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                                   bool is_clat, struct buf_output *out)
 204 {
 205         unsigned int divisor, len, i, j = 0;
 206         unsigned long long minv, maxv;
 207         unsigned long long *ovals;
 208         int is_last, per_line, scale_down, time_width;
 209         const char *pre = is_clat ? "clat" : " lat";
 210         char fmt[32];
 211
 212         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 213         if (!len)
 214                 goto out;
 215
 216         /*
 217          * We default to nsecs, but if the value range is such that we
 218          * should scale down to usecs or msecs, do that.
 219          */
 220         if (minv > 2000000 && maxv > 99999999ULL) {
 221                 scale_down = 2;
 222                 divisor = 1000000;
 223                 log_buf(out, "    %s percentiles (msec):\n     |", pre);
 224         } else if (minv > 2000 && maxv > 99999) {
 225                 scale_down = 1;
 226                 divisor = 1000;
 227                 log_buf(out, "    %s percentiles (usec):\n     |", pre);
 228         } else {
 229                 scale_down = 0;
 230                 divisor = 1;
 231                 log_buf(out, "    %s percentiles (nsec):\n     |", pre);
 232         }
 233
 234
 235         time_width = max(5, (int) (log10(maxv / divisor) + 1));
 236         snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
 237                         precision, time_width);
 238         /* fmt will be something like " %5.2fth=[%4llu]%c" */
 239         per_line = (80 - 7) / (precision + 10 + time_width);
 240
 241         for (j = 0; j < len; j++) {
 242                 /* for formatting */
 243                 if (j != 0 && (j % per_line) == 0)
 244                         log_buf(out, "     |");
 245
 246                 /* end of the list */
 247                 is_last = (j == len - 1);
 248
 249                 for (i = 0; i < scale_down; i++)
 250                         ovals[j] = (ovals[j] + 999) / 1000;
 251
 252                 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
 253
 254                 if (is_last)
 255                         break;
 256
 257                 if ((j % per_line) == per_line - 1)     /* for formatting */
 258                         log_buf(out, "\n");
 259         }
 260
 261 out:
 262         if (ovals)
 263                 free(ovals);
 264 }
 265
 266 bool calc_lat(struct io_stat *is, unsigned long long *min,
 267               unsigned long long *max, double *mean, double *dev)
 268 {
 269         double n = (double) is->samples;
 270
 271         if (n == 0)
 272                 return false;
 273
 274         *min = is->min_val;
 275         *max = is->max_val;
 276         *mean = is->mean.u.f;
 277
 278         if (n > 1.0)
 279                 *dev = sqrt(is->S.u.f / (n - 1.0));
 280         else
 281                 *dev = 0;
 282
 283         return true;
 284 }
 285
 286 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 287 {
 288         char *io, *agg, *min, *max;
 289         char *ioalt, *aggalt, *minalt, *maxalt;
 290         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 291         int i;
 292
 293         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 294
 295         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 296                 const int i2p = is_power_of_2(rs->kb_base);
 297
 298                 if (!rs->max_run[i])
 299                         continue;
 300
 301                 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
 302                 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
 303                 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
 304                 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
 305                 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
 306                 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
 307                 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
 308                 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
 309                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 310                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 311                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 312                                 (unsigned long long) rs->min_run[i],
 313                                 (unsigned long long) rs->max_run[i]);
 314
 315                 free(io);
 316                 free(agg);
 317                 free(min);
 318                 free(max);
 319                 free(ioalt);
 320                 free(aggalt);
 321                 free(minalt);
 322                 free(maxalt);
 323         }
 324 }
 325
 326 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 327 {
 328         int i;
 329
 330         /*
 331          * Do depth distribution calculations
 332          */
 333         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 334                 if (total) {
 335                         io_u_dist[i] = (double) map[i] / (double) total;
 336                         io_u_dist[i] *= 100.0;
 337                         if (io_u_dist[i] < 0.1 && map[i])
 338                                 io_u_dist[i] = 0.1;
 339                 } else
 340                         io_u_dist[i] = 0.0;
 341         }
 342 }
 343
 344 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 345                           unsigned int *src, int nr)
 346 {
 347         unsigned long total = ddir_rw_sum(ts->total_io_u);
 348         int i;
 349
 350         /*
 351          * Do latency distribution calculations
 352          */
 353         for (i = 0; i < nr; i++) {
 354                 if (total) {
 355                         dst[i] = (double) src[i] / (double) total;
 356                         dst[i] *= 100.0;
 357                         if (dst[i] < 0.01 && src[i])
 358                                 dst[i] = 0.01;
 359                 } else
 360                         dst[i] = 0.0;
 361         }
 362 }
 363
 364 /*
 365  * To keep the terse format unaltered, add all of the ns latency
 366  * buckets to the first us latency bucket
 367  */
 368 void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
 369 {
 370         unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
 371         int i;
 372
 373         stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 374
 375         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
 376                 ntotal += ts->io_u_lat_n[i];
 377
 378         io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
 379 }
 380
 381 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
 382 {
 383         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
 384 }
 385
 386 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 387 {
 388         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 389 }
 390
 391 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 392 {
 393         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 394 }
 395
 396 static void display_lat(const char *name, unsigned long long min,
 397                         unsigned long long max, double mean, double dev,
 398                         struct buf_output *out)
 399 {
 400         const char *base = "(nsec)";
 401         char *minp, *maxp;
 402
 403         if (nsec_to_msec(&min, &max, &mean, &dev))
 404                 base = "(msec)";
 405         else if (nsec_to_usec(&min, &max, &mean, &dev))
 406                 base = "(usec)";
 407
 408         minp = num2str(min, 6, 1, 0, N2S_NONE);
 409         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 410
 411         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 412                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 413
 414         free(minp);
 415         free(maxp);
 416 }
 417
 418 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 419                              int ddir, struct buf_output *out)
 420 {
 421         const char *str[] = { " read", "write", " trim" };
 422         unsigned long runt;
 423         unsigned long long min, max, bw, iops;
 424         double mean, dev;
 425         char *io_p, *bw_p, *bw_p_alt, *iops_p;
 426         int i2p;
 427
 428         assert(ddir_rw(ddir));
 429
 430         if (!ts->runtime[ddir])
 431                 return;
 432
 433         i2p = is_power_of_2(rs->kb_base);
 434         runt = ts->runtime[ddir];
 435
 436         bw = (1000 * ts->io_bytes[ddir]) / runt;
 437         io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
 438         bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
 439         bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
 440
 441         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 442         iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
 443
 444         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
 445                         rs->unified_rw_rep ? "mixed" : str[ddir],
 446                         iops_p, bw_p, bw_p_alt, io_p,
 447                         (unsigned long long) ts->runtime[ddir]);
 448
 449         free(io_p);
 450         free(bw_p);
 451         free(bw_p_alt);
 452         free(iops_p);
 453
 454         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 455                 display_lat("slat", min, max, mean, dev, out);
 456         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 457                 display_lat("clat", min, max, mean, dev, out);
 458         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 459                 display_lat(" lat", min, max, mean, dev, out);
 460
 461         if (ts->clat_percentiles || ts->lat_percentiles) {
 462                 show_clat_percentiles(ts->io_u_plat[ddir],
 463                                         ts->clat_stat[ddir].samples,
 464                                         ts->percentile_list,
 465                                         ts->percentile_precision,
 466                                         ts->clat_percentiles, out);
 467         }
 468         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 469                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 470                 const char *bw_str;
 471
 472                 if ((rs->unit_base == 1) && i2p)
 473                         bw_str = "Kibit";
 474                 else if (rs->unit_base == 1)
 475                         bw_str = "kbit";
 476                 else if (i2p)
 477                         bw_str = "KiB";
 478                 else
 479                         bw_str = "kB";
 480
 481                 if (rs->agg[ddir]) {
 482                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 483                         if (p_of_agg > 100.0)
 484                                 p_of_agg = 100.0;
 485                 }
 486
 487                 if (rs->unit_base == 1) {
 488                         min *= 8.0;
 489                         max *= 8.0;
 490                         mean *= 8.0;
 491                         dev *= 8.0;
 492                 }
 493
 494                 if (mean > fkb_base * fkb_base) {
 495                         min /= fkb_base;
 496                         max /= fkb_base;
 497                         mean /= fkb_base;
 498                         dev /= fkb_base;
 499                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 500                 }
 501
 502                 log_buf(out, "   bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
 503                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 504                         bw_str, min, max, p_of_agg, mean, dev,
 505                         (&ts->bw_stat[ddir])->samples);
 506         }
 507         if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
 508                 log_buf(out, "   iops        : min=%5llu, max=%5llu, "
 509                         "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
 510                         min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
 511         }
 512 }
 513
 514 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 515                     const char *msg, struct buf_output *out)
 516 {
 517         int new_line = 1, i, line = 0, shown = 0;
 518
 519         for (i = 0; i < nr; i++) {
 520                 if (io_u_lat[i] <= 0.0)
 521                         continue;
 522                 shown = 1;
 523                 if (new_line) {
 524                         if (line)
 525                                 log_buf(out, "\n");
 526                         log_buf(out, "  lat (%s)   : ", msg);
 527                         new_line = 0;
 528                         line = 0;
 529                 }
 530                 if (line)
 531                         log_buf(out, ", ");
 532                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 533                 line++;
 534                 if (line == 5)
 535                         new_line = 1;
 536         }
 537
 538         if (shown)
 539                 log_buf(out, "\n");
 540
 541         return shown;
 542 }
 543
 544 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
 545 {
 546         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 547                                  "250=", "500=", "750=", "1000=", };
 548
 549         show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
 550 }
 551
 552 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 553 {
 554         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 555                                  "250=", "500=", "750=", "1000=", };
 556
 557         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 558 }
 559
 560 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 561 {
 562         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 563                                  "250=", "500=", "750=", "1000=", "2000=",
 564                                  ">=2000=", };
 565
 566         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 567 }
 568
 569 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 570 {
 571         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
 572         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 573         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 574
 575         stat_calc_lat_n(ts, io_u_lat_n);
 576         stat_calc_lat_u(ts, io_u_lat_u);
 577         stat_calc_lat_m(ts, io_u_lat_m);
 578
 579         show_lat_n(io_u_lat_n, out);
 580         show_lat_u(io_u_lat_u, out);
 581         show_lat_m(io_u_lat_m, out);
 582 }
 583
 584 static int block_state_category(int block_state)
 585 {
 586         switch (block_state) {
 587         case BLOCK_STATE_UNINIT:
 588                 return 0;
 589         case BLOCK_STATE_TRIMMED:
 590         case BLOCK_STATE_WRITTEN:
 591                 return 1;
 592         case BLOCK_STATE_WRITE_FAILURE:
 593         case BLOCK_STATE_TRIM_FAILURE:
 594                 return 2;
 595         default:
 596                 /* Silence compile warning on some BSDs and have a return */
 597                 assert(0);
 598                 return -1;
 599         }
 600 }
 601
 602 static int compare_block_infos(const void *bs1, const void *bs2)
 603 {
 604         uint32_t block1 = *(uint32_t *)bs1;
 605         uint32_t block2 = *(uint32_t *)bs2;
 606         int state1 = BLOCK_INFO_STATE(block1);
 607         int state2 = BLOCK_INFO_STATE(block2);
 608         int bscat1 = block_state_category(state1);
 609         int bscat2 = block_state_category(state2);
 610         int cycles1 = BLOCK_INFO_TRIMS(block1);
 611         int cycles2 = BLOCK_INFO_TRIMS(block2);
 612
 613         if (bscat1 < bscat2)
 614                 return -1;
 615         if (bscat1 > bscat2)
 616                 return 1;
 617
 618         if (cycles1 < cycles2)
 619                 return -1;
 620         if (cycles1 > cycles2)
 621                 return 1;
 622
 623         if (state1 < state2)
 624                 return -1;
 625         if (state1 > state2)
 626                 return 1;
 627
 628         assert(block1 == block2);
 629         return 0;
 630 }
 631
 632 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 633                                   fio_fp64_t *plist, unsigned int **percentiles,
 634                                   unsigned int *types)
 635 {
 636         int len = 0;
 637         int i, nr_uninit;
 638
 639         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 640
 641         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 642                 len++;
 643
 644         if (!len)
 645                 return 0;
 646
 647         /*
 648          * Sort the percentile list. Note that it may already be sorted if
 649          * we are using the default values, but since it's a short list this
 650          * isn't a worry. Also note that this does not work for NaN values.
 651          */
 652         if (len > 1)
 653                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 654
 655         nr_uninit = 0;
 656         /* Start only after the uninit entries end */
 657         for (nr_uninit = 0;
 658              nr_uninit < nr_block_infos
 659                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 660              nr_uninit ++)
 661                 ;
 662
 663         if (nr_uninit == nr_block_infos)
 664                 return 0;
 665
 666         *percentiles = calloc(len, sizeof(**percentiles));
 667
 668         for (i = 0; i < len; i++) {
 669                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 670                                 + nr_uninit;
 671                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 672         }
 673
 674         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 675         for (i = 0; i < nr_block_infos; i++)
 676                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 677
 678         return len;
 679 }
 680
 681 static const char *block_state_names[] = {
 682         [BLOCK_STATE_UNINIT] = "unwritten",
 683         [BLOCK_STATE_TRIMMED] = "trimmed",
 684         [BLOCK_STATE_WRITTEN] = "written",
 685         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 686         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 687 };
 688
 689 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 690                              fio_fp64_t *plist, struct buf_output *out)
 691 {
 692         int len, pos, i;
 693         unsigned int *percentiles = NULL;
 694         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 695
 696         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 697                                      &percentiles, block_state_counts);
 698
 699         log_buf(out, "  block lifetime percentiles :\n   |");
 700         pos = 0;
 701         for (i = 0; i < len; i++) {
 702                 uint32_t block_info = percentiles[i];
 703 #define LINE_LENGTH     75
 704                 char str[LINE_LENGTH];
 705                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 706                                      plist[i].u.f, block_info,
 707                                      i == len - 1 ? '\n' : ',');
 708                 assert(strln < LINE_LENGTH);
 709                 if (pos + strln > LINE_LENGTH) {
 710                         pos = 0;
 711                         log_buf(out, "\n   |");
 712                 }
 713                 log_buf(out, "%s", str);
 714                 pos += strln;
 715 #undef LINE_LENGTH
 716         }
 717         if (percentiles)
 718                 free(percentiles);
 719
 720         log_buf(out, "        states               :");
 721         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 722                 log_buf(out, " %s=%u%c",
 723                          block_state_names[i], block_state_counts[i],
 724                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 725 }
 726
 727 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 728 {
 729         char *p1, *p1alt, *p2;
 730         unsigned long long bw_mean, iops_mean;
 731         const int i2p = is_power_of_2(ts->kb_base);
 732
 733         if (!ts->ss_dur)
 734                 return;
 735
 736         bw_mean = steadystate_bw_mean(ts);
 737         iops_mean = steadystate_iops_mean(ts);
 738
 739         p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
 740         p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
 741         p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
 742
 743         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 744                 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
 745                 p1, p1alt, p2,
 746                 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
 747                 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
 748                 ts->ss_criterion.u.f,
 749                 ts->ss_state & __FIO_SS_PCT ? "%" : "");
 750
 751         free(p1);
 752         free(p1alt);
 753         free(p2);
 754 }
 755
 756 static void show_thread_status_normal(struct thread_stat *ts,
 757                                       struct group_run_stats *rs,
 758                                       struct buf_output *out)
 759 {
 760         double usr_cpu, sys_cpu;
 761         unsigned long runtime;
 762         double io_u_dist[FIO_IO_U_MAP_NR];
 763         time_t time_p;
 764         char time_buf[32];
 765
 766         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 767                 return;
 768
 769         memset(time_buf, 0, sizeof(time_buf));
 770
 771         time(&time_p);
 772         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 773
 774         if (!ts->error) {
 775                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 776                                         ts->name, ts->groupid, ts->members,
 777                                         ts->error, (int) ts->pid, time_buf);
 778         } else {
 779                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 780                                         ts->name, ts->groupid, ts->members,
 781                                         ts->error, ts->verror, (int) ts->pid,
 782                                         time_buf);
 783         }
 784
 785         if (strlen(ts->description))
 786                 log_buf(out, "  Description  : [%s]\n", ts->description);
 787
 788         if (ts->io_bytes[DDIR_READ])
 789                 show_ddir_status(rs, ts, DDIR_READ, out);
 790         if (ts->io_bytes[DDIR_WRITE])
 791                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 792         if (ts->io_bytes[DDIR_TRIM])
 793                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 794
 795         show_latencies(ts, out);
 796
 797         runtime = ts->total_run_time;
 798         if (runtime) {
 799                 double runt = (double) runtime;
 800
 801                 usr_cpu = (double) ts->usr_time * 100 / runt;
 802                 sys_cpu = (double) ts->sys_time * 100 / runt;
 803         } else {
 804                 usr_cpu = 0;
 805                 sys_cpu = 0;
 806         }
 807
 808         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 809                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 810                         (unsigned long long) ts->ctx,
 811                         (unsigned long long) ts->majf,
 812                         (unsigned long long) ts->minf);
 813
 814         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 815         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 816                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 817                                         io_u_dist[1], io_u_dist[2],
 818                                         io_u_dist[3], io_u_dist[4],
 819                                         io_u_dist[5], io_u_dist[6]);
 820
 821         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 822         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 823                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 824                                         io_u_dist[1], io_u_dist[2],
 825                                         io_u_dist[3], io_u_dist[4],
 826                                         io_u_dist[5], io_u_dist[6]);
 827         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 828         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 829                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 830                                         io_u_dist[1], io_u_dist[2],
 831                                         io_u_dist[3], io_u_dist[4],
 832                                         io_u_dist[5], io_u_dist[6]);
 833         log_buf(out, "     issued rwt: total=%llu,%llu,%llu,"
 834                                  " short=%llu,%llu,%llu,"
 835                                  " dropped=%llu,%llu,%llu\n",
 836                                         (unsigned long long) ts->total_io_u[0],
 837                                         (unsigned long long) ts->total_io_u[1],
 838                                         (unsigned long long) ts->total_io_u[2],
 839                                         (unsigned long long) ts->short_io_u[0],
 840                                         (unsigned long long) ts->short_io_u[1],
 841                                         (unsigned long long) ts->short_io_u[2],
 842                                         (unsigned long long) ts->drop_io_u[0],
 843                                         (unsigned long long) ts->drop_io_u[1],
 844                                         (unsigned long long) ts->drop_io_u[2]);
 845         if (ts->continue_on_error) {
 846                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 847                                         (unsigned long long)ts->total_err_count,
 848                                         ts->first_error,
 849                                         strerror(ts->first_error));
 850         }
 851         if (ts->latency_depth) {
 852                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 853                                         (unsigned long long)ts->latency_target,
 854                                         (unsigned long long)ts->latency_window,
 855                                         ts->latency_percentile.u.f,
 856                                         ts->latency_depth);
 857         }
 858
 859         if (ts->nr_block_infos)
 860                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 861                                   ts->percentile_list, out);
 862
 863         if (ts->ss_dur)
 864                 show_ss_normal(ts, out);
 865 }
 866
 867 static void show_ddir_status_terse(struct thread_stat *ts,
 868                                    struct group_run_stats *rs, int ddir,
 869                                    int ver, struct buf_output *out)
 870 {
 871         unsigned long long min, max, minv, maxv, bw, iops;
 872         unsigned long long *ovals = NULL;
 873         double mean, dev;
 874         unsigned int len;
 875         int i, bw_stat;
 876
 877         assert(ddir_rw(ddir));
 878
 879         iops = bw = 0;
 880         if (ts->runtime[ddir]) {
 881                 uint64_t runt = ts->runtime[ddir];
 882
 883                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 884                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 885         }
 886
 887         log_buf(out, ";%llu;%llu;%llu;%llu",
 888                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 889                                         (unsigned long long) ts->runtime[ddir]);
 890
 891         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 892                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 893         else
 894                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 895
 896         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 897                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 898         else
 899                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 900
 901         if (ts->clat_percentiles || ts->lat_percentiles) {
 902                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 903                                         ts->clat_stat[ddir].samples,
 904                                         ts->percentile_list, &ovals, &maxv,
 905                                         &minv);
 906         } else
 907                 len = 0;
 908
 909         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 910                 if (i >= len) {
 911                         log_buf(out, ";0%%=0");
 912                         continue;
 913                 }
 914                 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
 915         }
 916
 917         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 918                 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
 919         else
 920                 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
 921
 922         if (ovals)
 923                 free(ovals);
 924
 925         bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
 926         if (bw_stat) {
 927                 double p_of_agg = 100.0;
 928
 929                 if (rs->agg[ddir]) {
 930                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
 931                         if (p_of_agg > 100.0)
 932                                 p_of_agg = 100.0;
 933                 }
 934
 935                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 936         } else
 937                 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
 938
 939         if (ver == 5) {
 940                 if (bw_stat)
 941                         log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
 942                 else
 943                         log_buf(out, ";%lu", 0UL);
 944
 945                 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
 946                         log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
 947                                 mean, dev, (&ts->iops_stat[ddir])->samples);
 948                 else
 949                         log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
 950         }
 951 }
 952
 953 static void add_ddir_status_json(struct thread_stat *ts,
 954                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 955 {
 956         unsigned long long min, max, minv, maxv;
 957         unsigned long long bw;
 958         unsigned long long *ovals = NULL;
 959         double mean, dev, iops;
 960         unsigned int len;
 961         int i;
 962         const char *ddirname[] = {"read", "write", "trim"};
 963         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 964         char buf[120];
 965         double p_of_agg = 100.0;
 966
 967         assert(ddir_rw(ddir));
 968
 969         if (ts->unified_rw_rep && ddir != DDIR_READ)
 970                 return;
 971
 972         dir_object = json_create_object();
 973         json_object_add_value_object(parent,
 974                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 975
 976         bw = 0;
 977         iops = 0.0;
 978         if (ts->runtime[ddir]) {
 979                 uint64_t runt = ts->runtime[ddir];
 980
 981                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 982                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 983         }
 984
 985         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
 986         json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
 987         json_object_add_value_int(dir_object, "bw", bw);
 988         json_object_add_value_float(dir_object, "iops", iops);
 989         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 990         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 991         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 992         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 993
 994         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 995                 min = max = 0;
 996                 mean = dev = 0.0;
 997         }
 998         tmp_object = json_create_object();
 999         json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1000         json_object_add_value_int(tmp_object, "min", min);
1001         json_object_add_value_int(tmp_object, "max", max);
1002         json_object_add_value_float(tmp_object, "mean", mean);
1003         json_object_add_value_float(tmp_object, "stddev", dev);
1004
1005         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1006                 min = max = 0;
1007                 mean = dev = 0.0;
1008         }
1009         tmp_object = json_create_object();
1010         json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1011         json_object_add_value_int(tmp_object, "min", min);
1012         json_object_add_value_int(tmp_object, "max", max);
1013         json_object_add_value_float(tmp_object, "mean", mean);
1014         json_object_add_value_float(tmp_object, "stddev", dev);
1015
1016         if (ts->clat_percentiles || ts->lat_percentiles) {
1017                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1018                                         ts->clat_stat[ddir].samples,
1019                                         ts->percentile_list, &ovals, &maxv,
1020                                         &minv);
1021         } else
1022                 len = 0;
1023
1024         percentile_object = json_create_object();
1025         json_object_add_value_object(tmp_object, "percentile", percentile_object);
1026         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1027                 if (i >= len) {
1028                         json_object_add_value_int(percentile_object, "0.00", 0);
1029                         continue;
1030                 }
1031                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1032                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1033         }
1034
1035         if (output_format & FIO_OUTPUT_JSON_PLUS) {
1036                 clat_bins_object = json_create_object();
1037                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1038                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1039                         if (ts->io_u_plat[ddir][i]) {
1040                                 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1041                                 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1042                         }
1043                 }
1044         }
1045
1046         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1047                 min = max = 0;
1048                 mean = dev = 0.0;
1049         }
1050         tmp_object = json_create_object();
1051         json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1052         json_object_add_value_int(tmp_object, "min", min);
1053         json_object_add_value_int(tmp_object, "max", max);
1054         json_object_add_value_float(tmp_object, "mean", mean);
1055         json_object_add_value_float(tmp_object, "stddev", dev);
1056         if (ovals)
1057                 free(ovals);
1058
1059         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1060                 if (rs->agg[ddir]) {
1061                         p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1062                         if (p_of_agg > 100.0)
1063                                 p_of_agg = 100.0;
1064                 }
1065         } else {
1066                 min = max = 0;
1067                 p_of_agg = mean = dev = 0.0;
1068         }
1069         json_object_add_value_int(dir_object, "bw_min", min);
1070         json_object_add_value_int(dir_object, "bw_max", max);
1071         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1072         json_object_add_value_float(dir_object, "bw_mean", mean);
1073         json_object_add_value_float(dir_object, "bw_dev", dev);
1074         json_object_add_value_int(dir_object, "bw_samples",
1075                                 (&ts->bw_stat[ddir])->samples);
1076
1077         if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1078                 min = max = 0;
1079                 mean = dev = 0.0;
1080         }
1081         json_object_add_value_int(dir_object, "iops_min", min);
1082         json_object_add_value_int(dir_object, "iops_max", max);
1083         json_object_add_value_float(dir_object, "iops_mean", mean);
1084         json_object_add_value_float(dir_object, "iops_stddev", dev);
1085         json_object_add_value_int(dir_object, "iops_samples",
1086                                 (&ts->iops_stat[ddir])->samples);
1087 }
1088
1089 static void show_thread_status_terse_all(struct thread_stat *ts,
1090                                          struct group_run_stats *rs, int ver,
1091                                          struct buf_output *out)
1092 {
1093         double io_u_dist[FIO_IO_U_MAP_NR];
1094         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1095         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1096         double usr_cpu, sys_cpu;
1097         int i;
1098
1099         /* General Info */
1100         if (ver == 2)
1101                 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1102         else
1103                 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1104                         ts->name, ts->groupid, ts->error);
1105
1106         /* Log Read Status */
1107         show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1108         /* Log Write Status */
1109         show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1110         /* Log Trim Status */
1111         if (ver == 2 || ver == 4 || ver == 5)
1112                 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1113
1114         /* CPU Usage */
1115         if (ts->total_run_time) {
1116                 double runt = (double) ts->total_run_time;
1117
1118                 usr_cpu = (double) ts->usr_time * 100 / runt;
1119                 sys_cpu = (double) ts->sys_time * 100 / runt;
1120         } else {
1121                 usr_cpu = 0;
1122                 sys_cpu = 0;
1123         }
1124
1125         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1126                                                 (unsigned long long) ts->ctx,
1127                                                 (unsigned long long) ts->majf,
1128                                                 (unsigned long long) ts->minf);
1129
1130         /* Calc % distribution of IO depths, usecond, msecond latency */
1131         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1132         stat_calc_lat_nu(ts, io_u_lat_u);
1133         stat_calc_lat_m(ts, io_u_lat_m);
1134
1135         /* Only show fixed 7 I/O depth levels*/
1136         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1137                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1138                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1139
1140         /* Microsecond latency */
1141         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1142                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1143         /* Millisecond latency */
1144         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1145                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1146
1147         /* disk util stats, if any */
1148         if (ver >= 3)
1149                 show_disk_util(1, NULL, out);
1150
1151         /* Additional output if continue_on_error set - default off*/
1152         if (ts->continue_on_error)
1153                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1154         if (ver == 2)
1155                 log_buf(out, "\n");
1156
1157         /* Additional output if description is set */
1158         if (strlen(ts->description))
1159                 log_buf(out, ";%s", ts->description);
1160
1161         log_buf(out, "\n");
1162 }
1163
1164 static void json_add_job_opts(struct json_object *root, const char *name,
1165                               struct flist_head *opt_list, bool num_jobs)
1166 {
1167         struct json_object *dir_object;
1168         struct flist_head *entry;
1169         struct print_option *p;
1170
1171         if (flist_empty(opt_list))
1172                 return;
1173
1174         dir_object = json_create_object();
1175         json_object_add_value_object(root, name, dir_object);
1176
1177         flist_for_each(entry, opt_list) {
1178                 const char *pos = "";
1179
1180                 p = flist_entry(entry, struct print_option, list);
1181                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1182                         continue;
1183                 if (p->value)
1184                         pos = p->value;
1185                 json_object_add_value_string(dir_object, p->name, pos);
1186         }
1187 }
1188
1189 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1190                                                    struct group_run_stats *rs,
1191                                                    struct flist_head *opt_list)
1192 {
1193         struct json_object *root, *tmp;
1194         struct jobs_eta *je;
1195         double io_u_dist[FIO_IO_U_MAP_NR];
1196         double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1197         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1198         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1199         double usr_cpu, sys_cpu;
1200         int i;
1201         size_t size;
1202
1203         root = json_create_object();
1204         json_object_add_value_string(root, "jobname", ts->name);
1205         json_object_add_value_int(root, "groupid", ts->groupid);
1206         json_object_add_value_int(root, "error", ts->error);
1207
1208         /* ETA Info */
1209         je = get_jobs_eta(true, &size);
1210         if (je) {
1211                 json_object_add_value_int(root, "eta", je->eta_sec);
1212                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1213         }
1214
1215         if (opt_list)
1216                 json_add_job_opts(root, "job options", opt_list, true);
1217
1218         add_ddir_status_json(ts, rs, DDIR_READ, root);
1219         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1220         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1221
1222         /* CPU Usage */
1223         if (ts->total_run_time) {
1224                 double runt = (double) ts->total_run_time;
1225
1226                 usr_cpu = (double) ts->usr_time * 100 / runt;
1227                 sys_cpu = (double) ts->sys_time * 100 / runt;
1228         } else {
1229                 usr_cpu = 0;
1230                 sys_cpu = 0;
1231         }
1232         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1233         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1234         json_object_add_value_int(root, "ctx", ts->ctx);
1235         json_object_add_value_int(root, "majf", ts->majf);
1236         json_object_add_value_int(root, "minf", ts->minf);
1237
1238
1239         /* Calc % distribution of IO depths, usecond, msecond latency */
1240         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1241         stat_calc_lat_n(ts, io_u_lat_n);
1242         stat_calc_lat_u(ts, io_u_lat_u);
1243         stat_calc_lat_m(ts, io_u_lat_m);
1244
1245         tmp = json_create_object();
1246         json_object_add_value_object(root, "iodepth_level", tmp);
1247         /* Only show fixed 7 I/O depth levels*/
1248         for (i = 0; i < 7; i++) {
1249                 char name[20];
1250                 if (i < 6)
1251                         snprintf(name, 20, "%d", 1 << i);
1252                 else
1253                         snprintf(name, 20, ">=%d", 1 << i);
1254                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1255         }
1256
1257         /* Nanosecond latency */
1258         tmp = json_create_object();
1259         json_object_add_value_object(root, "latency_ns", tmp);
1260         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1261                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1262                                  "250", "500", "750", "1000", };
1263                 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1264         }
1265         /* Microsecond latency */
1266         tmp = json_create_object();
1267         json_object_add_value_object(root, "latency_us", tmp);
1268         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1269                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1270                                  "250", "500", "750", "1000", };
1271                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1272         }
1273         /* Millisecond latency */
1274         tmp = json_create_object();
1275         json_object_add_value_object(root, "latency_ms", tmp);
1276         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1277                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1278                                  "250", "500", "750", "1000", "2000",
1279                                  ">=2000", };
1280                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1281         }
1282
1283         /* Additional output if continue_on_error set - default off*/
1284         if (ts->continue_on_error) {
1285                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1286                 json_object_add_value_int(root, "first_error", ts->first_error);
1287         }
1288
1289         if (ts->latency_depth) {
1290                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1291                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1292                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1293                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1294         }
1295
1296         /* Additional output if description is set */
1297         if (strlen(ts->description))
1298                 json_object_add_value_string(root, "desc", ts->description);
1299
1300         if (ts->nr_block_infos) {
1301                 /* Block error histogram and types */
1302                 int len;
1303                 unsigned int *percentiles = NULL;
1304                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1305
1306                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1307                                              ts->percentile_list,
1308                                              &percentiles, block_state_counts);
1309
1310                 if (len) {
1311                         struct json_object *block, *percentile_object, *states;
1312                         int state;
1313                         block = json_create_object();
1314                         json_object_add_value_object(root, "block", block);
1315
1316                         percentile_object = json_create_object();
1317                         json_object_add_value_object(block, "percentiles",
1318                                                      percentile_object);
1319                         for (i = 0; i < len; i++) {
1320                                 char buf[20];
1321                                 snprintf(buf, sizeof(buf), "%f",
1322                                          ts->percentile_list[i].u.f);
1323                                 json_object_add_value_int(percentile_object,
1324                                                           (const char *)buf,
1325                                                           percentiles[i]);
1326                         }
1327
1328                         states = json_create_object();
1329                         json_object_add_value_object(block, "states", states);
1330                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1331                                 json_object_add_value_int(states,
1332                                         block_state_names[state],
1333                                         block_state_counts[state]);
1334                         }
1335                         free(percentiles);
1336                 }
1337         }
1338
1339         if (ts->ss_dur) {
1340                 struct json_object *data;
1341                 struct json_array *iops, *bw;
1342                 int i, j, k;
1343                 char ss_buf[64];
1344
1345                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1346                         ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1347                         ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1348                         (float) ts->ss_limit.u.f,
1349                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1350
1351                 tmp = json_create_object();
1352                 json_object_add_value_object(root, "steadystate", tmp);
1353                 json_object_add_value_string(tmp, "ss", ss_buf);
1354                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1355                 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1356
1357                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1358                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1359                 json_object_add_value_string(tmp, "criterion", ss_buf);
1360                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1361                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1362
1363                 data = json_create_object();
1364                 json_object_add_value_object(tmp, "data", data);
1365                 bw = json_create_array();
1366                 iops = json_create_array();
1367
1368                 /*
1369                 ** if ss was attained or the buffer is not full,
1370                 ** ss->head points to the first element in the list.
1371                 ** otherwise it actually points to the second element
1372                 ** in the list
1373                 */
1374                 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1375                         j = ts->ss_head;
1376                 else
1377                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1378                 for (i = 0; i < ts->ss_dur; i++) {
1379                         k = (j + i) % ts->ss_dur;
1380                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1381                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1382                 }
1383                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1384                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1385                 json_object_add_value_array(data, "iops", iops);
1386                 json_object_add_value_array(data, "bw", bw);
1387         }
1388
1389         return root;
1390 }
1391
1392 static void show_thread_status_terse(struct thread_stat *ts,
1393                                      struct group_run_stats *rs,
1394                                      struct buf_output *out)
1395 {
1396         if (terse_version >= 2 && terse_version <= 5)
1397                 show_thread_status_terse_all(ts, rs, terse_version, out);
1398         else
1399                 log_err("fio: bad terse version!? %d\n", terse_version);
1400 }
1401
1402 struct json_object *show_thread_status(struct thread_stat *ts,
1403                                        struct group_run_stats *rs,
1404                                        struct flist_head *opt_list,
1405                                        struct buf_output *out)
1406 {
1407         struct json_object *ret = NULL;
1408
1409         if (output_format & FIO_OUTPUT_TERSE)
1410                 show_thread_status_terse(ts, rs,  out);
1411         if (output_format & FIO_OUTPUT_JSON)
1412                 ret = show_thread_status_json(ts, rs, opt_list);
1413         if (output_format & FIO_OUTPUT_NORMAL)
1414                 show_thread_status_normal(ts, rs,  out);
1415
1416         return ret;
1417 }
1418
1419 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1420 {
1421         double mean, S;
1422
1423         if (src->samples == 0)
1424                 return;
1425
1426         dst->min_val = min(dst->min_val, src->min_val);
1427         dst->max_val = max(dst->max_val, src->max_val);
1428
1429         /*
1430          * Compute new mean and S after the merge
1431          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1432          *  #Parallel_algorithm>
1433          */
1434         if (first) {
1435                 mean = src->mean.u.f;
1436                 S = src->S.u.f;
1437         } else {
1438                 double delta = src->mean.u.f - dst->mean.u.f;
1439
1440                 mean = ((src->mean.u.f * src->samples) +
1441                         (dst->mean.u.f * dst->samples)) /
1442                         (dst->samples + src->samples);
1443
1444                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1445                         (dst->samples * src->samples) /
1446                         (dst->samples + src->samples);
1447         }
1448
1449         dst->samples += src->samples;
1450         dst->mean.u.f = mean;
1451         dst->S.u.f = S;
1452 }
1453
1454 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1455 {
1456         int i;
1457
1458         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1459                 if (dst->max_run[i] < src->max_run[i])
1460                         dst->max_run[i] = src->max_run[i];
1461                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1462                         dst->min_run[i] = src->min_run[i];
1463                 if (dst->max_bw[i] < src->max_bw[i])
1464                         dst->max_bw[i] = src->max_bw[i];
1465                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1466                         dst->min_bw[i] = src->min_bw[i];
1467
1468                 dst->iobytes[i] += src->iobytes[i];
1469                 dst->agg[i] += src->agg[i];
1470         }
1471
1472         if (!dst->kb_base)
1473                 dst->kb_base = src->kb_base;
1474         if (!dst->unit_base)
1475                 dst->unit_base = src->unit_base;
1476 }
1477
1478 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1479                       bool first)
1480 {
1481         int l, k;
1482
1483         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1484                 if (!dst->unified_rw_rep) {
1485                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1486                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1487                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1488                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1489                         sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1490
1491                         dst->io_bytes[l] += src->io_bytes[l];
1492
1493                         if (dst->runtime[l] < src->runtime[l])
1494                                 dst->runtime[l] = src->runtime[l];
1495                 } else {
1496                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1497                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1498                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1499                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1500                         sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1501
1502                         dst->io_bytes[0] += src->io_bytes[l];
1503
1504                         if (dst->runtime[0] < src->runtime[l])
1505                                 dst->runtime[0] = src->runtime[l];
1506
1507                         /*
1508                          * We're summing to the same destination, so override
1509                          * 'first' after the first iteration of the loop
1510                          */
1511                         first = false;
1512                 }
1513         }
1514
1515         dst->usr_time += src->usr_time;
1516         dst->sys_time += src->sys_time;
1517         dst->ctx += src->ctx;
1518         dst->majf += src->majf;
1519         dst->minf += src->minf;
1520
1521         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1522                 dst->io_u_map[k] += src->io_u_map[k];
1523         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1524                 dst->io_u_submit[k] += src->io_u_submit[k];
1525         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1526                 dst->io_u_complete[k] += src->io_u_complete[k];
1527         for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1528                 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1529         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1530                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1531         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1532                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1533
1534         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1535                 if (!dst->unified_rw_rep) {
1536                         dst->total_io_u[k] += src->total_io_u[k];
1537                         dst->short_io_u[k] += src->short_io_u[k];
1538                         dst->drop_io_u[k] += src->drop_io_u[k];
1539                 } else {
1540                         dst->total_io_u[0] += src->total_io_u[k];
1541                         dst->short_io_u[0] += src->short_io_u[k];
1542                         dst->drop_io_u[0] += src->drop_io_u[k];
1543                 }
1544         }
1545
1546         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1547                 int m;
1548
1549                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1550                         if (!dst->unified_rw_rep)
1551                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1552                         else
1553                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1554                 }
1555         }
1556
1557         dst->total_run_time += src->total_run_time;
1558         dst->total_submit += src->total_submit;
1559         dst->total_complete += src->total_complete;
1560 }
1561
1562 void init_group_run_stat(struct group_run_stats *gs)
1563 {
1564         int i;
1565         memset(gs, 0, sizeof(*gs));
1566
1567         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1568                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1569 }
1570
1571 void init_thread_stat(struct thread_stat *ts)
1572 {
1573         int j;
1574
1575         memset(ts, 0, sizeof(*ts));
1576
1577         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1578                 ts->lat_stat[j].min_val = -1UL;
1579                 ts->clat_stat[j].min_val = -1UL;
1580                 ts->slat_stat[j].min_val = -1UL;
1581                 ts->bw_stat[j].min_val = -1UL;
1582                 ts->iops_stat[j].min_val = -1UL;
1583         }
1584         ts->groupid = -1;
1585 }
1586
1587 void __show_run_stats(void)
1588 {
1589         struct group_run_stats *runstats, *rs;
1590         struct thread_data *td;
1591         struct thread_stat *threadstats, *ts;
1592         int i, j, k, nr_ts, last_ts, idx;
1593         int kb_base_warned = 0;
1594         int unit_base_warned = 0;
1595         struct json_object *root = NULL;
1596         struct json_array *array = NULL;
1597         struct buf_output output[FIO_OUTPUT_NR];
1598         struct flist_head **opt_lists;
1599
1600         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1601
1602         for (i = 0; i < groupid + 1; i++)
1603                 init_group_run_stat(&runstats[i]);
1604
1605         /*
1606          * find out how many threads stats we need. if group reporting isn't
1607          * enabled, it's one-per-td.
1608          */
1609         nr_ts = 0;
1610         last_ts = -1;
1611         for_each_td(td, i) {
1612                 if (!td->o.group_reporting) {
1613                         nr_ts++;
1614                         continue;
1615                 }
1616                 if (last_ts == td->groupid)
1617                         continue;
1618                 if (!td->o.stats)
1619                         continue;
1620
1621                 last_ts = td->groupid;
1622                 nr_ts++;
1623         }
1624
1625         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1626         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1627
1628         for (i = 0; i < nr_ts; i++) {
1629                 init_thread_stat(&threadstats[i]);
1630                 opt_lists[i] = NULL;
1631         }
1632
1633         j = 0;
1634         last_ts = -1;
1635         idx = 0;
1636         for_each_td(td, i) {
1637                 if (!td->o.stats)
1638                         continue;
1639                 if (idx && (!td->o.group_reporting ||
1640                     (td->o.group_reporting && last_ts != td->groupid))) {
1641                         idx = 0;
1642                         j++;
1643                 }
1644
1645                 last_ts = td->groupid;
1646
1647                 ts = &threadstats[j];
1648
1649                 ts->clat_percentiles = td->o.clat_percentiles;
1650                 ts->lat_percentiles = td->o.lat_percentiles;
1651                 ts->percentile_precision = td->o.percentile_precision;
1652                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1653                 opt_lists[j] = &td->opt_list;
1654
1655                 idx++;
1656                 ts->members++;
1657
1658                 if (ts->groupid == -1) {
1659                         /*
1660                          * These are per-group shared already
1661                          */
1662                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1663                         if (td->o.description)
1664                                 strncpy(ts->description, td->o.description,
1665                                                 FIO_JOBDESC_SIZE - 1);
1666                         else
1667                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1668
1669                         /*
1670                          * If multiple entries in this group, this is
1671                          * the first member.
1672                          */
1673                         ts->thread_number = td->thread_number;
1674                         ts->groupid = td->groupid;
1675
1676                         /*
1677                          * first pid in group, not very useful...
1678                          */
1679                         ts->pid = td->pid;
1680
1681                         ts->kb_base = td->o.kb_base;
1682                         ts->unit_base = td->o.unit_base;
1683                         ts->unified_rw_rep = td->o.unified_rw_rep;
1684                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1685                         log_info("fio: kb_base differs for jobs in group, using"
1686                                  " %u as the base\n", ts->kb_base);
1687                         kb_base_warned = 1;
1688                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1689                         log_info("fio: unit_base differs for jobs in group, using"
1690                                  " %u as the base\n", ts->unit_base);
1691                         unit_base_warned = 1;
1692                 }
1693
1694                 ts->continue_on_error = td->o.continue_on_error;
1695                 ts->total_err_count += td->total_err_count;
1696                 ts->first_error = td->first_error;
1697                 if (!ts->error) {
1698                         if (!td->error && td->o.continue_on_error &&
1699                             td->first_error) {
1700                                 ts->error = td->first_error;
1701                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1702                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1703                         } else  if (td->error) {
1704                                 ts->error = td->error;
1705                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1706                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1707                         }
1708                 }
1709
1710                 ts->latency_depth = td->latency_qd;
1711                 ts->latency_target = td->o.latency_target;
1712                 ts->latency_percentile = td->o.latency_percentile;
1713                 ts->latency_window = td->o.latency_window;
1714
1715                 ts->nr_block_infos = td->ts.nr_block_infos;
1716                 for (k = 0; k < ts->nr_block_infos; k++)
1717                         ts->block_infos[k] = td->ts.block_infos[k];
1718
1719                 sum_thread_stats(ts, &td->ts, idx == 1);
1720
1721                 if (td->o.ss_dur) {
1722                         ts->ss_state = td->ss.state;
1723                         ts->ss_dur = td->ss.dur;
1724                         ts->ss_head = td->ss.head;
1725                         ts->ss_bw_data = td->ss.bw_data;
1726                         ts->ss_iops_data = td->ss.iops_data;
1727                         ts->ss_limit.u.f = td->ss.limit;
1728                         ts->ss_slope.u.f = td->ss.slope;
1729                         ts->ss_deviation.u.f = td->ss.deviation;
1730                         ts->ss_criterion.u.f = td->ss.criterion;
1731                 }
1732                 else
1733                         ts->ss_dur = ts->ss_state = 0;
1734         }
1735
1736         for (i = 0; i < nr_ts; i++) {
1737                 unsigned long long bw;
1738
1739                 ts = &threadstats[i];
1740                 if (ts->groupid == -1)
1741                         continue;
1742                 rs = &runstats[ts->groupid];
1743                 rs->kb_base = ts->kb_base;
1744                 rs->unit_base = ts->unit_base;
1745                 rs->unified_rw_rep += ts->unified_rw_rep;
1746
1747                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1748                         if (!ts->runtime[j])
1749                                 continue;
1750                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1751                                 rs->min_run[j] = ts->runtime[j];
1752                         if (ts->runtime[j] > rs->max_run[j])
1753                                 rs->max_run[j] = ts->runtime[j];
1754
1755                         bw = 0;
1756                         if (ts->runtime[j])
1757                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1758                         if (bw < rs->min_bw[j])
1759                                 rs->min_bw[j] = bw;
1760                         if (bw > rs->max_bw[j])
1761                                 rs->max_bw[j] = bw;
1762
1763                         rs->iobytes[j] += ts->io_bytes[j];
1764                 }
1765         }
1766
1767         for (i = 0; i < groupid + 1; i++) {
1768                 int ddir;
1769
1770                 rs = &runstats[i];
1771
1772                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1773                         if (rs->max_run[ddir])
1774                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1775                                                 rs->max_run[ddir];
1776                 }
1777         }
1778
1779         for (i = 0; i < FIO_OUTPUT_NR; i++)
1780                 buf_output_init(&output[i]);
1781
1782         /*
1783          * don't overwrite last signal output
1784          */
1785         if (output_format & FIO_OUTPUT_NORMAL)
1786                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1787         if (output_format & FIO_OUTPUT_JSON) {
1788                 struct thread_data *global;
1789                 char time_buf[32];
1790                 struct timeval now;
1791                 unsigned long long ms_since_epoch;
1792
1793                 gettimeofday(&now, NULL);
1794                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1795                                  (unsigned long long)(now.tv_usec) / 1000;
1796
1797                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1798                                 sizeof(time_buf));
1799                 if (time_buf[strlen(time_buf) - 1] == '\n')
1800                         time_buf[strlen(time_buf) - 1] = '\0';
1801
1802                 root = json_create_object();
1803                 json_object_add_value_string(root, "fio version", fio_version_string);
1804                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1805                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1806                 json_object_add_value_string(root, "time", time_buf);
1807                 global = get_global_options();
1808                 json_add_job_opts(root, "global options", &global->opt_list, false);
1809                 array = json_create_array();
1810                 json_object_add_value_array(root, "jobs", array);
1811         }
1812
1813         if (is_backend)
1814                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1815
1816         for (i = 0; i < nr_ts; i++) {
1817                 ts = &threadstats[i];
1818                 rs = &runstats[ts->groupid];
1819
1820                 if (is_backend) {
1821                         fio_server_send_job_options(opt_lists[i], i);
1822                         fio_server_send_ts(ts, rs);
1823                 } else {
1824                         if (output_format & FIO_OUTPUT_TERSE)
1825                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1826                         if (output_format & FIO_OUTPUT_JSON) {
1827                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1828                                 json_array_add_value_object(array, tmp);
1829                         }
1830                         if (output_format & FIO_OUTPUT_NORMAL)
1831                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1832                 }
1833         }
1834         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1835                 /* disk util stats, if any */
1836                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1837
1838                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1839
1840                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1841                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1842                 json_free_object(root);
1843         }
1844
1845         for (i = 0; i < groupid + 1; i++) {
1846                 rs = &runstats[i];
1847
1848                 rs->groupid = i;
1849                 if (is_backend)
1850                         fio_server_send_gs(rs);
1851                 else if (output_format & FIO_OUTPUT_NORMAL)
1852                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1853         }
1854
1855         if (is_backend)
1856                 fio_server_send_du();
1857         else if (output_format & FIO_OUTPUT_NORMAL) {
1858                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1859                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1860         }
1861
1862         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1863                 struct buf_output *out = &output[i];
1864
1865                 log_info_buf(out->buf, out->buflen);
1866                 buf_output_free(out);
1867         }
1868
1869         log_info_flush();
1870         free(runstats);
1871         free(threadstats);
1872         free(opt_lists);
1873 }
1874
1875 void show_run_stats(void)
1876 {
1877         fio_mutex_down(stat_mutex);
1878         __show_run_stats();
1879         fio_mutex_up(stat_mutex);
1880 }
1881
1882 void __show_running_run_stats(void)
1883 {
1884         struct thread_data *td;
1885         unsigned long long *rt;
1886         struct timespec ts;
1887         int i;
1888
1889         fio_mutex_down(stat_mutex);
1890
1891         rt = malloc(thread_number * sizeof(unsigned long long));
1892         fio_gettime(&ts, NULL);
1893
1894         for_each_td(td, i) {
1895                 td->update_rusage = 1;
1896                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1897                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1898                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1899                 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1900
1901                 rt[i] = mtime_since(&td->start, &ts);
1902                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1903                         td->ts.runtime[DDIR_READ] += rt[i];
1904                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1905                         td->ts.runtime[DDIR_WRITE] += rt[i];
1906                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1907                         td->ts.runtime[DDIR_TRIM] += rt[i];
1908         }
1909
1910         for_each_td(td, i) {
1911                 if (td->runstate >= TD_EXITED)
1912                         continue;
1913                 if (td->rusage_sem) {
1914                         td->update_rusage = 1;
1915                         fio_mutex_down(td->rusage_sem);
1916                 }
1917                 td->update_rusage = 0;
1918         }
1919
1920         __show_run_stats();
1921
1922         for_each_td(td, i) {
1923                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1924                         td->ts.runtime[DDIR_READ] -= rt[i];
1925                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1926                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1927                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1928                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1929         }
1930
1931         free(rt);
1932         fio_mutex_up(stat_mutex);
1933 }
1934
1935 static int status_interval_init;
1936 static struct timespec status_time;
1937 static int status_file_disabled;
1938
1939 #define FIO_STATUS_FILE         "fio-dump-status"
1940
1941 static int check_status_file(void)
1942 {
1943         struct stat sb;
1944         const char *temp_dir;
1945         char fio_status_file_path[PATH_MAX];
1946
1947         if (status_file_disabled)
1948                 return 0;
1949
1950         temp_dir = getenv("TMPDIR");
1951         if (temp_dir == NULL) {
1952                 temp_dir = getenv("TEMP");
1953                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1954                         temp_dir = NULL;
1955         }
1956         if (temp_dir == NULL)
1957                 temp_dir = "/tmp";
1958
1959         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1960
1961         if (stat(fio_status_file_path, &sb))
1962                 return 0;
1963
1964         if (unlink(fio_status_file_path) < 0) {
1965                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1966                                                         strerror(errno));
1967                 log_err("fio: disabling status file updates\n");
1968                 status_file_disabled = 1;
1969         }
1970
1971         return 1;
1972 }
1973
1974 void check_for_running_stats(void)
1975 {
1976         if (status_interval) {
1977                 if (!status_interval_init) {
1978                         fio_gettime(&status_time, NULL);
1979                         status_interval_init = 1;
1980                 } else if (mtime_since_now(&status_time) >= status_interval) {
1981                         show_running_run_stats();
1982                         fio_gettime(&status_time, NULL);
1983                         return;
1984                 }
1985         }
1986         if (check_status_file()) {
1987                 show_running_run_stats();
1988                 return;
1989         }
1990 }
1991
1992 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
1993 {
1994         double val = data;
1995         double delta;
1996
1997         if (data > is->max_val)
1998                 is->max_val = data;
1999         if (data < is->min_val)
2000                 is->min_val = data;
2001
2002         delta = val - is->mean.u.f;
2003         if (delta) {
2004                 is->mean.u.f += delta / (is->samples + 1.0);
2005                 is->S.u.f += delta * (val - is->mean.u.f);
2006         }
2007
2008         is->samples++;
2009 }
2010
2011 /*
2012  * Return a struct io_logs, which is added to the tail of the log
2013  * list for 'iolog'.
2014  */
2015 static struct io_logs *get_new_log(struct io_log *iolog)
2016 {
2017         size_t new_size, new_samples;
2018         struct io_logs *cur_log;
2019
2020         /*
2021          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2022          * forever
2023          */
2024         if (!iolog->cur_log_max)
2025                 new_samples = DEF_LOG_ENTRIES;
2026         else {
2027                 new_samples = iolog->cur_log_max * 2;
2028                 if (new_samples > MAX_LOG_ENTRIES)
2029                         new_samples = MAX_LOG_ENTRIES;
2030         }
2031
2032         new_size = new_samples * log_entry_sz(iolog);
2033
2034         cur_log = smalloc(sizeof(*cur_log));
2035         if (cur_log) {
2036                 INIT_FLIST_HEAD(&cur_log->list);
2037                 cur_log->log = malloc(new_size);
2038                 if (cur_log->log) {
2039                         cur_log->nr_samples = 0;
2040                         cur_log->max_samples = new_samples;
2041                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2042                         iolog->cur_log_max = new_samples;
2043                         return cur_log;
2044                 }
2045                 sfree(cur_log);
2046         }
2047
2048         return NULL;
2049 }
2050
2051 /*
2052  * Add and return a new log chunk, or return current log if big enough
2053  */
2054 static struct io_logs *regrow_log(struct io_log *iolog)
2055 {
2056         struct io_logs *cur_log;
2057         int i;
2058
2059         if (!iolog || iolog->disabled)
2060                 goto disable;
2061
2062         cur_log = iolog_cur_log(iolog);
2063         if (!cur_log) {
2064                 cur_log = get_new_log(iolog);
2065                 if (!cur_log)
2066                         return NULL;
2067         }
2068
2069         if (cur_log->nr_samples < cur_log->max_samples)
2070                 return cur_log;
2071
2072         /*
2073          * No room for a new sample. If we're compressing on the fly, flush
2074          * out the current chunk
2075          */
2076         if (iolog->log_gz) {
2077                 if (iolog_cur_flush(iolog, cur_log)) {
2078                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2079                         return NULL;
2080                 }
2081         }
2082
2083         /*
2084          * Get a new log array, and add to our list
2085          */
2086         cur_log = get_new_log(iolog);
2087         if (!cur_log) {
2088                 log_err("fio: failed extending iolog! Will stop logging.\n");
2089                 return NULL;
2090         }
2091
2092         if (!iolog->pending || !iolog->pending->nr_samples)
2093                 return cur_log;
2094
2095         /*
2096          * Flush pending items to new log
2097          */
2098         for (i = 0; i < iolog->pending->nr_samples; i++) {
2099                 struct io_sample *src, *dst;
2100
2101                 src = get_sample(iolog, iolog->pending, i);
2102                 dst = get_sample(iolog, cur_log, i);
2103                 memcpy(dst, src, log_entry_sz(iolog));
2104         }
2105         cur_log->nr_samples = iolog->pending->nr_samples;
2106
2107         iolog->pending->nr_samples = 0;
2108         return cur_log;
2109 disable:
2110         if (iolog)
2111                 iolog->disabled = true;
2112         return NULL;
2113 }
2114
2115 void regrow_logs(struct thread_data *td)
2116 {
2117         regrow_log(td->slat_log);
2118         regrow_log(td->clat_log);
2119         regrow_log(td->clat_hist_log);
2120         regrow_log(td->lat_log);
2121         regrow_log(td->bw_log);
2122         regrow_log(td->iops_log);
2123         td->flags &= ~TD_F_REGROW_LOGS;
2124 }
2125
2126 static struct io_logs *get_cur_log(struct io_log *iolog)
2127 {
2128         struct io_logs *cur_log;
2129
2130         cur_log = iolog_cur_log(iolog);
2131         if (!cur_log) {
2132                 cur_log = get_new_log(iolog);
2133                 if (!cur_log)
2134                         return NULL;
2135         }
2136
2137         if (cur_log->nr_samples < cur_log->max_samples)
2138                 return cur_log;
2139
2140         /*
2141          * Out of space. If we're in IO offload mode, or we're not doing
2142          * per unit logging (hence logging happens outside of the IO thread
2143          * as well), add a new log chunk inline. If we're doing inline
2144          * submissions, flag 'td' as needing a log regrow and we'll take
2145          * care of it on the submission side.
2146          */
2147         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2148             !per_unit_log(iolog))
2149                 return regrow_log(iolog);
2150
2151         iolog->td->flags |= TD_F_REGROW_LOGS;
2152         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2153         return iolog->pending;
2154 }
2155
2156 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2157                              enum fio_ddir ddir, unsigned int bs,
2158                              unsigned long t, uint64_t offset)
2159 {
2160         struct io_logs *cur_log;
2161
2162         if (iolog->disabled)
2163                 return;
2164         if (flist_empty(&iolog->io_logs))
2165                 iolog->avg_last[ddir] = t;
2166
2167         cur_log = get_cur_log(iolog);
2168         if (cur_log) {
2169                 struct io_sample *s;
2170
2171                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2172
2173                 s->data = data;
2174                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2175                 io_sample_set_ddir(iolog, s, ddir);
2176                 s->bs = bs;
2177
2178                 if (iolog->log_offset) {
2179                         struct io_sample_offset *so = (void *) s;
2180
2181                         so->offset = offset;
2182                 }
2183
2184                 cur_log->nr_samples++;
2185                 return;
2186         }
2187
2188         iolog->disabled = true;
2189 }
2190
2191 static inline void reset_io_stat(struct io_stat *ios)
2192 {
2193         ios->max_val = ios->min_val = ios->samples = 0;
2194         ios->mean.u.f = ios->S.u.f = 0;
2195 }
2196
2197 void reset_io_stats(struct thread_data *td)
2198 {
2199         struct thread_stat *ts = &td->ts;
2200         int i, j;
2201
2202         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2203                 reset_io_stat(&ts->clat_stat[i]);
2204                 reset_io_stat(&ts->slat_stat[i]);
2205                 reset_io_stat(&ts->lat_stat[i]);
2206                 reset_io_stat(&ts->bw_stat[i]);
2207                 reset_io_stat(&ts->iops_stat[i]);
2208
2209                 ts->io_bytes[i] = 0;
2210                 ts->runtime[i] = 0;
2211                 ts->total_io_u[i] = 0;
2212                 ts->short_io_u[i] = 0;
2213                 ts->drop_io_u[i] = 0;
2214
2215                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2216                         ts->io_u_plat[i][j] = 0;
2217         }
2218
2219         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2220                 ts->io_u_map[i] = 0;
2221                 ts->io_u_submit[i] = 0;
2222                 ts->io_u_complete[i] = 0;
2223         }
2224
2225         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2226                 ts->io_u_lat_n[i] = 0;
2227         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2228                 ts->io_u_lat_u[i] = 0;
2229         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2230                 ts->io_u_lat_m[i] = 0;
2231
2232         ts->total_submit = 0;
2233         ts->total_complete = 0;
2234 }
2235
2236 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2237                               unsigned long elapsed, bool log_max)
2238 {
2239         /*
2240          * Note an entry in the log. Use the mean from the logged samples,
2241          * making sure to properly round up. Only write a log entry if we
2242          * had actual samples done.
2243          */
2244         if (iolog->avg_window[ddir].samples) {
2245                 union io_sample_data data;
2246
2247                 if (log_max)
2248                         data.val = iolog->avg_window[ddir].max_val;
2249                 else
2250                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2251
2252                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2253         }
2254
2255         reset_io_stat(&iolog->avg_window[ddir]);
2256 }
2257
2258 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2259                              bool log_max)
2260 {
2261         int ddir;
2262
2263         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2264                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2265 }
2266
2267 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2268                            union io_sample_data data, enum fio_ddir ddir,
2269                            unsigned int bs, uint64_t offset)
2270 {
2271         unsigned long elapsed, this_window;
2272
2273         if (!ddir_rw(ddir))
2274                 return 0;
2275
2276         elapsed = mtime_since_now(&td->epoch);
2277
2278         /*
2279          * If no time averaging, just add the log sample.
2280          */
2281         if (!iolog->avg_msec) {
2282                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2283                 return 0;
2284         }
2285
2286         /*
2287          * Add the sample. If the time period has passed, then
2288          * add that entry to the log and clear.
2289          */
2290         add_stat_sample(&iolog->avg_window[ddir], data.val);
2291
2292         /*
2293          * If period hasn't passed, adding the above sample is all we
2294          * need to do.
2295          */
2296         this_window = elapsed - iolog->avg_last[ddir];
2297         if (elapsed < iolog->avg_last[ddir])
2298                 return iolog->avg_last[ddir] - elapsed;
2299         else if (this_window < iolog->avg_msec) {
2300                 int diff = iolog->avg_msec - this_window;
2301
2302                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2303                         return diff;
2304         }
2305
2306         __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2307
2308         iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2309         return iolog->avg_msec;
2310 }
2311
2312 void finalize_logs(struct thread_data *td, bool unit_logs)
2313 {
2314         unsigned long elapsed;
2315
2316         elapsed = mtime_since_now(&td->epoch);
2317
2318         if (td->clat_log && unit_logs)
2319                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2320         if (td->slat_log && unit_logs)
2321                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2322         if (td->lat_log && unit_logs)
2323                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2324         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2325                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2326         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2327                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2328 }
2329
2330 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2331 {
2332         struct io_log *iolog;
2333
2334         if (!ddir_rw(ddir))
2335                 return;
2336
2337         iolog = agg_io_log[ddir];
2338         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2339 }
2340
2341 static void add_clat_percentile_sample(struct thread_stat *ts,
2342                                 unsigned long long nsec, enum fio_ddir ddir)
2343 {
2344         unsigned int idx = plat_val_to_idx(nsec);
2345         assert(idx < FIO_IO_U_PLAT_NR);
2346
2347         ts->io_u_plat[ddir][idx]++;
2348 }
2349
2350 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2351                      unsigned long long nsec, unsigned int bs, uint64_t offset)
2352 {
2353         unsigned long elapsed, this_window;
2354         struct thread_stat *ts = &td->ts;
2355         struct io_log *iolog = td->clat_hist_log;
2356
2357         td_io_u_lock(td);
2358
2359         add_stat_sample(&ts->clat_stat[ddir], nsec);
2360
2361         if (td->clat_log)
2362                 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2363                                offset);
2364
2365         if (ts->clat_percentiles)
2366                 add_clat_percentile_sample(ts, nsec, ddir);
2367
2368         if (iolog && iolog->hist_msec) {
2369                 struct io_hist *hw = &iolog->hist_window[ddir];
2370
2371                 hw->samples++;
2372                 elapsed = mtime_since_now(&td->epoch);
2373                 if (!hw->hist_last)
2374                         hw->hist_last = elapsed;
2375                 this_window = elapsed - hw->hist_last;
2376
2377                 if (this_window >= iolog->hist_msec) {
2378                         unsigned int *io_u_plat;
2379                         struct io_u_plat_entry *dst;
2380
2381                         /*
2382                          * Make a byte-for-byte copy of the latency histogram
2383                          * stored in td->ts.io_u_plat[ddir], recording it in a
2384                          * log sample. Note that the matching call to free() is
2385                          * located in iolog.c after printing this sample to the
2386                          * log file.
2387                          */
2388                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2389                         dst = malloc(sizeof(struct io_u_plat_entry));
2390                         memcpy(&(dst->io_u_plat), io_u_plat,
2391                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2392                         flist_add(&dst->list, &hw->list);
2393                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2394                                                 elapsed, offset);
2395
2396                         /*
2397                          * Update the last time we recorded as being now, minus
2398                          * any drift in time we encountered before actually
2399                          * making the record.
2400                          */
2401                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2402                         hw->samples = 0;
2403                 }
2404         }
2405
2406         td_io_u_unlock(td);
2407 }
2408
2409 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2410                      unsigned long usec, unsigned int bs, uint64_t offset)
2411 {
2412         struct thread_stat *ts = &td->ts;
2413
2414         if (!ddir_rw(ddir))
2415                 return;
2416
2417         td_io_u_lock(td);
2418
2419         add_stat_sample(&ts->slat_stat[ddir], usec);
2420
2421         if (td->slat_log)
2422                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2423
2424         td_io_u_unlock(td);
2425 }
2426
2427 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2428                     unsigned long long nsec, unsigned int bs, uint64_t offset)
2429 {
2430         struct thread_stat *ts = &td->ts;
2431
2432         if (!ddir_rw(ddir))
2433                 return;
2434
2435         td_io_u_lock(td);
2436
2437         add_stat_sample(&ts->lat_stat[ddir], nsec);
2438
2439         if (td->lat_log)
2440                 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2441                                offset);
2442
2443         if (ts->lat_percentiles)
2444                 add_clat_percentile_sample(ts, nsec, ddir);
2445
2446         td_io_u_unlock(td);
2447 }
2448
2449 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2450                    unsigned int bytes, unsigned long long spent)
2451 {
2452         struct thread_stat *ts = &td->ts;
2453         unsigned long rate;
2454
2455         if (spent)
2456                 rate = (unsigned long) (bytes * 1000000ULL / spent);
2457         else
2458                 rate = 0;
2459
2460         td_io_u_lock(td);
2461
2462         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2463
2464         if (td->bw_log)
2465                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2466                                bytes, io_u->offset);
2467
2468         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2469         td_io_u_unlock(td);
2470 }
2471
2472 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2473                          struct timespec *t, unsigned int avg_time,
2474                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2475                          struct io_stat *stat, struct io_log *log,
2476                          bool is_kb)
2477 {
2478         unsigned long spent, rate;
2479         enum fio_ddir ddir;
2480         unsigned int next, next_log;
2481
2482         next_log = avg_time;
2483
2484         spent = mtime_since(parent_tv, t);
2485         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2486                 return avg_time - spent;
2487
2488         td_io_u_lock(td);
2489
2490         /*
2491          * Compute both read and write rates for the interval.
2492          */
2493         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2494                 uint64_t delta;
2495
2496                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2497                 if (!delta)
2498                         continue; /* No entries for interval */
2499
2500                 if (spent) {
2501                         if (is_kb)
2502                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2503                         else
2504                                 rate = (delta * 1000) / spent;
2505                 } else
2506                         rate = 0;
2507
2508                 add_stat_sample(&stat[ddir], rate);
2509
2510                 if (log) {
2511                         unsigned int bs = 0;
2512
2513                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2514                                 bs = td->o.min_bs[ddir];
2515
2516                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2517                         next_log = min(next_log, next);
2518                 }
2519
2520                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2521         }
2522
2523         timespec_add_msec(parent_tv, avg_time);
2524
2525         td_io_u_unlock(td);
2526
2527         if (spent <= avg_time)
2528                 next = avg_time;
2529         else
2530                 next = avg_time - (1 + spent - avg_time);
2531
2532         return min(next, next_log);
2533 }
2534
2535 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2536 {
2537         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2538                                 td->this_io_bytes, td->stat_io_bytes,
2539                                 td->ts.bw_stat, td->bw_log, true);
2540 }
2541
2542 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2543                      unsigned int bytes)
2544 {
2545         struct thread_stat *ts = &td->ts;
2546
2547         td_io_u_lock(td);
2548
2549         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2550
2551         if (td->iops_log)
2552                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2553                                bytes, io_u->offset);
2554
2555         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2556         td_io_u_unlock(td);
2557 }
2558
2559 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2560 {
2561         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2562                                 td->this_io_blocks, td->stat_io_blocks,
2563                                 td->ts.iops_stat, td->iops_log, false);
2564 }
2565
2566 /*
2567  * Returns msecs to next event
2568  */
2569 int calc_log_samples(void)
2570 {
2571         struct thread_data *td;
2572         unsigned int next = ~0U, tmp;
2573         struct timespec now;
2574         int i;
2575
2576         fio_gettime(&now, NULL);
2577
2578         for_each_td(td, i) {
2579                 if (!td->o.stats)
2580                         continue;
2581                 if (in_ramp_time(td) ||
2582                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2583                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2584                         continue;
2585                 }
2586                 if (!td->bw_log ||
2587                         (td->bw_log && !per_unit_log(td->bw_log))) {
2588                         tmp = add_bw_samples(td, &now);
2589                         if (tmp < next)
2590                                 next = tmp;
2591                 }
2592                 if (!td->iops_log ||
2593                         (td->iops_log && !per_unit_log(td->iops_log))) {
2594                         tmp = add_iops_samples(td, &now);
2595                         if (tmp < next)
2596                                 next = tmp;
2597                 }
2598         }
2599
2600         return next == ~0U ? 0 : next;
2601 }
2602
2603 void stat_init(void)
2604 {
2605         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2606 }
2607
2608 void stat_exit(void)
2609 {
2610         /*
2611          * When we have the mutex, we know out-of-band access to it
2612          * have ended.
2613          */
2614         fio_mutex_down(stat_mutex);
2615         fio_mutex_remove(stat_mutex);
2616 }
2617
2618 /*
2619  * Called from signal handler. Wake up status thread.
2620  */
2621 void show_running_run_stats(void)
2622 {
2623         helper_do_stat();
2624 }
2625
2626 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2627 {
2628         /* Ignore io_u's which span multiple blocks--they will just get
2629          * inaccurate counts. */
2630         int idx = (io_u->offset - io_u->file->file_offset)
2631                         / td->o.bs[DDIR_TRIM];
2632         uint32_t *info = &td->ts.block_infos[idx];
2633         assert(idx < td->ts.nr_block_infos);
2634         return info;
2635 }