n");
1194 }
1195
1196 static void json_add_job_opts(struct json_object *root, const char *name,
1197                               struct flist_head *opt_list, bool num_jobs)
1198 {
1199         struct json_object *dir_object;
1200         struct flist_head *entry;
1201         struct print_option *p;
1202
1203         if (flist_empty(opt_list))
1204                 return;
1205
1206         dir_object = json_create_object();
1207         json_object_add_value_object(root, name, dir_object);
1208
1209         flist_for_each(entry, opt_list) {
1210                 const char *pos = "";
1211
1212                 p = flist_entry(entry, struct print_option, list);
1213                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1214                         continue;
1215                 if (p->value)
1216                         pos = p->value;
1217                 json_object_add_value_string(dir_object, p->name, pos);
1218         }
1219 }
1220
1221 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1222                                                    struct group_run_stats *rs,
1223                                                    struct flist_head *opt_list)
1224 {
1225         struct json_object *root, *tmp;
1226         struct jobs_eta *je;
1227         double io_u_dist[FIO_IO_U_MAP_NR];
1228         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1229         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1230         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1231         double usr_cpu, sys_cpu;
1232         int i;
1233         size_t size;
1234
1235         root = json_create_object();
1236         json_object_add_value_string(root, "jobname", ts->name);
1237         json_object_add_value_int(root, "groupid", ts->groupid);
1238         json_object_add_value_int(root, "error", ts->error);
1239
1240         /* ETA Info */
1241         je = get_jobs_eta(true, &size);
1242         if (je) {
1243                 json_object_add_value_int(root, "eta", je->eta_sec);
1244                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1245         }
1246
1247         if (opt_list)
1248                 json_add_job_opts(root, "job options", opt_list, true);
1249
1250         add_ddir_status_json(ts, rs, DDIR_READ, root);
1251         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1252         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1253
1254         /* CPU Usage */
1255         if (ts->total_run_time) {
1256                 double runt = (double) ts->total_run_time;
1257
1258                 usr_cpu = (double) ts->usr_time * 100 / runt;
1259                 sys_cpu = (double) ts->sys_time * 100 / runt;
1260         } else {
1261                 usr_cpu = 0;
1262                 sys_cpu = 0;
1263         }
1264         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1265         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1266         json_object_add_value_int(root, "ctx", ts->ctx);
1267         json_object_add_value_int(root, "majf", ts->majf);
1268         json_object_add_value_int(root, "minf", ts->minf);
1269
1270
1271         /* Calc % distribution of IO depths, usecond, msecond latency */
1272         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1273         stat_calc_lat_n(ts, io_u_lat_n);
1274         stat_calc_lat_u(ts, io_u_lat_u);
1275         stat_calc_lat_m(ts, io_u_lat_m);
1276
1277         tmp = json_create_object();
1278         json_object_add_value_object(root, "iodepth_level", tmp);
1279         /* Only show fixed 7 I/O depth levels*/
1280         for (i = 0; i < 7; i++) {
1281                 char name[20];
1282                 if (i < 6)
1283                         snprintf(name, 20, "%d", 1 << i);
1284                 else
1285                         snprintf(name, 20, ">=%d", 1 << i);
1286                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1287         }
1288
1289         /* Nanosecond latency */
1290         tmp = json_create_object();
1291         json_object_add_value_object(root, "latency_ns", tmp);
1292         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1293                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1294                                  "250", "500", "750", "1000", };
1295                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1296         }
1297         /* Microsecond latency */
1298         tmp = json_create_object();
1299         json_object_add_value_object(root, "latency_us", tmp);
1300         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1301                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1302                                  "250", "500", "750", "1000", };
1303                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1304         }
1305         /* Millisecond latency */
1306         tmp = json_create_object();
1307         json_object_add_value_object(root, "latency_ms", tmp);
1308         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1309                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1310                                  "250", "500", "750", "1000", "2000",
1311                                  ">=2000", };
1312                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1313         }
1314
1315         /* Additional output if continue_on_error set - default off*/
1316         if (ts->continue_on_error) {
1317                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1318                 json_object_add_value_int(root, "first_error", ts->first_error);
1319         }
1320
1321         if (ts->latency_depth) {
1322                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1323                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1324                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1325                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1326         }
1327
1328         /* Additional output if description is set */
1329         if (strlen(ts->description))
1330                 json_object_add_value_string(root, "desc", ts->description);
1331
1332         if (ts->nr_block_infos) {
1333                 /* Block error histogram and types */
1334                 int len;
1335                 unsigned int *percentiles = NULL;
1336                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1337
1338                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1339                                              ts->percentile_list,
1340                                              &percentiles, block_state_counts);
1341
1342                 if (len) {
1343                         struct json_object *block, *percentile_object, *states;
1344                         int state;
1345                         block = json_create_object();
1346                         json_object_add_value_object(root, "block", block);
1347
1348                         percentile_object = json_create_object();
1349                         json_object_add_value_object(block, "percentiles",
1350                                                      percentile_object);
1351                         for (i = 0; i < len; i++) {
1352                                 char buf[20];
1353                                 snprintf(buf, sizeof(buf), "%f",
1354                                          ts->percentile_list[i].u.f);
1355                                 json_object_add_value_int(percentile_object,
1356                                                           (const char *)buf,
1357                                                           percentiles[i]);
1358                         }
1359
1360                         states = json_create_object();
1361                         json_object_add_value_object(block, "states", states);
1362                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1363                                 json_object_add_value_int(states,
1364                                         block_state_names[state],
1365                                         block_state_counts[state]);
1366                         }
1367                         free(percentiles);
1368                 }
1369         }
1370
1371         if (ts->ss_dur) {
1372                 struct json_object *data;
1373                 struct json_array *iops, *bw;
1374                 int i, j, k;
1375                 char ss_buf[64];
1376
1377                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1378                         ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1379                         ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1380                         (float) ts->ss_limit.u.f,
1381                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1382
1383                 tmp = json_create_object();
1384                 json_object_add_value_object(root, "steadystate", tmp);
1385                 json_object_add_value_string(tmp, "ss", ss_buf);
1386                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1387                 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1388
1389                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1390                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1391                 json_object_add_value_string(tmp, "criterion", ss_buf);
1392                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1393                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1394
1395                 data = json_create_object();
1396                 json_object_add_value_object(tmp, "data", data);
1397                 bw = json_create_array();
1398                 iops = json_create_array();
1399
1400                 /*
1401                 ** if ss was attained or the buffer is not full,
1402                 ** ss->head points to the first element in the list.
1403                 ** otherwise it actually points to the second element
1404                 ** in the list
1405                 */
1406                 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1407                         j = ts->ss_head;
1408                 else
1409                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1410                 for (i = 0; i < ts->ss_dur; i++) {
1411                         k = (j + i) % ts->ss_dur;
1412                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1413                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1414                 }
1415                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1416                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1417                 json_object_add_value_array(data, "iops", iops);
1418                 json_object_add_value_array(data, "bw", bw);
1419         }
1420
1421         return root;
1422 }
1423
1424 static void show_thread_status_terse(struct thread_stat *ts,
1425                                      struct group_run_stats *rs,
1426                                      struct buf_output *out)
1427 {
1428         if (terse_version == 2)
1429                 show_thread_status_terse_v2(ts, rs, out);
1430         else if (terse_version == 3 || terse_version == 4)
1431                 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1432         else
1433                 log_err("fio: bad terse version!? %d\n", terse_version);
1434 }
1435
1436 struct json_object *show_thread_status(struct thread_stat *ts,
1437                                        struct group_run_stats *rs,
1438                                        struct flist_head *opt_list,
1439                                        struct buf_output *out)
1440 {
1441         struct json_object *ret = NULL;
1442
1443         if (output_format & FIO_OUTPUT_TERSE)
1444                 show_thread_status_terse(ts, rs,  out);
1445         if (output_format & FIO_OUTPUT_JSON)
1446                 ret = show_thread_status_json(ts, rs, opt_list);
1447         if (output_format & FIO_OUTPUT_NORMAL)
1448                 show_thread_status_normal(ts, rs,  out);
1449
1450         return ret;
1451 }
1452
1453 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1454 {
1455         double mean, S;
1456
1457         if (src->samples == 0)
1458                 return;
1459
1460         dst->min_val = min(dst->min_val, src->min_val);
1461         dst->max_val = max(dst->max_val, src->max_val);
1462
1463         /*
1464          * Compute new mean and S after the merge
1465          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1466          *  #Parallel_algorithm>
1467          */
1468         if (first) {
1469                 mean = src->mean.u.f;
1470                 S = src->S.u.f;
1471         } else {
1472                 double delta = src->mean.u.f - dst->mean.u.f;
1473
1474                 mean = ((src->mean.u.f * src->samples) +
1475                         (dst->mean.u.f * dst->samples)) /
1476                         (dst->samples + src->samples);
1477
1478                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1479                         (dst->samples * src->samples) /
1480                         (dst->samples + src->samples);
1481         }
1482
1483         dst->samples += src->samples;
1484         dst->mean.u.f = mean;
1485         dst->S.u.f = S;
1486 }
1487
1488 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1489 {
1490         int i;
1491
1492         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1493                 if (dst->max_run[i] < src->max_run[i])
1494                         dst->max_run[i] = src->max_run[i];
1495                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1496                         dst->min_run[i] = src->min_run[i];
1497                 if (dst->max_bw[i] < src->max_bw[i])
1498                         dst->max_bw[i] = src->max_bw[i];
1499                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1500                         dst->min_bw[i] = src->min_bw[i];
1501
1502                 dst->iobytes[i] += src->iobytes[i];
1503                 dst->agg[i] += src->agg[i];
1504         }
1505
1506         if (!dst->kb_base)
1507                 dst->kb_base = src->kb_base;
1508         if (!dst->unit_base)
1509                 dst->unit_base = src->unit_base;
1510 }
1511
1512 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1513                       bool first)
1514 {
1515         int l, k;
1516
1517         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1518                 if (!dst->unified_rw_rep) {
1519                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1520                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1521                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1522                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1523                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1524
1525                         dst->io_bytes[l] += src->io_bytes[l];
1526
1527                         if (dst->runtime[l] < src->runtime[l])
1528                                 dst->runtime[l] = src->runtime[l];
1529                 } else {
1530                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1531                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1532                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1533                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1534                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1535
1536                         dst->io_bytes[0] += src->io_bytes[l];
1537
1538                         if (dst->runtime[0] < src->runtime[l])
1539                                 dst->runtime[0] = src->runtime[l];
1540
1541                         /*
1542                          * We're summing to the same destination, so override
1543                          * 'first' after the first iteration of the loop
1544                          */
1545                         first = false;
1546                 }
1547         }
1548
1549         dst->usr_time += src->usr_time;
1550         dst->sys_time += src->sys_time;
1551         dst->ctx += src->ctx;
1552         dst->majf += src->majf;
1553         dst->minf += src->minf;
1554
1555         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1556                 dst->io_u_map[k] += src->io_u_map[k];
1557         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1558                 dst->io_u_submit[k] += src->io_u_submit[k];
1559         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1560                 dst->io_u_complete[k] += src->io_u_complete[k];
1561         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1562                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1563         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1564                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1565         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1566                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1567
1568         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1569                 if (!dst->unified_rw_rep) {
1570                         dst->total_io_u[k] += src->total_io_u[k];
1571                         dst->short_io_u[k] += src->short_io_u[k];
1572                         dst->drop_io_u[k] += src->drop_io_u[k];
1573                 } else {
1574                         dst->total_io_u[0] += src->total_io_u[k];
1575                         dst->short_io_u[0] += src->short_io_u[k];
1576                         dst->drop_io_u[0] += src->drop_io_u[k];
1577                 }
1578         }
1579
1580         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1581                 int m;
1582
1583                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1584                         if (!dst->unified_rw_rep)
1585                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1586                         else
1587                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1588                 }
1589         }
1590
1591         dst->total_run_time += src->total_run_time;
1592         dst->total_submit += src->total_submit;
1593         dst->total_complete += src->total_complete;
1594 }
1595
1596 void init_group_run_stat(struct group_run_stats *gs)
1597 {
1598         int i;
1599         memset(gs, 0, sizeof(*gs));
1600
1601         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1602                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1603 }
1604
1605 void init_thread_stat(struct thread_stat *ts)
1606 {
1607         int j;
1608
1609         memset(ts, 0, sizeof(*ts));
1610
1611         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1612                 ts->lat_stat[j].min_val = -1UL;
1613                 ts->clat_stat[j].min_val = -1UL;
1614                 ts->slat_stat[j].min_val = -1UL;
1615                 ts->bw_stat[j].min_val = -1UL;
1616                 ts->iops_stat[j].min_val = -1UL;
1617         }
1618         ts->groupid = -1;
1619 }
1620
1621 void __show_run_stats(void)
1622 {
1623         struct group_run_stats *runstats, *rs;
1624         struct thread_data *td;
1625         struct thread_stat *threadstats, *ts;
1626         int i, j, k, nr_ts, last_ts, idx;
1627         int kb_base_warned = 0;
1628         int unit_base_warned = 0;
1629         struct json_object *root = NULL;
1630         struct json_array *array = NULL;
1631         struct buf_output output[FIO_OUTPUT_NR];
1632         struct flist_head **opt_lists;
1633
1634         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1635
1636         for (i = 0; i < groupid + 1; i++)
1637                 init_group_run_stat(&runstats[i]);
1638
1639         /*
1640          * find out how many threads stats we need. if group reporting isn't
1641          * enabled, it's one-per-td.
1642          */
1643         nr_ts = 0;
1644         last_ts = -1;
1645         for_each_td(td, i) {
1646                 if (!td->o.group_reporting) {
1647                         nr_ts++;
1648                         continue;
1649                 }
1650                 if (last_ts == td->groupid)
1651                         continue;
1652                 if (!td->o.stats)
1653                         continue;
1654
1655                 last_ts = td->groupid;
1656                 nr_ts++;
1657         }
1658
1659         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1660         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1661
1662         for (i = 0; i < nr_ts; i++) {
1663                 init_thread_stat(&threadstats[i]);
1664                 opt_lists[i] = NULL;
1665         }
1666
1667         j = 0;
1668         last_ts = -1;
1669         idx = 0;
1670         for_each_td(td, i) {
1671                 if (!td->o.stats)
1672                         continue;
1673                 if (idx && (!td->o.group_reporting ||
1674                     (td->o.group_reporting && last_ts != td->groupid))) {
1675                         idx = 0;
1676                         j++;
1677                 }
1678
1679                 last_ts = td->groupid;
1680
1681                 ts = &threadstats[j];
1682
1683                 ts->clat_percentiles = td->o.clat_percentiles;
1684                 ts->percentile_precision = td->o.percentile_precision;
1685                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1686                 opt_lists[j] = &td->opt_list;
1687
1688                 idx++;
1689                 ts->members++;
1690
1691                 if (ts->groupid == -1) {
1692                         /*
1693                          * These are per-group shared already
1694                          */
1695                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1696                         if (td->o.description)
1697                                 strncpy(ts->description, td->o.description,
1698                                                 FIO_JOBDESC_SIZE - 1);
1699                         else
1700                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1701
1702                         /*
1703                          * If multiple entries in this group, this is
1704                          * the first member.
1705                          */
1706                         ts->thread_number = td->thread_number;
1707                         ts->groupid = td->groupid;
1708
1709                         /*
1710                          * first pid in group, not very useful...
1711                          */
1712                         ts->pid = td->pid;
1713
1714                         ts->kb_base = td->o.kb_base;
1715                         ts->unit_base = td->o.unit_base;
1716                         ts->unified_rw_rep = td->o.unified_rw_rep;
1717                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1718                         log_info("fio: kb_base differs for jobs in group, using"
1719                                  " %u as the base\n", ts->kb_base);
1720                         kb_base_warned = 1;
1721                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1722                         log_info("fio: unit_base differs for jobs in group, using"
1723                                  " %u as the base\n", ts->unit_base);
1724                         unit_base_warned = 1;
1725                 }
1726
1727                 ts->continue_on_error = td->o.continue_on_error;
1728                 ts->total_err_count += td->total_err_count;
1729                 ts->first_error = td->first_error;
1730                 if (!ts->error) {
1731                         if (!td->error && td->o.continue_on_error &&
1732                             td->first_error) {
1733                                 ts->error = td->first_error;
1734                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1735                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1736                         } else  if (td->error) {
1737                                 ts->error = td->error;
1738                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1739                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1740                         }
1741                 }
1742
1743                 ts->latency_depth = td->latency_qd;
1744                 ts->latency_target = td->o.latency_target;
1745                 ts->latency_percentile = td->o.latency_percentile;
1746                 ts->latency_window = td->o.latency_window;
1747
1748                 ts->nr_block_infos = td->ts.nr_block_infos;
1749                 for (k = 0; k < ts->nr_block_infos; k++)
1750                         ts->block_infos[k] = td->ts.block_infos[k];
1751
1752                 sum_thread_stats(ts, &td->ts, idx == 1);
1753
1754                 if (td->o.ss_dur) {
1755                         ts->ss_state = td->ss.state;
1756                         ts->ss_dur = td->ss.dur;
1757                         ts->ss_head = td->ss.head;
1758                         ts->ss_bw_data = td->ss.bw_data;
1759                         ts->ss_iops_data = td->ss.iops_data;
1760                         ts->ss_limit.u.f = td->ss.limit;
1761                         ts->ss_slope.u.f = td->ss.slope;
1762                         ts->ss_deviation.u.f = td->ss.deviation;
1763                         ts->ss_criterion.u.f = td->ss.criterion;
1764                 }
1765                 else
1766                         ts->ss_dur = ts->ss_state = 0;
1767         }
1768
1769         for (i = 0; i < nr_ts; i++) {
1770                 unsigned long long bw;
1771
1772                 ts = &threadstats[i];
1773                 if (ts->groupid == -1)
1774                         continue;
1775                 rs = &runstats[ts->groupid];
1776                 rs->kb_base = ts->kb_base;
1777                 rs->unit_base = ts->unit_base;
1778                 rs->unified_rw_rep += ts->unified_rw_rep;
1779
1780                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1781                         if (!ts->runtime[j])
1782                                 continue;
1783                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1784                                 rs->min_run[j] = ts->runtime[j];
1785                         if (ts->runtime[j] > rs->max_run[j])
1786                                 rs->max_run[j] = ts->runtime[j];
1787
1788                         bw = 0;
1789                         if (ts->runtime[j])
1790                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1791                         if (bw < rs->min_bw[j])
1792                                 rs->min_bw[j] = bw;
1793                         if (bw > rs->max_bw[j])
1794                                 rs->max_bw[j] = bw;
1795
1796                         rs->iobytes[j] += ts->io_bytes[j];
1797                 }
1798         }
1799
1800         for (i = 0; i < groupid + 1; i++) {
1801                 int ddir;
1802
1803                 rs = &runstats[i];
1804
1805                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1806                         if (rs->max_run[ddir])
1807                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1808                                                 rs->max_run[ddir];
1809                 }
1810         }
1811
1812         for (i = 0; i < FIO_OUTPUT_NR; i++)
1813                 buf_output_init(&output[i]);
1814
1815         /*
1816          * don't overwrite last signal output
1817          */
1818         if (output_format & FIO_OUTPUT_NORMAL)
1819                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1820         if (output_format & FIO_OUTPUT_JSON) {
1821                 struct thread_data *global;
1822                 char time_buf[32];
1823                 struct timeval now;
1824                 unsigned long long ms_since_epoch;
1825
1826                 gettimeofday(&now, NULL);
1827                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1828                                  (unsigned long long)(now.tv_usec) / 1000;
1829
1830                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1831                                 sizeof(time_buf));
1832                 if (time_buf[strlen(time_buf) - 1] == '\n')
1833                         time_buf[strlen(time_buf) - 1] = '\0';
1834
1835                 root = json_create_object();
1836                 json_object_add_value_string(root, "fio version", fio_version_string);
1837                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1838                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1839                 json_object_add_value_string(root, "time", time_buf);
1840                 global = get_global_options();
1841                 json_add_job_opts(root, "global options", &global->opt_list, false);
1842                 array = json_create_array();
1843                 json_object_add_value_array(root, "jobs", array);
1844         }
1845
1846         if (is_backend)
1847                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1848
1849         for (i = 0; i < nr_ts; i++) {
1850                 ts = &threadstats[i];
1851                 rs = &runstats[ts->groupid];
1852
1853                 if (is_backend) {
1854                         fio_server_send_job_options(opt_lists[i], i);
1855                         fio_server_send_ts(ts, rs);
1856                 } else {
1857                         if (output_format & FIO_OUTPUT_TERSE)
1858                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1859                         if (output_format & FIO_OUTPUT_JSON) {
1860                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1861                                 json_array_add_value_object(array, tmp);
1862                         }
1863                         if (output_format & FIO_OUTPUT_NORMAL)
1864                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1865                 }
1866         }
1867         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1868                 /* disk util stats, if any */
1869                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1870
1871                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1872
1873                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1874                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1875                 json_free_object(root);
1876         }
1877
1878         for (i = 0; i < groupid + 1; i++) {
1879                 rs = &runstats[i];
1880
1881                 rs->groupid = i;
1882                 if (is_backend)
1883                         fio_server_send_gs(rs);
1884                 else if (output_format & FIO_OUTPUT_NORMAL)
1885                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1886         }
1887
1888         if (is_backend)
1889                 fio_server_send_du();
1890         else if (output_format & FIO_OUTPUT_NORMAL) {
1891                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1892                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1893         }
1894
1895         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1896                 struct buf_output *out = &output[i];
1897
1898                 log_info_buf(out->buf, out->buflen);
1899                 buf_output_free(out);
1900         }
1901
1902         log_info_flush();
1903         free(runstats);
1904         free(threadstats);
1905         free(opt_lists);
1906 }
1907
1908 void show_run_stats(void)
1909 {
1910         fio_mutex_down(stat_mutex);
1911         __show_run_stats();
1912         fio_mutex_up(stat_mutex);
1913 }
1914
1915 void __show_running_run_stats(void)
1916 {
1917         struct thread_data *td;
1918         unsigned long long *rt;
1919         struct timespec ts;
1920         int i;
1921
1922         fio_mutex_down(stat_mutex);
1923
1924         rt = malloc(thread_number * sizeof(unsigned long long));
1925         fio_gettime(&ts, NULL);
1926
1927         for_each_td(td, i) {
1928                 td->update_rusage = 1;
1929                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1930                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1931                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1932                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1933
1934                 rt[i] = mtime_since(&td->start, &ts);
1935                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1936                         td->ts.runtime[DDIR_READ] += rt[i];
1937                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1938                         td->ts.runtime[DDIR_WRITE] += rt[i];
1939                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1940                         td->ts.runtime[DDIR_TRIM] += rt[i];
1941         }
1942
1943         for_each_td(td, i) {
1944                 if (td->runstate >= TD_EXITED)
1945                         continue;
1946                 if (td->rusage_sem) {
1947                         td->update_rusage = 1;
1948                         fio_mutex_down(td->rusage_sem);
1949                 }
1950                 td->update_rusage = 0;
1951         }
1952
1953         __show_run_stats();
1954
1955         for_each_td(td, i) {
1956                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1957                         td->ts.runtime[DDIR_READ] -= rt[i];
1958                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1959                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1960                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1961                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1962         }
1963
1964         free(rt);
1965         fio_mutex_up(stat_mutex);
1966 }
1967
1968 static int status_interval_init;
1969 static struct timespec status_time;
1970 static int status_file_disabled;
1971
1972 #define FIO_STATUS_FILE         "fio-dump-status"
1973
1974 static int check_status_file(void)
1975 {
1976         struct stat sb;
1977         const char *temp_dir;
1978         char fio_status_file_path[PATH_MAX];
1979
1980         if (status_file_disabled)
1981                 return 0;
1982
1983         temp_dir = getenv("TMPDIR");
1984         if (temp_dir == NULL) {
1985                 temp_dir = getenv("TEMP");
1986                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1987                         temp_dir = NULL;
1988         }
1989         if (temp_dir == NULL)
1990                 temp_dir = "/tmp";
1991
1992         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1993
1994         if (stat(fio_status_file_path, &sb))
1995                 return 0;
1996
1997         if (unlink(fio_status_file_path) < 0) {
1998                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1999                                                         strerror(errno));
2000                 log_err("fio: disabling status file updates\n");
2001                 status_file_disabled = 1;
2002         }
2003
2004         return 1;
2005 }
2006
2007 void check_for_running_stats(void)
2008 {
2009         if (status_interval) {
2010                 if (!status_interval_init) {
2011                         fio_gettime(&status_time, NULL);
2012                         status_interval_init = 1;
2013                 } else if (mtime_since_now(&status_time) >= status_interval) {
2014                         show_running_run_stats();
2015                         fio_gettime(&status_time, NULL);
2016                         return;
2017                 }
2018         }
2019         if (check_status_file()) {
2020                 show_running_run_stats();
2021                 return;
2022         }
2023 }
2024
2025 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2026 {
2027         double val = data;
2028         double delta;
2029
2030         if (data > is->max_val)
2031                 is->max_val = data;
2032         if (data < is->min_val)
2033                 is->min_val = data;
2034
2035         delta = val - is->mean.u.f;
2036         if (delta) {
2037                 is->mean.u.f += delta / (is->samples + 1.0);
2038                 is->S.u.f += delta * (val - is->mean.u.f);
2039         }
2040
2041         is->samples++;
2042 }
2043
2044 /*
2045  * Return a struct io_logs, which is added to the tail of the log
2046  * list for 'iolog'.
2047  */
2048 static struct io_logs *get_new_log(struct io_log *iolog)
2049 {
2050         size_t new_size, new_samples;
2051         struct io_logs *cur_log;
2052
2053         /*
2054          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2055          * forever
2056          */
2057         if (!iolog->cur_log_max)
2058                 new_samples = DEF_LOG_ENTRIES;
2059         else {
2060                 new_samples = iolog->cur_log_max * 2;
2061                 if (new_samples > MAX_LOG_ENTRIES)
2062                         new_samples = MAX_LOG_ENTRIES;
2063         }
2064
2065         new_size = new_samples * log_entry_sz(iolog);
2066
2067         cur_log = smalloc(sizeof(*cur_log));
2068         if (cur_log) {
2069                 INIT_FLIST_HEAD(&cur_log->list);
2070                 cur_log->log = malloc(new_size);
2071                 if (cur_log->log) {
2072                         cur_log->nr_samples = 0;
2073                         cur_log->max_samples = new_samples;
2074                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2075                         iolog->cur_log_max = new_samples;
2076                         return cur_log;
2077                 }
2078                 sfree(cur_log);
2079         }
2080
2081         return NULL;
2082 }
2083
2084 /*
2085  * Add and return a new log chunk, or return current log if big enough
2086  */
2087 static struct io_logs *regrow_log(struct io_log *iolog)
2088 {
2089         struct io_logs *cur_log;
2090         int i;
2091
2092         if (!iolog || iolog->disabled)
2093                 goto disable;
2094
2095         cur_log = iolog_cur_log(iolog);
2096         if (!cur_log) {
2097                 cur_log = get_new_log(iolog);
2098                 if (!cur_log)
2099                         return NULL;
2100         }
2101
2102         if (cur_log->nr_samples < cur_log->max_samples)
2103                 return cur_log;
2104
2105         /*
2106          * No room for a new sample. If we're compressing on the fly, flush
2107          * out the current chunk
2108          */
2109         if (iolog->log_gz) {
2110                 if (iolog_cur_flush(iolog, cur_log)) {
2111                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2112                         return NULL;
2113                 }
2114         }
2115
2116         /*
2117          * Get a new log array, and add to our list
2118          */
2119         cur_log = get_new_log(iolog);
2120         if (!cur_log) {
2121                 log_err("fio: failed extending iolog! Will stop logging.\n");
2122                 return NULL;
2123         }
2124
2125         if (!iolog->pending || !iolog->pending->nr_samples)
2126                 return cur_log;
2127
2128         /*
2129          * Flush pending items to new log
2130          */
2131         for (i = 0; i < iolog->pending->nr_samples; i++) {
2132                 struct io_sample *src, *dst;
2133
2134                 src = get_sample(iolog, iolog->pending, i);
2135                 dst = get_sample(iolog, cur_log, i);
2136                 memcpy(dst, src, log_entry_sz(iolog));
2137         }
2138         cur_log->nr_samples = iolog->pending->nr_samples;
2139
2140         iolog->pending->nr_samples = 0;
2141         return cur_log;
2142 disable:
2143         if (iolog)
2144                 iolog->disabled = true;
2145         return NULL;
2146 }
2147
2148 void regrow_logs(struct thread_data *td)
2149 {
2150         regrow_log(td->slat_log);
2151         regrow_log(td->clat_log);
2152         regrow_log(td->clat_hist_log);
2153         regrow_log(td->lat_log);
2154         regrow_log(td->bw_log);
2155         regrow_log(td->iops_log);
2156         td->flags &= ~TD_F_REGROW_LOGS;
2157 }
2158
2159 static struct io_logs *get_cur_log(struct io_log *iolog)
2160 {
2161         struct io_logs *cur_log;
2162
2163         cur_log = iolog_cur_log(iolog);
2164         if (!cur_log) {
2165                 cur_log = get_new_log(iolog);
2166                 if (!cur_log)
2167                         return NULL;
2168         }
2169
2170         if (cur_log->nr_samples < cur_log->max_samples)
2171                 return cur_log;
2172
2173         /*
2174          * Out of space. If we're in IO offload mode, or we're not doing
2175          * per unit logging (hence logging happens outside of the IO thread
2176          * as well), add a new log chunk inline. If we're doing inline
2177          * submissions, flag 'td' as needing a log regrow and we'll take
2178          * care of it on the submission side.
2179          */
2180         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2181             !per_unit_log(iolog))
2182                 return regrow_log(iolog);
2183
2184         iolog->td->flags |= TD_F_REGROW_LOGS;
2185         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2186         return iolog->pending;
2187 }
2188
2189 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2190                              enum fio_ddir ddir, unsigned int bs,
2191                              unsigned long t, uint64_t offset)
2192 {
2193         struct io_logs *cur_log;
2194
2195         if (iolog->disabled)
2196                 return;
2197         if (flist_empty(&iolog->io_logs))
2198                 iolog->avg_last = t;
2199
2200         cur_log = get_cur_log(iolog);
2201         if (cur_log) {
2202                 struct io_sample *s;
2203
2204                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2205
2206                 s->data = data;
2207                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2208                 io_sample_set_ddir(iolog, s, ddir);
2209                 s->bs = bs;
2210
2211                 if (iolog->log_offset) {
2212                         struct io_sample_offset *so = (void *) s;
2213
2214                         so->offset = offset;
2215                 }
2216
2217                 cur_log->nr_samples++;
2218                 return;
2219         }
2220
2221         iolog->disabled = true;
2222 }
2223
2224 static inline void reset_io_stat(struct io_stat *ios)
2225 {
2226         ios->max_val = ios->min_val = ios->samples = 0;
2227         ios->mean.u.f = ios->S.u.f = 0;
2228 }
2229
2230 void reset_io_stats(struct thread_data *td)
2231 {
2232         struct thread_stat *ts = &td->ts;
2233         int i, j;
2234
2235         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2236                 reset_io_stat(&ts->clat_stat[i]);
2237                 reset_io_stat(&ts->slat_stat[i]);
2238                 reset_io_stat(&ts->lat_stat[i]);
2239                 reset_io_stat(&ts->bw_stat[i]);
2240                 reset_io_stat(&ts->iops_stat[i]);
2241
2242                 ts->io_bytes[i] = 0;
2243                 ts->runtime[i] = 0;
2244                 ts->total_io_u[i] = 0;
2245                 ts->short_io_u[i] = 0;
2246                 ts->drop_io_u[i] = 0;
2247
2248                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2249                         ts->io_u_plat[i][j] = 0;
2250         }
2251
2252         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2253                 ts->io_u_map[i] = 0;
2254                 ts->io_u_submit[i] = 0;
2255                 ts->io_u_complete[i] = 0;
2256         }
2257
2258         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2259                 ts->io_u_lat_n[i] = 0;
2260         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2261                 ts->io_u_lat_u[i] = 0;
2262         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2263                 ts->io_u_lat_m[i] = 0;
2264
2265         ts->total_submit = 0;
2266         ts->total_complete = 0;
2267 }
2268
2269 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2270                               unsigned long elapsed, bool log_max)
2271 {
2272         /*
2273          * Note an entry in the log. Use the mean from the logged samples,
2274          * making sure to properly round up. Only write a log entry if we
2275          * had actual samples done.
2276          */
2277         if (iolog->avg_window[ddir].samples) {
2278                 union io_sample_data data;
2279
2280                 if (log_max)
2281                         data.val = iolog->avg_window[ddir].max_val;
2282                 else
2283                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2284
2285                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2286         }
2287
2288         reset_io_stat(&iolog->avg_window[ddir]);
2289 }
2290
2291 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2292                              bool log_max)
2293 {
2294         int ddir;
2295
2296         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2297                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2298 }
2299
2300 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2301                            union io_sample_data data, enum fio_ddir ddir,
2302                            unsigned int bs, uint64_t offset)
2303 {
2304         unsigned long elapsed, this_window;
2305
2306         if (!ddir_rw(ddir))
2307                 return 0;
2308
2309         elapsed = mtime_since_now(&td->epoch);
2310
2311         /*
2312          * If no time averaging, just add the log sample.
2313          */
2314         if (!iolog->avg_msec) {
2315                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2316                 return 0;
2317         }
2318
2319         /*
2320          * Add the sample. If the time period has passed, then
2321          * add that entry to the log and clear.
2322          */
2323         add_stat_sample(&iolog->avg_window[ddir], data.val);
2324
2325         /*
2326          * If period hasn't passed, adding the above sample is all we
2327          * need to do.
2328          */
2329         this_window = elapsed - iolog->avg_last;
2330         if (elapsed < iolog->avg_last)
2331                 return iolog->avg_last - elapsed;
2332         else if (this_window < iolog->avg_msec) {
2333                 int diff = iolog->avg_msec - this_window;
2334
2335                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2336                         return diff;
2337         }
2338
2339         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2340
2341         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2342         return iolog->avg_msec;
2343 }
2344
2345 void finalize_logs(struct thread_data *td, bool unit_logs)
2346 {
2347         unsigned long elapsed;
2348
2349         elapsed = mtime_since_now(&td->epoch);
2350
2351         if (td->clat_log && unit_logs)
2352                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2353         if (td->slat_log && unit_logs)
2354                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2355         if (td->lat_log && unit_logs)
2356                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2357         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2358                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2359         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2360                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2361 }
2362
2363 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2364 {
2365         struct io_log *iolog;
2366
2367         if (!ddir_rw(ddir))
2368                 return;
2369
2370         iolog = agg_io_log[ddir];
2371         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2372 }
2373
2374 static void add_clat_percentile_sample(struct thread_stat *ts,
2375                                 unsigned long long nsec, enum fio_ddir ddir)
2376 {
2377         unsigned int idx = plat_val_to_idx(nsec);
2378         assert(idx < FIO_IO_U_PLAT_NR);
2379
2380         ts->io_u_plat[ddir][idx]++;
2381 }
2382
2383 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2384                      unsigned long long nsec, unsigned int bs, uint64_t offset)
2385 {
2386         unsigned long elapsed, this_window;
2387         struct thread_stat *ts = &td->ts;
2388         struct io_log *iolog = td->clat_hist_log;
2389
2390         td_io_u_lock(td);
2391
2392         add_stat_sample(&ts->clat_stat[ddir], nsec);
2393
2394         if (td->clat_log)
2395                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2396                                offset);
2397
2398         if (ts->clat_percentiles)
2399                 add_clat_percentile_sample(ts, nsec, ddir);
2400
2401         if (iolog && iolog->hist_msec) {
2402                 struct io_hist *hw = &iolog->hist_window[ddir];
2403
2404                 hw->samples++;
2405                 elapsed = mtime_since_now(&td->epoch);
2406                 if (!hw->hist_last)
2407                         hw->hist_last = elapsed;
2408                 this_window = elapsed - hw->hist_last;
2409
2410                 if (this_window >= iolog->hist_msec) {
2411                         unsigned int *io_u_plat;
2412                         struct io_u_plat_entry *dst;
2413
2414                         /*
2415                          * Make a byte-for-byte copy of the latency histogram
2416                          * stored in td->ts.io_u_plat[ddir], recording it in a
2417                          * log sample. Note that the matching call to free() is
2418                          * located in iolog.c after printing this sample to the
2419                          * log file.
2420                          */
2421                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2422                         dst = malloc(sizeof(struct io_u_plat_entry));
2423                         memcpy(&(dst->io_u_plat), io_u_plat,
2424                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2425                         flist_add(&dst->list, &hw->list);
2426                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2427                                                 elapsed, offset);
2428
2429                         /*
2430                          * Update the last time we recorded as being now, minus
2431                          * any drift in time we encountered before actually
2432                          * making the record.
2433                          */
2434                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2435                         hw->samples = 0;
2436                 }
2437         }
2438
2439         td_io_u_unlock(td);
2440 }
2441
2442 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2443                      unsigned long usec, unsigned int bs, uint64_t offset)
2444 {
2445         struct thread_stat *ts = &td->ts;
2446
2447         if (!ddir_rw(ddir))
2448                 return;
2449
2450         td_io_u_lock(td);
2451
2452         add_stat_sample(&ts->slat_stat[ddir], usec);
2453
2454         if (td->slat_log)
2455                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2456
2457         td_io_u_unlock(td);
2458 }
2459
2460 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2461                     unsigned long long nsec, unsigned int bs, uint64_t offset)
2462 {
2463         struct thread_stat *ts = &td->ts;
2464
2465         if (!ddir_rw(ddir))
2466                 return;
2467
2468         td_io_u_lock(td);
2469
2470         add_stat_sample(&ts->lat_stat[ddir], nsec);
2471
2472         if (td->lat_log)
2473                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2474                                offset);
2475
2476         td_io_u_unlock(td);
2477 }
2478
2479 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2480                    unsigned int bytes, unsigned long long spent)
2481 {
2482         struct thread_stat *ts = &td->ts;
2483         unsigned long rate;
2484
2485         if (spent)
2486                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2487         else
2488                 rate = 0;
2489
2490         td_io_u_lock(td);
2491
2492         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2493
2494         if (td->bw_log)
2495                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2496                                bytes, io_u->offset);
2497
2498         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2499         td_io_u_unlock(td);
2500 }
2501
2502 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2503                          struct timespec *t, unsigned int avg_time,
2504                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2505                          struct io_stat *stat, struct io_log *log,
2506                          bool is_kb)
2507 {
2508         unsigned long spent, rate;
2509         enum fio_ddir ddir;
2510         unsigned int next, next_log;
2511
2512         next_log = avg_time;
2513
2514         spent = mtime_since(parent_tv, t);
2515         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2516                 return avg_time - spent;
2517
2518         td_io_u_lock(td);
2519
2520         /*
2521          * Compute both read and write rates for the interval.
2522          */
2523         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2524                 uint64_t delta;
2525
2526                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2527                 if (!delta)
2528                         continue; /* No entries for interval */
2529
2530                 if (spent) {
2531                         if (is_kb)
2532                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2533                         else
2534                                 rate = (delta * 1000) / spent;
2535                 } else
2536                         rate = 0;
2537
2538                 add_stat_sample(&stat[ddir], rate);
2539
2540                 if (log) {
2541                         unsigned int bs = 0;
2542
2543                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2544                                 bs = td->o.min_bs[ddir];
2545
2546                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2547                         next_log = min(next_log, next);
2548                 }
2549
2550                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2551         }
2552
2553         timespec_add_msec(parent_tv, avg_time);
2554
2555         td_io_u_unlock(td);
2556
2557         if (spent <= avg_time)
2558                 next = avg_time;
2559         else
2560                 next = avg_time - (1 + spent - avg_time);
2561
2562         return min(next, next_log);
2563 }
2564
2565 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2566 {
2567         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2568                                 td->this_io_bytes, td->stat_io_bytes,
2569                                 td->ts.bw_stat, td->bw_log, true);
2570 }
2571
2572 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2573                      unsigned int bytes)
2574 {
2575         struct thread_stat *ts = &td->ts;
2576
2577         td_io_u_lock(td);
2578
2579         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2580
2581         if (td->iops_log)
2582                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2583                                bytes, io_u->offset);
2584
2585         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2586         td_io_u_unlock(td);
2587 }
2588
2589 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2590 {
2591         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2592                                 td->this_io_blocks, td->stat_io_blocks,
2593                                 td->ts.iops_stat, td->iops_log, false);
2594 }
2595
2596 /*
2597  * Returns msecs to next event
2598  */
2599 int calc_log_samples(void)
2600 {
2601         struct thread_data *td;
2602         unsigned int next = ~0U, tmp;
2603         struct timespec now;
2604         int i;
2605
2606         fio_gettime(&now, NULL);
2607
2608         for_each_td(td, i) {
2609                 if (!td->o.stats)
2610                         continue;
2611                 if (in_ramp_time(td) ||
2612                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2613                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2614                         continue;
2615                 }
2616                 if (!td->bw_log ||
2617                         (td->bw_log && !per_unit_log(td->bw_log))) {
2618                         tmp = add_bw_samples(td, &now);
2619                         if (tmp < next)
2620                                 next = tmp;
2621                 }
2622                 if (!td->iops_log ||
2623                         (td->iops_log && !per_unit_log(td->iops_log))) {
2624                         tmp = add_iops_samples(td, &now);
2625                         if (tmp < next)
2626                                 next = tmp;
2627                 }
2628         }
2629
2630         return next == ~0U ? 0 : next;
2631 }
2632
2633 void stat_init(void)
2634 {
2635         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2636 }
2637
2638 void stat_exit(void)
2639 {
2640         /*
2641          * When we have the mutex, we know out-of-band access to it
2642          * have ended.
2643          */
2644         fio_mutex_down(stat_mutex);
2645         fio_mutex_remove(stat_mutex);
2646 }
2647
2648 /*
2649  * Called from signal handler. Wake up status thread.
2650  */
2651 void show_running_run_stats(void)
2652 {
2653         helper_do_stat();
2654 }
2655
2656 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2657 {
2658         /* Ignore io_u's which span multiple blocks--they will just get
2659          * inaccurate counts. */
2660         int idx = (io_u->offset - io_u->file->file_offset)
2661                         / td->o.bs[DDIR_TRIM];
2662         uint32_t *info = &td->ts.block_infos[idx];
2663         assert(idx < td->ts.nr_block_infos);
2664         return info;
2665 }