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         int 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 = 0;
 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);
 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                                   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 is_last, per_line, scale_down, time_width;
 209         char fmt[32];
 210
 211         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 212         if (!len)
 213                 goto out;
 214
 215         /*
 216          * We default to nsecs, but if the value range is such that we
 217          * should scale down to usecs or msecs, do that.
 218          */
 219         if (minv > 2000000 && maxv > 99999999ULL) {
 220                 scale_down = 2;
 221                 divisor = 1000000;
 222                 log_buf(out, "    clat percentiles (msec):\n     |");
 223         } else if (minv > 2000 && maxv > 99999) {
 224                 scale_down = 1;
 225                 divisor = 1000;
 226                 log_buf(out, "    clat percentiles (usec):\n     |");
 227         } else {
 228                 scale_down = 0;
 229                 divisor = 1;
 230                 log_buf(out, "    clat percentiles (nsec):\n     |");
 231         }
 232
 233
 234         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 235         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 236                         precision, time_width);
 237         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 238         per_line = (80 - 7) / (precision + 10 + time_width);
 239
 240         for (j = 0; j < len; j++) {
 241                 /* for formatting */
 242                 if (j != 0 && (j % per_line) == 0)
 243                         log_buf(out, "     |");
 244
 245                 /* end of the list */
 246                 is_last = (j == len - 1);
 247
 248                 for (i = 0; i < scale_down; i++)
 249                         ovals[j] = (ovals[j] + 999) / 1000;
 250
 251                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 252
 253                 if (is_last)
 254                         break;
 255
 256                 if ((j % per_line) == per_line - 1)     /* for formatting */
 257                         log_buf(out, "\n");
 258         }
 259
 260 out:
 261         if (ovals)
 262                 free(ovals);
 263 }
 264
 265 bool calc_lat(struct io_stat *is, unsigned long long *min,
 266               unsigned long long *max, double *mean, double *dev)
 267 {
 268         double n = (double) is->samples;
 269
 270         if (n == 0)
 271                 return false;
 272
 273         *min = is->min_val;
 274         *max = is->max_val;
 275         *mean = is->mean.u.f;
 276
 277         if (n > 1.0)
 278                 *dev = sqrt(is->S.u.f / (n - 1.0));
 279         else
 280                 *dev = 0;
 281
 282         return true;
 283 }
 284
 285 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 286 {
 287         char *io, *agg, *min, *max;
 288         char *ioalt, *aggalt, *minalt, *maxalt;
 289         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 290         int i;
 291
 292         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 293
 294         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 295                 const int i2p = is_power_of_2(rs->kb_base);
 296
 297                 if (!rs->max_run[i])
 298                         continue;
 299
 300                 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
 301                 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
 302                 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
 303                 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
 304                 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
 305                 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
 306                 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
 307                 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
 308                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 309                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 310                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 311                                 (unsigned long long) rs->min_run[i],
 312                                 (unsigned long long) rs->max_run[i]);
 313
 314                 free(io);
 315                 free(agg);
 316                 free(min);
 317                 free(max);
 318                 free(ioalt);
 319                 free(aggalt);
 320                 free(minalt);
 321                 free(maxalt);
 322         }
 323 }
 324
 325 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 326 {
 327         int i;
 328
 329         /*
 330          * Do depth distribution calculations
 331          */
 332         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 333                 if (total) {
 334                         io_u_dist[i] = (double) map[i] / (double) total;
 335                         io_u_dist[i] *= 100.0;
 336                         if (io_u_dist[i] < 0.1 && map[i])
 337                                 io_u_dist[i] = 0.1;
 338                 } else
 339                         io_u_dist[i] = 0.0;
 340         }
 341 }
 342
 343 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 344                           unsigned int *src, int nr)
 345 {
 346         unsigned long total = ddir_rw_sum(ts->total_io_u);
 347         int i;
 348
 349         /*
 350          * Do latency distribution calculations
 351          */
 352         for (i = 0; i < nr; i++) {
 353                 if (total) {
 354                         dst[i] = (double) src[i] / (double) total;
 355                         dst[i] *= 100.0;
 356                         if (dst[i] < 0.01 && src[i])
 357                                 dst[i] = 0.01;
 358                 } else
 359                         dst[i] = 0.0;
 360         }
 361 }
 362
 363 /*
 364  * To keep the terse format unaltered, add all of the ns latency
 365  * buckets to the first us latency bucket
 366  */
 367 void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 368 {
 369         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 370         int i;
 371
 372         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 373
 374         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 375                 ntotal += ts->io_u_lat_n[i];
 376
 377         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 378 }
 379
 380 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 381 {
 382         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 383 }
 384
 385 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 386 {
 387         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 388 }
 389
 390 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 391 {
 392         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 393 }
 394
 395 static void display_lat(const char *name, unsigned long long min,
 396                         unsigned long long max, double mean, double dev,
 397                         struct buf_output *out)
 398 {
 399         const char *base = "(nsec)";
 400         char *minp, *maxp;
 401
 402         if (nsec_to_msec(&min, &max, &mean, &dev))
 403                 base = "(msec)";
 404         else if (nsec_to_usec(&min, &max, &mean, &dev))
 405                 base = "(usec)";
 406
 407         minp = num2str(min, 6, 1, 0, N2S_NONE);
 408         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 409
 410         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 411                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 412
 413         free(minp);
 414         free(maxp);
 415 }
 416
 417 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 418                              int ddir, struct buf_output *out)
 419 {
 420         const char *str[] = { " read", "write", " trim" };
 421         unsigned long runt;
 422         unsigned long long min, max, bw, iops;
 423         double mean, dev;
 424         char *io_p, *bw_p, *bw_p_alt, *iops_p;
 425         int i2p;
 426
 427         assert(ddir_rw(ddir));
 428
 429         if (!ts->runtime[ddir])
 430                 return;
 431
 432         i2p = is_power_of_2(rs->kb_base);
 433         runt = ts->runtime[ddir];
 434
 435         bw = (1000 * ts->io_bytes[ddir]) / runt;
 436         io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
 437         bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
 438         bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
 439
 440         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 441         iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
 442
 443         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
 444                         rs->unified_rw_rep ? "mixed" : str[ddir],
 445                         iops_p, bw_p, bw_p_alt, io_p,
 446                         (unsigned long long) ts->runtime[ddir]);
 447
 448         free(io_p);
 449         free(bw_p);
 450         free(bw_p_alt);
 451         free(iops_p);
 452
 453         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 454                 display_lat("slat", min, max, mean, dev, out);
 455         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 456                 display_lat("clat", min, max, mean, dev, out);
 457         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 458                 display_lat(" lat", min, max, mean, dev, out);
 459
 460         if (ts->clat_percentiles) {
 461                 show_clat_percentiles(ts->io_u_plat[ddir],
 462                                         ts->clat_stat[ddir].samples,
 463                                         ts->percentile_list,
 464                                         ts->percentile_precision, out);
 465         }
 466         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 467                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 468                 const char *bw_str;
 469
 470                 if ((rs->unit_base == 1) && i2p)
 471                         bw_str = "Kibit";
 472                 else if (rs->unit_base == 1)
 473                         bw_str = "kbit";
 474                 else if (i2p)
 475                         bw_str = "KiB";
 476                 else
 477                         bw_str = "kB";
 478
 479                 if (rs->agg[ddir]) {
 480                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 481                         if (p_of_agg > 100.0)
 482                                 p_of_agg = 100.0;
 483                 }
 484
 485                 if (rs->unit_base == 1) {
 486                         min *= 8.0;
 487                         max *= 8.0;
 488                         mean *= 8.0;
 489                         dev *= 8.0;
 490                 }
 491
 492                 if (mean > fkb_base * fkb_base) {
 493                         min /= fkb_base;
 494                         max /= fkb_base;
 495                         mean /= fkb_base;
 496                         dev /= fkb_base;
 497                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 498                 }
 499
 500                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 501                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 502                         bw_str, min, max, p_of_agg, mean, dev,
 503                         (&ts->bw_stat[ddir])->samples);
 504         }
 505         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 506                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 507                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 508                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 509         }
 510 }
 511
 512 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 513                     const char *msg, struct buf_output *out)
 514 {
 515         int new_line = 1, i, line = 0, shown = 0;
 516
 517         for (i = 0; i < nr; i++) {
 518                 if (io_u_lat[i] <= 0.0)
 519                         continue;
 520                 shown = 1;
 521                 if (new_line) {
 522                         if (line)
 523                                 log_buf(out, "\n");
 524                         log_buf(out, "  lat (%s)   : ", msg);
 525                         new_line = 0;
 526                         line = 0;
 527                 }
 528                 if (line)
 529                         log_buf(out, ", ");
 530                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 531                 line++;
 532                 if (line == 5)
 533                         new_line = 1;
 534         }
 535
 536         if (shown)
 537                 log_buf(out, "\n");
 538
 539         return shown;
 540 }
 541
 542 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 543 {
 544         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 545                                  "250=", "500=", "750=", "1000=", };
 546
 547         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 548 }
 549
 550 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 551 {
 552         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 553                                  "250=", "500=", "750=", "1000=", };
 554
 555         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 556 }
 557
 558 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 559 {
 560         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 561                                  "250=", "500=", "750=", "1000=", "2000=",
 562                                  ">=2000=", };
 563
 564         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 565 }
 566
 567 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 568 {
 569         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 570         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 571         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 572
 573         stat_calc_lat_n(ts, io_u_lat_n);
 574         stat_calc_lat_u(ts, io_u_lat_u);
 575         stat_calc_lat_m(ts, io_u_lat_m);
 576
 577         show_lat_n(io_u_lat_n, out);
 578         show_lat_u(io_u_lat_u, out);
 579         show_lat_m(io_u_lat_m, out);
 580 }
 581
 582 static int block_state_category(int block_state)
 583 {
 584         switch (block_state) {
 585         case BLOCK_STATE_UNINIT:
 586                 return 0;
 587         case BLOCK_STATE_TRIMMED:
 588         case BLOCK_STATE_WRITTEN:
 589                 return 1;
 590         case BLOCK_STATE_WRITE_FAILURE:
 591         case BLOCK_STATE_TRIM_FAILURE:
 592                 return 2;
 593         default:
 594                 /* Silence compile warning on some BSDs and have a return */
 595                 assert(0);
 596                 return -1;
 597         }
 598 }
 599
 600 static int compare_block_infos(const void *bs1, const void *bs2)
 601 {
 602         uint32_t block1 = *(uint32_t *)bs1;
 603         uint32_t block2 = *(uint32_t *)bs2;
 604         int state1 = BLOCK_INFO_STATE(block1);
 605         int state2 = BLOCK_INFO_STATE(block2);
 606         int bscat1 = block_state_category(state1);
 607         int bscat2 = block_state_category(state2);
 608         int cycles1 = BLOCK_INFO_TRIMS(block1);
 609         int cycles2 = BLOCK_INFO_TRIMS(block2);
 610
 611         if (bscat1 < bscat2)
 612                 return -1;
 613         if (bscat1 > bscat2)
 614                 return 1;
 615
 616         if (cycles1 < cycles2)
 617                 return -1;
 618         if (cycles1 > cycles2)
 619                 return 1;
 620
 621         if (state1 < state2)
 622                 return -1;
 623         if (state1 > state2)
 624                 return 1;
 625
 626         assert(block1 == block2);
 627         return 0;
 628 }
 629
 630 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 631                                   fio_fp64_t *plist, unsigned int **percentiles,
 632                                   unsigned int *types)
 633 {
 634         int len = 0;
 635         int i, nr_uninit;
 636
 637         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 638
 639         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 640                 len++;
 641
 642         if (!len)
 643                 return 0;
 644
 645         /*
 646          * Sort the percentile list. Note that it may already be sorted if
 647          * we are using the default values, but since it's a short list this
 648          * isn't a worry. Also note that this does not work for NaN values.
 649          */
 650         if (len > 1)
 651                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 652
 653         nr_uninit = 0;
 654         /* Start only after the uninit entries end */
 655         for (nr_uninit = 0;
 656              nr_uninit < nr_block_infos
 657                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 658              nr_uninit ++)
 659                 ;
 660
 661         if (nr_uninit == nr_block_infos)
 662                 return 0;
 663
 664         *percentiles = calloc(len, sizeof(**percentiles));
 665
 666         for (i = 0; i < len; i++) {
 667                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 668                                 + nr_uninit;
 669                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 670         }
 671
 672         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 673         for (i = 0; i < nr_block_infos; i++)
 674                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 675
 676         return len;
 677 }
 678
 679 static const char *block_state_names[] = {
 680         [BLOCK_STATE_UNINIT] = "unwritten",
 681         [BLOCK_STATE_TRIMMED] = "trimmed",
 682         [BLOCK_STATE_WRITTEN] = "written",
 683         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 684         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 685 };
 686
 687 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 688                              fio_fp64_t *plist, struct buf_output *out)
 689 {
 690         int len, pos, i;
 691         unsigned int *percentiles = NULL;
 692         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 693
 694         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 695                                      &percentiles, block_state_counts);
 696
 697         log_buf(out, "  block lifetime percentiles :\n   |");
 698         pos = 0;
 699         for (i = 0; i < len; i++) {
 700                 uint32_t block_info = percentiles[i];
 701 #define LINE_LENGTH     75
 702                 char str[LINE_LENGTH];
 703                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 704                                      plist[i].u.f, block_info,
 705                                      i == len - 1 ? '\n' : ',');
 706                 assert(strln < LINE_LENGTH);
 707                 if (pos + strln > LINE_LENGTH) {
 708                         pos = 0;
 709                         log_buf(out, "\n   |");
 710                 }
 711                 log_buf(out, "%s", str);
 712                 pos += strln;
 713 #undef LINE_LENGTH
 714         }
 715         if (percentiles)
 716                 free(percentiles);
 717
 718         log_buf(out, "        states               :");
 719         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 720                 log_buf(out, " %s=%u%c",
 721                          block_state_names[i], block_state_counts[i],
 722                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 723 }
 724
 725 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 726 {
 727         char *p1, *p1alt, *p2;
 728         unsigned long long bw_mean, iops_mean;
 729         const int i2p = is_power_of_2(ts->kb_base);
 730
 731         if (!ts->ss_dur)
 732                 return;
 733
 734         bw_mean = steadystate_bw_mean(ts);
 735         iops_mean = steadystate_iops_mean(ts);
 736
 737         p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
 738         p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
 739         p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
 740
 741         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 742                 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
 743                 p1, p1alt, p2,
 744                 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
 745                 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
 746                 ts->ss_criterion.u.f,
 747                 ts->ss_state & __FIO_SS_PCT ? "%" : "");
 748
 749         free(p1);
 750         free(p1alt);
 751         free(p2);
 752 }
 753
 754 static void show_thread_status_normal(struct thread_stat *ts,
 755                                       struct group_run_stats *rs,
 756                                       struct buf_output *out)
 757 {
 758         double usr_cpu, sys_cpu;
 759         unsigned long runtime;
 760         double io_u_dist[FIO_IO_U_MAP_NR];
 761         time_t time_p;
 762         char time_buf[32];
 763
 764         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 765                 return;
 766
 767         memset(time_buf, 0, sizeof(time_buf));
 768
 769         time(&time_p);
 770         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 771
 772         if (!ts->error) {
 773                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 774                                         ts->name, ts->groupid, ts->members,
 775                                         ts->error, (int) ts->pid, time_buf);
 776         } else {
 777                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 778                                         ts->name, ts->groupid, ts->members,
 779                                         ts->error, ts->verror, (int) ts->pid,
 780                                         time_buf);
 781         }
 782
 783         if (strlen(ts->description))
 784                 log_buf(out, "  Description  : [%s]\n", ts->description);
 785
 786         if (ts->io_bytes[DDIR_READ])
 787                 show_ddir_status(rs, ts, DDIR_READ, out);
 788         if (ts->io_bytes[DDIR_WRITE])
 789                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 790         if (ts->io_bytes[DDIR_TRIM])
 791                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 792
 793         show_latencies(ts, out);
 794
 795         runtime = ts->total_run_time;
 796         if (runtime) {
 797                 double runt = (double) runtime;
 798
 799                 usr_cpu = (double) ts->usr_time * 100 / runt;
 800                 sys_cpu = (double) ts->sys_time * 100 / runt;
 801         } else {
 802                 usr_cpu = 0;
 803                 sys_cpu = 0;
 804         }
 805
 806         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 807                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 808                         (unsigned long long) ts->ctx,
 809                         (unsigned long long) ts->majf,
 810                         (unsigned long long) ts->minf);
 811
 812         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 813         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 814                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 815                                         io_u_dist[1], io_u_dist[2],
 816                                         io_u_dist[3], io_u_dist[4],
 817                                         io_u_dist[5], io_u_dist[6]);
 818
 819         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 820         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 821                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 822                                         io_u_dist[1], io_u_dist[2],
 823                                         io_u_dist[3], io_u_dist[4],
 824                                         io_u_dist[5], io_u_dist[6]);
 825         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 826         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 827                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 828                                         io_u_dist[1], io_u_dist[2],
 829                                         io_u_dist[3], io_u_dist[4],
 830                                         io_u_dist[5], io_u_dist[6]);
 831         log_buf(out, "     issued rwt: total=%llu,%llu,%llu,"
 832                                  " short=%llu,%llu,%llu,"
 833                                  " dropped=%llu,%llu,%llu\n",
 834                                         (unsigned long long) ts->total_io_u[0],
 835                                         (unsigned long long) ts->total_io_u[1],
 836                                         (unsigned long long) ts->total_io_u[2],
 837                                         (unsigned long long) ts->short_io_u[0],
 838                                         (unsigned long long) ts->short_io_u[1],
 839                                         (unsigned long long) ts->short_io_u[2],
 840                                         (unsigned long long) ts->drop_io_u[0],
 841                                         (unsigned long long) ts->drop_io_u[1],
 842                                         (unsigned long long) ts->drop_io_u[2]);
 843         if (ts->continue_on_error) {
 844                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 845                                         (unsigned long long)ts->total_err_count,
 846                                         ts->first_error,
 847                                         strerror(ts->first_error));
 848         }
 849         if (ts->latency_depth) {
 850                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 851                                         (unsigned long long)ts->latency_target,
 852                                         (unsigned long long)ts->latency_window,
 853                                         ts->latency_percentile.u.f,
 854                                         ts->latency_depth);
 855         }
 856
 857         if (ts->nr_block_infos)
 858                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 859                                   ts->percentile_list, out);
 860
 861         if (ts->ss_dur)
 862                 show_ss_normal(ts, out);
 863 }
 864
 865 static void show_ddir_status_terse(struct thread_stat *ts,
 866                                    struct group_run_stats *rs, int ddir,
 867                                    int ver, struct buf_output *out)
 868 {
 869         unsigned long long min, max, minv, maxv, bw, iops;
 870         unsigned long long *ovals = NULL;
 871         double mean, dev;
 872         unsigned int len;
 873         int i, bw_stat;
 874
 875         assert(ddir_rw(ddir));
 876
 877         iops = bw = 0;
 878         if (ts->runtime[ddir]) {
 879                 uint64_t runt = ts->runtime[ddir];
 880
 881                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 882                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 883         }
 884
 885         log_buf(out, ";%llu;%llu;%llu;%llu",
 886                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 887                                         (unsigned long long) ts->runtime[ddir]);
 888
 889         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 890                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 891         else
 892                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 893
 894         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 895                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 896         else
 897                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 898
 899         if (ts->clat_percentiles) {
 900                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 901                                         ts->clat_stat[ddir].samples,
 902                                         ts->percentile_list, &ovals, &maxv,
 903                                         &minv);
 904         } else
 905                 len = 0;
 906
 907         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 908                 if (i >= len) {
 909                         log_buf(out, ";0%%=0");
 910                         continue;
 911                 }
 912                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
 913         }
 914
 915         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 916                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 917         else
 918                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 919
 920         if (ovals)
 921                 free(ovals);
 922
 923         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
 924         if (bw_stat) {
 925                 double p_of_agg = 100.0;
 926
 927                 if (rs->agg[ddir]) {
 928                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 929                         if (p_of_agg > 100.0)
 930                                 p_of_agg = 100.0;
 931                 }
 932
 933                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 934         } else
 935                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
 936
 937         if (ver == 5) {
 938                 if (bw_stat)
 939                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
 940                 else
 941                         log_buf(out, ";%lu", 0UL);
 942
 943                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
 944                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
 945                                 mean, dev, (&ts->iops_stat[ddir])->samples);
 946                 else
 947                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
 948         }
 949 }
 950
 951 static void add_ddir_status_json(struct thread_stat *ts,
 952                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 953 {
 954         unsigned long long min, max, minv, maxv;
 955         unsigned long long bw;
 956         unsigned long long *ovals = NULL;
 957         double mean, dev, iops;
 958         unsigned int len;
 959         int i;
 960         const char *ddirname[] = {"read", "write", "trim"};
 961         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 962         char buf[120];
 963         double p_of_agg = 100.0;
 964
 965         assert(ddir_rw(ddir));
 966
 967         if (ts->unified_rw_rep && ddir != DDIR_READ)
 968                 return;
 969
 970         dir_object = json_create_object();
 971         json_object_add_value_object(parent,
 972                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 973
 974         bw = 0;
 975         iops = 0.0;
 976         if (ts->runtime[ddir]) {
 977                 uint64_t runt = ts->runtime[ddir];
 978
 979                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 980                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 981         }
 982
 983         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
 984         json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
 985         json_object_add_value_int(dir_object, "bw", bw);
 986         json_object_add_value_float(dir_object, "iops", iops);
 987         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 988         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 989         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 990         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 991
 992         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 993                 min = max = 0;
 994                 mean = dev = 0.0;
 995         }
 996         tmp_object = json_create_object();
 997         json_object_add_value_object(dir_object, "slat_ns", tmp_object);
 998         json_object_add_value_int(tmp_object, "min", min);
 999         json_object_add_value_int(tmp_object, "max", max);
1000         json_object_add_value_float(tmp_object, "mean", mean);
1001         json_object_add_value_float(tmp_object, "stddev", dev);
1002
1003         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1004                 min = max = 0;
1005                 mean = dev = 0.0;
1006         }
1007         tmp_object = json_create_object();
1008         json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1009         json_object_add_value_int(tmp_object, "min", min);
1010         json_object_add_value_int(tmp_object, "max", max);
1011         json_object_add_value_float(tmp_object, "mean", mean);
1012         json_object_add_value_float(tmp_object, "stddev", dev);
1013
1014         if (ts->clat_percentiles) {
1015                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1016                                         ts->clat_stat[ddir].samples,
1017                                         ts->percentile_list, &ovals, &maxv,
1018                                         &minv);
1019         } else
1020                 len = 0;
1021
1022         percentile_object = json_create_object();
1023         json_object_add_value_object(tmp_object, "percentile", percentile_object);
1024         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1025                 if (i >= len) {
1026                         json_object_add_value_int(percentile_object, "0.00", 0);
1027                         continue;
1028                 }
1029                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1030                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1031         }
1032
1033         if (output_format & FIO_OUTPUT_JSON_PLUS) {
1034                 clat_bins_object = json_create_object();
1035                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1036                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1037                         if (ts->io_u_plat[ddir][i]) {
1038                                 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1039                                 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1040                         }
1041                 }
1042         }
1043
1044         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1045                 min = max = 0;
1046                 mean = dev = 0.0;
1047         }
1048         tmp_object = json_create_object();
1049         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1050         json_object_add_value_int(tmp_object, "min", min);
1051         json_object_add_value_int(tmp_object, "max", max);
1052         json_object_add_value_float(tmp_object, "mean", mean);
1053         json_object_add_value_float(tmp_object, "stddev", dev);
1054         if (ovals)
1055                 free(ovals);
1056
1057         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1058                 if (rs->agg[ddir]) {
1059                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1060                         if (p_of_agg > 100.0)
1061                                 p_of_agg = 100.0;
1062                 }
1063         } else {
1064                 min = max = 0;
1065                 p_of_agg = mean = dev = 0.0;
1066         }
1067         json_object_add_value_int(dir_object, "bw_min", min);
1068         json_object_add_value_int(dir_object, "bw_max", max);
1069         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1070         json_object_add_value_float(dir_object, "bw_mean", mean);
1071         json_object_add_value_float(dir_object, "bw_dev", dev);
1072         json_object_add_value_int(dir_object, "bw_samples",
1073                                 (&ts->bw_stat[ddir])->samples);
1074
1075         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1076                 min = max = 0;
1077                 mean = dev = 0.0;
1078         }
1079         json_object_add_value_int(dir_object, "iops_min", min);
1080         json_object_add_value_int(dir_object, "iops_max", max);
1081         json_object_add_value_float(dir_object, "iops_mean", mean);
1082         json_object_add_value_float(dir_object, "iops_stddev", dev);
1083         json_object_add_value_int(dir_object, "iops_samples",
1084                                 (&ts->iops_stat[ddir])->samples);
1085 }
1086
1087 static void show_thread_status_terse_all(struct thread_stat *ts,
1088                                          struct group_run_stats *rs, int ver,
1089                                          struct buf_output *out)
1090 {
1091         double io_u_dist[FIO_IO_U_MAP_NR];
1092         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1093         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1094         double usr_cpu, sys_cpu;
1095         int i;
1096
1097         /* General Info */
1098         if (ver == 2)
1099                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1100         else
1101                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1102                         ts->name, ts->groupid, ts->error);
1103
1104         /* Log Read Status */
1105         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1106         /* Log Write Status */
1107         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1108         /* Log Trim Status */
1109         if (ver == 2 || ver == 4 || ver == 5)
1110                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1111
1112         /* CPU Usage */
1113         if (ts->total_run_time) {
1114                 double runt = (double) ts->total_run_time;
1115
1116                 usr_cpu = (double) ts->usr_time * 100 / runt;
1117                 sys_cpu = (double) ts->sys_time * 100 / runt;
1118         } else {
1119                 usr_cpu = 0;
1120                 sys_cpu = 0;
1121         }
1122
1123         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1124                                                 (unsigned long long) ts->ctx,
1125                                                 (unsigned long long) ts->majf,
1126                                                 (unsigned long long) ts->minf);
1127
1128         /* Calc % distribution of IO depths, usecond, msecond latency */
1129         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1130         stat_calc_lat_nu(ts, io_u_lat_u);
1131         stat_calc_lat_m(ts, io_u_lat_m);
1132
1133         /* Only show fixed 7 I/O depth levels*/
1134         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1135                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1136                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1137
1138         /* Microsecond latency */
1139         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1140                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1141         /* Millisecond latency */
1142         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1143                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1144
1145         /* disk util stats, if any */
1146         if (ver >= 3)
1147                 show_disk_util(1, NULL, out);
1148
1149         /* Additional output if continue_on_error set - default off*/
1150         if (ts->continue_on_error)
1151                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1152         if (ver == 2)
1153                 log_buf(out, "\n");
1154
1155         /* Additional output if description is set */
1156         if (strlen(ts->description))
1157                 log_buf(out, ";%s", ts->description);
1158
1159         log_buf(out, "\n");
1160 }
1161
1162 static void json_add_job_opts(struct json_object *root, const char *name,
1163                               struct flist_head *opt_list, bool num_jobs)
1164 {
1165         struct json_object *dir_object;
1166         struct flist_head *entry;
1167         struct print_option *p;
1168
1169         if (flist_empty(opt_list))
1170                 return;
1171
1172         dir_object = json_create_object();
1173         json_object_add_value_object(root, name, dir_object);
1174
1175         flist_for_each(entry, opt_list) {
1176                 const char *pos = "";
1177
1178                 p = flist_entry(entry, struct print_option, list);
1179                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1180                         continue;
1181                 if (p->value)
1182                         pos = p->value;
1183                 json_object_add_value_string(dir_object, p->name, pos);
1184         }
1185 }
1186
1187 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1188                                                    struct group_run_stats *rs,
1189                                                    struct flist_head *opt_list)
1190 {
1191         struct json_object *root, *tmp;
1192         struct jobs_eta *je;
1193         double io_u_dist[FIO_IO_U_MAP_NR];
1194         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1195         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1196         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1197         double usr_cpu, sys_cpu;
1198         int i;
1199         size_t size;
1200
1201         root = json_create_object();
1202         json_object_add_value_string(root, "jobname", ts->name);
1203         json_object_add_value_int(root, "groupid", ts->groupid);
1204         json_object_add_value_int(root, "error", ts->error);
1205
1206         /* ETA Info */
1207         je = get_jobs_eta(true, &size);
1208         if (je) {
1209                 json_object_add_value_int(root, "eta", je->eta_sec);
1210                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1211         }
1212
1213         if (opt_list)
1214                 json_add_job_opts(root, "job options", opt_list, true);
1215
1216         add_ddir_status_json(ts, rs, DDIR_READ, root);
1217         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1218         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1219
1220         /* CPU Usage */
1221         if (ts->total_run_time) {
1222                 double runt = (double) ts->total_run_time;
1223
1224                 usr_cpu = (double) ts->usr_time * 100 / runt;
1225                 sys_cpu = (double) ts->sys_time * 100 / runt;
1226         } else {
1227                 usr_cpu = 0;
1228                 sys_cpu = 0;
1229         }
1230         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1231         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1232         json_object_add_value_int(root, "ctx", ts->ctx);
1233         json_object_add_value_int(root, "majf", ts->majf);
1234         json_object_add_value_int(root, "minf", ts->minf);
1235
1236
1237         /* Calc % distribution of IO depths, usecond, msecond latency */
1238         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1239         stat_calc_lat_n(ts, io_u_lat_n);
1240         stat_calc_lat_u(ts, io_u_lat_u);
1241         stat_calc_lat_m(ts, io_u_lat_m);
1242
1243         tmp = json_create_object();
1244         json_object_add_value_object(root, "iodepth_level", tmp);
1245         /* Only show fixed 7 I/O depth levels*/
1246         for (i = 0; i < 7; i++) {
1247                 char name[20];
1248                 if (i < 6)
1249                         snprintf(name, 20, "%d", 1 << i);
1250                 else
1251                         snprintf(name, 20, ">=%d", 1 << i);
1252                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1253         }
1254
1255         /* Nanosecond latency */
1256         tmp = json_create_object();
1257         json_object_add_value_object(root, "latency_ns", tmp);
1258         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1259                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1260                                  "250", "500", "750", "1000", };
1261                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1262         }
1263         /* Microsecond latency */
1264         tmp = json_create_object();
1265         json_object_add_value_object(root, "latency_us", tmp);
1266         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1267                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1268                                  "250", "500", "750", "1000", };
1269                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1270         }
1271         /* Millisecond latency */
1272         tmp = json_create_object();
1273         json_object_add_value_object(root, "latency_ms", tmp);
1274         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1275                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1276                                  "250", "500", "750", "1000", "2000",
1277                                  ">=2000", };
1278                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1279         }
1280
1281         /* Additional output if continue_on_error set - default off*/
1282         if (ts->continue_on_error) {
1283                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1284                 json_object_add_value_int(root, "first_error", ts->first_error);
1285         }
1286
1287         if (ts->latency_depth) {
1288                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1289                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1290                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1291                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1292         }
1293
1294         /* Additional output if description is set */
1295         if (strlen(ts->description))
1296                 json_object_add_value_string(root, "desc", ts->description);
1297
1298         if (ts->nr_block_infos) {
1299                 /* Block error histogram and types */
1300                 int len;
1301                 unsigned int *percentiles = NULL;
1302                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1303
1304                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1305                                              ts->percentile_list,
1306                                              &percentiles, block_state_counts);
1307
1308                 if (len) {
1309                         struct json_object *block, *percentile_object, *states;
1310                         int state;
1311                         block = json_create_object();
1312                         json_object_add_value_object(root, "block", block);
1313
1314                         percentile_object = json_create_object();
1315                         json_object_add_value_object(block, "percentiles",
1316                                                      percentile_object);
1317                         for (i = 0; i < len; i++) {
1318                                 char buf[20];
1319                                 snprintf(buf, sizeof(buf), "%f",
1320                                          ts->percentile_list[i].u.f);
1321                                 json_object_add_value_int(percentile_object,
1322                                                           (const char *)buf,
1323                                                           percentiles[i]);
1324                         }
1325
1326                         states = json_create_object();
1327                         json_object_add_value_object(block, "states", states);
1328                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1329                                 json_object_add_value_int(states,
1330                                         block_state_names[state],
1331                                         block_state_counts[state]);
1332                         }
1333                         free(percentiles);
1334                 }
1335         }
1336
1337         if (ts->ss_dur) {
1338                 struct json_object *data;
1339                 struct json_array *iops, *bw;
1340                 int i, j, k;
1341                 char ss_buf[64];
1342
1343                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1344                         ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1345                         ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1346                         (float) ts->ss_limit.u.f,
1347                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1348
1349                 tmp = json_create_object();
1350                 json_object_add_value_object(root, "steadystate", tmp);
1351                 json_object_add_value_string(tmp, "ss", ss_buf);
1352                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1353                 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1354
1355                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1356                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1357                 json_object_add_value_string(tmp, "criterion", ss_buf);
1358                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1359                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1360
1361                 data = json_create_object();
1362                 json_object_add_value_object(tmp, "data", data);
1363                 bw = json_create_array();
1364                 iops = json_create_array();
1365
1366                 /*
1367                 ** if ss was attained or the buffer is not full,
1368                 ** ss->head points to the first element in the list.
1369                 ** otherwise it actually points to the second element
1370                 ** in the list
1371                 */
1372                 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1373                         j = ts->ss_head;
1374                 else
1375                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1376                 for (i = 0; i < ts->ss_dur; i++) {
1377                         k = (j + i) % ts->ss_dur;
1378                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1379                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1380                 }
1381                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1382                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1383                 json_object_add_value_array(data, "iops", iops);
1384                 json_object_add_value_array(data, "bw", bw);
1385         }
1386
1387         return root;
1388 }
1389
1390 static void show_thread_status_terse(struct thread_stat *ts,
1391                                      struct group_run_stats *rs,
1392                                      struct buf_output *out)
1393 {
1394         if (terse_version >= 2 && terse_version <= 5)
1395                 show_thread_status_terse_all(ts, rs, terse_version, out);
1396         else
1397                 log_err("fio: bad terse version!? %d\n", terse_version);
1398 }
1399
1400 struct json_object *show_thread_status(struct thread_stat *ts,
1401                                        struct group_run_stats *rs,
1402                                        struct flist_head *opt_list,
1403                                        struct buf_output *out)
1404 {
1405         struct json_object *ret = NULL;
1406
1407         if (output_format & FIO_OUTPUT_TERSE)
1408                 show_thread_status_terse(ts, rs,  out);
1409         if (output_format & FIO_OUTPUT_JSON)
1410                 ret = show_thread_status_json(ts, rs, opt_list);
1411         if (output_format & FIO_OUTPUT_NORMAL)
1412                 show_thread_status_normal(ts, rs,  out);
1413
1414         return ret;
1415 }
1416
1417 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1418 {
1419         double mean, S;
1420
1421         if (src->samples == 0)
1422                 return;
1423
1424         dst->min_val = min(dst->min_val, src->min_val);
1425         dst->max_val = max(dst->max_val, src->max_val);
1426
1427         /*
1428          * Compute new mean and S after the merge
1429          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1430          *  #Parallel_algorithm>
1431          */
1432         if (first) {
1433                 mean = src->mean.u.f;
1434                 S = src->S.u.f;
1435         } else {
1436                 double delta = src->mean.u.f - dst->mean.u.f;
1437
1438                 mean = ((src->mean.u.f * src->samples) +
1439                         (dst->mean.u.f * dst->samples)) /
1440                         (dst->samples + src->samples);
1441
1442                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1443                         (dst->samples * src->samples) /
1444                         (dst->samples + src->samples);
1445         }
1446
1447         dst->samples += src->samples;
1448         dst->mean.u.f = mean;
1449         dst->S.u.f = S;
1450 }
1451
1452 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1453 {
1454         int i;
1455
1456         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1457                 if (dst->max_run[i] < src->max_run[i])
1458                         dst->max_run[i] = src->max_run[i];
1459                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1460                         dst->min_run[i] = src->min_run[i];
1461                 if (dst->max_bw[i] < src->max_bw[i])
1462                         dst->max_bw[i] = src->max_bw[i];
1463                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1464                         dst->min_bw[i] = src->min_bw[i];
1465
1466                 dst->iobytes[i] += src->iobytes[i];
1467                 dst->agg[i] += src->agg[i];
1468         }
1469
1470         if (!dst->kb_base)
1471                 dst->kb_base = src->kb_base;
1472         if (!dst->unit_base)
1473                 dst->unit_base = src->unit_base;
1474 }
1475
1476 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1477                       bool first)
1478 {
1479         int l, k;
1480
1481         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1482                 if (!dst->unified_rw_rep) {
1483                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1484                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1485                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1486                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1487                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1488
1489                         dst->io_bytes[l] += src->io_bytes[l];
1490
1491                         if (dst->runtime[l] < src->runtime[l])
1492                                 dst->runtime[l] = src->runtime[l];
1493                 } else {
1494                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1495                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1496                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1497                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1498                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1499
1500                         dst->io_bytes[0] += src->io_bytes[l];
1501
1502                         if (dst->runtime[0] < src->runtime[l])
1503                                 dst->runtime[0] = src->runtime[l];
1504
1505                         /*
1506                          * We're summing to the same destination, so override
1507                          * 'first' after the first iteration of the loop
1508                          */
1509                         first = false;
1510                 }
1511         }
1512
1513         dst->usr_time += src->usr_time;
1514         dst->sys_time += src->sys_time;
1515         dst->ctx += src->ctx;
1516         dst->majf += src->majf;
1517         dst->minf += src->minf;
1518
1519         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1520                 dst->io_u_map[k] += src->io_u_map[k];
1521         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1522                 dst->io_u_submit[k] += src->io_u_submit[k];
1523         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1524                 dst->io_u_complete[k] += src->io_u_complete[k];
1525         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1526                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1527         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1528                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1529         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1530                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1531
1532         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1533                 if (!dst->unified_rw_rep) {
1534                         dst->total_io_u[k] += src->total_io_u[k];
1535                         dst->short_io_u[k] += src->short_io_u[k];
1536                         dst->drop_io_u[k] += src->drop_io_u[k];
1537                 } else {
1538                         dst->total_io_u[0] += src->total_io_u[k];
1539                         dst->short_io_u[0] += src->short_io_u[k];
1540                         dst->drop_io_u[0] += src->drop_io_u[k];
1541                 }
1542         }
1543
1544         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1545                 int m;
1546
1547                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1548                         if (!dst->unified_rw_rep)
1549                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1550                         else
1551                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1552                 }
1553         }
1554
1555         dst->total_run_time += src->total_run_time;
1556         dst->total_submit += src->total_submit;
1557         dst->total_complete += src->total_complete;
1558 }
1559
1560 void init_group_run_stat(struct group_run_stats *gs)
1561 {
1562         int i;
1563         memset(gs, 0, sizeof(*gs));
1564
1565         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1566                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1567 }
1568
1569 void init_thread_stat(struct thread_stat *ts)
1570 {
1571         int j;
1572
1573         memset(ts, 0, sizeof(*ts));
1574
1575         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1576                 ts->lat_stat[j].min_val = -1UL;
1577                 ts->clat_stat[j].min_val = -1UL;
1578                 ts->slat_stat[j].min_val = -1UL;
1579                 ts->bw_stat[j].min_val = -1UL;
1580                 ts->iops_stat[j].min_val = -1UL;
1581         }
1582         ts->groupid = -1;
1583 }
1584
1585 void __show_run_stats(void)
1586 {
1587         struct group_run_stats *runstats, *rs;
1588         struct thread_data *td;
1589         struct thread_stat *threadstats, *ts;
1590         int i, j, k, nr_ts, last_ts, idx;
1591         int kb_base_warned = 0;
1592         int unit_base_warned = 0;
1593         struct json_object *root = NULL;
1594         struct json_array *array = NULL;
1595         struct buf_output output[FIO_OUTPUT_NR];
1596         struct flist_head **opt_lists;
1597
1598         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1599
1600         for (i = 0; i < groupid + 1; i++)
1601                 init_group_run_stat(&runstats[i]);
1602
1603         /*
1604          * find out how many threads stats we need. if group reporting isn't
1605          * enabled, it's one-per-td.
1606          */
1607         nr_ts = 0;
1608         last_ts = -1;
1609         for_each_td(td, i) {
1610                 if (!td->o.group_reporting) {
1611                         nr_ts++;
1612                         continue;
1613                 }
1614                 if (last_ts == td->groupid)
1615                         continue;
1616                 if (!td->o.stats)
1617                         continue;
1618
1619                 last_ts = td->groupid;
1620                 nr_ts++;
1621         }
1622
1623         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1624         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1625
1626         for (i = 0; i < nr_ts; i++) {
1627                 init_thread_stat(&threadstats[i]);
1628                 opt_lists[i] = NULL;
1629         }
1630
1631         j = 0;
1632         last_ts = -1;
1633         idx = 0;
1634         for_each_td(td, i) {
1635                 if (!td->o.stats)
1636                         continue;
1637                 if (idx && (!td->o.group_reporting ||
1638                     (td->o.group_reporting && last_ts != td->groupid))) {
1639                         idx = 0;
1640                         j++;
1641                 }
1642
1643                 last_ts = td->groupid;
1644
1645                 ts = &threadstats[j];
1646
1647                 ts->clat_percentiles = td->o.clat_percentiles;
1648                 ts->percentile_precision = td->o.percentile_precision;
1649                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1650                 opt_lists[j] = &td->opt_list;
1651
1652                 idx++;
1653                 ts->members++;
1654
1655                 if (ts->groupid == -1) {
1656                         /*
1657                          * These are per-group shared already
1658                          */
1659                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1660                         if (td->o.description)
1661                                 strncpy(ts->description, td->o.description,
1662                                                 FIO_JOBDESC_SIZE - 1);
1663                         else
1664                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1665
1666                         /*
1667                          * If multiple entries in this group, this is
1668                          * the first member.
1669                          */
1670                         ts->thread_number = td->thread_number;
1671                         ts->groupid = td->groupid;
1672
1673                         /*
1674                          * first pid in group, not very useful...
1675                          */
1676                         ts->pid = td->pid;
1677
1678                         ts->kb_base = td->o.kb_base;
1679                         ts->unit_base = td->o.unit_base;
1680                         ts->unified_rw_rep = td->o.unified_rw_rep;
1681                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1682                         log_info("fio: kb_base differs for jobs in group, using"
1683                                  " %u as the base\n", ts->kb_base);
1684                         kb_base_warned = 1;
1685                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1686                         log_info("fio: unit_base differs for jobs in group, using"
1687                                  " %u as the base\n", ts->unit_base);
1688                         unit_base_warned = 1;
1689                 }
1690
1691                 ts->continue_on_error = td->o.continue_on_error;
1692                 ts->total_err_count += td->total_err_count;
1693                 ts->first_error = td->first_error;
1694                 if (!ts->error) {
1695                         if (!td->error && td->o.continue_on_error &&
1696                             td->first_error) {
1697                                 ts->error = td->first_error;
1698                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1699                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1700                         } else  if (td->error) {
1701                                 ts->error = td->error;
1702                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1703                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1704                         }
1705                 }
1706
1707                 ts->latency_depth = td->latency_qd;
1708                 ts->latency_target = td->o.latency_target;
1709                 ts->latency_percentile = td->o.latency_percentile;
1710                 ts->latency_window = td->o.latency_window;
1711
1712                 ts->nr_block_infos = td->ts.nr_block_infos;
1713                 for (k = 0; k < ts->nr_block_infos; k++)
1714                         ts->block_infos[k] = td->ts.block_infos[k];
1715
1716                 sum_thread_stats(ts, &td->ts, idx == 1);
1717
1718                 if (td->o.ss_dur) {
1719                         ts->ss_state = td->ss.state;
1720                         ts->ss_dur = td->ss.dur;
1721                         ts->ss_head = td->ss.head;
1722                         ts->ss_bw_data = td->ss.bw_data;
1723                         ts->ss_iops_data = td->ss.iops_data;
1724                         ts->ss_limit.u.f = td->ss.limit;
1725                         ts->ss_slope.u.f = td->ss.slope;
1726                         ts->ss_deviation.u.f = td->ss.deviation;
1727                         ts->ss_criterion.u.f = td->ss.criterion;
1728                 }
1729                 else
1730                         ts->ss_dur = ts->ss_state = 0;
1731         }
1732
1733         for (i = 0; i < nr_ts; i++) {
1734                 unsigned long long bw;
1735
1736                 ts = &threadstats[i];
1737                 if (ts->groupid == -1)
1738                         continue;
1739                 rs = &runstats[ts->groupid];
1740                 rs->kb_base = ts->kb_base;
1741                 rs->unit_base = ts->unit_base;
1742                 rs->unified_rw_rep += ts->unified_rw_rep;
1743
1744                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1745                         if (!ts->runtime[j])
1746                                 continue;
1747                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1748                                 rs->min_run[j] = ts->runtime[j];
1749                         if (ts->runtime[j] > rs->max_run[j])
1750                                 rs->max_run[j] = ts->runtime[j];
1751
1752                         bw = 0;
1753                         if (ts->runtime[j])
1754                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1755                         if (bw < rs->min_bw[j])
1756                                 rs->min_bw[j] = bw;
1757                         if (bw > rs->max_bw[j])
1758                                 rs->max_bw[j] = bw;
1759
1760                         rs->iobytes[j] += ts->io_bytes[j];
1761                 }
1762         }
1763
1764         for (i = 0; i < groupid + 1; i++) {
1765                 int ddir;
1766
1767                 rs = &runstats[i];
1768
1769                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1770                         if (rs->max_run[ddir])
1771                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1772                                                 rs->max_run[ddir];
1773                 }
1774         }
1775
1776         for (i = 0; i < FIO_OUTPUT_NR; i++)
1777                 buf_output_init(&output[i]);
1778
1779         /*
1780          * don't overwrite last signal output
1781          */
1782         if (output_format & FIO_OUTPUT_NORMAL)
1783                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1784         if (output_format & FIO_OUTPUT_JSON) {
1785                 struct thread_data *global;
1786                 char time_buf[32];
1787                 struct timeval now;
1788                 unsigned long long ms_since_epoch;
1789
1790                 gettimeofday(&now, NULL);
1791                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1792                                  (unsigned long long)(now.tv_usec) / 1000;
1793
1794                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1795                                 sizeof(time_buf));
1796                 if (time_buf[strlen(time_buf) - 1] == '\n')
1797                         time_buf[strlen(time_buf) - 1] = '\0';
1798
1799                 root = json_create_object();
1800                 json_object_add_value_string(root, "fio version", fio_version_string);
1801                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1802                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1803                 json_object_add_value_string(root, "time", time_buf);
1804                 global = get_global_options();
1805                 json_add_job_opts(root, "global options", &global->opt_list, false);
1806                 array = json_create_array();
1807                 json_object_add_value_array(root, "jobs", array);
1808         }
1809
1810         if (is_backend)
1811                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1812
1813         for (i = 0; i < nr_ts; i++) {
1814                 ts = &threadstats[i];
1815                 rs = &runstats[ts->groupid];
1816
1817                 if (is_backend) {
1818                         fio_server_send_job_options(opt_lists[i], i);
1819                         fio_server_send_ts(ts, rs);
1820                 } else {
1821                         if (output_format & FIO_OUTPUT_TERSE)
1822                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1823                         if (output_format & FIO_OUTPUT_JSON) {
1824                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1825                                 json_array_add_value_object(array, tmp);
1826                         }
1827                         if (output_format & FIO_OUTPUT_NORMAL)
1828                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1829                 }
1830         }
1831         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1832                 /* disk util stats, if any */
1833                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1834
1835                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1836
1837                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1838                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1839                 json_free_object(root);
1840         }
1841
1842         for (i = 0; i < groupid + 1; i++) {
1843                 rs = &runstats[i];
1844
1845                 rs->groupid = i;
1846                 if (is_backend)
1847                         fio_server_send_gs(rs);
1848                 else if (output_format & FIO_OUTPUT_NORMAL)
1849                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1850         }
1851
1852         if (is_backend)
1853                 fio_server_send_du();
1854         else if (output_format & FIO_OUTPUT_NORMAL) {
1855                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1856                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1857         }
1858
1859         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1860                 struct buf_output *out = &output[i];
1861
1862                 log_info_buf(out->buf, out->buflen);
1863                 buf_output_free(out);
1864         }
1865
1866         log_info_flush();
1867         free(runstats);
1868         free(threadstats);
1869         free(opt_lists);
1870 }
1871
1872 void show_run_stats(void)
1873 {
1874         fio_mutex_down(stat_mutex);
1875         __show_run_stats();
1876         fio_mutex_up(stat_mutex);
1877 }
1878
1879 void __show_running_run_stats(void)
1880 {
1881         struct thread_data *td;
1882         unsigned long long *rt;
1883         struct timespec ts;
1884         int i;
1885
1886         fio_mutex_down(stat_mutex);
1887
1888         rt = malloc(thread_number * sizeof(unsigned long long));
1889         fio_gettime(&ts, NULL);
1890
1891         for_each_td(td, i) {
1892                 td->update_rusage = 1;
1893                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1894                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1895                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1896                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1897
1898                 rt[i] = mtime_since(&td->start, &ts);
1899                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1900                         td->ts.runtime[DDIR_READ] += rt[i];
1901                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1902                         td->ts.runtime[DDIR_WRITE] += rt[i];
1903                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1904                         td->ts.runtime[DDIR_TRIM] += rt[i];
1905         }
1906
1907         for_each_td(td, i) {
1908                 if (td->runstate >= TD_EXITED)
1909                         continue;
1910                 if (td->rusage_sem) {
1911                         td->update_rusage = 1;
1912                         fio_mutex_down(td->rusage_sem);
1913                 }
1914                 td->update_rusage = 0;
1915         }
1916
1917         __show_run_stats();
1918
1919         for_each_td(td, i) {
1920                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1921                         td->ts.runtime[DDIR_READ] -= rt[i];
1922                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1923                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1924                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1925                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1926         }
1927
1928         free(rt);
1929         fio_mutex_up(stat_mutex);
1930 }
1931
1932 static int status_interval_init;
1933 static struct timespec status_time;
1934 static int status_file_disabled;
1935
1936 #define FIO_STATUS_FILE         "fio-dump-status"
1937
1938 static int check_status_file(void)
1939 {
1940         struct stat sb;
1941         const char *temp_dir;
1942         char fio_status_file_path[PATH_MAX];
1943
1944         if (status_file_disabled)
1945                 return 0;
1946
1947         temp_dir = getenv("TMPDIR");
1948         if (temp_dir == NULL) {
1949                 temp_dir = getenv("TEMP");
1950                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1951                         temp_dir = NULL;
1952         }
1953         if (temp_dir == NULL)
1954                 temp_dir = "/tmp";
1955
1956         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1957
1958         if (stat(fio_status_file_path, &sb))
1959                 return 0;
1960
1961         if (unlink(fio_status_file_path) < 0) {
1962                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1963                                                         strerror(errno));
1964                 log_err("fio: disabling status file updates\n");
1965                 status_file_disabled = 1;
1966         }
1967
1968         return 1;
1969 }
1970
1971 void check_for_running_stats(void)
1972 {
1973         if (status_interval) {
1974                 if (!status_interval_init) {
1975                         fio_gettime(&status_time, NULL);
1976                         status_interval_init = 1;
1977                 } else if (mtime_since_now(&status_time) >= status_interval) {
1978                         show_running_run_stats();
1979                         fio_gettime(&status_time, NULL);
1980                         return;
1981                 }
1982         }
1983         if (check_status_file()) {
1984                 show_running_run_stats();
1985                 return;
1986         }
1987 }
1988
1989 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
1990 {
1991         double val = data;
1992         double delta;
1993
1994         if (data > is->max_val)
1995                 is->max_val = data;
1996         if (data < is->min_val)
1997                 is->min_val = data;
1998
1999         delta = val - is->mean.u.f;
2000         if (delta) {
2001                 is->mean.u.f += delta / (is->samples + 1.0);
2002                 is->S.u.f += delta * (val - is->mean.u.f);
2003         }
2004
2005         is->samples++;
2006 }
2007
2008 /*
2009  * Return a struct io_logs, which is added to the tail of the log
2010  * list for 'iolog'.
2011  */
2012 static struct io_logs *get_new_log(struct io_log *iolog)
2013 {
2014         size_t new_size, new_samples;
2015         struct io_logs *cur_log;
2016
2017         /*
2018          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2019          * forever
2020          */
2021         if (!iolog->cur_log_max)
2022                 new_samples = DEF_LOG_ENTRIES;
2023         else {
2024                 new_samples = iolog->cur_log_max * 2;
2025                 if (new_samples > MAX_LOG_ENTRIES)
2026                         new_samples = MAX_LOG_ENTRIES;
2027         }
2028
2029         new_size = new_samples * log_entry_sz(iolog);
2030
2031         cur_log = smalloc(sizeof(*cur_log));
2032         if (cur_log) {
2033                 INIT_FLIST_HEAD(&cur_log->list);
2034                 cur_log->log = malloc(new_size);
2035                 if (cur_log->log) {
2036                         cur_log->nr_samples = 0;
2037                         cur_log->max_samples = new_samples;
2038                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2039                         iolog->cur_log_max = new_samples;
2040                         return cur_log;
2041                 }
2042                 sfree(cur_log);
2043         }
2044
2045         return NULL;
2046 }
2047
2048 /*
2049  * Add and return a new log chunk, or return current log if big enough
2050  */
2051 static struct io_logs *regrow_log(struct io_log *iolog)
2052 {
2053         struct io_logs *cur_log;
2054         int i;
2055
2056         if (!iolog || iolog->disabled)
2057                 goto disable;
2058
2059         cur_log = iolog_cur_log(iolog);
2060         if (!cur_log) {
2061                 cur_log = get_new_log(iolog);
2062                 if (!cur_log)
2063                         return NULL;
2064         }
2065
2066         if (cur_log->nr_samples < cur_log->max_samples)
2067                 return cur_log;
2068
2069         /*
2070          * No room for a new sample. If we're compressing on the fly, flush
2071          * out the current chunk
2072          */
2073         if (iolog->log_gz) {
2074                 if (iolog_cur_flush(iolog, cur_log)) {
2075                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2076                         return NULL;
2077                 }
2078         }
2079
2080         /*
2081          * Get a new log array, and add to our list
2082          */
2083         cur_log = get_new_log(iolog);
2084         if (!cur_log) {
2085                 log_err("fio: failed extending iolog! Will stop logging.\n");
2086                 return NULL;
2087         }
2088
2089         if (!iolog->pending || !iolog->pending->nr_samples)
2090                 return cur_log;
2091
2092         /*
2093          * Flush pending items to new log
2094          */
2095         for (i = 0; i < iolog->pending->nr_samples; i++) {
2096                 struct io_sample *src, *dst;
2097
2098                 src = get_sample(iolog, iolog->pending, i);
2099                 dst = get_sample(iolog, cur_log, i);
2100                 memcpy(dst, src, log_entry_sz(iolog));
2101         }
2102         cur_log->nr_samples = iolog->pending->nr_samples;
2103
2104         iolog->pending->nr_samples = 0;
2105         return cur_log;
2106 disable:
2107         if (iolog)
2108                 iolog->disabled = true;
2109         return NULL;
2110 }
2111
2112 void regrow_logs(struct thread_data *td)
2113 {
2114         regrow_log(td->slat_log);
2115         regrow_log(td->clat_log);
2116         regrow_log(td->clat_hist_log);
2117         regrow_log(td->lat_log);
2118         regrow_log(td->bw_log);
2119         regrow_log(td->iops_log);
2120         td->flags &= ~TD_F_REGROW_LOGS;
2121 }
2122
2123 static struct io_logs *get_cur_log(struct io_log *iolog)
2124 {
2125         struct io_logs *cur_log;
2126
2127         cur_log = iolog_cur_log(iolog);
2128         if (!cur_log) {
2129                 cur_log = get_new_log(iolog);
2130                 if (!cur_log)
2131                         return NULL;
2132         }
2133
2134         if (cur_log->nr_samples < cur_log->max_samples)
2135                 return cur_log;
2136
2137         /*
2138          * Out of space. If we're in IO offload mode, or we're not doing
2139          * per unit logging (hence logging happens outside of the IO thread
2140          * as well), add a new log chunk inline. If we're doing inline
2141          * submissions, flag 'td' as needing a log regrow and we'll take
2142          * care of it on the submission side.
2143          */
2144         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2145             !per_unit_log(iolog))
2146                 return regrow_log(iolog);
2147
2148         iolog->td->flags |= TD_F_REGROW_LOGS;
2149         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2150         return iolog->pending;
2151 }
2152
2153 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2154                              enum fio_ddir ddir, unsigned int bs,
2155                              unsigned long t, uint64_t offset)
2156 {
2157         struct io_logs *cur_log;
2158
2159         if (iolog->disabled)
2160                 return;
2161         if (flist_empty(&iolog->io_logs))
2162                 iolog->avg_last = t;
2163
2164         cur_log = get_cur_log(iolog);
2165         if (cur_log) {
2166                 struct io_sample *s;
2167
2168                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2169
2170                 s->data = data;
2171                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2172                 io_sample_set_ddir(iolog, s, ddir);
2173                 s->bs = bs;
2174
2175                 if (iolog->log_offset) {
2176                         struct io_sample_offset *so = (void *) s;
2177
2178                         so->offset = offset;
2179                 }
2180
2181                 cur_log->nr_samples++;
2182                 return;
2183         }
2184
2185         iolog->disabled = true;
2186 }
2187
2188 static inline void reset_io_stat(struct io_stat *ios)
2189 {
2190         ios->max_val = ios->min_val = ios->samples = 0;
2191         ios->mean.u.f = ios->S.u.f = 0;
2192 }
2193
2194 void reset_io_stats(struct thread_data *td)
2195 {
2196         struct thread_stat *ts = &td->ts;
2197         int i, j;
2198
2199         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2200                 reset_io_stat(&ts->clat_stat[i]);
2201                 reset_io_stat(&ts->slat_stat[i]);
2202                 reset_io_stat(&ts->lat_stat[i]);
2203                 reset_io_stat(&ts->bw_stat[i]);
2204                 reset_io_stat(&ts->iops_stat[i]);
2205
2206                 ts->io_bytes[i] = 0;
2207                 ts->runtime[i] = 0;
2208                 ts->total_io_u[i] = 0;
2209                 ts->short_io_u[i] = 0;
2210                 ts->drop_io_u[i] = 0;
2211
2212                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2213                         ts->io_u_plat[i][j] = 0;
2214         }
2215
2216         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2217                 ts->io_u_map[i] = 0;
2218                 ts->io_u_submit[i] = 0;
2219                 ts->io_u_complete[i] = 0;
2220         }
2221
2222         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2223                 ts->io_u_lat_n[i] = 0;
2224         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2225                 ts->io_u_lat_u[i] = 0;
2226         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2227                 ts->io_u_lat_m[i] = 0;
2228
2229         ts->total_submit = 0;
2230         ts->total_complete = 0;
2231 }
2232
2233 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2234                               unsigned long elapsed, bool log_max)
2235 {
2236         /*
2237          * Note an entry in the log. Use the mean from the logged samples,
2238          * making sure to properly round up. Only write a log entry if we
2239          * had actual samples done.
2240          */
2241         if (iolog->avg_window[ddir].samples) {
2242                 union io_sample_data data;
2243
2244                 if (log_max)
2245                         data.val = iolog->avg_window[ddir].max_val;
2246                 else
2247                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2248
2249                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2250         }
2251
2252         reset_io_stat(&iolog->avg_window[ddir]);
2253 }
2254
2255 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2256                              bool log_max)
2257 {
2258         int ddir;
2259
2260         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2261                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2262 }
2263
2264 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2265                            union io_sample_data data, enum fio_ddir ddir,
2266                            unsigned int bs, uint64_t offset)
2267 {
2268         unsigned long elapsed, this_window;
2269
2270         if (!ddir_rw(ddir))
2271                 return 0;
2272
2273         elapsed = mtime_since_now(&td->epoch);
2274
2275         /*
2276          * If no time averaging, just add the log sample.
2277          */
2278         if (!iolog->avg_msec) {
2279                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2280                 return 0;
2281         }
2282
2283         /*
2284          * Add the sample. If the time period has passed, then
2285          * add that entry to the log and clear.
2286          */
2287         add_stat_sample(&iolog->avg_window[ddir], data.val);
2288
2289         /*
2290          * If period hasn't passed, adding the above sample is all we
2291          * need to do.
2292          */
2293         this_window = elapsed - iolog->avg_last;
2294         if (elapsed < iolog->avg_last)
2295                 return iolog->avg_last - elapsed;
2296         else if (this_window < iolog->avg_msec) {
2297                 int diff = iolog->avg_msec - this_window;
2298
2299                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2300                         return diff;
2301         }
2302
2303         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2304
2305         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2306         return iolog->avg_msec;
2307 }
2308
2309 void finalize_logs(struct thread_data *td, bool unit_logs)
2310 {
2311         unsigned long elapsed;
2312
2313         elapsed = mtime_since_now(&td->epoch);
2314
2315         if (td->clat_log && unit_logs)
2316                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2317         if (td->slat_log && unit_logs)
2318                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2319         if (td->lat_log && unit_logs)
2320                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2321         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2322                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2323         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2324                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2325 }
2326
2327 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2328 {
2329         struct io_log *iolog;
2330
2331         if (!ddir_rw(ddir))
2332                 return;
2333
2334         iolog = agg_io_log[ddir];
2335         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2336 }
2337
2338 static void add_clat_percentile_sample(struct thread_stat *ts,
2339                                 unsigned long long nsec, enum fio_ddir ddir)
2340 {
2341         unsigned int idx = plat_val_to_idx(nsec);
2342         assert(idx < FIO_IO_U_PLAT_NR);
2343
2344         ts->io_u_plat[ddir][idx]++;
2345 }
2346
2347 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2348                      unsigned long long nsec, unsigned int bs, uint64_t offset)
2349 {
2350         unsigned long elapsed, this_window;
2351         struct thread_stat *ts = &td->ts;
2352         struct io_log *iolog = td->clat_hist_log;
2353
2354         td_io_u_lock(td);
2355
2356         add_stat_sample(&ts->clat_stat[ddir], nsec);
2357
2358         if (td->clat_log)
2359                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2360                                offset);
2361
2362         if (ts->clat_percentiles)
2363                 add_clat_percentile_sample(ts, nsec, ddir);
2364
2365         if (iolog && iolog->hist_msec) {
2366                 struct io_hist *hw = &iolog->hist_window[ddir];
2367
2368                 hw->samples++;
2369                 elapsed = mtime_since_now(&td->epoch);
2370                 if (!hw->hist_last)
2371                         hw->hist_last = elapsed;
2372                 this_window = elapsed - hw->hist_last;
2373
2374                 if (this_window >= iolog->hist_msec) {
2375                         unsigned int *io_u_plat;
2376                         struct io_u_plat_entry *dst;
2377
2378                         /*
2379                          * Make a byte-for-byte copy of the latency histogram
2380                          * stored in td->ts.io_u_plat[ddir], recording it in a
2381                          * log sample. Note that the matching call to free() is
2382                          * located in iolog.c after printing this sample to the
2383                          * log file.
2384                          */
2385                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2386                         dst = malloc(sizeof(struct io_u_plat_entry));
2387                         memcpy(&(dst->io_u_plat), io_u_plat,
2388                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2389                         flist_add(&dst->list, &hw->list);
2390                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2391                                                 elapsed, offset);
2392
2393                         /*
2394                          * Update the last time we recorded as being now, minus
2395                          * any drift in time we encountered before actually
2396                          * making the record.
2397                          */
2398                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2399                         hw->samples = 0;
2400                 }
2401         }
2402
2403         td_io_u_unlock(td);
2404 }
2405
2406 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2407                      unsigned long usec, unsigned int bs, uint64_t offset)
2408 {
2409         struct thread_stat *ts = &td->ts;
2410
2411         if (!ddir_rw(ddir))
2412                 return;
2413
2414         td_io_u_lock(td);
2415
2416         add_stat_sample(&ts->slat_stat[ddir], usec);
2417
2418         if (td->slat_log)
2419                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2420
2421         td_io_u_unlock(td);
2422 }
2423
2424 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2425                     unsigned long long nsec, unsigned int bs, uint64_t offset)
2426 {
2427         struct thread_stat *ts = &td->ts;
2428
2429         if (!ddir_rw(ddir))
2430                 return;
2431
2432         td_io_u_lock(td);
2433
2434         add_stat_sample(&ts->lat_stat[ddir], nsec);
2435
2436         if (td->lat_log)
2437                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2438                                offset);
2439
2440         td_io_u_unlock(td);
2441 }
2442
2443 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2444                    unsigned int bytes, unsigned long long spent)
2445 {
2446         struct thread_stat *ts = &td->ts;
2447         unsigned long rate;
2448
2449         if (spent)
2450                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2451         else
2452                 rate = 0;
2453
2454         td_io_u_lock(td);
2455
2456         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2457
2458         if (td->bw_log)
2459                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2460                                bytes, io_u->offset);
2461
2462         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2463         td_io_u_unlock(td);
2464 }
2465
2466 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2467                          struct timespec *t, unsigned int avg_time,
2468                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2469                          struct io_stat *stat, struct io_log *log,
2470                          bool is_kb)
2471 {
2472         unsigned long spent, rate;
2473         enum fio_ddir ddir;
2474         unsigned int next, next_log;
2475
2476         next_log = avg_time;
2477
2478         spent = mtime_since(parent_tv, t);
2479         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2480                 return avg_time - spent;
2481
2482         td_io_u_lock(td);
2483
2484         /*
2485          * Compute both read and write rates for the interval.
2486          */
2487         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2488                 uint64_t delta;
2489
2490                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2491                 if (!delta)
2492                         continue; /* No entries for interval */
2493
2494                 if (spent) {
2495                         if (is_kb)
2496                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2497                         else
2498                                 rate = (delta * 1000) / spent;
2499                 } else
2500                         rate = 0;
2501
2502                 add_stat_sample(&stat[ddir], rate);
2503
2504                 if (log) {
2505                         unsigned int bs = 0;
2506
2507                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2508                                 bs = td->o.min_bs[ddir];
2509
2510                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2511                         next_log = min(next_log, next);
2512                 }
2513
2514                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2515         }
2516
2517         timespec_add_msec(parent_tv, avg_time);
2518
2519         td_io_u_unlock(td);
2520
2521         if (spent <= avg_time)
2522                 next = avg_time;
2523         else
2524                 next = avg_time - (1 + spent - avg_time);
2525
2526         return min(next, next_log);
2527 }
2528
2529 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2530 {
2531         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2532                                 td->this_io_bytes, td->stat_io_bytes,
2533                                 td->ts.bw_stat, td->bw_log, true);
2534 }
2535
2536 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2537                      unsigned int bytes)
2538 {
2539         struct thread_stat *ts = &td->ts;
2540
2541         td_io_u_lock(td);
2542
2543         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2544
2545         if (td->iops_log)
2546                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2547                                bytes, io_u->offset);
2548
2549         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2550         td_io_u_unlock(td);
2551 }
2552
2553 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2554 {
2555         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2556                                 td->this_io_blocks, td->stat_io_blocks,
2557                                 td->ts.iops_stat, td->iops_log, false);
2558 }
2559
2560 /*
2561  * Returns msecs to next event
2562  */
2563 int calc_log_samples(void)
2564 {
2565         struct thread_data *td;
2566         unsigned int next = ~0U, tmp;
2567         struct timespec now;
2568         int i;
2569
2570         fio_gettime(&now, NULL);
2571
2572         for_each_td(td, i) {
2573                 if (!td->o.stats)
2574                         continue;
2575                 if (in_ramp_time(td) ||
2576                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2577                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2578                         continue;
2579                 }
2580                 if (!td->bw_log ||
2581                         (td->bw_log && !per_unit_log(td->bw_log))) {
2582                         tmp = add_bw_samples(td, &now);
2583                         if (tmp < next)
2584                                 next = tmp;
2585                 }
2586                 if (!td->iops_log ||
2587                         (td->iops_log && !per_unit_log(td->iops_log))) {
2588                         tmp = add_iops_samples(td, &now);
2589                         if (tmp < next)
2590                                 next = tmp;
2591                 }
2592         }
2593
2594         return next == ~0U ? 0 : next;
2595 }
2596
2597 void stat_init(void)
2598 {
2599         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2600 }
2601
2602 void stat_exit(void)
2603 {
2604         /*
2605          * When we have the mutex, we know out-of-band access to it
2606          * have ended.
2607          */
2608         fio_mutex_down(stat_mutex);
2609         fio_mutex_remove(stat_mutex);
2610 }
2611
2612 /*
2613  * Called from signal handler. Wake up status thread.
2614  */
2615 void show_running_run_stats(void)
2616 {
2617         helper_do_stat();
2618 }
2619
2620 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2621 {
2622         /* Ignore io_u's which span multiple blocks--they will just get
2623          * inaccurate counts. */
2624         int idx = (io_u->offset - io_u->file->file_offset)
2625                         / td->o.bs[DDIR_TRIM];
2626         uint32_t *info = &td->ts.block_infos[idx];
2627         assert(idx < td->ts.nr_block_infos);
2628         return info;
2629 }