stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/types.h>
   5 #include <sys/stat.h>
   6 #include <dirent.h>
   7 #include <libgen.h>
   8 #include <math.h>
   9
  10 #include "fio.h"
  11 #include "diskutil.h"
  12 #include "lib/ieee754.h"
  13 #include "json.h"
  14 #include "lib/getrusage.h"
  15 #include "idletime.h"
  16 #include "lib/pow2.h"
  17 #include "lib/output_buffer.h"
  18 #include "helper_thread.h"
  19 #include "smalloc.h"
  20
  21 #define LOG_MSEC_SLACK  10
  22
  23 struct fio_mutex *stat_mutex;
  24
  25 void clear_rusage_stat(struct thread_data *td)
  26 {
  27         struct thread_stat *ts = &td->ts;
  28
  29         fio_getrusage(&td->ru_start);
  30         ts->usr_time = ts->sys_time = 0;
  31         ts->ctx = 0;
  32         ts->minf = ts->majf = 0;
  33 }
  34
  35 void update_rusage_stat(struct thread_data *td)
  36 {
  37         struct thread_stat *ts = &td->ts;
  38
  39         fio_getrusage(&td->ru_end);
  40         ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
  41                                         &td->ru_end.ru_utime);
  42         ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
  43                                         &td->ru_end.ru_stime);
  44         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  45                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  46         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  47         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  48
  49         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  50 }
  51
  52 /*
  53  * Given a latency, return the index of the corresponding bucket in
  54  * the structure tracking percentiles.
  55  *
  56  * (1) find the group (and error bits) that the value (latency)
  57  * belongs to by looking at its MSB. (2) find the bucket number in the
  58  * group by looking at the index bits.
  59  *
  60  */
  61 static unsigned int plat_val_to_idx(unsigned long long val)
  62 {
  63         unsigned int msb, error_bits, base, offset, idx;
  64
  65         /* Find MSB starting from bit 0 */
  66         if (val == 0)
  67                 msb = 0;
  68         else
  69                 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
  70
  71         /*
  72          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  73          * all bits of the sample as index
  74          */
  75         if (msb <= FIO_IO_U_PLAT_BITS)
  76                 return val;
  77
  78         /* Compute the number of error bits to discard*/
  79         error_bits = msb - FIO_IO_U_PLAT_BITS;
  80
  81         /* Compute the number of buckets before the group */
  82         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  83
  84         /*
  85          * Discard the error bits and apply the mask to find the
  86          * index for the buckets in the group
  87          */
  88         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  89
  90         /* Make sure the index does not exceed (array size - 1) */
  91         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  92                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  93
  94         return idx;
  95 }
  96
  97 /*
  98  * Convert the given index of the bucket array to the value
  99  * represented by the bucket
 100  */
 101 static unsigned long long plat_idx_to_val(unsigned int idx)
 102 {
 103         unsigned int error_bits;
 104         unsigned long long k, base;
 105
 106         assert(idx < FIO_IO_U_PLAT_NR);
 107
 108         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 109          * all bits of the sample as index */
 110         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 111                 return idx;
 112
 113         /* Find the group and compute the minimum value of that group */
 114         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 115         base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
 116
 117         /* Find its bucket number of the group */
 118         k = idx % FIO_IO_U_PLAT_VAL;
 119
 120         /* Return the mean of the range of the bucket */
 121         return base + ((k + 0.5) * (1 << error_bits));
 122 }
 123
 124 static int double_cmp(const void *a, const void *b)
 125 {
 126         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 127         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 128         int cmp = 0;
 129
 130         if (fa.u.f > fb.u.f)
 131                 cmp = 1;
 132         else if (fa.u.f < fb.u.f)
 133                 cmp = -1;
 134
 135         return cmp;
 136 }
 137
 138 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 139                                    fio_fp64_t *plist, unsigned long long **output,
 140                                    unsigned long long *maxv, unsigned long long *minv)
 141 {
 142         unsigned long sum = 0;
 143         unsigned int len, i, j = 0;
 144         unsigned int oval_len = 0;
 145         unsigned long long *ovals = NULL;
 146         bool is_last;
 147
 148         *minv = -1ULL;
 149         *maxv = 0;
 150
 151         len = 0;
 152         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 153                 len++;
 154
 155         if (!len)
 156                 return 0;
 157
 158         /*
 159          * Sort the percentile list. Note that it may already be sorted if
 160          * we are using the default values, but since it's a short list this
 161          * isn't a worry. Also note that this does not work for NaN values.
 162          */
 163         if (len > 1)
 164                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 165
 166         /*
 167          * Calculate bucket values, note down max and min values
 168          */
 169         is_last = false;
 170         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 171                 sum += io_u_plat[i];
 172                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 173                         assert(plist[j].u.f <= 100.0);
 174
 175                         if (j == oval_len) {
 176                                 oval_len += 100;
 177                                 ovals = realloc(ovals, oval_len * sizeof(*ovals));
 178                         }
 179
 180                         ovals[j] = plat_idx_to_val(i);
 181                         if (ovals[j] < *minv)
 182                                 *minv = ovals[j];
 183                         if (ovals[j] > *maxv)
 184                                 *maxv = ovals[j];
 185
 186                         is_last = (j == len - 1) != 0;
 187                         if (is_last)
 188                                 break;
 189
 190                         j++;
 191                 }
 192         }
 193
 194         *output = ovals;
 195         return len;
 196 }
 197
 198 /*
 199  * Find and display the p-th percentile of clat
 200  */
 201 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 202                                   fio_fp64_t *plist, unsigned int precision,
 203                                   bool is_clat, struct buf_output *out)
 204 {
 205         unsigned int divisor, len, i, j = 0;
 206         unsigned long long minv, maxv;
 207         unsigned long long *ovals;
 208         int per_line, scale_down, time_width;
 209         const char *pre = is_clat ? "clat" : " lat";
 210         bool is_last;
 211         char fmt[32];
 212
 213         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 214         if (!len)
 215                 goto out;
 216
 217         /*
 218          * We default to nsecs, but if the value range is such that we
 219          * should scale down to usecs or msecs, do that.
 220          */
 221         if (minv > 2000000 && maxv > 99999999ULL) {
 222                 scale_down = 2;
 223                 divisor = 1000000;
 224                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 225         } else if (minv > 2000 && maxv > 99999) {
 226                 scale_down = 1;
 227                 divisor = 1000;
 228                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 229         } else {
 230                 scale_down = 0;
 231                 divisor = 1;
 232                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 233         }
 234
 235
 236         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 237         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 238                         precision, time_width);
 239         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 240         per_line = (80 - 7) / (precision + 10 + time_width);
 241
 242         for (j = 0; j < len; j++) {
 243                 /* for formatting */
 244                 if (j != 0 && (j % per_line) == 0)
 245                         log_buf(out, "     |");
 246
 247                 /* end of the list */
 248                 is_last = (j == len - 1) != 0;
 249
 250                 for (i = 0; i < scale_down; i++)
 251                         ovals[j] = (ovals[j] + 999) / 1000;
 252
 253                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 254
 255                 if (is_last)
 256                         break;
 257
 258                 if ((j % per_line) == per_line - 1)     /* for formatting */
 259                         log_buf(out, "\n");
 260         }
 261
 262 out:
 263         if (ovals)
 264                 free(ovals);
 265 }
 266
 267 bool calc_lat(struct io_stat *is, unsigned long long *min,
 268               unsigned long long *max, double *mean, double *dev)
 269 {
 270         double n = (double) is->samples;
 271
 272         if (n == 0)
 273                 return false;
 274
 275         *min = is->min_val;
 276         *max = is->max_val;
 277         *mean = is->mean.u.f;
 278
 279         if (n > 1.0)
 280                 *dev = sqrt(is->S.u.f / (n - 1.0));
 281         else
 282                 *dev = 0;
 283
 284         return true;
 285 }
 286
 287 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 288 {
 289         char *io, *agg, *min, *max;
 290         char *ioalt, *aggalt, *minalt, *maxalt;
 291         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 292         int i;
 293
 294         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 295
 296         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 297                 const int i2p = is_power_of_2(rs->kb_base);
 298
 299                 if (!rs->max_run[i])
 300                         continue;
 301
 302                 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
 303                 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
 304                 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
 305                 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
 306                 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
 307                 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
 308                 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
 309                 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
 310                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 311                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 312                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 313                                 (unsigned long long) rs->min_run[i],
 314                                 (unsigned long long) rs->max_run[i]);
 315
 316                 free(io);
 317                 free(agg);
 318                 free(min);
 319                 free(max);
 320                 free(ioalt);
 321                 free(aggalt);
 322                 free(minalt);
 323                 free(maxalt);
 324         }
 325 }
 326
 327 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 328 {
 329         int i;
 330
 331         /*
 332          * Do depth distribution calculations
 333          */
 334         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 335                 if (total) {
 336                         io_u_dist[i] = (double) map[i] / (double) total;
 337                         io_u_dist[i] *= 100.0;
 338                         if (io_u_dist[i] < 0.1 && map[i])
 339                                 io_u_dist[i] = 0.1;
 340                 } else
 341                         io_u_dist[i] = 0.0;
 342         }
 343 }
 344
 345 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 346                           unsigned int *src, int nr)
 347 {
 348         unsigned long total = ddir_rw_sum(ts->total_io_u);
 349         int i;
 350
 351         /*
 352          * Do latency distribution calculations
 353          */
 354         for (i = 0; i < nr; i++) {
 355                 if (total) {
 356                         dst[i] = (double) src[i] / (double) total;
 357                         dst[i] *= 100.0;
 358                         if (dst[i] < 0.01 && src[i])
 359                                 dst[i] = 0.01;
 360                 } else
 361                         dst[i] = 0.0;
 362         }
 363 }
 364
 365 /*
 366  * To keep the terse format unaltered, add all of the ns latency
 367  * buckets to the first us latency bucket
 368  */
 369 void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 370 {
 371         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 372         int i;
 373
 374         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 375
 376         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 377                 ntotal += ts->io_u_lat_n[i];
 378
 379         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 380 }
 381
 382 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 383 {
 384         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 385 }
 386
 387 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 388 {
 389         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 390 }
 391
 392 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 393 {
 394         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 395 }
 396
 397 static void display_lat(const char *name, unsigned long long min,
 398                         unsigned long long max, double mean, double dev,
 399                         struct buf_output *out)
 400 {
 401         const char *base = "(nsec)";
 402         char *minp, *maxp;
 403
 404         if (nsec_to_msec(&min, &max, &mean, &dev))
 405                 base = "(msec)";
 406         else if (nsec_to_usec(&min, &max, &mean, &dev))
 407                 base = "(usec)";
 408
 409         minp = num2str(min, 6, 1, 0, N2S_NONE);
 410         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 411
 412         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 413                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 414
 415         free(minp);
 416         free(maxp);
 417 }
 418
 419 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 420                              int ddir, struct buf_output *out)
 421 {
 422         const char *str[] = { " read", "write", " trim" };
 423         unsigned long runt;
 424         unsigned long long min, max, bw, iops;
 425         double mean, dev;
 426         char *io_p, *bw_p, *bw_p_alt, *iops_p;
 427         int i2p;
 428
 429         assert(ddir_rw(ddir));
 430
 431         if (!ts->runtime[ddir])
 432                 return;
 433
 434         i2p = is_power_of_2(rs->kb_base);
 435         runt = ts->runtime[ddir];
 436
 437         bw = (1000 * ts->io_bytes[ddir]) / runt;
 438         io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
 439         bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
 440         bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
 441
 442         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 443         iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
 444
 445         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
 446                         rs->unified_rw_rep ? "mixed" : str[ddir],
 447                         iops_p, bw_p, bw_p_alt, io_p,
 448                         (unsigned long long) ts->runtime[ddir]);
 449
 450         free(io_p);
 451         free(bw_p);
 452         free(bw_p_alt);
 453         free(iops_p);
 454
 455         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 456                 display_lat("slat", min, max, mean, dev, out);
 457         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 458                 display_lat("clat", min, max, mean, dev, out);
 459         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 460                 display_lat(" lat", min, max, mean, dev, out);
 461
 462         if (ts->clat_percentiles || ts->lat_percentiles) {
 463                 show_clat_percentiles(ts->io_u_plat[ddir],
 464                                         ts->clat_stat[ddir].samples,
 465                                         ts->percentile_list,
 466                                         ts->percentile_precision,
 467                                         ts->clat_percentiles, out);
 468         }
 469         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 470                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 471                 const char *bw_str;
 472
 473                 if ((rs->unit_base == 1) && i2p)
 474                         bw_str = "Kibit";
 475                 else if (rs->unit_base == 1)
 476                         bw_str = "kbit";
 477                 else if (i2p)
 478                         bw_str = "KiB";
 479                 else
 480                         bw_str = "kB";
 481
 482                 if (rs->agg[ddir]) {
 483                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 484                         if (p_of_agg > 100.0)
 485                                 p_of_agg = 100.0;
 486                 }
 487
 488                 if (rs->unit_base == 1) {
 489                         min *= 8.0;
 490                         max *= 8.0;
 491                         mean *= 8.0;
 492                         dev *= 8.0;
 493                 }
 494
 495                 if (mean > fkb_base * fkb_base) {
 496                         min /= fkb_base;
 497                         max /= fkb_base;
 498                         mean /= fkb_base;
 499                         dev /= fkb_base;
 500                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 501                 }
 502
 503                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 504                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 505                         bw_str, min, max, p_of_agg, mean, dev,
 506                         (&ts->bw_stat[ddir])->samples);
 507         }
 508         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 509                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 510                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 511                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 512         }
 513 }
 514
 515 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
 516                      const char *msg, struct buf_output *out)
 517 {
 518         bool new_line = true, shown = false;
 519         int i, line = 0;
 520
 521         for (i = 0; i < nr; i++) {
 522                 if (io_u_lat[i] <= 0.0)
 523                         continue;
 524                 shown = true;
 525                 if (new_line) {
 526                         if (line)
 527                                 log_buf(out, "\n");
 528                         log_buf(out, "  lat (%s)   : ", msg);
 529                         new_line = false;
 530                         line = 0;
 531                 }
 532                 if (line)
 533                         log_buf(out, ", ");
 534                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 535                 line++;
 536                 if (line == 5)
 537                         new_line = true;
 538         }
 539
 540         if (shown)
 541                 log_buf(out, "\n");
 542
 543         return true;
 544 }
 545
 546 static void show_lat_n(double *io_u_lat_n, 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_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 552 }
 553
 554 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 555 {
 556         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 557                                  "250=", "500=", "750=", "1000=", };
 558
 559         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 560 }
 561
 562 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 563 {
 564         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 565                                  "250=", "500=", "750=", "1000=", "2000=",
 566                                  ">=2000=", };
 567
 568         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 569 }
 570
 571 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 572 {
 573         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 574         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 575         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 576
 577         stat_calc_lat_n(ts, io_u_lat_n);
 578         stat_calc_lat_u(ts, io_u_lat_u);
 579         stat_calc_lat_m(ts, io_u_lat_m);
 580
 581         show_lat_n(io_u_lat_n, out);
 582         show_lat_u(io_u_lat_u, out);
 583         show_lat_m(io_u_lat_m, out);
 584 }
 585
 586 static int block_state_category(int block_state)
 587 {
 588         switch (block_state) {
 589         case BLOCK_STATE_UNINIT:
 590                 return 0;
 591         case BLOCK_STATE_TRIMMED:
 592         case BLOCK_STATE_WRITTEN:
 593                 return 1;
 594         case BLOCK_STATE_WRITE_FAILURE:
 595         case BLOCK_STATE_TRIM_FAILURE:
 596                 return 2;
 597         default:
 598                 /* Silence compile warning on some BSDs and have a return */
 599                 assert(0);
 600                 return -1;
 601         }
 602 }
 603
 604 static int compare_block_infos(const void *bs1, const void *bs2)
 605 {
 606         uint32_t block1 = *(uint32_t *)bs1;
 607         uint32_t block2 = *(uint32_t *)bs2;
 608         int state1 = BLOCK_INFO_STATE(block1);
 609         int state2 = BLOCK_INFO_STATE(block2);
 610         int bscat1 = block_state_category(state1);
 611         int bscat2 = block_state_category(state2);
 612         int cycles1 = BLOCK_INFO_TRIMS(block1);
 613         int cycles2 = BLOCK_INFO_TRIMS(block2);
 614
 615         if (bscat1 < bscat2)
 616                 return -1;
 617         if (bscat1 > bscat2)
 618                 return 1;
 619
 620         if (cycles1 < cycles2)
 621                 return -1;
 622         if (cycles1 > cycles2)
 623                 return 1;
 624
 625         if (state1 < state2)
 626                 return -1;
 627         if (state1 > state2)
 628                 return 1;
 629
 630         assert(block1 == block2);
 631         return 0;
 632 }
 633
 634 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 635                                   fio_fp64_t *plist, unsigned int **percentiles,
 636                                   unsigned int *types)
 637 {
 638         int len = 0;
 639         int i, nr_uninit;
 640
 641         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 642
 643         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 644                 len++;
 645
 646         if (!len)
 647                 return 0;
 648
 649         /*
 650          * Sort the percentile list. Note that it may already be sorted if
 651          * we are using the default values, but since it's a short list this
 652          * isn't a worry. Also note that this does not work for NaN values.
 653          */
 654         if (len > 1)
 655                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 656
 657         nr_uninit = 0;
 658         /* Start only after the uninit entries end */
 659         for (nr_uninit = 0;
 660              nr_uninit < nr_block_infos
 661                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 662              nr_uninit ++)
 663                 ;
 664
 665         if (nr_uninit == nr_block_infos)
 666                 return 0;
 667
 668         *percentiles = calloc(len, sizeof(**percentiles));
 669
 670         for (i = 0; i < len; i++) {
 671                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 672                                 + nr_uninit;
 673                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 674         }
 675
 676         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 677         for (i = 0; i < nr_block_infos; i++)
 678                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 679
 680         return len;
 681 }
 682
 683 static const char *block_state_names[] = {
 684         [BLOCK_STATE_UNINIT] = "unwritten",
 685         [BLOCK_STATE_TRIMMED] = "trimmed",
 686         [BLOCK_STATE_WRITTEN] = "written",
 687         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 688         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 689 };
 690
 691 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 692                              fio_fp64_t *plist, struct buf_output *out)
 693 {
 694         int len, pos, i;
 695         unsigned int *percentiles = NULL;
 696         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 697
 698         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 699                                      &percentiles, block_state_counts);
 700
 701         log_buf(out, "  block lifetime percentiles :\n   |");
 702         pos = 0;
 703         for (i = 0; i < len; i++) {
 704                 uint32_t block_info = percentiles[i];
 705 #define LINE_LENGTH     75
 706                 char str[LINE_LENGTH];
 707                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 708                                      plist[i].u.f, block_info,
 709                                      i == len - 1 ? '\n' : ',');
 710                 assert(strln < LINE_LENGTH);
 711                 if (pos + strln > LINE_LENGTH) {
 712                         pos = 0;
 713                         log_buf(out, "\n   |");
 714                 }
 715                 log_buf(out, "%s", str);
 716                 pos += strln;
 717 #undef LINE_LENGTH
 718         }
 719         if (percentiles)
 720                 free(percentiles);
 721
 722         log_buf(out, "        states               :");
 723         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 724                 log_buf(out, " %s=%u%c",
 725                          block_state_names[i], block_state_counts[i],
 726                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 727 }
 728
 729 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 730 {
 731         char *p1, *p1alt, *p2;
 732         unsigned long long bw_mean, iops_mean;
 733         const int i2p = is_power_of_2(ts->kb_base);
 734
 735         if (!ts->ss_dur)
 736                 return;
 737
 738         bw_mean = steadystate_bw_mean(ts);
 739         iops_mean = steadystate_iops_mean(ts);
 740
 741         p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
 742         p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
 743         p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
 744
 745         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 746                 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
 747                 p1, p1alt, p2,
 748                 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
 749                 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
 750                 ts->ss_criterion.u.f,
 751                 ts->ss_state & __FIO_SS_PCT ? "%" : "");
 752
 753         free(p1);
 754         free(p1alt);
 755         free(p2);
 756 }
 757
 758 static void show_thread_status_normal(struct thread_stat *ts,
 759                                       struct group_run_stats *rs,
 760                                       struct buf_output *out)
 761 {
 762         double usr_cpu, sys_cpu;
 763         unsigned long runtime;
 764         double io_u_dist[FIO_IO_U_MAP_NR];
 765         time_t time_p;
 766         char time_buf[32];
 767
 768         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 769                 return;
 770
 771         memset(time_buf, 0, sizeof(time_buf));
 772
 773         time(&time_p);
 774         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 775
 776         if (!ts->error) {
 777                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 778                                         ts->name, ts->groupid, ts->members,
 779                                         ts->error, (int) ts->pid, time_buf);
 780         } else {
 781                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 782                                         ts->name, ts->groupid, ts->members,
 783                                         ts->error, ts->verror, (int) ts->pid,
 784                                         time_buf);
 785         }
 786
 787         if (strlen(ts->description))
 788                 log_buf(out, "  Description  : [%s]\n", ts->description);
 789
 790         if (ts->io_bytes[DDIR_READ])
 791                 show_ddir_status(rs, ts, DDIR_READ, out);
 792         if (ts->io_bytes[DDIR_WRITE])
 793                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 794         if (ts->io_bytes[DDIR_TRIM])
 795                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 796
 797         show_latencies(ts, out);
 798
 799         runtime = ts->total_run_time;
 800         if (runtime) {
 801                 double runt = (double) runtime;
 802
 803                 usr_cpu = (double) ts->usr_time * 100 / runt;
 804                 sys_cpu = (double) ts->sys_time * 100 / runt;
 805         } else {
 806                 usr_cpu = 0;
 807                 sys_cpu = 0;
 808         }
 809
 810         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 811                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 812                         (unsigned long long) ts->ctx,
 813                         (unsigned long long) ts->majf,
 814                         (unsigned long long) ts->minf);
 815
 816         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 817         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 818                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 819                                         io_u_dist[1], io_u_dist[2],
 820                                         io_u_dist[3], io_u_dist[4],
 821                                         io_u_dist[5], io_u_dist[6]);
 822
 823         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 824         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 825                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 826                                         io_u_dist[1], io_u_dist[2],
 827                                         io_u_dist[3], io_u_dist[4],
 828                                         io_u_dist[5], io_u_dist[6]);
 829         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 830         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 831                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 832                                         io_u_dist[1], io_u_dist[2],
 833                                         io_u_dist[3], io_u_dist[4],
 834                                         io_u_dist[5], io_u_dist[6]);
 835         log_buf(out, "     issued rwt: total=%llu,%llu,%llu,"
 836                                  " short=%llu,%llu,%llu,"
 837                                  " dropped=%llu,%llu,%llu\n",
 838                                         (unsigned long long) ts->total_io_u[0],
 839                                         (unsigned long long) ts->total_io_u[1],
 840                                         (unsigned long long) ts->total_io_u[2],
 841                                         (unsigned long long) ts->short_io_u[0],
 842                                         (unsigned long long) ts->short_io_u[1],
 843                                         (unsigned long long) ts->short_io_u[2],
 844                                         (unsigned long long) ts->drop_io_u[0],
 845                                         (unsigned long long) ts->drop_io_u[1],
 846                                         (unsigned long long) ts->drop_io_u[2]);
 847         if (ts->continue_on_error) {
 848                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 849                                         (unsigned long long)ts->total_err_count,
 850                                         ts->first_error,
 851                                         strerror(ts->first_error));
 852         }
 853         if (ts->latency_depth) {
 854                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 855                                         (unsigned long long)ts->latency_target,
 856                                         (unsigned long long)ts->latency_window,
 857                                         ts->latency_percentile.u.f,
 858                                         ts->latency_depth);
 859         }
 860
 861         if (ts->nr_block_infos)
 862                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 863                                   ts->percentile_list, out);
 864
 865         if (ts->ss_dur)
 866                 show_ss_normal(ts, out);
 867 }
 868
 869 static void show_ddir_status_terse(struct thread_stat *ts,
 870                                    struct group_run_stats *rs, int ddir,
 871                                    int ver, struct buf_output *out)
 872 {
 873         unsigned long long min, max, minv, maxv, bw, iops;
 874         unsigned long long *ovals = NULL;
 875         double mean, dev;
 876         unsigned int len;
 877         int i, bw_stat;
 878
 879         assert(ddir_rw(ddir));
 880
 881         iops = bw = 0;
 882         if (ts->runtime[ddir]) {
 883                 uint64_t runt = ts->runtime[ddir];
 884
 885                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 886                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 887         }
 888
 889         log_buf(out, ";%llu;%llu;%llu;%llu",
 890                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 891                                         (unsigned long long) ts->runtime[ddir]);
 892
 893         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 894                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 895         else
 896                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 897
 898         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 899                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 900         else
 901                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 902
 903         if (ts->clat_percentiles || ts->lat_percentiles) {
 904                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 905                                         ts->clat_stat[ddir].samples,
 906                                         ts->percentile_list, &ovals, &maxv,
 907                                         &minv);
 908         } else
 909                 len = 0;
 910
 911         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 912                 if (i >= len) {
 913                         log_buf(out, ";0%%=0");
 914                         continue;
 915                 }
 916                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
 917         }
 918
 919         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 920                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 921         else
 922                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 923
 924         if (ovals)
 925                 free(ovals);
 926
 927         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
 928         if (bw_stat) {
 929                 double p_of_agg = 100.0;
 930
 931                 if (rs->agg[ddir]) {
 932                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 933                         if (p_of_agg > 100.0)
 934                                 p_of_agg = 100.0;
 935                 }
 936
 937                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 938         } else
 939                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
 940
 941         if (ver == 5) {
 942                 if (bw_stat)
 943                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
 944                 else
 945                         log_buf(out, ";%lu", 0UL);
 946
 947                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
 948                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
 949                                 mean, dev, (&ts->iops_stat[ddir])->samples);
 950                 else
 951                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
 952         }
 953 }
 954
 955 static void add_ddir_status_json(struct thread_stat *ts,
 956                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 957 {
 958         unsigned long long min, max, minv, maxv;
 959         unsigned long long bw;
 960         unsigned long long *ovals = NULL;
 961         double mean, dev, iops;
 962         unsigned int len;
 963         int i;
 964         const char *ddirname[] = {"read", "write", "trim"};
 965         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 966         char buf[120];
 967         double p_of_agg = 100.0;
 968
 969         assert(ddir_rw(ddir));
 970
 971         if (ts->unified_rw_rep && ddir != DDIR_READ)
 972                 return;
 973
 974         dir_object = json_create_object();
 975         json_object_add_value_object(parent,
 976                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 977
 978         bw = 0;
 979         iops = 0.0;
 980         if (ts->runtime[ddir]) {
 981                 uint64_t runt = ts->runtime[ddir];
 982
 983                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 984                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 985         }
 986
 987         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
 988         json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
 989         json_object_add_value_int(dir_object, "bw", bw);
 990         json_object_add_value_float(dir_object, "iops", iops);
 991         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 992         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 993         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 994         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 995
 996         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 997                 min = max = 0;
 998                 mean = dev = 0.0;
 999         }
1000         tmp_object = json_create_object();
1001         json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1002         json_object_add_value_int(tmp_object, "min", min);
1003         json_object_add_value_int(tmp_object, "max", max);
1004         json_object_add_value_float(tmp_object, "mean", mean);
1005         json_object_add_value_float(tmp_object, "stddev", dev);
1006
1007         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1008                 min = max = 0;
1009                 mean = dev = 0.0;
1010         }
1011         tmp_object = json_create_object();
1012         json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1013         json_object_add_value_int(tmp_object, "min", min);
1014         json_object_add_value_int(tmp_object, "max", max);
1015         json_object_add_value_float(tmp_object, "mean", mean);
1016         json_object_add_value_float(tmp_object, "stddev", dev);
1017
1018         if (ts->clat_percentiles || ts->lat_percentiles) {
1019                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1020                                         ts->clat_stat[ddir].samples,
1021                                         ts->percentile_list, &ovals, &maxv,
1022                                         &minv);
1023         } else
1024                 len = 0;
1025
1026         percentile_object = json_create_object();
1027         json_object_add_value_object(tmp_object, "percentile", percentile_object);
1028         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1029                 if (i >= len) {
1030                         json_object_add_value_int(percentile_object, "0.00", 0);
1031                         continue;
1032                 }
1033                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1034                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1035         }
1036
1037         if (output_format & FIO_OUTPUT_JSON_PLUS) {
1038                 clat_bins_object = json_create_object();
1039                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1040                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1041                         if (ts->io_u_plat[ddir][i]) {
1042                                 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1043                                 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1044                         }
1045                 }
1046         }
1047
1048         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1049                 min = max = 0;
1050                 mean = dev = 0.0;
1051         }
1052         tmp_object = json_create_object();
1053         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1054         json_object_add_value_int(tmp_object, "min", min);
1055         json_object_add_value_int(tmp_object, "max", max);
1056         json_object_add_value_float(tmp_object, "mean", mean);
1057         json_object_add_value_float(tmp_object, "stddev", dev);
1058         if (ovals)
1059                 free(ovals);
1060
1061         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1062                 if (rs->agg[ddir]) {
1063                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1064                         if (p_of_agg > 100.0)
1065                                 p_of_agg = 100.0;
1066                 }
1067         } else {
1068                 min = max = 0;
1069                 p_of_agg = mean = dev = 0.0;
1070         }
1071         json_object_add_value_int(dir_object, "bw_min", min);
1072         json_object_add_value_int(dir_object, "bw_max", max);
1073         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1074         json_object_add_value_float(dir_object, "bw_mean", mean);
1075         json_object_add_value_float(dir_object, "bw_dev", dev);
1076         json_object_add_value_int(dir_object, "bw_samples",
1077                                 (&ts->bw_stat[ddir])->samples);
1078
1079         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1080                 min = max = 0;
1081                 mean = dev = 0.0;
1082         }
1083         json_object_add_value_int(dir_object, "iops_min", min);
1084         json_object_add_value_int(dir_object, "iops_max", max);
1085         json_object_add_value_float(dir_object, "iops_mean", mean);
1086         json_object_add_value_float(dir_object, "iops_stddev", dev);
1087         json_object_add_value_int(dir_object, "iops_samples",
1088                                 (&ts->iops_stat[ddir])->samples);
1089 }
1090
1091 static void show_thread_status_terse_all(struct thread_stat *ts,
1092                                          struct group_run_stats *rs, int ver,
1093                                          struct buf_output *out)
1094 {
1095         double io_u_dist[FIO_IO_U_MAP_NR];
1096         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1097         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1098         double usr_cpu, sys_cpu;
1099         int i;
1100
1101         /* General Info */
1102         if (ver == 2)
1103                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1104         else
1105                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1106                         ts->name, ts->groupid, ts->error);
1107
1108         /* Log Read Status */
1109         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1110         /* Log Write Status */
1111         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1112         /* Log Trim Status */
1113         if (ver == 2 || ver == 4 || ver == 5)
1114                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1115
1116         /* CPU Usage */
1117         if (ts->total_run_time) {
1118                 double runt = (double) ts->total_run_time;
1119
1120                 usr_cpu = (double) ts->usr_time * 100 / runt;
1121                 sys_cpu = (double) ts->sys_time * 100 / runt;
1122         } else {
1123                 usr_cpu = 0;
1124                 sys_cpu = 0;
1125         }
1126
1127         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1128                                                 (unsigned long long) ts->ctx,
1129                                                 (unsigned long long) ts->majf,
1130                                                 (unsigned long long) ts->minf);
1131
1132         /* Calc % distribution of IO depths, usecond, msecond latency */
1133         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1134         stat_calc_lat_nu(ts, io_u_lat_u);
1135         stat_calc_lat_m(ts, io_u_lat_m);
1136
1137         /* Only show fixed 7 I/O depth levels*/
1138         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1139                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1140                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1141
1142         /* Microsecond latency */
1143         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1144                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1145         /* Millisecond latency */
1146         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1147                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1148
1149         /* disk util stats, if any */
1150         if (ver >= 3)
1151                 show_disk_util(1, NULL, out);
1152
1153         /* Additional output if continue_on_error set - default off*/
1154         if (ts->continue_on_error)
1155                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1156         if (ver == 2)
1157                 log_buf(out, "\n");
1158
1159         /* Additional output if description is set */
1160         if (strlen(ts->description))
1161                 log_buf(out, ";%s", ts->description);
1162
1163         log_buf(out, "\n");
1164 }
1165
1166 static void json_add_job_opts(struct json_object *root, const char *name,
1167                               struct flist_head *opt_list, bool num_jobs)
1168 {
1169         struct json_object *dir_object;
1170         struct flist_head *entry;
1171         struct print_option *p;
1172
1173         if (flist_empty(opt_list))
1174                 return;
1175
1176         dir_object = json_create_object();
1177         json_object_add_value_object(root, name, dir_object);
1178
1179         flist_for_each(entry, opt_list) {
1180                 const char *pos = "";
1181
1182                 p = flist_entry(entry, struct print_option, list);
1183                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1184                         continue;
1185                 if (p->value)
1186                         pos = p->value;
1187                 json_object_add_value_string(dir_object, p->name, pos);
1188         }
1189 }
1190
1191 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1192                                                    struct group_run_stats *rs,
1193                                                    struct flist_head *opt_list)
1194 {
1195         struct json_object *root, *tmp;
1196         struct jobs_eta *je;
1197         double io_u_dist[FIO_IO_U_MAP_NR];
1198         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1199         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1200         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1201         double usr_cpu, sys_cpu;
1202         int i;
1203         size_t size;
1204
1205         root = json_create_object();
1206         json_object_add_value_string(root, "jobname", ts->name);
1207         json_object_add_value_int(root, "groupid", ts->groupid);
1208         json_object_add_value_int(root, "error", ts->error);
1209
1210         /* ETA Info */
1211         je = get_jobs_eta(true, &size);
1212         if (je) {
1213                 json_object_add_value_int(root, "eta", je->eta_sec);
1214                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1215         }
1216
1217         if (opt_list)
1218                 json_add_job_opts(root, "job options", opt_list, true);
1219
1220         add_ddir_status_json(ts, rs, DDIR_READ, root);
1221         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1222         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1223
1224         /* CPU Usage */
1225         if (ts->total_run_time) {
1226                 double runt = (double) ts->total_run_time;
1227
1228                 usr_cpu = (double) ts->usr_time * 100 / runt;
1229                 sys_cpu = (double) ts->sys_time * 100 / runt;
1230         } else {
1231                 usr_cpu = 0;
1232                 sys_cpu = 0;
1233         }
1234         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1235         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1236         json_object_add_value_int(root, "ctx", ts->ctx);
1237         json_object_add_value_int(root, "majf", ts->majf);
1238         json_object_add_value_int(root, "minf", ts->minf);
1239
1240
1241         /* Calc % distribution of IO depths, usecond, msecond latency */
1242         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1243         stat_calc_lat_n(ts, io_u_lat_n);
1244         stat_calc_lat_u(ts, io_u_lat_u);
1245         stat_calc_lat_m(ts, io_u_lat_m);
1246
1247         tmp = json_create_object();
1248         json_object_add_value_object(root, "iodepth_level", tmp);
1249         /* Only show fixed 7 I/O depth levels*/
1250         for (i = 0; i < 7; i++) {
1251                 char name[20];
1252                 if (i < 6)
1253                         snprintf(name, 20, "%d", 1 << i);
1254                 else
1255                         snprintf(name, 20, ">=%d", 1 << i);
1256                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1257         }
1258
1259         /* Nanosecond latency */
1260         tmp = json_create_object();
1261         json_object_add_value_object(root, "latency_ns", tmp);
1262         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1263                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1264                                  "250", "500", "750", "1000", };
1265                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1266         }
1267         /* Microsecond latency */
1268         tmp = json_create_object();
1269         json_object_add_value_object(root, "latency_us", tmp);
1270         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1271                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1272                                  "250", "500", "750", "1000", };
1273                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1274         }
1275         /* Millisecond latency */
1276         tmp = json_create_object();
1277         json_object_add_value_object(root, "latency_ms", tmp);
1278         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1279                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1280                                  "250", "500", "750", "1000", "2000",
1281                                  ">=2000", };
1282                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1283         }
1284
1285         /* Additional output if continue_on_error set - default off*/
1286         if (ts->continue_on_error) {
1287                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1288                 json_object_add_value_int(root, "first_error", ts->first_error);
1289         }
1290
1291         if (ts->latency_depth) {
1292                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1293                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1294                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1295                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1296         }
1297
1298         /* Additional output if description is set */
1299         if (strlen(ts->description))
1300                 json_object_add_value_string(root, "desc", ts->description);
1301
1302         if (ts->nr_block_infos) {
1303                 /* Block error histogram and types */
1304                 int len;
1305                 unsigned int *percentiles = NULL;
1306                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1307
1308                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1309                                              ts->percentile_list,
1310                                              &percentiles, block_state_counts);
1311
1312                 if (len) {
1313                         struct json_object *block, *percentile_object, *states;
1314                         int state;
1315                         block = json_create_object();
1316                         json_object_add_value_object(root, "block", block);
1317
1318                         percentile_object = json_create_object();
1319                         json_object_add_value_object(block, "percentiles",
1320                                                      percentile_object);
1321                         for (i = 0; i < len; i++) {
1322                                 char buf[20];
1323                                 snprintf(buf, sizeof(buf), "%f",
1324                                          ts->percentile_list[i].u.f);
1325                                 json_object_add_value_int(percentile_object,
1326                                                           (const char *)buf,
1327                                                           percentiles[i]);
1328                         }
1329
1330                         states = json_create_object();
1331                         json_object_add_value_object(block, "states", states);
1332                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1333                                 json_object_add_value_int(states,
1334                                         block_state_names[state],
1335                                         block_state_counts[state]);
1336                         }
1337                         free(percentiles);
1338                 }
1339         }
1340
1341         if (ts->ss_dur) {
1342                 struct json_object *data;
1343                 struct json_array *iops, *bw;
1344                 int i, j, k;
1345                 char ss_buf[64];
1346
1347                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1348                         ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1349                         ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1350                         (float) ts->ss_limit.u.f,
1351                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1352
1353                 tmp = json_create_object();
1354                 json_object_add_value_object(root, "steadystate", tmp);
1355                 json_object_add_value_string(tmp, "ss", ss_buf);
1356                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1357                 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1358
1359                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1360                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1361                 json_object_add_value_string(tmp, "criterion", ss_buf);
1362                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1363                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1364
1365                 data = json_create_object();
1366                 json_object_add_value_object(tmp, "data", data);
1367                 bw = json_create_array();
1368                 iops = json_create_array();
1369
1370                 /*
1371                 ** if ss was attained or the buffer is not full,
1372                 ** ss->head points to the first element in the list.
1373                 ** otherwise it actually points to the second element
1374                 ** in the list
1375                 */
1376                 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1377                         j = ts->ss_head;
1378                 else
1379                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1380                 for (i = 0; i < ts->ss_dur; i++) {
1381                         k = (j + i) % ts->ss_dur;
1382                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1383                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1384                 }
1385                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1386                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1387                 json_object_add_value_array(data, "iops", iops);
1388                 json_object_add_value_array(data, "bw", bw);
1389         }
1390
1391         return root;
1392 }
1393
1394 static void show_thread_status_terse(struct thread_stat *ts,
1395                                      struct group_run_stats *rs,
1396                                      struct buf_output *out)
1397 {
1398         if (terse_version >= 2 && terse_version <= 5)
1399                 show_thread_status_terse_all(ts, rs, terse_version, out);
1400         else
1401                 log_err("fio: bad terse version!? %d\n", terse_version);
1402 }
1403
1404 struct json_object *show_thread_status(struct thread_stat *ts,
1405                                        struct group_run_stats *rs,
1406                                        struct flist_head *opt_list,
1407                                        struct buf_output *out)
1408 {
1409         struct json_object *ret = NULL;
1410
1411         if (output_format & FIO_OUTPUT_TERSE)
1412                 show_thread_status_terse(ts, rs,  out);
1413         if (output_format & FIO_OUTPUT_JSON)
1414                 ret = show_thread_status_json(ts, rs, opt_list);
1415         if (output_format & FIO_OUTPUT_NORMAL)
1416                 show_thread_status_normal(ts, rs,  out);
1417
1418         return ret;
1419 }
1420
1421 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1422 {
1423         double mean, S;
1424
1425         if (src->samples == 0)
1426                 return;
1427
1428         dst->min_val = min(dst->min_val, src->min_val);
1429         dst->max_val = max(dst->max_val, src->max_val);
1430
1431         /*
1432          * Compute new mean and S after the merge
1433          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1434          *  #Parallel_algorithm>
1435          */
1436         if (first) {
1437                 mean = src->mean.u.f;
1438                 S = src->S.u.f;
1439         } else {
1440                 double delta = src->mean.u.f - dst->mean.u.f;
1441
1442                 mean = ((src->mean.u.f * src->samples) +
1443                         (dst->mean.u.f * dst->samples)) /
1444                         (dst->samples + src->samples);
1445
1446                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1447                         (dst->samples * src->samples) /
1448                         (dst->samples + src->samples);
1449         }
1450
1451         dst->samples += src->samples;
1452         dst->mean.u.f = mean;
1453         dst->S.u.f = S;
1454 }
1455
1456 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1457 {
1458         int i;
1459
1460         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1461                 if (dst->max_run[i] < src->max_run[i])
1462                         dst->max_run[i] = src->max_run[i];
1463                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1464                         dst->min_run[i] = src->min_run[i];
1465                 if (dst->max_bw[i] < src->max_bw[i])
1466                         dst->max_bw[i] = src->max_bw[i];
1467                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1468                         dst->min_bw[i] = src->min_bw[i];
1469
1470                 dst->iobytes[i] += src->iobytes[i];
1471                 dst->agg[i] += src->agg[i];
1472         }
1473
1474         if (!dst->kb_base)
1475                 dst->kb_base = src->kb_base;
1476         if (!dst->unit_base)
1477                 dst->unit_base = src->unit_base;
1478 }
1479
1480 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1481                       bool first)
1482 {
1483         int l, k;
1484
1485         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1486                 if (!dst->unified_rw_rep) {
1487                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1488                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1489                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1490                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1491                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1492
1493                         dst->io_bytes[l] += src->io_bytes[l];
1494
1495                         if (dst->runtime[l] < src->runtime[l])
1496                                 dst->runtime[l] = src->runtime[l];
1497                 } else {
1498                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1499                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1500                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1501                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1502                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1503
1504                         dst->io_bytes[0] += src->io_bytes[l];
1505
1506                         if (dst->runtime[0] < src->runtime[l])
1507                                 dst->runtime[0] = src->runtime[l];
1508
1509                         /*
1510                          * We're summing to the same destination, so override
1511                          * 'first' after the first iteration of the loop
1512                          */
1513                         first = false;
1514                 }
1515         }
1516
1517         dst->usr_time += src->usr_time;
1518         dst->sys_time += src->sys_time;
1519         dst->ctx += src->ctx;
1520         dst->majf += src->majf;
1521         dst->minf += src->minf;
1522
1523         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1524                 dst->io_u_map[k] += src->io_u_map[k];
1525         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1526                 dst->io_u_submit[k] += src->io_u_submit[k];
1527         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1528                 dst->io_u_complete[k] += src->io_u_complete[k];
1529         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1530                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1531         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1532                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1533         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1534                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1535
1536         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1537                 if (!dst->unified_rw_rep) {
1538                         dst->total_io_u[k] += src->total_io_u[k];
1539                         dst->short_io_u[k] += src->short_io_u[k];
1540                         dst->drop_io_u[k] += src->drop_io_u[k];
1541                 } else {
1542                         dst->total_io_u[0] += src->total_io_u[k];
1543                         dst->short_io_u[0] += src->short_io_u[k];
1544                         dst->drop_io_u[0] += src->drop_io_u[k];
1545                 }
1546         }
1547
1548         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1549                 int m;
1550
1551                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1552                         if (!dst->unified_rw_rep)
1553                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1554                         else
1555                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1556                 }
1557         }
1558
1559         dst->total_run_time += src->total_run_time;
1560         dst->total_submit += src->total_submit;
1561         dst->total_complete += src->total_complete;
1562 }
1563
1564 void init_group_run_stat(struct group_run_stats *gs)
1565 {
1566         int i;
1567         memset(gs, 0, sizeof(*gs));
1568
1569         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1570                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1571 }
1572
1573 void init_thread_stat(struct thread_stat *ts)
1574 {
1575         int j;
1576
1577         memset(ts, 0, sizeof(*ts));
1578
1579         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1580                 ts->lat_stat[j].min_val = -1UL;
1581                 ts->clat_stat[j].min_val = -1UL;
1582                 ts->slat_stat[j].min_val = -1UL;
1583                 ts->bw_stat[j].min_val = -1UL;
1584                 ts->iops_stat[j].min_val = -1UL;
1585         }
1586         ts->groupid = -1;
1587 }
1588
1589 void __show_run_stats(void)
1590 {
1591         struct group_run_stats *runstats, *rs;
1592         struct thread_data *td;
1593         struct thread_stat *threadstats, *ts;
1594         int i, j, k, nr_ts, last_ts, idx;
1595         bool kb_base_warned = false;
1596         bool unit_base_warned = false;
1597         struct json_object *root = NULL;
1598         struct json_array *array = NULL;
1599         struct buf_output output[FIO_OUTPUT_NR];
1600         struct flist_head **opt_lists;
1601
1602         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1603
1604         for (i = 0; i < groupid + 1; i++)
1605                 init_group_run_stat(&runstats[i]);
1606
1607         /*
1608          * find out how many threads stats we need. if group reporting isn't
1609          * enabled, it's one-per-td.
1610          */
1611         nr_ts = 0;
1612         last_ts = -1;
1613         for_each_td(td, i) {
1614                 if (!td->o.group_reporting) {
1615                         nr_ts++;
1616                         continue;
1617                 }
1618                 if (last_ts == td->groupid)
1619                         continue;
1620                 if (!td->o.stats)
1621                         continue;
1622
1623                 last_ts = td->groupid;
1624                 nr_ts++;
1625         }
1626
1627         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1628         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1629
1630         for (i = 0; i < nr_ts; i++) {
1631                 init_thread_stat(&threadstats[i]);
1632                 opt_lists[i] = NULL;
1633         }
1634
1635         j = 0;
1636         last_ts = -1;
1637         idx = 0;
1638         for_each_td(td, i) {
1639                 if (!td->o.stats)
1640                         continue;
1641                 if (idx && (!td->o.group_reporting ||
1642                     (td->o.group_reporting && last_ts != td->groupid))) {
1643                         idx = 0;
1644                         j++;
1645                 }
1646
1647                 last_ts = td->groupid;
1648
1649                 ts = &threadstats[j];
1650
1651                 ts->clat_percentiles = td->o.clat_percentiles;
1652                 ts->lat_percentiles = td->o.lat_percentiles;
1653                 ts->percentile_precision = td->o.percentile_precision;
1654                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1655                 opt_lists[j] = &td->opt_list;
1656
1657                 idx++;
1658                 ts->members++;
1659
1660                 if (ts->groupid == -1) {
1661                         /*
1662                          * These are per-group shared already
1663                          */
1664                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1665                         if (td->o.description)
1666                                 strncpy(ts->description, td->o.description,
1667                                                 FIO_JOBDESC_SIZE - 1);
1668                         else
1669                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1670
1671                         /*
1672                          * If multiple entries in this group, this is
1673                          * the first member.
1674                          */
1675                         ts->thread_number = td->thread_number;
1676                         ts->groupid = td->groupid;
1677
1678                         /*
1679                          * first pid in group, not very useful...
1680                          */
1681                         ts->pid = td->pid;
1682
1683                         ts->kb_base = td->o.kb_base;
1684                         ts->unit_base = td->o.unit_base;
1685                         ts->unified_rw_rep = td->o.unified_rw_rep;
1686                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1687                         log_info("fio: kb_base differs for jobs in group, using"
1688                                  " %u as the base\n", ts->kb_base);
1689                         kb_base_warned = true;
1690                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1691                         log_info("fio: unit_base differs for jobs in group, using"
1692                                  " %u as the base\n", ts->unit_base);
1693                         unit_base_warned = true;
1694                 }
1695
1696                 ts->continue_on_error = td->o.continue_on_error;
1697                 ts->total_err_count += td->total_err_count;
1698                 ts->first_error = td->first_error;
1699                 if (!ts->error) {
1700                         if (!td->error && td->o.continue_on_error &&
1701                             td->first_error) {
1702                                 ts->error = td->first_error;
1703                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1704                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1705                         } else  if (td->error) {
1706                                 ts->error = td->error;
1707                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1708                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1709                         }
1710                 }
1711
1712                 ts->latency_depth = td->latency_qd;
1713                 ts->latency_target = td->o.latency_target;
1714                 ts->latency_percentile = td->o.latency_percentile;
1715                 ts->latency_window = td->o.latency_window;
1716
1717                 ts->nr_block_infos = td->ts.nr_block_infos;
1718                 for (k = 0; k < ts->nr_block_infos; k++)
1719                         ts->block_infos[k] = td->ts.block_infos[k];
1720
1721                 sum_thread_stats(ts, &td->ts, idx == 1);
1722
1723                 if (td->o.ss_dur) {
1724                         ts->ss_state = td->ss.state;
1725                         ts->ss_dur = td->ss.dur;
1726                         ts->ss_head = td->ss.head;
1727                         ts->ss_bw_data = td->ss.bw_data;
1728                         ts->ss_iops_data = td->ss.iops_data;
1729                         ts->ss_limit.u.f = td->ss.limit;
1730                         ts->ss_slope.u.f = td->ss.slope;
1731                         ts->ss_deviation.u.f = td->ss.deviation;
1732                         ts->ss_criterion.u.f = td->ss.criterion;
1733                 }
1734                 else
1735                         ts->ss_dur = ts->ss_state = 0;
1736         }
1737
1738         for (i = 0; i < nr_ts; i++) {
1739                 unsigned long long bw;
1740
1741                 ts = &threadstats[i];
1742                 if (ts->groupid == -1)
1743                         continue;
1744                 rs = &runstats[ts->groupid];
1745                 rs->kb_base = ts->kb_base;
1746                 rs->unit_base = ts->unit_base;
1747                 rs->unified_rw_rep += ts->unified_rw_rep;
1748
1749                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1750                         if (!ts->runtime[j])
1751                                 continue;
1752                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1753                                 rs->min_run[j] = ts->runtime[j];
1754                         if (ts->runtime[j] > rs->max_run[j])
1755                                 rs->max_run[j] = ts->runtime[j];
1756
1757                         bw = 0;
1758                         if (ts->runtime[j])
1759                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1760                         if (bw < rs->min_bw[j])
1761                                 rs->min_bw[j] = bw;
1762                         if (bw > rs->max_bw[j])
1763                                 rs->max_bw[j] = bw;
1764
1765                         rs->iobytes[j] += ts->io_bytes[j];
1766                 }
1767         }
1768
1769         for (i = 0; i < groupid + 1; i++) {
1770                 int ddir;
1771
1772                 rs = &runstats[i];
1773
1774                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1775                         if (rs->max_run[ddir])
1776                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1777                                                 rs->max_run[ddir];
1778                 }
1779         }
1780
1781         for (i = 0; i < FIO_OUTPUT_NR; i++)
1782                 buf_output_init(&output[i]);
1783
1784         /*
1785          * don't overwrite last signal output
1786          */
1787         if (output_format & FIO_OUTPUT_NORMAL)
1788                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1789         if (output_format & FIO_OUTPUT_JSON) {
1790                 struct thread_data *global;
1791                 char time_buf[32];
1792                 struct timeval now;
1793                 unsigned long long ms_since_epoch;
1794
1795                 gettimeofday(&now, NULL);
1796                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1797                                  (unsigned long long)(now.tv_usec) / 1000;
1798
1799                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1800                                 sizeof(time_buf));
1801                 if (time_buf[strlen(time_buf) - 1] == '\n')
1802                         time_buf[strlen(time_buf) - 1] = '\0';
1803
1804                 root = json_create_object();
1805                 json_object_add_value_string(root, "fio version", fio_version_string);
1806                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1807                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1808                 json_object_add_value_string(root, "time", time_buf);
1809                 global = get_global_options();
1810                 json_add_job_opts(root, "global options", &global->opt_list, false);
1811                 array = json_create_array();
1812                 json_object_add_value_array(root, "jobs", array);
1813         }
1814
1815         if (is_backend)
1816                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1817
1818         for (i = 0; i < nr_ts; i++) {
1819                 ts = &threadstats[i];
1820                 rs = &runstats[ts->groupid];
1821
1822                 if (is_backend) {
1823                         fio_server_send_job_options(opt_lists[i], i);
1824                         fio_server_send_ts(ts, rs);
1825                 } else {
1826                         if (output_format & FIO_OUTPUT_TERSE)
1827                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1828                         if (output_format & FIO_OUTPUT_JSON) {
1829                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1830                                 json_array_add_value_object(array, tmp);
1831                         }
1832                         if (output_format & FIO_OUTPUT_NORMAL)
1833                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1834                 }
1835         }
1836         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1837                 /* disk util stats, if any */
1838                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1839
1840                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1841
1842                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1843                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1844                 json_free_object(root);
1845         }
1846
1847         for (i = 0; i < groupid + 1; i++) {
1848                 rs = &runstats[i];
1849
1850                 rs->groupid = i;
1851                 if (is_backend)
1852                         fio_server_send_gs(rs);
1853                 else if (output_format & FIO_OUTPUT_NORMAL)
1854                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1855         }
1856
1857         if (is_backend)
1858                 fio_server_send_du();
1859         else if (output_format & FIO_OUTPUT_NORMAL) {
1860                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1861                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1862         }
1863
1864         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1865                 struct buf_output *out = &output[i];
1866
1867                 log_info_buf(out->buf, out->buflen);
1868                 buf_output_free(out);
1869         }
1870
1871         log_info_flush();
1872         free(runstats);
1873         free(threadstats);
1874         free(opt_lists);
1875 }
1876
1877 void show_run_stats(void)
1878 {
1879         fio_mutex_down(stat_mutex);
1880         __show_run_stats();
1881         fio_mutex_up(stat_mutex);
1882 }
1883
1884 void __show_running_run_stats(void)
1885 {
1886         struct thread_data *td;
1887         unsigned long long *rt;
1888         struct timespec ts;
1889         int i;
1890
1891         fio_mutex_down(stat_mutex);
1892
1893         rt = malloc(thread_number * sizeof(unsigned long long));
1894         fio_gettime(&ts, NULL);
1895
1896         for_each_td(td, i) {
1897                 td->update_rusage = 1;
1898                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1899                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1900                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1901                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1902
1903                 rt[i] = mtime_since(&td->start, &ts);
1904                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1905                         td->ts.runtime[DDIR_READ] += rt[i];
1906                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1907                         td->ts.runtime[DDIR_WRITE] += rt[i];
1908                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1909                         td->ts.runtime[DDIR_TRIM] += rt[i];
1910         }
1911
1912         for_each_td(td, i) {
1913                 if (td->runstate >= TD_EXITED)
1914                         continue;
1915                 if (td->rusage_sem) {
1916                         td->update_rusage = 1;
1917                         fio_mutex_down(td->rusage_sem);
1918                 }
1919                 td->update_rusage = 0;
1920         }
1921
1922         __show_run_stats();
1923
1924         for_each_td(td, i) {
1925                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1926                         td->ts.runtime[DDIR_READ] -= rt[i];
1927                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1928                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1929                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1930                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1931         }
1932
1933         free(rt);
1934         fio_mutex_up(stat_mutex);
1935 }
1936
1937 static bool status_interval_init;
1938 static struct timespec status_time;
1939 static bool status_file_disabled;
1940
1941 #define FIO_STATUS_FILE         "fio-dump-status"
1942
1943 static int check_status_file(void)
1944 {
1945         struct stat sb;
1946         const char *temp_dir;
1947         char fio_status_file_path[PATH_MAX];
1948
1949         if (status_file_disabled)
1950                 return 0;
1951
1952         temp_dir = getenv("TMPDIR");
1953         if (temp_dir == NULL) {
1954                 temp_dir = getenv("TEMP");
1955                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1956                         temp_dir = NULL;
1957         }
1958         if (temp_dir == NULL)
1959                 temp_dir = "/tmp";
1960
1961         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1962
1963         if (stat(fio_status_file_path, &sb))
1964                 return 0;
1965
1966         if (unlink(fio_status_file_path) < 0) {
1967                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1968                                                         strerror(errno));
1969                 log_err("fio: disabling status file updates\n");
1970                 status_file_disabled = true;
1971         }
1972
1973         return 1;
1974 }
1975
1976 void check_for_running_stats(void)
1977 {
1978         if (status_interval) {
1979                 if (!status_interval_init) {
1980                         fio_gettime(&status_time, NULL);
1981                         status_interval_init = true;
1982                 } else if (mtime_since_now(&status_time) >= status_interval) {
1983                         show_running_run_stats();
1984                         fio_gettime(&status_time, NULL);
1985                         return;
1986                 }
1987         }
1988         if (check_status_file()) {
1989                 show_running_run_stats();
1990                 return;
1991         }
1992 }
1993
1994 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
1995 {
1996         double val = data;
1997         double delta;
1998
1999         if (data > is->max_val)
2000                 is->max_val = data;
2001         if (data < is->min_val)
2002                 is->min_val = data;
2003
2004         delta = val - is->mean.u.f;
2005         if (delta) {
2006                 is->mean.u.f += delta / (is->samples + 1.0);
2007                 is->S.u.f += delta * (val - is->mean.u.f);
2008         }
2009
2010         is->samples++;
2011 }
2012
2013 /*
2014  * Return a struct io_logs, which is added to the tail of the log
2015  * list for 'iolog'.
2016  */
2017 static struct io_logs *get_new_log(struct io_log *iolog)
2018 {
2019         size_t new_size, new_samples;
2020         struct io_logs *cur_log;
2021
2022         /*
2023          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2024          * forever
2025          */
2026         if (!iolog->cur_log_max)
2027                 new_samples = DEF_LOG_ENTRIES;
2028         else {
2029                 new_samples = iolog->cur_log_max * 2;
2030                 if (new_samples > MAX_LOG_ENTRIES)
2031                         new_samples = MAX_LOG_ENTRIES;
2032         }
2033
2034         new_size = new_samples * log_entry_sz(iolog);
2035
2036         cur_log = smalloc(sizeof(*cur_log));
2037         if (cur_log) {
2038                 INIT_FLIST_HEAD(&cur_log->list);
2039                 cur_log->log = malloc(new_size);
2040                 if (cur_log->log) {
2041                         cur_log->nr_samples = 0;
2042                         cur_log->max_samples = new_samples;
2043                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2044                         iolog->cur_log_max = new_samples;
2045                         return cur_log;
2046                 }
2047                 sfree(cur_log);
2048         }
2049
2050         return NULL;
2051 }
2052
2053 /*
2054  * Add and return a new log chunk, or return current log if big enough
2055  */
2056 static struct io_logs *regrow_log(struct io_log *iolog)
2057 {
2058         struct io_logs *cur_log;
2059         int i;
2060
2061         if (!iolog || iolog->disabled)
2062                 goto disable;
2063
2064         cur_log = iolog_cur_log(iolog);
2065         if (!cur_log) {
2066                 cur_log = get_new_log(iolog);
2067                 if (!cur_log)
2068                         return NULL;
2069         }
2070
2071         if (cur_log->nr_samples < cur_log->max_samples)
2072                 return cur_log;
2073
2074         /*
2075          * No room for a new sample. If we're compressing on the fly, flush
2076          * out the current chunk
2077          */
2078         if (iolog->log_gz) {
2079                 if (iolog_cur_flush(iolog, cur_log)) {
2080                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2081                         return NULL;
2082                 }
2083         }
2084
2085         /*
2086          * Get a new log array, and add to our list
2087          */
2088         cur_log = get_new_log(iolog);
2089         if (!cur_log) {
2090                 log_err("fio: failed extending iolog! Will stop logging.\n");
2091                 return NULL;
2092         }
2093
2094         if (!iolog->pending || !iolog->pending->nr_samples)
2095                 return cur_log;
2096
2097         /*
2098          * Flush pending items to new log
2099          */
2100         for (i = 0; i < iolog->pending->nr_samples; i++) {
2101                 struct io_sample *src, *dst;
2102
2103                 src = get_sample(iolog, iolog->pending, i);
2104                 dst = get_sample(iolog, cur_log, i);
2105                 memcpy(dst, src, log_entry_sz(iolog));
2106         }
2107         cur_log->nr_samples = iolog->pending->nr_samples;
2108
2109         iolog->pending->nr_samples = 0;
2110         return cur_log;
2111 disable:
2112         if (iolog)
2113                 iolog->disabled = true;
2114         return NULL;
2115 }
2116
2117 void regrow_logs(struct thread_data *td)
2118 {
2119         regrow_log(td->slat_log);
2120         regrow_log(td->clat_log);
2121         regrow_log(td->clat_hist_log);
2122         regrow_log(td->lat_log);
2123         regrow_log(td->bw_log);
2124         regrow_log(td->iops_log);
2125         td->flags &= ~TD_F_REGROW_LOGS;
2126 }
2127
2128 static struct io_logs *get_cur_log(struct io_log *iolog)
2129 {
2130         struct io_logs *cur_log;
2131
2132         cur_log = iolog_cur_log(iolog);
2133         if (!cur_log) {
2134                 cur_log = get_new_log(iolog);
2135                 if (!cur_log)
2136                         return NULL;
2137         }
2138
2139         if (cur_log->nr_samples < cur_log->max_samples)
2140                 return cur_log;
2141
2142         /*
2143          * Out of space. If we're in IO offload mode, or we're not doing
2144          * per unit logging (hence logging happens outside of the IO thread
2145          * as well), add a new log chunk inline. If we're doing inline
2146          * submissions, flag 'td' as needing a log regrow and we'll take
2147          * care of it on the submission side.
2148          */
2149         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2150             !per_unit_log(iolog))
2151                 return regrow_log(iolog);
2152
2153         iolog->td->flags |= TD_F_REGROW_LOGS;
2154         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2155         return iolog->pending;
2156 }
2157
2158 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2159                              enum fio_ddir ddir, unsigned int bs,
2160                              unsigned long t, uint64_t offset)
2161 {
2162         struct io_logs *cur_log;
2163
2164         if (iolog->disabled)
2165                 return;
2166         if (flist_empty(&iolog->io_logs))
2167                 iolog->avg_last[ddir] = t;
2168
2169         cur_log = get_cur_log(iolog);
2170         if (cur_log) {
2171                 struct io_sample *s;
2172
2173                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2174
2175                 s->data = data;
2176                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2177                 io_sample_set_ddir(iolog, s, ddir);
2178                 s->bs = bs;
2179
2180                 if (iolog->log_offset) {
2181                         struct io_sample_offset *so = (void *) s;
2182
2183                         so->offset = offset;
2184                 }
2185
2186                 cur_log->nr_samples++;
2187                 return;
2188         }
2189
2190         iolog->disabled = true;
2191 }
2192
2193 static inline void reset_io_stat(struct io_stat *ios)
2194 {
2195         ios->max_val = ios->min_val = ios->samples = 0;
2196         ios->mean.u.f = ios->S.u.f = 0;
2197 }
2198
2199 void reset_io_stats(struct thread_data *td)
2200 {
2201         struct thread_stat *ts = &td->ts;
2202         int i, j;
2203
2204         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2205                 reset_io_stat(&ts->clat_stat[i]);
2206                 reset_io_stat(&ts->slat_stat[i]);
2207                 reset_io_stat(&ts->lat_stat[i]);
2208                 reset_io_stat(&ts->bw_stat[i]);
2209                 reset_io_stat(&ts->iops_stat[i]);
2210
2211                 ts->io_bytes[i] = 0;
2212                 ts->runtime[i] = 0;
2213                 ts->total_io_u[i] = 0;
2214                 ts->short_io_u[i] = 0;
2215                 ts->drop_io_u[i] = 0;
2216
2217                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2218                         ts->io_u_plat[i][j] = 0;
2219         }
2220
2221         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2222                 ts->io_u_map[i] = 0;
2223                 ts->io_u_submit[i] = 0;
2224                 ts->io_u_complete[i] = 0;
2225         }
2226
2227         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2228                 ts->io_u_lat_n[i] = 0;
2229         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2230                 ts->io_u_lat_u[i] = 0;
2231         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2232                 ts->io_u_lat_m[i] = 0;
2233
2234         ts->total_submit = 0;
2235         ts->total_complete = 0;
2236 }
2237
2238 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2239                               unsigned long elapsed, bool log_max)
2240 {
2241         /*
2242          * Note an entry in the log. Use the mean from the logged samples,
2243          * making sure to properly round up. Only write a log entry if we
2244          * had actual samples done.
2245          */
2246         if (iolog->avg_window[ddir].samples) {
2247                 union io_sample_data data;
2248
2249                 if (log_max)
2250                         data.val = iolog->avg_window[ddir].max_val;
2251                 else
2252                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2253
2254                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2255         }
2256
2257         reset_io_stat(&iolog->avg_window[ddir]);
2258 }
2259
2260 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2261                              bool log_max)
2262 {
2263         int ddir;
2264
2265         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2266                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2267 }
2268
2269 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2270                            union io_sample_data data, enum fio_ddir ddir,
2271                            unsigned int bs, uint64_t offset)
2272 {
2273         unsigned long elapsed, this_window;
2274
2275         if (!ddir_rw(ddir))
2276                 return 0;
2277
2278         elapsed = mtime_since_now(&td->epoch);
2279
2280         /*
2281          * If no time averaging, just add the log sample.
2282          */
2283         if (!iolog->avg_msec) {
2284                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2285                 return 0;
2286         }
2287
2288         /*
2289          * Add the sample. If the time period has passed, then
2290          * add that entry to the log and clear.
2291          */
2292         add_stat_sample(&iolog->avg_window[ddir], data.val);
2293
2294         /*
2295          * If period hasn't passed, adding the above sample is all we
2296          * need to do.
2297          */
2298         this_window = elapsed - iolog->avg_last[ddir];
2299         if (elapsed < iolog->avg_last[ddir])
2300                 return iolog->avg_last[ddir] - elapsed;
2301         else if (this_window < iolog->avg_msec) {
2302                 int diff = iolog->avg_msec - this_window;
2303
2304                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2305                         return diff;
2306         }
2307
2308         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2309
2310         iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2311         return iolog->avg_msec;
2312 }
2313
2314 void finalize_logs(struct thread_data *td, bool unit_logs)
2315 {
2316         unsigned long elapsed;
2317
2318         elapsed = mtime_since_now(&td->epoch);
2319
2320         if (td->clat_log && unit_logs)
2321                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2322         if (td->slat_log && unit_logs)
2323                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2324         if (td->lat_log && unit_logs)
2325                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2326         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2327                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2328         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2329                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2330 }
2331
2332 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2333 {
2334         struct io_log *iolog;
2335
2336         if (!ddir_rw(ddir))
2337                 return;
2338
2339         iolog = agg_io_log[ddir];
2340         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2341 }
2342
2343 static void add_clat_percentile_sample(struct thread_stat *ts,
2344                                 unsigned long long nsec, enum fio_ddir ddir)
2345 {
2346         unsigned int idx = plat_val_to_idx(nsec);
2347         assert(idx < FIO_IO_U_PLAT_NR);
2348
2349         ts->io_u_plat[ddir][idx]++;
2350 }
2351
2352 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2353                      unsigned long long nsec, unsigned int bs, uint64_t offset)
2354 {
2355         unsigned long elapsed, this_window;
2356         struct thread_stat *ts = &td->ts;
2357         struct io_log *iolog = td->clat_hist_log;
2358
2359         td_io_u_lock(td);
2360
2361         add_stat_sample(&ts->clat_stat[ddir], nsec);
2362
2363         if (td->clat_log)
2364                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2365                                offset);
2366
2367         if (ts->clat_percentiles)
2368                 add_clat_percentile_sample(ts, nsec, ddir);
2369
2370         if (iolog && iolog->hist_msec) {
2371                 struct io_hist *hw = &iolog->hist_window[ddir];
2372
2373                 hw->samples++;
2374                 elapsed = mtime_since_now(&td->epoch);
2375                 if (!hw->hist_last)
2376                         hw->hist_last = elapsed;
2377                 this_window = elapsed - hw->hist_last;
2378
2379                 if (this_window >= iolog->hist_msec) {
2380                         unsigned int *io_u_plat;
2381                         struct io_u_plat_entry *dst;
2382
2383                         /*
2384                          * Make a byte-for-byte copy of the latency histogram
2385                          * stored in td->ts.io_u_plat[ddir], recording it in a
2386                          * log sample. Note that the matching call to free() is
2387                          * located in iolog.c after printing this sample to the
2388                          * log file.
2389                          */
2390                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2391                         dst = malloc(sizeof(struct io_u_plat_entry));
2392                         memcpy(&(dst->io_u_plat), io_u_plat,
2393                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2394                         flist_add(&dst->list, &hw->list);
2395                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2396                                                 elapsed, offset);
2397
2398                         /*
2399                          * Update the last time we recorded as being now, minus
2400                          * any drift in time we encountered before actually
2401                          * making the record.
2402                          */
2403                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2404                         hw->samples = 0;
2405                 }
2406         }
2407
2408         td_io_u_unlock(td);
2409 }
2410
2411 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2412                      unsigned long usec, unsigned int bs, uint64_t offset)
2413 {
2414         struct thread_stat *ts = &td->ts;
2415
2416         if (!ddir_rw(ddir))
2417                 return;
2418
2419         td_io_u_lock(td);
2420
2421         add_stat_sample(&ts->slat_stat[ddir], usec);
2422
2423         if (td->slat_log)
2424                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2425
2426         td_io_u_unlock(td);
2427 }
2428
2429 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2430                     unsigned long long nsec, unsigned int bs, uint64_t offset)
2431 {
2432         struct thread_stat *ts = &td->ts;
2433
2434         if (!ddir_rw(ddir))
2435                 return;
2436
2437         td_io_u_lock(td);
2438
2439         add_stat_sample(&ts->lat_stat[ddir], nsec);
2440
2441         if (td->lat_log)
2442                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2443                                offset);
2444
2445         if (ts->lat_percentiles)
2446                 add_clat_percentile_sample(ts, nsec, ddir);
2447
2448         td_io_u_unlock(td);
2449 }
2450
2451 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2452                    unsigned int bytes, unsigned long long spent)
2453 {
2454         struct thread_stat *ts = &td->ts;
2455         unsigned long rate;
2456
2457         if (spent)
2458                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2459         else
2460                 rate = 0;
2461
2462         td_io_u_lock(td);
2463
2464         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2465
2466         if (td->bw_log)
2467                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2468                                bytes, io_u->offset);
2469
2470         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2471         td_io_u_unlock(td);
2472 }
2473
2474 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2475                          struct timespec *t, unsigned int avg_time,
2476                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2477                          struct io_stat *stat, struct io_log *log,
2478                          bool is_kb)
2479 {
2480         unsigned long spent, rate;
2481         enum fio_ddir ddir;
2482         unsigned int next, next_log;
2483
2484         next_log = avg_time;
2485
2486         spent = mtime_since(parent_tv, t);
2487         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2488                 return avg_time - spent;
2489
2490         td_io_u_lock(td);
2491
2492         /*
2493          * Compute both read and write rates for the interval.
2494          */
2495         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2496                 uint64_t delta;
2497
2498                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2499                 if (!delta)
2500                         continue; /* No entries for interval */
2501
2502                 if (spent) {
2503                         if (is_kb)
2504                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2505                         else
2506                                 rate = (delta * 1000) / spent;
2507                 } else
2508                         rate = 0;
2509
2510                 add_stat_sample(&stat[ddir], rate);
2511
2512                 if (log) {
2513                         unsigned int bs = 0;
2514
2515                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2516                                 bs = td->o.min_bs[ddir];
2517
2518                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2519                         next_log = min(next_log, next);
2520                 }
2521
2522                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2523         }
2524
2525         timespec_add_msec(parent_tv, avg_time);
2526
2527         td_io_u_unlock(td);
2528
2529         if (spent <= avg_time)
2530                 next = avg_time;
2531         else
2532                 next = avg_time - (1 + spent - avg_time);
2533
2534         return min(next, next_log);
2535 }
2536
2537 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2538 {
2539         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2540                                 td->this_io_bytes, td->stat_io_bytes,
2541                                 td->ts.bw_stat, td->bw_log, true);
2542 }
2543
2544 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2545                      unsigned int bytes)
2546 {
2547         struct thread_stat *ts = &td->ts;
2548
2549         td_io_u_lock(td);
2550
2551         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2552
2553         if (td->iops_log)
2554                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2555                                bytes, io_u->offset);
2556
2557         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2558         td_io_u_unlock(td);
2559 }
2560
2561 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2562 {
2563         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2564                                 td->this_io_blocks, td->stat_io_blocks,
2565                                 td->ts.iops_stat, td->iops_log, false);
2566 }
2567
2568 /*
2569  * Returns msecs to next event
2570  */
2571 int calc_log_samples(void)
2572 {
2573         struct thread_data *td;
2574         unsigned int next = ~0U, tmp;
2575         struct timespec now;
2576         int i;
2577
2578         fio_gettime(&now, NULL);
2579
2580         for_each_td(td, i) {
2581                 if (!td->o.stats)
2582                         continue;
2583                 if (in_ramp_time(td) ||
2584                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2585                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2586                         continue;
2587                 }
2588                 if (!td->bw_log ||
2589                         (td->bw_log && !per_unit_log(td->bw_log))) {
2590                         tmp = add_bw_samples(td, &now);
2591                         if (tmp < next)
2592                                 next = tmp;
2593                 }
2594                 if (!td->iops_log ||
2595                         (td->iops_log && !per_unit_log(td->iops_log))) {
2596                         tmp = add_iops_samples(td, &now);
2597                         if (tmp < next)
2598                                 next = tmp;
2599                 }
2600         }
2601
2602         return next == ~0U ? 0 : next;
2603 }
2604
2605 void stat_init(void)
2606 {
2607         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2608 }
2609
2610 void stat_exit(void)
2611 {
2612         /*
2613          * When we have the mutex, we know out-of-band access to it
2614          * have ended.
2615          */
2616         fio_mutex_down(stat_mutex);
2617         fio_mutex_remove(stat_mutex);
2618 }
2619
2620 /*
2621  * Called from signal handler. Wake up status thread.
2622  */
2623 void show_running_run_stats(void)
2624 {
2625         helper_do_stat();
2626 }
2627
2628 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2629 {
2630         /* Ignore io_u's which span multiple blocks--they will just get
2631          * inaccurate counts. */
2632         int idx = (io_u->offset - io_u->file->file_offset)
2633                         / td->o.bs[DDIR_TRIM];
2634         uint32_t *info = &td->ts.block_infos[idx];
2635         assert(idx < td->ts.nr_block_infos);
2636         return info;
2637 }