9 #include "lib/ieee754.h"
11 #include "lib/getrusage.h"
14 #include "lib/output_buffer.h"
15 #include "helper_thread.h"
19 #define LOG_MSEC_SLACK 1
21 struct fio_sem *stat_sem;
23 void clear_rusage_stat(struct thread_data *td)
25 struct thread_stat *ts = &td->ts;
27 fio_getrusage(&td->ru_start);
28 ts->usr_time = ts->sys_time = 0;
30 ts->minf = ts->majf = 0;
33 void update_rusage_stat(struct thread_data *td)
35 struct thread_stat *ts = &td->ts;
37 fio_getrusage(&td->ru_end);
38 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
39 &td->ru_end.ru_utime);
40 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
41 &td->ru_end.ru_stime);
42 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
43 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
44 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
45 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
47 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
51 * Given a latency, return the index of the corresponding bucket in
52 * the structure tracking percentiles.
54 * (1) find the group (and error bits) that the value (latency)
55 * belongs to by looking at its MSB. (2) find the bucket number in the
56 * group by looking at the index bits.
59 static unsigned int plat_val_to_idx(unsigned long long val)
61 unsigned int msb, error_bits, base, offset, idx;
63 /* Find MSB starting from bit 0 */
67 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
70 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
71 * all bits of the sample as index
73 if (msb <= FIO_IO_U_PLAT_BITS)
76 /* Compute the number of error bits to discard*/
77 error_bits = msb - FIO_IO_U_PLAT_BITS;
79 /* Compute the number of buckets before the group */
80 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
83 * Discard the error bits and apply the mask to find the
84 * index for the buckets in the group
86 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
88 /* Make sure the index does not exceed (array size - 1) */
89 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
90 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
96 * Convert the given index of the bucket array to the value
97 * represented by the bucket
99 static unsigned long long plat_idx_to_val(unsigned int idx)
101 unsigned int error_bits;
102 unsigned long long k, base;
104 assert(idx < FIO_IO_U_PLAT_NR);
106 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
107 * all bits of the sample as index */
108 if (idx < (FIO_IO_U_PLAT_VAL << 1))
111 /* Find the group and compute the minimum value of that group */
112 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
113 base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
115 /* Find its bucket number of the group */
116 k = idx % FIO_IO_U_PLAT_VAL;
118 /* Return the mean of the range of the bucket */
119 return base + ((k + 0.5) * (1 << error_bits));
122 static int double_cmp(const void *a, const void *b)
124 const fio_fp64_t fa = *(const fio_fp64_t *) a;
125 const fio_fp64_t fb = *(const fio_fp64_t *) b;
130 else if (fa.u.f < fb.u.f)
136 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
137 fio_fp64_t *plist, unsigned long long **output,
138 unsigned long long *maxv, unsigned long long *minv)
140 unsigned long long sum = 0;
141 unsigned int len, i, j = 0;
142 unsigned int oval_len = 0;
143 unsigned long long *ovals = NULL;
150 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
157 * Sort the percentile list. Note that it may already be sorted if
158 * we are using the default values, but since it's a short list this
159 * isn't a worry. Also note that this does not work for NaN values.
162 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
165 * Calculate bucket values, note down max and min values
168 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
170 while (sum >= (plist[j].u.f / 100.0 * nr)) {
171 assert(plist[j].u.f <= 100.0);
175 ovals = realloc(ovals, oval_len * sizeof(*ovals));
178 ovals[j] = plat_idx_to_val(i);
179 if (ovals[j] < *minv)
181 if (ovals[j] > *maxv)
184 is_last = (j == len - 1) != 0;
197 * Find and display the p-th percentile of clat
199 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
200 fio_fp64_t *plist, unsigned int precision,
201 const char *pre, struct buf_output *out)
203 unsigned int divisor, len, i, j = 0;
204 unsigned long long minv, maxv;
205 unsigned long long *ovals;
206 int per_line, scale_down, time_width;
210 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
215 * We default to nsecs, but if the value range is such that we
216 * should scale down to usecs or msecs, do that.
218 if (minv > 2000000 && maxv > 99999999ULL) {
221 log_buf(out, " %s percentiles (msec):\n |", pre);
222 } else if (minv > 2000 && maxv > 99999) {
225 log_buf(out, " %s percentiles (usec):\n |", pre);
229 log_buf(out, " %s percentiles (nsec):\n |", pre);
233 time_width = max(5, (int) (log10(maxv / divisor) + 1));
234 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
235 precision, time_width);
236 /* fmt will be something like " %5.2fth=[%4llu]%c" */
237 per_line = (80 - 7) / (precision + 10 + time_width);
239 for (j = 0; j < len; j++) {
241 if (j != 0 && (j % per_line) == 0)
244 /* end of the list */
245 is_last = (j == len - 1) != 0;
247 for (i = 0; i < scale_down; i++)
248 ovals[j] = (ovals[j] + 999) / 1000;
250 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
255 if ((j % per_line) == per_line - 1) /* for formatting */
264 bool calc_lat(struct io_stat *is, unsigned long long *min,
265 unsigned long long *max, double *mean, double *dev)
267 double n = (double) is->samples;
274 *mean = is->mean.u.f;
277 *dev = sqrt(is->S.u.f / (n - 1.0));
284 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
286 char *io, *agg, *min, *max;
287 char *ioalt, *aggalt, *minalt, *maxalt;
288 const char *str[] = { " READ", " WRITE" , " TRIM"};
291 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
293 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
294 const int i2p = is_power_of_2(rs->kb_base);
299 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
300 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
301 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
302 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
303 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
304 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
305 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
306 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
307 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
308 rs->unified_rw_rep ? " MIXED" : str[i],
309 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
310 (unsigned long long) rs->min_run[i],
311 (unsigned long long) rs->max_run[i]);
324 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
329 * Do depth distribution calculations
331 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
333 io_u_dist[i] = (double) map[i] / (double) total;
334 io_u_dist[i] *= 100.0;
335 if (io_u_dist[i] < 0.1 && map[i])
342 static void stat_calc_lat(struct thread_stat *ts, double *dst,
343 uint64_t *src, int nr)
345 unsigned long total = ddir_rw_sum(ts->total_io_u);
349 * Do latency distribution calculations
351 for (i = 0; i < nr; i++) {
353 dst[i] = (double) src[i] / (double) total;
355 if (dst[i] < 0.01 && src[i])
363 * To keep the terse format unaltered, add all of the ns latency
364 * buckets to the first us latency bucket
366 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
368 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
371 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
373 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
374 ntotal += ts->io_u_lat_n[i];
376 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
379 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
381 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
384 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
386 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
389 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
391 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
394 static void display_lat(const char *name, unsigned long long min,
395 unsigned long long max, double mean, double dev,
396 struct buf_output *out)
398 const char *base = "(nsec)";
401 if (nsec_to_msec(&min, &max, &mean, &dev))
403 else if (nsec_to_usec(&min, &max, &mean, &dev))
406 minp = num2str(min, 6, 1, 0, N2S_NONE);
407 maxp = num2str(max, 6, 1, 0, N2S_NONE);
409 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
410 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
416 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
417 int ddir, struct buf_output *out)
420 unsigned long long min, max, bw, iops;
422 char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
425 if (ddir_sync(ddir)) {
426 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
427 log_buf(out, " %s:\n", "fsync/fdatasync/sync_file_range");
428 display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
429 show_clat_percentiles(ts->io_u_sync_plat,
430 ts->sync_stat.samples,
432 ts->percentile_precision,
433 io_ddir_name(ddir), out);
438 assert(ddir_rw(ddir));
440 if (!ts->runtime[ddir])
443 i2p = is_power_of_2(rs->kb_base);
444 runt = ts->runtime[ddir];
446 bw = (1000 * ts->io_bytes[ddir]) / runt;
447 io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
448 bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
449 bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
451 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
452 iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
453 if (ddir == DDIR_WRITE)
454 post_st = zbd_write_status(ts);
455 else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
459 total = ts->cachehit + ts->cachemiss;
460 hit = (double) ts->cachehit / (double) total;
462 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
466 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
467 rs->unified_rw_rep ? "mixed" : io_ddir_name(ddir),
468 iops_p, bw_p, bw_p_alt, io_p,
469 (unsigned long long) ts->runtime[ddir],
478 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
479 display_lat("slat", min, max, mean, dev, out);
480 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
481 display_lat("clat", min, max, mean, dev, out);
482 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
483 display_lat(" lat", min, max, mean, dev, out);
485 if (ts->clat_percentiles || ts->lat_percentiles) {
486 const char *name = ts->clat_percentiles ? "clat" : " lat";
489 if (ts->clat_percentiles)
490 samples = ts->clat_stat[ddir].samples;
492 samples = ts->lat_stat[ddir].samples;
494 show_clat_percentiles(ts->io_u_plat[ddir],
497 ts->percentile_precision, name, out);
499 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
500 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
503 if ((rs->unit_base == 1) && i2p)
505 else if (rs->unit_base == 1)
513 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
514 if (p_of_agg > 100.0)
518 if (rs->unit_base == 1) {
525 if (mean > fkb_base * fkb_base) {
530 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
533 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
534 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
535 bw_str, min, max, p_of_agg, mean, dev,
536 (&ts->bw_stat[ddir])->samples);
538 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
539 log_buf(out, " iops : min=%5llu, max=%5llu, "
540 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
541 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
545 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
546 const char *msg, struct buf_output *out)
548 bool new_line = true, shown = false;
551 for (i = 0; i < nr; i++) {
552 if (io_u_lat[i] <= 0.0)
558 log_buf(out, " lat (%s) : ", msg);
564 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
576 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
578 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
579 "250=", "500=", "750=", "1000=", };
581 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
584 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
586 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
587 "250=", "500=", "750=", "1000=", };
589 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
592 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
594 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
595 "250=", "500=", "750=", "1000=", "2000=",
598 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
601 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
603 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
604 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
605 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
607 stat_calc_lat_n(ts, io_u_lat_n);
608 stat_calc_lat_u(ts, io_u_lat_u);
609 stat_calc_lat_m(ts, io_u_lat_m);
611 show_lat_n(io_u_lat_n, out);
612 show_lat_u(io_u_lat_u, out);
613 show_lat_m(io_u_lat_m, out);
616 static int block_state_category(int block_state)
618 switch (block_state) {
619 case BLOCK_STATE_UNINIT:
621 case BLOCK_STATE_TRIMMED:
622 case BLOCK_STATE_WRITTEN:
624 case BLOCK_STATE_WRITE_FAILURE:
625 case BLOCK_STATE_TRIM_FAILURE:
628 /* Silence compile warning on some BSDs and have a return */
634 static int compare_block_infos(const void *bs1, const void *bs2)
636 uint64_t block1 = *(uint64_t *)bs1;
637 uint64_t block2 = *(uint64_t *)bs2;
638 int state1 = BLOCK_INFO_STATE(block1);
639 int state2 = BLOCK_INFO_STATE(block2);
640 int bscat1 = block_state_category(state1);
641 int bscat2 = block_state_category(state2);
642 int cycles1 = BLOCK_INFO_TRIMS(block1);
643 int cycles2 = BLOCK_INFO_TRIMS(block2);
650 if (cycles1 < cycles2)
652 if (cycles1 > cycles2)
660 assert(block1 == block2);
664 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
665 fio_fp64_t *plist, unsigned int **percentiles,
671 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
673 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
680 * Sort the percentile list. Note that it may already be sorted if
681 * we are using the default values, but since it's a short list this
682 * isn't a worry. Also note that this does not work for NaN values.
685 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
687 /* Start only after the uninit entries end */
689 nr_uninit < nr_block_infos
690 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
694 if (nr_uninit == nr_block_infos)
697 *percentiles = calloc(len, sizeof(**percentiles));
699 for (i = 0; i < len; i++) {
700 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
702 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
705 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
706 for (i = 0; i < nr_block_infos; i++)
707 types[BLOCK_INFO_STATE(block_infos[i])]++;
712 static const char *block_state_names[] = {
713 [BLOCK_STATE_UNINIT] = "unwritten",
714 [BLOCK_STATE_TRIMMED] = "trimmed",
715 [BLOCK_STATE_WRITTEN] = "written",
716 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
717 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
720 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
721 fio_fp64_t *plist, struct buf_output *out)
724 unsigned int *percentiles = NULL;
725 unsigned int block_state_counts[BLOCK_STATE_COUNT];
727 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
728 &percentiles, block_state_counts);
730 log_buf(out, " block lifetime percentiles :\n |");
732 for (i = 0; i < len; i++) {
733 uint32_t block_info = percentiles[i];
734 #define LINE_LENGTH 75
735 char str[LINE_LENGTH];
736 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
737 plist[i].u.f, block_info,
738 i == len - 1 ? '\n' : ',');
739 assert(strln < LINE_LENGTH);
740 if (pos + strln > LINE_LENGTH) {
742 log_buf(out, "\n |");
744 log_buf(out, "%s", str);
751 log_buf(out, " states :");
752 for (i = 0; i < BLOCK_STATE_COUNT; i++)
753 log_buf(out, " %s=%u%c",
754 block_state_names[i], block_state_counts[i],
755 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
758 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
760 char *p1, *p1alt, *p2;
761 unsigned long long bw_mean, iops_mean;
762 const int i2p = is_power_of_2(ts->kb_base);
767 bw_mean = steadystate_bw_mean(ts);
768 iops_mean = steadystate_iops_mean(ts);
770 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
771 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
772 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
774 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
775 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
777 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
778 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
779 ts->ss_criterion.u.f,
780 ts->ss_state & FIO_SS_PCT ? "%" : "");
787 static void show_thread_status_normal(struct thread_stat *ts,
788 struct group_run_stats *rs,
789 struct buf_output *out)
791 double usr_cpu, sys_cpu;
792 unsigned long runtime;
793 double io_u_dist[FIO_IO_U_MAP_NR];
797 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
800 memset(time_buf, 0, sizeof(time_buf));
803 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
806 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
807 ts->name, ts->groupid, ts->members,
808 ts->error, (int) ts->pid, time_buf);
810 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
811 ts->name, ts->groupid, ts->members,
812 ts->error, ts->verror, (int) ts->pid,
816 if (strlen(ts->description))
817 log_buf(out, " Description : [%s]\n", ts->description);
819 if (ts->io_bytes[DDIR_READ])
820 show_ddir_status(rs, ts, DDIR_READ, out);
821 if (ts->io_bytes[DDIR_WRITE])
822 show_ddir_status(rs, ts, DDIR_WRITE, out);
823 if (ts->io_bytes[DDIR_TRIM])
824 show_ddir_status(rs, ts, DDIR_TRIM, out);
826 show_latencies(ts, out);
828 if (ts->sync_stat.samples)
829 show_ddir_status(rs, ts, DDIR_SYNC, out);
831 runtime = ts->total_run_time;
833 double runt = (double) runtime;
835 usr_cpu = (double) ts->usr_time * 100 / runt;
836 sys_cpu = (double) ts->sys_time * 100 / runt;
842 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
843 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
844 (unsigned long long) ts->ctx,
845 (unsigned long long) ts->majf,
846 (unsigned long long) ts->minf);
848 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
849 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
850 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
851 io_u_dist[1], io_u_dist[2],
852 io_u_dist[3], io_u_dist[4],
853 io_u_dist[5], io_u_dist[6]);
855 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
856 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
857 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
858 io_u_dist[1], io_u_dist[2],
859 io_u_dist[3], io_u_dist[4],
860 io_u_dist[5], io_u_dist[6]);
861 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
862 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
863 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
864 io_u_dist[1], io_u_dist[2],
865 io_u_dist[3], io_u_dist[4],
866 io_u_dist[5], io_u_dist[6]);
867 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
868 " short=%llu,%llu,%llu,0"
869 " dropped=%llu,%llu,%llu,0\n",
870 (unsigned long long) ts->total_io_u[0],
871 (unsigned long long) ts->total_io_u[1],
872 (unsigned long long) ts->total_io_u[2],
873 (unsigned long long) ts->total_io_u[3],
874 (unsigned long long) ts->short_io_u[0],
875 (unsigned long long) ts->short_io_u[1],
876 (unsigned long long) ts->short_io_u[2],
877 (unsigned long long) ts->drop_io_u[0],
878 (unsigned long long) ts->drop_io_u[1],
879 (unsigned long long) ts->drop_io_u[2]);
880 if (ts->continue_on_error) {
881 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
882 (unsigned long long)ts->total_err_count,
884 strerror(ts->first_error));
886 if (ts->latency_depth) {
887 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
888 (unsigned long long)ts->latency_target,
889 (unsigned long long)ts->latency_window,
890 ts->latency_percentile.u.f,
894 if (ts->nr_block_infos)
895 show_block_infos(ts->nr_block_infos, ts->block_infos,
896 ts->percentile_list, out);
899 show_ss_normal(ts, out);
902 static void show_ddir_status_terse(struct thread_stat *ts,
903 struct group_run_stats *rs, int ddir,
904 int ver, struct buf_output *out)
906 unsigned long long min, max, minv, maxv, bw, iops;
907 unsigned long long *ovals = NULL;
912 assert(ddir_rw(ddir));
915 if (ts->runtime[ddir]) {
916 uint64_t runt = ts->runtime[ddir];
918 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
919 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
922 log_buf(out, ";%llu;%llu;%llu;%llu",
923 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
924 (unsigned long long) ts->runtime[ddir]);
926 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
927 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
929 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
931 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
932 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
934 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
936 if (ts->clat_percentiles || ts->lat_percentiles) {
937 len = calc_clat_percentiles(ts->io_u_plat[ddir],
938 ts->clat_stat[ddir].samples,
939 ts->percentile_list, &ovals, &maxv,
944 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
946 log_buf(out, ";0%%=0");
949 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
952 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
953 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
955 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
960 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
962 double p_of_agg = 100.0;
965 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
966 if (p_of_agg > 100.0)
970 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
972 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
976 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
978 log_buf(out, ";%lu", 0UL);
980 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
981 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
982 mean, dev, (&ts->iops_stat[ddir])->samples);
984 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
988 static void add_ddir_status_json(struct thread_stat *ts,
989 struct group_run_stats *rs, int ddir, struct json_object *parent)
991 unsigned long long min, max, minv, maxv;
992 unsigned long long bw_bytes, bw;
993 unsigned long long *ovals = NULL;
994 double mean, dev, iops;
997 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object = NULL;
999 double p_of_agg = 100.0;
1001 assert(ddir_rw(ddir) || ddir_sync(ddir));
1003 if (ts->unified_rw_rep && ddir != DDIR_READ)
1006 dir_object = json_create_object();
1007 json_object_add_value_object(parent,
1008 ts->unified_rw_rep ? "mixed" : io_ddir_name(ddir), dir_object);
1010 if (ddir_rw(ddir)) {
1014 if (ts->runtime[ddir]) {
1015 uint64_t runt = ts->runtime[ddir];
1017 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1018 bw = bw_bytes / 1024; /* KiB/s */
1019 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1022 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1023 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1024 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1025 json_object_add_value_int(dir_object, "bw", bw);
1026 json_object_add_value_float(dir_object, "iops", iops);
1027 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1028 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1029 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1030 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1032 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
1036 tmp_object = json_create_object();
1037 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1038 json_object_add_value_int(tmp_object, "min", min);
1039 json_object_add_value_int(tmp_object, "max", max);
1040 json_object_add_value_float(tmp_object, "mean", mean);
1041 json_object_add_value_float(tmp_object, "stddev", dev);
1043 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1047 tmp_object = json_create_object();
1048 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1049 json_object_add_value_int(tmp_object, "min", min);
1050 json_object_add_value_int(tmp_object, "max", max);
1051 json_object_add_value_float(tmp_object, "mean", mean);
1052 json_object_add_value_float(tmp_object, "stddev", dev);
1054 if (!calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
1059 tmp_object = json_create_object();
1060 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1061 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1062 json_object_add_value_int(tmp_object, "min", min);
1063 json_object_add_value_int(tmp_object, "max", max);
1064 json_object_add_value_float(tmp_object, "mean", mean);
1065 json_object_add_value_float(tmp_object, "stddev", dev);
1068 if (ts->clat_percentiles || ts->lat_percentiles) {
1069 if (ddir_rw(ddir)) {
1072 if (ts->clat_percentiles)
1073 samples = ts->clat_stat[ddir].samples;
1075 samples = ts->lat_stat[ddir].samples;
1077 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1078 samples, ts->percentile_list, &ovals,
1081 len = calc_clat_percentiles(ts->io_u_sync_plat,
1082 ts->sync_stat.samples,
1083 ts->percentile_list, &ovals, &maxv,
1087 if (len > FIO_IO_U_LIST_MAX_LEN)
1088 len = FIO_IO_U_LIST_MAX_LEN;
1092 percentile_object = json_create_object();
1093 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1094 for (i = 0; i < len; i++) {
1095 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1096 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1099 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1100 clat_bins_object = json_create_object();
1101 if (ts->clat_percentiles)
1102 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1104 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1105 if (ddir_rw(ddir)) {
1106 if (ts->io_u_plat[ddir][i]) {
1107 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1108 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1111 if (ts->io_u_sync_plat[i]) {
1112 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1113 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_sync_plat[i]);
1122 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1126 tmp_object = json_create_object();
1127 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1128 json_object_add_value_int(tmp_object, "min", min);
1129 json_object_add_value_int(tmp_object, "max", max);
1130 json_object_add_value_float(tmp_object, "mean", mean);
1131 json_object_add_value_float(tmp_object, "stddev", dev);
1132 if (output_format & FIO_OUTPUT_JSON_PLUS && ts->lat_percentiles)
1133 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1138 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1139 if (rs->agg[ddir]) {
1140 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1141 if (p_of_agg > 100.0)
1146 p_of_agg = mean = dev = 0.0;
1148 json_object_add_value_int(dir_object, "bw_min", min);
1149 json_object_add_value_int(dir_object, "bw_max", max);
1150 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1151 json_object_add_value_float(dir_object, "bw_mean", mean);
1152 json_object_add_value_float(dir_object, "bw_dev", dev);
1153 json_object_add_value_int(dir_object, "bw_samples",
1154 (&ts->bw_stat[ddir])->samples);
1156 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1160 json_object_add_value_int(dir_object, "iops_min", min);
1161 json_object_add_value_int(dir_object, "iops_max", max);
1162 json_object_add_value_float(dir_object, "iops_mean", mean);
1163 json_object_add_value_float(dir_object, "iops_stddev", dev);
1164 json_object_add_value_int(dir_object, "iops_samples",
1165 (&ts->iops_stat[ddir])->samples);
1167 if (ts->cachehit + ts->cachemiss) {
1171 total = ts->cachehit + ts->cachemiss;
1172 hit = (double) ts->cachehit / (double) total;
1174 json_object_add_value_float(dir_object, "cachehit", hit);
1178 static void show_thread_status_terse_all(struct thread_stat *ts,
1179 struct group_run_stats *rs, int ver,
1180 struct buf_output *out)
1182 double io_u_dist[FIO_IO_U_MAP_NR];
1183 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1184 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1185 double usr_cpu, sys_cpu;
1190 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1192 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1193 ts->name, ts->groupid, ts->error);
1195 /* Log Read Status */
1196 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1197 /* Log Write Status */
1198 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1199 /* Log Trim Status */
1200 if (ver == 2 || ver == 4 || ver == 5)
1201 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1204 if (ts->total_run_time) {
1205 double runt = (double) ts->total_run_time;
1207 usr_cpu = (double) ts->usr_time * 100 / runt;
1208 sys_cpu = (double) ts->sys_time * 100 / runt;
1214 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1215 (unsigned long long) ts->ctx,
1216 (unsigned long long) ts->majf,
1217 (unsigned long long) ts->minf);
1219 /* Calc % distribution of IO depths, usecond, msecond latency */
1220 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1221 stat_calc_lat_nu(ts, io_u_lat_u);
1222 stat_calc_lat_m(ts, io_u_lat_m);
1224 /* Only show fixed 7 I/O depth levels*/
1225 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1226 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1227 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1229 /* Microsecond latency */
1230 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1231 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1232 /* Millisecond latency */
1233 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1234 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1236 /* disk util stats, if any */
1237 if (ver >= 3 && is_running_backend())
1238 show_disk_util(1, NULL, out);
1240 /* Additional output if continue_on_error set - default off*/
1241 if (ts->continue_on_error)
1242 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1246 /* Additional output if description is set */
1247 if (strlen(ts->description))
1248 log_buf(out, ";%s", ts->description);
1253 static void json_add_job_opts(struct json_object *root, const char *name,
1254 struct flist_head *opt_list)
1256 struct json_object *dir_object;
1257 struct flist_head *entry;
1258 struct print_option *p;
1260 if (flist_empty(opt_list))
1263 dir_object = json_create_object();
1264 json_object_add_value_object(root, name, dir_object);
1266 flist_for_each(entry, opt_list) {
1267 const char *pos = "";
1269 p = flist_entry(entry, struct print_option, list);
1272 json_object_add_value_string(dir_object, p->name, pos);
1276 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1277 struct group_run_stats *rs,
1278 struct flist_head *opt_list)
1280 struct json_object *root, *tmp;
1281 struct jobs_eta *je;
1282 double io_u_dist[FIO_IO_U_MAP_NR];
1283 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1284 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1285 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1286 double usr_cpu, sys_cpu;
1290 root = json_create_object();
1291 json_object_add_value_string(root, "jobname", ts->name);
1292 json_object_add_value_int(root, "groupid", ts->groupid);
1293 json_object_add_value_int(root, "error", ts->error);
1296 je = get_jobs_eta(true, &size);
1298 json_object_add_value_int(root, "eta", je->eta_sec);
1299 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1303 json_add_job_opts(root, "job options", opt_list);
1305 add_ddir_status_json(ts, rs, DDIR_READ, root);
1306 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1307 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1308 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1311 if (ts->total_run_time) {
1312 double runt = (double) ts->total_run_time;
1314 usr_cpu = (double) ts->usr_time * 100 / runt;
1315 sys_cpu = (double) ts->sys_time * 100 / runt;
1320 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1321 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1322 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1323 json_object_add_value_int(root, "ctx", ts->ctx);
1324 json_object_add_value_int(root, "majf", ts->majf);
1325 json_object_add_value_int(root, "minf", ts->minf);
1327 /* Calc % distribution of IO depths */
1328 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1329 tmp = json_create_object();
1330 json_object_add_value_object(root, "iodepth_level", tmp);
1331 /* Only show fixed 7 I/O depth levels*/
1332 for (i = 0; i < 7; i++) {
1335 snprintf(name, 20, "%d", 1 << i);
1337 snprintf(name, 20, ">=%d", 1 << i);
1338 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1341 /* Calc % distribution of submit IO depths */
1342 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1343 tmp = json_create_object();
1344 json_object_add_value_object(root, "iodepth_submit", tmp);
1345 /* Only show fixed 7 I/O depth levels*/
1346 for (i = 0; i < 7; i++) {
1349 snprintf(name, 20, "0");
1351 snprintf(name, 20, "%d", 1 << (i+1));
1353 snprintf(name, 20, ">=%d", 1 << i);
1354 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1357 /* Calc % distribution of completion IO depths */
1358 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1359 tmp = json_create_object();
1360 json_object_add_value_object(root, "iodepth_complete", tmp);
1361 /* Only show fixed 7 I/O depth levels*/
1362 for (i = 0; i < 7; i++) {
1365 snprintf(name, 20, "0");
1367 snprintf(name, 20, "%d", 1 << (i+1));
1369 snprintf(name, 20, ">=%d", 1 << i);
1370 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1373 /* Calc % distribution of nsecond, usecond, msecond latency */
1374 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1375 stat_calc_lat_n(ts, io_u_lat_n);
1376 stat_calc_lat_u(ts, io_u_lat_u);
1377 stat_calc_lat_m(ts, io_u_lat_m);
1379 /* Nanosecond latency */
1380 tmp = json_create_object();
1381 json_object_add_value_object(root, "latency_ns", tmp);
1382 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1383 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1384 "250", "500", "750", "1000", };
1385 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1387 /* Microsecond latency */
1388 tmp = json_create_object();
1389 json_object_add_value_object(root, "latency_us", tmp);
1390 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1391 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1392 "250", "500", "750", "1000", };
1393 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1395 /* Millisecond latency */
1396 tmp = json_create_object();
1397 json_object_add_value_object(root, "latency_ms", tmp);
1398 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1399 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1400 "250", "500", "750", "1000", "2000",
1402 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1405 /* Additional output if continue_on_error set - default off*/
1406 if (ts->continue_on_error) {
1407 json_object_add_value_int(root, "total_err", ts->total_err_count);
1408 json_object_add_value_int(root, "first_error", ts->first_error);
1411 if (ts->latency_depth) {
1412 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1413 json_object_add_value_int(root, "latency_target", ts->latency_target);
1414 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1415 json_object_add_value_int(root, "latency_window", ts->latency_window);
1418 /* Additional output if description is set */
1419 if (strlen(ts->description))
1420 json_object_add_value_string(root, "desc", ts->description);
1422 if (ts->nr_block_infos) {
1423 /* Block error histogram and types */
1425 unsigned int *percentiles = NULL;
1426 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1428 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1429 ts->percentile_list,
1430 &percentiles, block_state_counts);
1433 struct json_object *block, *percentile_object, *states;
1435 block = json_create_object();
1436 json_object_add_value_object(root, "block", block);
1438 percentile_object = json_create_object();
1439 json_object_add_value_object(block, "percentiles",
1441 for (i = 0; i < len; i++) {
1443 snprintf(buf, sizeof(buf), "%f",
1444 ts->percentile_list[i].u.f);
1445 json_object_add_value_int(percentile_object,
1450 states = json_create_object();
1451 json_object_add_value_object(block, "states", states);
1452 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1453 json_object_add_value_int(states,
1454 block_state_names[state],
1455 block_state_counts[state]);
1462 struct json_object *data;
1463 struct json_array *iops, *bw;
1467 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1468 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1469 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1470 (float) ts->ss_limit.u.f,
1471 ts->ss_state & FIO_SS_PCT ? "%" : "");
1473 tmp = json_create_object();
1474 json_object_add_value_object(root, "steadystate", tmp);
1475 json_object_add_value_string(tmp, "ss", ss_buf);
1476 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1477 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1479 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1480 ts->ss_state & FIO_SS_PCT ? "%" : "");
1481 json_object_add_value_string(tmp, "criterion", ss_buf);
1482 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1483 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1485 data = json_create_object();
1486 json_object_add_value_object(tmp, "data", data);
1487 bw = json_create_array();
1488 iops = json_create_array();
1491 ** if ss was attained or the buffer is not full,
1492 ** ss->head points to the first element in the list.
1493 ** otherwise it actually points to the second element
1496 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1499 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1500 for (l = 0; l < ts->ss_dur; l++) {
1501 k = (j + l) % ts->ss_dur;
1502 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1503 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1505 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1506 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1507 json_object_add_value_array(data, "iops", iops);
1508 json_object_add_value_array(data, "bw", bw);
1514 static void show_thread_status_terse(struct thread_stat *ts,
1515 struct group_run_stats *rs,
1516 struct buf_output *out)
1518 if (terse_version >= 2 && terse_version <= 5)
1519 show_thread_status_terse_all(ts, rs, terse_version, out);
1521 log_err("fio: bad terse version!? %d\n", terse_version);
1524 struct json_object *show_thread_status(struct thread_stat *ts,
1525 struct group_run_stats *rs,
1526 struct flist_head *opt_list,
1527 struct buf_output *out)
1529 struct json_object *ret = NULL;
1531 if (output_format & FIO_OUTPUT_TERSE)
1532 show_thread_status_terse(ts, rs, out);
1533 if (output_format & FIO_OUTPUT_JSON)
1534 ret = show_thread_status_json(ts, rs, opt_list);
1535 if (output_format & FIO_OUTPUT_NORMAL)
1536 show_thread_status_normal(ts, rs, out);
1541 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1545 dst->min_val = min(dst->min_val, src->min_val);
1546 dst->max_val = max(dst->max_val, src->max_val);
1549 * Compute new mean and S after the merge
1550 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1551 * #Parallel_algorithm>
1554 mean = src->mean.u.f;
1557 double delta = src->mean.u.f - dst->mean.u.f;
1559 mean = ((src->mean.u.f * src->samples) +
1560 (dst->mean.u.f * dst->samples)) /
1561 (dst->samples + src->samples);
1563 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1564 (dst->samples * src->samples) /
1565 (dst->samples + src->samples);
1568 dst->samples += src->samples;
1569 dst->mean.u.f = mean;
1575 * We sum two kinds of stats - one that is time based, in which case we
1576 * apply the proper summing technique, and then one that is iops/bw
1577 * numbers. For group_reporting, we should just add those up, not make
1578 * them the mean of everything.
1580 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first,
1583 if (src->samples == 0)
1587 __sum_stat(dst, src, first);
1592 dst->min_val = src->min_val;
1593 dst->max_val = src->max_val;
1594 dst->samples = src->samples;
1595 dst->mean.u.f = src->mean.u.f;
1596 dst->S.u.f = src->S.u.f;
1598 dst->min_val += src->min_val;
1599 dst->max_val += src->max_val;
1600 dst->samples += src->samples;
1601 dst->mean.u.f += src->mean.u.f;
1602 dst->S.u.f += src->S.u.f;
1606 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1610 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1611 if (dst->max_run[i] < src->max_run[i])
1612 dst->max_run[i] = src->max_run[i];
1613 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1614 dst->min_run[i] = src->min_run[i];
1615 if (dst->max_bw[i] < src->max_bw[i])
1616 dst->max_bw[i] = src->max_bw[i];
1617 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1618 dst->min_bw[i] = src->min_bw[i];
1620 dst->iobytes[i] += src->iobytes[i];
1621 dst->agg[i] += src->agg[i];
1625 dst->kb_base = src->kb_base;
1626 if (!dst->unit_base)
1627 dst->unit_base = src->unit_base;
1629 dst->sig_figs = src->sig_figs;
1632 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1637 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1638 if (!dst->unified_rw_rep) {
1639 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first, false);
1640 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first, false);
1641 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first, false);
1642 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first, true);
1643 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first, true);
1645 dst->io_bytes[l] += src->io_bytes[l];
1647 if (dst->runtime[l] < src->runtime[l])
1648 dst->runtime[l] = src->runtime[l];
1650 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first, false);
1651 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first, false);
1652 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first, false);
1653 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first, true);
1654 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first, true);
1656 dst->io_bytes[0] += src->io_bytes[l];
1658 if (dst->runtime[0] < src->runtime[l])
1659 dst->runtime[0] = src->runtime[l];
1662 * We're summing to the same destination, so override
1663 * 'first' after the first iteration of the loop
1669 sum_stat(&dst->sync_stat, &src->sync_stat, first, false);
1670 dst->usr_time += src->usr_time;
1671 dst->sys_time += src->sys_time;
1672 dst->ctx += src->ctx;
1673 dst->majf += src->majf;
1674 dst->minf += src->minf;
1676 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
1677 dst->io_u_map[k] += src->io_u_map[k];
1678 dst->io_u_submit[k] += src->io_u_submit[k];
1679 dst->io_u_complete[k] += src->io_u_complete[k];
1681 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++) {
1682 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1683 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1684 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1686 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1687 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
1689 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1690 if (!dst->unified_rw_rep) {
1691 dst->total_io_u[k] += src->total_io_u[k];
1692 dst->short_io_u[k] += src->short_io_u[k];
1693 dst->drop_io_u[k] += src->drop_io_u[k];
1695 dst->total_io_u[0] += src->total_io_u[k];
1696 dst->short_io_u[0] += src->short_io_u[k];
1697 dst->drop_io_u[0] += src->drop_io_u[k];
1701 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
1703 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1706 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1707 if (!dst->unified_rw_rep)
1708 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1710 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1714 dst->total_run_time += src->total_run_time;
1715 dst->total_submit += src->total_submit;
1716 dst->total_complete += src->total_complete;
1717 dst->nr_zone_resets += src->nr_zone_resets;
1718 dst->cachehit += src->cachehit;
1719 dst->cachemiss += src->cachemiss;
1722 void init_group_run_stat(struct group_run_stats *gs)
1725 memset(gs, 0, sizeof(*gs));
1727 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1728 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1731 void init_thread_stat(struct thread_stat *ts)
1735 memset(ts, 0, sizeof(*ts));
1737 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1738 ts->lat_stat[j].min_val = -1UL;
1739 ts->clat_stat[j].min_val = -1UL;
1740 ts->slat_stat[j].min_val = -1UL;
1741 ts->bw_stat[j].min_val = -1UL;
1742 ts->iops_stat[j].min_val = -1UL;
1744 ts->sync_stat.min_val = -1UL;
1748 void __show_run_stats(void)
1750 struct group_run_stats *runstats, *rs;
1751 struct thread_data *td;
1752 struct thread_stat *threadstats, *ts;
1753 int i, j, k, nr_ts, last_ts, idx;
1754 bool kb_base_warned = false;
1755 bool unit_base_warned = false;
1756 struct json_object *root = NULL;
1757 struct json_array *array = NULL;
1758 struct buf_output output[FIO_OUTPUT_NR];
1759 struct flist_head **opt_lists;
1761 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1763 for (i = 0; i < groupid + 1; i++)
1764 init_group_run_stat(&runstats[i]);
1767 * find out how many threads stats we need. if group reporting isn't
1768 * enabled, it's one-per-td.
1772 for_each_td(td, i) {
1773 if (!td->o.group_reporting) {
1777 if (last_ts == td->groupid)
1782 last_ts = td->groupid;
1786 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1787 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1789 for (i = 0; i < nr_ts; i++) {
1790 init_thread_stat(&threadstats[i]);
1791 opt_lists[i] = NULL;
1797 for_each_td(td, i) {
1800 if (idx && (!td->o.group_reporting ||
1801 (td->o.group_reporting && last_ts != td->groupid))) {
1806 last_ts = td->groupid;
1808 ts = &threadstats[j];
1810 ts->clat_percentiles = td->o.clat_percentiles;
1811 ts->lat_percentiles = td->o.lat_percentiles;
1812 ts->percentile_precision = td->o.percentile_precision;
1813 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1814 opt_lists[j] = &td->opt_list;
1819 if (ts->groupid == -1) {
1821 * These are per-group shared already
1823 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1824 if (td->o.description)
1825 strncpy(ts->description, td->o.description,
1826 FIO_JOBDESC_SIZE - 1);
1828 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1831 * If multiple entries in this group, this is
1834 ts->thread_number = td->thread_number;
1835 ts->groupid = td->groupid;
1838 * first pid in group, not very useful...
1842 ts->kb_base = td->o.kb_base;
1843 ts->unit_base = td->o.unit_base;
1844 ts->sig_figs = td->o.sig_figs;
1845 ts->unified_rw_rep = td->o.unified_rw_rep;
1846 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1847 log_info("fio: kb_base differs for jobs in group, using"
1848 " %u as the base\n", ts->kb_base);
1849 kb_base_warned = true;
1850 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1851 log_info("fio: unit_base differs for jobs in group, using"
1852 " %u as the base\n", ts->unit_base);
1853 unit_base_warned = true;
1856 ts->continue_on_error = td->o.continue_on_error;
1857 ts->total_err_count += td->total_err_count;
1858 ts->first_error = td->first_error;
1860 if (!td->error && td->o.continue_on_error &&
1862 ts->error = td->first_error;
1863 ts->verror[sizeof(ts->verror) - 1] = '\0';
1864 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1865 } else if (td->error) {
1866 ts->error = td->error;
1867 ts->verror[sizeof(ts->verror) - 1] = '\0';
1868 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1872 ts->latency_depth = td->latency_qd;
1873 ts->latency_target = td->o.latency_target;
1874 ts->latency_percentile = td->o.latency_percentile;
1875 ts->latency_window = td->o.latency_window;
1877 ts->nr_block_infos = td->ts.nr_block_infos;
1878 for (k = 0; k < ts->nr_block_infos; k++)
1879 ts->block_infos[k] = td->ts.block_infos[k];
1881 sum_thread_stats(ts, &td->ts, idx == 1);
1884 ts->ss_state = td->ss.state;
1885 ts->ss_dur = td->ss.dur;
1886 ts->ss_head = td->ss.head;
1887 ts->ss_bw_data = td->ss.bw_data;
1888 ts->ss_iops_data = td->ss.iops_data;
1889 ts->ss_limit.u.f = td->ss.limit;
1890 ts->ss_slope.u.f = td->ss.slope;
1891 ts->ss_deviation.u.f = td->ss.deviation;
1892 ts->ss_criterion.u.f = td->ss.criterion;
1895 ts->ss_dur = ts->ss_state = 0;
1898 for (i = 0; i < nr_ts; i++) {
1899 unsigned long long bw;
1901 ts = &threadstats[i];
1902 if (ts->groupid == -1)
1904 rs = &runstats[ts->groupid];
1905 rs->kb_base = ts->kb_base;
1906 rs->unit_base = ts->unit_base;
1907 rs->sig_figs = ts->sig_figs;
1908 rs->unified_rw_rep += ts->unified_rw_rep;
1910 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1911 if (!ts->runtime[j])
1913 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1914 rs->min_run[j] = ts->runtime[j];
1915 if (ts->runtime[j] > rs->max_run[j])
1916 rs->max_run[j] = ts->runtime[j];
1920 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1921 if (bw < rs->min_bw[j])
1923 if (bw > rs->max_bw[j])
1926 rs->iobytes[j] += ts->io_bytes[j];
1930 for (i = 0; i < groupid + 1; i++) {
1935 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1936 if (rs->max_run[ddir])
1937 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1942 for (i = 0; i < FIO_OUTPUT_NR; i++)
1943 buf_output_init(&output[i]);
1946 * don't overwrite last signal output
1948 if (output_format & FIO_OUTPUT_NORMAL)
1949 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1950 if (output_format & FIO_OUTPUT_JSON) {
1951 struct thread_data *global;
1954 unsigned long long ms_since_epoch;
1957 gettimeofday(&now, NULL);
1958 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1959 (unsigned long long)(now.tv_usec) / 1000;
1961 tv_sec = now.tv_sec;
1962 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
1963 if (time_buf[strlen(time_buf) - 1] == '\n')
1964 time_buf[strlen(time_buf) - 1] = '\0';
1966 root = json_create_object();
1967 json_object_add_value_string(root, "fio version", fio_version_string);
1968 json_object_add_value_int(root, "timestamp", now.tv_sec);
1969 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1970 json_object_add_value_string(root, "time", time_buf);
1971 global = get_global_options();
1972 json_add_job_opts(root, "global options", &global->opt_list);
1973 array = json_create_array();
1974 json_object_add_value_array(root, "jobs", array);
1978 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1980 for (i = 0; i < nr_ts; i++) {
1981 ts = &threadstats[i];
1982 rs = &runstats[ts->groupid];
1985 fio_server_send_job_options(opt_lists[i], i);
1986 fio_server_send_ts(ts, rs);
1988 if (output_format & FIO_OUTPUT_TERSE)
1989 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1990 if (output_format & FIO_OUTPUT_JSON) {
1991 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1992 json_array_add_value_object(array, tmp);
1994 if (output_format & FIO_OUTPUT_NORMAL)
1995 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1998 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1999 /* disk util stats, if any */
2000 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2002 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2004 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2005 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2006 json_free_object(root);
2009 for (i = 0; i < groupid + 1; i++) {
2014 fio_server_send_gs(rs);
2015 else if (output_format & FIO_OUTPUT_NORMAL)
2016 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2020 fio_server_send_du();
2021 else if (output_format & FIO_OUTPUT_NORMAL) {
2022 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2023 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2026 for (i = 0; i < FIO_OUTPUT_NR; i++) {
2027 struct buf_output *out = &output[i];
2029 log_info_buf(out->buf, out->buflen);
2030 buf_output_free(out);
2033 fio_idle_prof_cleanup();
2041 void __show_running_run_stats(void)
2043 struct thread_data *td;
2044 unsigned long long *rt;
2048 fio_sem_down(stat_sem);
2050 rt = malloc(thread_number * sizeof(unsigned long long));
2051 fio_gettime(&ts, NULL);
2053 for_each_td(td, i) {
2054 td->update_rusage = 1;
2055 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
2056 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
2057 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2058 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2060 rt[i] = mtime_since(&td->start, &ts);
2061 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2062 td->ts.runtime[DDIR_READ] += rt[i];
2063 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2064 td->ts.runtime[DDIR_WRITE] += rt[i];
2065 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2066 td->ts.runtime[DDIR_TRIM] += rt[i];
2069 for_each_td(td, i) {
2070 if (td->runstate >= TD_EXITED)
2072 if (td->rusage_sem) {
2073 td->update_rusage = 1;
2074 fio_sem_down(td->rusage_sem);
2076 td->update_rusage = 0;
2081 for_each_td(td, i) {
2082 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2083 td->ts.runtime[DDIR_READ] -= rt[i];
2084 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2085 td->ts.runtime[DDIR_WRITE] -= rt[i];
2086 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2087 td->ts.runtime[DDIR_TRIM] -= rt[i];
2091 fio_sem_up(stat_sem);
2094 static bool status_interval_init;
2095 static struct timespec status_time;
2096 static bool status_file_disabled;
2098 #define FIO_STATUS_FILE "fio-dump-status"
2100 static int check_status_file(void)
2103 const char *temp_dir;
2104 char fio_status_file_path[PATH_MAX];
2106 if (status_file_disabled)
2109 temp_dir = getenv("TMPDIR");
2110 if (temp_dir == NULL) {
2111 temp_dir = getenv("TEMP");
2112 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2115 if (temp_dir == NULL)
2118 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2120 if (stat(fio_status_file_path, &sb))
2123 if (unlink(fio_status_file_path) < 0) {
2124 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2126 log_err("fio: disabling status file updates\n");
2127 status_file_disabled = true;
2133 void check_for_running_stats(void)
2135 if (status_interval) {
2136 if (!status_interval_init) {
2137 fio_gettime(&status_time, NULL);
2138 status_interval_init = true;
2139 } else if (mtime_since_now(&status_time) >= status_interval) {
2140 show_running_run_stats();
2141 fio_gettime(&status_time, NULL);
2145 if (check_status_file()) {
2146 show_running_run_stats();
2151 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2156 if (data > is->max_val)
2158 if (data < is->min_val)
2161 delta = val - is->mean.u.f;
2163 is->mean.u.f += delta / (is->samples + 1.0);
2164 is->S.u.f += delta * (val - is->mean.u.f);
2171 * Return a struct io_logs, which is added to the tail of the log
2174 static struct io_logs *get_new_log(struct io_log *iolog)
2176 size_t new_size, new_samples;
2177 struct io_logs *cur_log;
2180 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2183 if (!iolog->cur_log_max)
2184 new_samples = DEF_LOG_ENTRIES;
2186 new_samples = iolog->cur_log_max * 2;
2187 if (new_samples > MAX_LOG_ENTRIES)
2188 new_samples = MAX_LOG_ENTRIES;
2191 new_size = new_samples * log_entry_sz(iolog);
2193 cur_log = smalloc(sizeof(*cur_log));
2195 INIT_FLIST_HEAD(&cur_log->list);
2196 cur_log->log = malloc(new_size);
2198 cur_log->nr_samples = 0;
2199 cur_log->max_samples = new_samples;
2200 flist_add_tail(&cur_log->list, &iolog->io_logs);
2201 iolog->cur_log_max = new_samples;
2211 * Add and return a new log chunk, or return current log if big enough
2213 static struct io_logs *regrow_log(struct io_log *iolog)
2215 struct io_logs *cur_log;
2218 if (!iolog || iolog->disabled)
2221 cur_log = iolog_cur_log(iolog);
2223 cur_log = get_new_log(iolog);
2228 if (cur_log->nr_samples < cur_log->max_samples)
2232 * No room for a new sample. If we're compressing on the fly, flush
2233 * out the current chunk
2235 if (iolog->log_gz) {
2236 if (iolog_cur_flush(iolog, cur_log)) {
2237 log_err("fio: failed flushing iolog! Will stop logging.\n");
2243 * Get a new log array, and add to our list
2245 cur_log = get_new_log(iolog);
2247 log_err("fio: failed extending iolog! Will stop logging.\n");
2251 if (!iolog->pending || !iolog->pending->nr_samples)
2255 * Flush pending items to new log
2257 for (i = 0; i < iolog->pending->nr_samples; i++) {
2258 struct io_sample *src, *dst;
2260 src = get_sample(iolog, iolog->pending, i);
2261 dst = get_sample(iolog, cur_log, i);
2262 memcpy(dst, src, log_entry_sz(iolog));
2264 cur_log->nr_samples = iolog->pending->nr_samples;
2266 iolog->pending->nr_samples = 0;
2270 iolog->disabled = true;
2274 void regrow_logs(struct thread_data *td)
2276 regrow_log(td->slat_log);
2277 regrow_log(td->clat_log);
2278 regrow_log(td->clat_hist_log);
2279 regrow_log(td->lat_log);
2280 regrow_log(td->bw_log);
2281 regrow_log(td->iops_log);
2282 td->flags &= ~TD_F_REGROW_LOGS;
2285 static struct io_logs *get_cur_log(struct io_log *iolog)
2287 struct io_logs *cur_log;
2289 cur_log = iolog_cur_log(iolog);
2291 cur_log = get_new_log(iolog);
2296 if (cur_log->nr_samples < cur_log->max_samples)
2300 * Out of space. If we're in IO offload mode, or we're not doing
2301 * per unit logging (hence logging happens outside of the IO thread
2302 * as well), add a new log chunk inline. If we're doing inline
2303 * submissions, flag 'td' as needing a log regrow and we'll take
2304 * care of it on the submission side.
2306 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2307 !per_unit_log(iolog))
2308 return regrow_log(iolog);
2311 iolog->td->flags |= TD_F_REGROW_LOGS;
2313 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2314 return iolog->pending;
2317 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2318 enum fio_ddir ddir, unsigned long long bs,
2319 unsigned long t, uint64_t offset)
2321 struct io_logs *cur_log;
2323 if (iolog->disabled)
2325 if (flist_empty(&iolog->io_logs))
2326 iolog->avg_last[ddir] = t;
2328 cur_log = get_cur_log(iolog);
2330 struct io_sample *s;
2332 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2335 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2336 io_sample_set_ddir(iolog, s, ddir);
2339 if (iolog->log_offset) {
2340 struct io_sample_offset *so = (void *) s;
2342 so->offset = offset;
2345 cur_log->nr_samples++;
2349 iolog->disabled = true;
2352 static inline void reset_io_stat(struct io_stat *ios)
2354 ios->min_val = -1ULL;
2355 ios->max_val = ios->samples = 0;
2356 ios->mean.u.f = ios->S.u.f = 0;
2359 void reset_io_stats(struct thread_data *td)
2361 struct thread_stat *ts = &td->ts;
2364 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2365 reset_io_stat(&ts->clat_stat[i]);
2366 reset_io_stat(&ts->slat_stat[i]);
2367 reset_io_stat(&ts->lat_stat[i]);
2368 reset_io_stat(&ts->bw_stat[i]);
2369 reset_io_stat(&ts->iops_stat[i]);
2371 ts->io_bytes[i] = 0;
2373 ts->total_io_u[i] = 0;
2374 ts->short_io_u[i] = 0;
2375 ts->drop_io_u[i] = 0;
2377 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2378 ts->io_u_plat[i][j] = 0;
2380 ts->io_u_sync_plat[j] = 0;
2384 ts->total_io_u[DDIR_SYNC] = 0;
2386 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2387 ts->io_u_map[i] = 0;
2388 ts->io_u_submit[i] = 0;
2389 ts->io_u_complete[i] = 0;
2392 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2393 ts->io_u_lat_n[i] = 0;
2394 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2395 ts->io_u_lat_u[i] = 0;
2396 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2397 ts->io_u_lat_m[i] = 0;
2399 ts->total_submit = 0;
2400 ts->total_complete = 0;
2401 ts->nr_zone_resets = 0;
2402 ts->cachehit = ts->cachemiss = 0;
2405 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2406 unsigned long elapsed, bool log_max)
2409 * Note an entry in the log. Use the mean from the logged samples,
2410 * making sure to properly round up. Only write a log entry if we
2411 * had actual samples done.
2413 if (iolog->avg_window[ddir].samples) {
2414 union io_sample_data data;
2417 data.val = iolog->avg_window[ddir].max_val;
2419 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2421 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2424 reset_io_stat(&iolog->avg_window[ddir]);
2427 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2432 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2433 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2436 static unsigned long add_log_sample(struct thread_data *td,
2437 struct io_log *iolog,
2438 union io_sample_data data,
2439 enum fio_ddir ddir, unsigned long long bs,
2442 unsigned long elapsed, this_window;
2447 elapsed = mtime_since_now(&td->epoch);
2450 * If no time averaging, just add the log sample.
2452 if (!iolog->avg_msec) {
2453 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2458 * Add the sample. If the time period has passed, then
2459 * add that entry to the log and clear.
2461 add_stat_sample(&iolog->avg_window[ddir], data.val);
2464 * If period hasn't passed, adding the above sample is all we
2467 this_window = elapsed - iolog->avg_last[ddir];
2468 if (elapsed < iolog->avg_last[ddir])
2469 return iolog->avg_last[ddir] - elapsed;
2470 else if (this_window < iolog->avg_msec) {
2471 unsigned long diff = iolog->avg_msec - this_window;
2473 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2477 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2479 iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2480 return iolog->avg_msec;
2483 void finalize_logs(struct thread_data *td, bool unit_logs)
2485 unsigned long elapsed;
2487 elapsed = mtime_since_now(&td->epoch);
2489 if (td->clat_log && unit_logs)
2490 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2491 if (td->slat_log && unit_logs)
2492 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2493 if (td->lat_log && unit_logs)
2494 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2495 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2496 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2497 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2498 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2501 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned long long bs)
2503 struct io_log *iolog;
2508 iolog = agg_io_log[ddir];
2509 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2512 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2514 unsigned int idx = plat_val_to_idx(nsec);
2515 assert(idx < FIO_IO_U_PLAT_NR);
2517 ts->io_u_sync_plat[idx]++;
2518 add_stat_sample(&ts->sync_stat, nsec);
2521 static void add_clat_percentile_sample(struct thread_stat *ts,
2522 unsigned long long nsec, enum fio_ddir ddir)
2524 unsigned int idx = plat_val_to_idx(nsec);
2525 assert(idx < FIO_IO_U_PLAT_NR);
2527 ts->io_u_plat[ddir][idx]++;
2530 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2531 unsigned long long nsec, unsigned long long bs,
2534 const bool needs_lock = td_async_processing(td);
2535 unsigned long elapsed, this_window;
2536 struct thread_stat *ts = &td->ts;
2537 struct io_log *iolog = td->clat_hist_log;
2542 add_stat_sample(&ts->clat_stat[ddir], nsec);
2545 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2548 if (ts->clat_percentiles)
2549 add_clat_percentile_sample(ts, nsec, ddir);
2551 if (iolog && iolog->hist_msec) {
2552 struct io_hist *hw = &iolog->hist_window[ddir];
2555 elapsed = mtime_since_now(&td->epoch);
2557 hw->hist_last = elapsed;
2558 this_window = elapsed - hw->hist_last;
2560 if (this_window >= iolog->hist_msec) {
2561 uint64_t *io_u_plat;
2562 struct io_u_plat_entry *dst;
2565 * Make a byte-for-byte copy of the latency histogram
2566 * stored in td->ts.io_u_plat[ddir], recording it in a
2567 * log sample. Note that the matching call to free() is
2568 * located in iolog.c after printing this sample to the
2571 io_u_plat = (uint64_t *) td->ts.io_u_plat[ddir];
2572 dst = malloc(sizeof(struct io_u_plat_entry));
2573 memcpy(&(dst->io_u_plat), io_u_plat,
2574 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2575 flist_add(&dst->list, &hw->list);
2576 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2580 * Update the last time we recorded as being now, minus
2581 * any drift in time we encountered before actually
2582 * making the record.
2584 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2590 __td_io_u_unlock(td);
2593 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2594 unsigned long usec, unsigned long long bs, uint64_t offset)
2596 const bool needs_lock = td_async_processing(td);
2597 struct thread_stat *ts = &td->ts;
2605 add_stat_sample(&ts->slat_stat[ddir], usec);
2608 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2611 __td_io_u_unlock(td);
2614 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2615 unsigned long long nsec, unsigned long long bs,
2618 const bool needs_lock = td_async_processing(td);
2619 struct thread_stat *ts = &td->ts;
2627 add_stat_sample(&ts->lat_stat[ddir], nsec);
2630 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2633 if (ts->lat_percentiles)
2634 add_clat_percentile_sample(ts, nsec, ddir);
2637 __td_io_u_unlock(td);
2640 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2641 unsigned int bytes, unsigned long long spent)
2643 const bool needs_lock = td_async_processing(td);
2644 struct thread_stat *ts = &td->ts;
2648 rate = (unsigned long) (bytes * 1000000ULL / spent);
2655 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2658 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2659 bytes, io_u->offset);
2661 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2664 __td_io_u_unlock(td);
2667 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2668 struct timespec *t, unsigned int avg_time,
2669 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2670 struct io_stat *stat, struct io_log *log,
2673 const bool needs_lock = td_async_processing(td);
2674 unsigned long spent, rate;
2676 unsigned long next, next_log;
2678 next_log = avg_time;
2680 spent = mtime_since(parent_tv, t);
2681 if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2682 return avg_time - spent;
2688 * Compute both read and write rates for the interval.
2690 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2693 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2695 continue; /* No entries for interval */
2699 rate = delta * 1000 / spent / 1024; /* KiB/s */
2701 rate = (delta * 1000) / spent;
2705 add_stat_sample(&stat[ddir], rate);
2708 unsigned long long bs = 0;
2710 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2711 bs = td->o.min_bs[ddir];
2713 next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2714 next_log = min(next_log, next);
2717 stat_io_bytes[ddir] = this_io_bytes[ddir];
2720 timespec_add_msec(parent_tv, avg_time);
2723 __td_io_u_unlock(td);
2725 if (spent <= avg_time)
2728 next = avg_time - (1 + spent - avg_time);
2730 return min(next, next_log);
2733 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2735 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2736 td->this_io_bytes, td->stat_io_bytes,
2737 td->ts.bw_stat, td->bw_log, true);
2740 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2743 const bool needs_lock = td_async_processing(td);
2744 struct thread_stat *ts = &td->ts;
2749 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2752 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2753 bytes, io_u->offset);
2755 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2758 __td_io_u_unlock(td);
2761 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2763 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2764 td->this_io_blocks, td->stat_io_blocks,
2765 td->ts.iops_stat, td->iops_log, false);
2769 * Returns msecs to next event
2771 int calc_log_samples(void)
2773 struct thread_data *td;
2774 unsigned int next = ~0U, tmp;
2775 struct timespec now;
2778 fio_gettime(&now, NULL);
2780 for_each_td(td, i) {
2783 if (in_ramp_time(td) ||
2784 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2785 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2789 (td->bw_log && !per_unit_log(td->bw_log))) {
2790 tmp = add_bw_samples(td, &now);
2794 if (!td->iops_log ||
2795 (td->iops_log && !per_unit_log(td->iops_log))) {
2796 tmp = add_iops_samples(td, &now);
2802 return next == ~0U ? 0 : next;
2805 void stat_init(void)
2807 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
2810 void stat_exit(void)
2813 * When we have the mutex, we know out-of-band access to it
2816 fio_sem_down(stat_sem);
2817 fio_sem_remove(stat_sem);
2821 * Called from signal handler. Wake up status thread.
2823 void show_running_run_stats(void)
2828 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2830 /* Ignore io_u's which span multiple blocks--they will just get
2831 * inaccurate counts. */
2832 int idx = (io_u->offset - io_u->file->file_offset)
2833 / td->o.bs[DDIR_TRIM];
2834 uint32_t *info = &td->ts.block_infos[idx];
2835 assert(idx < td->ts.nr_block_infos);