12 #include "lib/ieee754.h"
14 #include "lib/getrusage.h"
17 #include "lib/output_buffer.h"
18 #include "helper_thread.h"
21 #define LOG_MSEC_SLACK 10
23 struct fio_mutex *stat_mutex;
25 void clear_rusage_stat(struct thread_data *td)
27 struct thread_stat *ts = &td->ts;
29 fio_getrusage(&td->ru_start);
30 ts->usr_time = ts->sys_time = 0;
32 ts->minf = ts->majf = 0;
35 void update_rusage_stat(struct thread_data *td)
37 struct thread_stat *ts = &td->ts;
39 fio_getrusage(&td->ru_end);
40 ts->usr_time += mtime_since(&td->ru_start.ru_utime,
41 &td->ru_end.ru_utime);
42 ts->sys_time += mtime_since(&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;
49 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
53 * Given a latency, return the index of the corresponding bucket in
54 * the structure tracking percentiles.
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.
61 static unsigned int plat_val_to_idx(unsigned int val)
63 unsigned int msb, error_bits, base, offset, idx;
65 /* Find MSB starting from bit 0 */
69 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
72 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
73 * all bits of the sample as index
75 if (msb <= FIO_IO_U_PLAT_BITS)
78 /* Compute the number of error bits to discard*/
79 error_bits = msb - FIO_IO_U_PLAT_BITS;
81 /* Compute the number of buckets before the group */
82 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
85 * Discard the error bits and apply the mask to find the
86 * index for the buckets in the group
88 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
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);
98 * Convert the given index of the bucket array to the value
99 * represented by the bucket
101 static unsigned int plat_idx_to_val(unsigned int idx)
103 unsigned int error_bits, k, base;
105 assert(idx < FIO_IO_U_PLAT_NR);
107 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
108 * all bits of the sample as index */
109 if (idx < (FIO_IO_U_PLAT_VAL << 1))
112 /* Find the group and compute the minimum value of that group */
113 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
114 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
116 /* Find its bucket number of the group */
117 k = idx % FIO_IO_U_PLAT_VAL;
119 /* Return the mean of the range of the bucket */
120 return base + ((k + 0.5) * (1 << error_bits));
123 static int double_cmp(const void *a, const void *b)
125 const fio_fp64_t fa = *(const fio_fp64_t *) a;
126 const fio_fp64_t fb = *(const fio_fp64_t *) b;
131 else if (fa.u.f < fb.u.f)
137 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
138 fio_fp64_t *plist, unsigned int **output,
139 unsigned int *maxv, unsigned int *minv)
141 unsigned long sum = 0;
142 unsigned int len, i, j = 0;
143 unsigned int oval_len = 0;
144 unsigned int *ovals = NULL;
151 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
158 * Sort the percentile list. Note that it may already be sorted if
159 * we are using the default values, but since it's a short list this
160 * isn't a worry. Also note that this does not work for NaN values.
163 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
166 * Calculate bucket values, note down max and min values
169 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
171 while (sum >= (plist[j].u.f / 100.0 * nr)) {
172 assert(plist[j].u.f <= 100.0);
176 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
179 ovals[j] = plat_idx_to_val(i);
180 if (ovals[j] < *minv)
182 if (ovals[j] > *maxv)
185 is_last = (j == len - 1);
198 * Find and display the p-th percentile of clat
200 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
201 fio_fp64_t *plist, unsigned int precision,
202 struct buf_output *out)
204 unsigned int len, j = 0, minv, maxv;
206 int is_last, per_line, scale_down;
209 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
214 * We default to usecs, but if the value range is such that we
215 * should scale down to msecs, do that.
217 if (minv > 2000 && maxv > 99999) {
219 log_buf(out, " clat percentiles (msec):\n |");
222 log_buf(out, " clat percentiles (usec):\n |");
225 snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
226 per_line = (80 - 7) / (precision + 14);
228 for (j = 0; j < len; j++) {
229 char fbuf[16], *ptr = fbuf;
232 if (j != 0 && (j % per_line) == 0)
235 /* end of the list */
236 is_last = (j == len - 1);
238 if (plist[j].u.f < 10.0)
239 ptr += sprintf(fbuf, " ");
241 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
244 ovals[j] = (ovals[j] + 999) / 1000;
246 log_buf(out, " %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
251 if ((j % per_line) == per_line - 1) /* for formatting */
260 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
261 double *mean, double *dev)
263 double n = (double) is->samples;
270 *mean = is->mean.u.f;
273 *dev = sqrt(is->S.u.f / (n - 1.0));
280 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
282 char *p1, *p2, *p3, *p4;
283 const char *str[] = { " READ", " WRITE" , " TRIM"};
286 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
288 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
289 const int i2p = is_power_of_2(rs->kb_base);
294 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
295 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
296 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
297 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
299 log_buf(out, "%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
300 " mint=%llumsec, maxt=%llumsec\n",
301 rs->unified_rw_rep ? " MIXED" : str[i],
303 (unsigned long long) rs->min_run[i],
304 (unsigned long long) rs->max_run[i]);
313 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
318 * Do depth distribution calculations
320 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
322 io_u_dist[i] = (double) map[i] / (double) total;
323 io_u_dist[i] *= 100.0;
324 if (io_u_dist[i] < 0.1 && map[i])
331 static void stat_calc_lat(struct thread_stat *ts, double *dst,
332 unsigned int *src, int nr)
334 unsigned long total = ddir_rw_sum(ts->total_io_u);
338 * Do latency distribution calculations
340 for (i = 0; i < nr; i++) {
342 dst[i] = (double) src[i] / (double) total;
344 if (dst[i] < 0.01 && src[i])
351 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
353 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
356 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
358 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
361 static void display_lat(const char *name, unsigned long min, unsigned long max,
362 double mean, double dev, struct buf_output *out)
364 const char *base = "(usec)";
367 if (!usec_to_msec(&min, &max, &mean, &dev))
370 minp = num2str(min, 6, 1, 0, 0);
371 maxp = num2str(max, 6, 1, 0, 0);
373 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
374 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
380 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
381 int ddir, struct buf_output *out)
383 const char *str[] = { "read ", "write", "trim" };
384 unsigned long min, max, runt;
385 unsigned long long bw, iops;
387 char *io_p, *bw_p, *iops_p;
390 assert(ddir_rw(ddir));
392 if (!ts->runtime[ddir])
395 i2p = is_power_of_2(rs->kb_base);
396 runt = ts->runtime[ddir];
398 bw = (1000 * ts->io_bytes[ddir]) / runt;
399 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
400 bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
402 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
403 iops_p = num2str(iops, 6, 1, 0, 0);
405 log_buf(out, " %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
406 rs->unified_rw_rep ? "mixed" : str[ddir],
408 (unsigned long long) ts->runtime[ddir]);
414 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
415 display_lat("slat", min, max, mean, dev, out);
416 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
417 display_lat("clat", min, max, mean, dev, out);
418 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
419 display_lat(" lat", min, max, mean, dev, out);
421 if (ts->clat_percentiles) {
422 show_clat_percentiles(ts->io_u_plat[ddir],
423 ts->clat_stat[ddir].samples,
425 ts->percentile_precision, out);
427 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
428 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
429 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
431 if (rs->unit_base == 1) {
439 p_of_agg = mean * 100 / (double) rs->agg[ddir];
440 if (p_of_agg > 100.0)
444 if (mean > fkb_base * fkb_base) {
449 bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
452 log_buf(out, " bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
453 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
454 p_of_agg, mean, dev);
458 static int show_lat(double *io_u_lat, int nr, const char **ranges,
459 const char *msg, struct buf_output *out)
461 int new_line = 1, i, line = 0, shown = 0;
463 for (i = 0; i < nr; i++) {
464 if (io_u_lat[i] <= 0.0)
470 log_buf(out, " lat (%s) : ", msg);
476 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
488 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
490 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
491 "250=", "500=", "750=", "1000=", };
493 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
496 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
498 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
499 "250=", "500=", "750=", "1000=", "2000=",
502 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
505 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
507 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
508 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
510 stat_calc_lat_u(ts, io_u_lat_u);
511 stat_calc_lat_m(ts, io_u_lat_m);
513 show_lat_u(io_u_lat_u, out);
514 show_lat_m(io_u_lat_m, out);
517 static int block_state_category(int block_state)
519 switch (block_state) {
520 case BLOCK_STATE_UNINIT:
522 case BLOCK_STATE_TRIMMED:
523 case BLOCK_STATE_WRITTEN:
525 case BLOCK_STATE_WRITE_FAILURE:
526 case BLOCK_STATE_TRIM_FAILURE:
529 /* Silence compile warning on some BSDs and have a return */
535 static int compare_block_infos(const void *bs1, const void *bs2)
537 uint32_t block1 = *(uint32_t *)bs1;
538 uint32_t block2 = *(uint32_t *)bs2;
539 int state1 = BLOCK_INFO_STATE(block1);
540 int state2 = BLOCK_INFO_STATE(block2);
541 int bscat1 = block_state_category(state1);
542 int bscat2 = block_state_category(state2);
543 int cycles1 = BLOCK_INFO_TRIMS(block1);
544 int cycles2 = BLOCK_INFO_TRIMS(block2);
551 if (cycles1 < cycles2)
553 if (cycles1 > cycles2)
561 assert(block1 == block2);
565 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
566 fio_fp64_t *plist, unsigned int **percentiles,
572 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
574 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
581 * Sort the percentile list. Note that it may already be sorted if
582 * we are using the default values, but since it's a short list this
583 * isn't a worry. Also note that this does not work for NaN values.
586 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
589 /* Start only after the uninit entries end */
591 nr_uninit < nr_block_infos
592 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
596 if (nr_uninit == nr_block_infos)
599 *percentiles = calloc(len, sizeof(**percentiles));
601 for (i = 0; i < len; i++) {
602 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
604 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
607 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
608 for (i = 0; i < nr_block_infos; i++)
609 types[BLOCK_INFO_STATE(block_infos[i])]++;
614 static const char *block_state_names[] = {
615 [BLOCK_STATE_UNINIT] = "unwritten",
616 [BLOCK_STATE_TRIMMED] = "trimmed",
617 [BLOCK_STATE_WRITTEN] = "written",
618 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
619 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
622 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
623 fio_fp64_t *plist, struct buf_output *out)
626 unsigned int *percentiles = NULL;
627 unsigned int block_state_counts[BLOCK_STATE_COUNT];
629 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
630 &percentiles, block_state_counts);
632 log_buf(out, " block lifetime percentiles :\n |");
634 for (i = 0; i < len; i++) {
635 uint32_t block_info = percentiles[i];
636 #define LINE_LENGTH 75
637 char str[LINE_LENGTH];
638 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
639 plist[i].u.f, block_info,
640 i == len - 1 ? '\n' : ',');
641 assert(strln < LINE_LENGTH);
642 if (pos + strln > LINE_LENGTH) {
644 log_buf(out, "\n |");
646 log_buf(out, "%s", str);
653 log_buf(out, " states :");
654 for (i = 0; i < BLOCK_STATE_COUNT; i++)
655 log_buf(out, " %s=%u%c",
656 block_state_names[i], block_state_counts[i],
657 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
660 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
663 unsigned long long bw_mean, iops_mean;
664 const int i2p = is_power_of_2(ts->kb_base);
669 bw_mean = steadystate_bw_mean(ts);
670 iops_mean = steadystate_iops_mean(ts);
672 p1 = num2str(bw_mean / ts->kb_base, 6, ts->kb_base, i2p, ts->unit_base);
673 p2 = num2str(iops_mean, 6, 1, 0, 0);
675 log_buf(out, " steadystate : attained=%s, bw=%s/s, iops=%s, %s%s=%.3f%s\n",
676 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
678 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
679 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
680 ts->ss_criterion.u.f,
681 ts->ss_state & __FIO_SS_PCT ? "%" : "");
687 static void show_thread_status_normal(struct thread_stat *ts,
688 struct group_run_stats *rs,
689 struct buf_output *out)
691 double usr_cpu, sys_cpu;
692 unsigned long runtime;
693 double io_u_dist[FIO_IO_U_MAP_NR];
697 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
701 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
704 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
705 ts->name, ts->groupid, ts->members,
706 ts->error, (int) ts->pid, time_buf);
708 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
709 ts->name, ts->groupid, ts->members,
710 ts->error, ts->verror, (int) ts->pid,
714 if (strlen(ts->description))
715 log_buf(out, " Description : [%s]\n", ts->description);
717 if (ts->io_bytes[DDIR_READ])
718 show_ddir_status(rs, ts, DDIR_READ, out);
719 if (ts->io_bytes[DDIR_WRITE])
720 show_ddir_status(rs, ts, DDIR_WRITE, out);
721 if (ts->io_bytes[DDIR_TRIM])
722 show_ddir_status(rs, ts, DDIR_TRIM, out);
724 show_latencies(ts, out);
726 runtime = ts->total_run_time;
728 double runt = (double) runtime;
730 usr_cpu = (double) ts->usr_time * 100 / runt;
731 sys_cpu = (double) ts->sys_time * 100 / runt;
737 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
738 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
739 (unsigned long long) ts->ctx,
740 (unsigned long long) ts->majf,
741 (unsigned long long) ts->minf);
743 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
744 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
745 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
746 io_u_dist[1], io_u_dist[2],
747 io_u_dist[3], io_u_dist[4],
748 io_u_dist[5], io_u_dist[6]);
750 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
751 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
752 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
753 io_u_dist[1], io_u_dist[2],
754 io_u_dist[3], io_u_dist[4],
755 io_u_dist[5], io_u_dist[6]);
756 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
757 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
758 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
759 io_u_dist[1], io_u_dist[2],
760 io_u_dist[3], io_u_dist[4],
761 io_u_dist[5], io_u_dist[6]);
762 log_buf(out, " issued : total=r=%llu/w=%llu/d=%llu,"
763 " short=r=%llu/w=%llu/d=%llu,"
764 " drop=r=%llu/w=%llu/d=%llu\n",
765 (unsigned long long) ts->total_io_u[0],
766 (unsigned long long) ts->total_io_u[1],
767 (unsigned long long) ts->total_io_u[2],
768 (unsigned long long) ts->short_io_u[0],
769 (unsigned long long) ts->short_io_u[1],
770 (unsigned long long) ts->short_io_u[2],
771 (unsigned long long) ts->drop_io_u[0],
772 (unsigned long long) ts->drop_io_u[1],
773 (unsigned long long) ts->drop_io_u[2]);
774 if (ts->continue_on_error) {
775 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
776 (unsigned long long)ts->total_err_count,
778 strerror(ts->first_error));
780 if (ts->latency_depth) {
781 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
782 (unsigned long long)ts->latency_target,
783 (unsigned long long)ts->latency_window,
784 ts->latency_percentile.u.f,
788 if (ts->nr_block_infos)
789 show_block_infos(ts->nr_block_infos, ts->block_infos,
790 ts->percentile_list, out);
793 show_ss_normal(ts, out);
796 static void show_ddir_status_terse(struct thread_stat *ts,
797 struct group_run_stats *rs, int ddir,
798 struct buf_output *out)
800 unsigned long min, max;
801 unsigned long long bw, iops;
802 unsigned int *ovals = NULL;
804 unsigned int len, minv, maxv;
807 assert(ddir_rw(ddir));
810 if (ts->runtime[ddir]) {
811 uint64_t runt = ts->runtime[ddir];
813 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
814 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
817 log_buf(out, ";%llu;%llu;%llu;%llu",
818 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
819 (unsigned long long) ts->runtime[ddir]);
821 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
822 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
824 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
826 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
827 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
829 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
831 if (ts->clat_percentiles) {
832 len = calc_clat_percentiles(ts->io_u_plat[ddir],
833 ts->clat_stat[ddir].samples,
834 ts->percentile_list, &ovals, &maxv,
839 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
841 log_buf(out, ";0%%=0");
844 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
847 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
848 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
850 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
855 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
856 double p_of_agg = 100.0;
859 p_of_agg = mean * 100 / (double) rs->agg[ddir];
860 if (p_of_agg > 100.0)
864 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
866 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
869 static void add_ddir_status_json(struct thread_stat *ts,
870 struct group_run_stats *rs, int ddir, struct json_object *parent)
872 unsigned long min, max;
873 unsigned long long bw;
874 unsigned int *ovals = NULL;
875 double mean, dev, iops;
876 unsigned int len, minv, maxv;
878 const char *ddirname[] = {"read", "write", "trim"};
879 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
881 double p_of_agg = 100.0;
883 assert(ddir_rw(ddir));
885 if (ts->unified_rw_rep && ddir != DDIR_READ)
888 dir_object = json_create_object();
889 json_object_add_value_object(parent,
890 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
894 if (ts->runtime[ddir]) {
895 uint64_t runt = ts->runtime[ddir];
897 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
898 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
901 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
902 json_object_add_value_int(dir_object, "bw", bw);
903 json_object_add_value_float(dir_object, "iops", iops);
904 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
905 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
906 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
907 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
909 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
913 tmp_object = json_create_object();
914 json_object_add_value_object(dir_object, "slat", tmp_object);
915 json_object_add_value_int(tmp_object, "min", min);
916 json_object_add_value_int(tmp_object, "max", max);
917 json_object_add_value_float(tmp_object, "mean", mean);
918 json_object_add_value_float(tmp_object, "stddev", dev);
920 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
924 tmp_object = json_create_object();
925 json_object_add_value_object(dir_object, "clat", tmp_object);
926 json_object_add_value_int(tmp_object, "min", min);
927 json_object_add_value_int(tmp_object, "max", max);
928 json_object_add_value_float(tmp_object, "mean", mean);
929 json_object_add_value_float(tmp_object, "stddev", dev);
931 if (ts->clat_percentiles) {
932 len = calc_clat_percentiles(ts->io_u_plat[ddir],
933 ts->clat_stat[ddir].samples,
934 ts->percentile_list, &ovals, &maxv,
939 percentile_object = json_create_object();
940 json_object_add_value_object(tmp_object, "percentile", percentile_object);
941 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
943 json_object_add_value_int(percentile_object, "0.00", 0);
946 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
947 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
950 if (output_format & FIO_OUTPUT_JSON_PLUS) {
951 clat_bins_object = json_create_object();
952 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
953 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
954 snprintf(buf, sizeof(buf), "%d", i);
955 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
957 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS);
958 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL);
959 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR);
962 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
966 tmp_object = json_create_object();
967 json_object_add_value_object(dir_object, "lat", tmp_object);
968 json_object_add_value_int(tmp_object, "min", min);
969 json_object_add_value_int(tmp_object, "max", max);
970 json_object_add_value_float(tmp_object, "mean", mean);
971 json_object_add_value_float(tmp_object, "stddev", dev);
975 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
977 p_of_agg = mean * 100 / (double) rs->agg[ddir];
978 if (p_of_agg > 100.0)
983 p_of_agg = mean = dev = 0.0;
985 json_object_add_value_int(dir_object, "bw_min", min);
986 json_object_add_value_int(dir_object, "bw_max", max);
987 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
988 json_object_add_value_float(dir_object, "bw_mean", mean);
989 json_object_add_value_float(dir_object, "bw_dev", dev);
992 static void show_thread_status_terse_v2(struct thread_stat *ts,
993 struct group_run_stats *rs,
994 struct buf_output *out)
996 double io_u_dist[FIO_IO_U_MAP_NR];
997 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
998 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
999 double usr_cpu, sys_cpu;
1003 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1004 /* Log Read Status */
1005 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1006 /* Log Write Status */
1007 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1008 /* Log Trim Status */
1009 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1012 if (ts->total_run_time) {
1013 double runt = (double) ts->total_run_time;
1015 usr_cpu = (double) ts->usr_time * 100 / runt;
1016 sys_cpu = (double) ts->sys_time * 100 / runt;
1022 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1023 (unsigned long long) ts->ctx,
1024 (unsigned long long) ts->majf,
1025 (unsigned long long) ts->minf);
1027 /* Calc % distribution of IO depths, usecond, msecond latency */
1028 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1029 stat_calc_lat_u(ts, io_u_lat_u);
1030 stat_calc_lat_m(ts, io_u_lat_m);
1032 /* Only show fixed 7 I/O depth levels*/
1033 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1034 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1035 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1037 /* Microsecond latency */
1038 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1039 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1040 /* Millisecond latency */
1041 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1042 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1043 /* Additional output if continue_on_error set - default off*/
1044 if (ts->continue_on_error)
1045 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1048 /* Additional output if description is set */
1049 if (strlen(ts->description))
1050 log_buf(out, ";%s", ts->description);
1055 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1056 struct group_run_stats *rs, int ver,
1057 struct buf_output *out)
1059 double io_u_dist[FIO_IO_U_MAP_NR];
1060 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1061 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1062 double usr_cpu, sys_cpu;
1066 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1067 ts->name, ts->groupid, ts->error);
1068 /* Log Read Status */
1069 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1070 /* Log Write Status */
1071 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1072 /* Log Trim Status */
1074 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1077 if (ts->total_run_time) {
1078 double runt = (double) ts->total_run_time;
1080 usr_cpu = (double) ts->usr_time * 100 / runt;
1081 sys_cpu = (double) ts->sys_time * 100 / runt;
1087 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1088 (unsigned long long) ts->ctx,
1089 (unsigned long long) ts->majf,
1090 (unsigned long long) ts->minf);
1092 /* Calc % distribution of IO depths, usecond, msecond latency */
1093 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1094 stat_calc_lat_u(ts, io_u_lat_u);
1095 stat_calc_lat_m(ts, io_u_lat_m);
1097 /* Only show fixed 7 I/O depth levels*/
1098 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1099 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1100 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1102 /* Microsecond latency */
1103 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1104 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1105 /* Millisecond latency */
1106 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1107 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1109 /* disk util stats, if any */
1110 show_disk_util(1, NULL, out);
1112 /* Additional output if continue_on_error set - default off*/
1113 if (ts->continue_on_error)
1114 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1116 /* Additional output if description is set */
1117 if (strlen(ts->description))
1118 log_buf(out, ";%s", ts->description);
1123 void json_add_job_opts(struct json_object *root, const char *name,
1124 struct flist_head *opt_list, bool num_jobs)
1126 struct json_object *dir_object;
1127 struct flist_head *entry;
1128 struct print_option *p;
1130 if (flist_empty(opt_list))
1133 dir_object = json_create_object();
1134 json_object_add_value_object(root, name, dir_object);
1136 flist_for_each(entry, opt_list) {
1137 const char *pos = "";
1139 p = flist_entry(entry, struct print_option, list);
1140 if (!num_jobs && !strcmp(p->name, "numjobs"))
1144 json_object_add_value_string(dir_object, p->name, pos);
1148 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1149 struct group_run_stats *rs,
1150 struct flist_head *opt_list)
1152 struct json_object *root, *tmp;
1153 struct jobs_eta *je;
1154 double io_u_dist[FIO_IO_U_MAP_NR];
1155 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1156 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1157 double usr_cpu, sys_cpu;
1161 root = json_create_object();
1162 json_object_add_value_string(root, "jobname", ts->name);
1163 json_object_add_value_int(root, "groupid", ts->groupid);
1164 json_object_add_value_int(root, "error", ts->error);
1167 je = get_jobs_eta(true, &size);
1169 json_object_add_value_int(root, "eta", je->eta_sec);
1170 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1174 json_add_job_opts(root, "job options", opt_list, true);
1176 add_ddir_status_json(ts, rs, DDIR_READ, root);
1177 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1178 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1181 if (ts->total_run_time) {
1182 double runt = (double) ts->total_run_time;
1184 usr_cpu = (double) ts->usr_time * 100 / runt;
1185 sys_cpu = (double) ts->sys_time * 100 / runt;
1190 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1191 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1192 json_object_add_value_int(root, "ctx", ts->ctx);
1193 json_object_add_value_int(root, "majf", ts->majf);
1194 json_object_add_value_int(root, "minf", ts->minf);
1197 /* Calc % distribution of IO depths, usecond, msecond latency */
1198 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1199 stat_calc_lat_u(ts, io_u_lat_u);
1200 stat_calc_lat_m(ts, io_u_lat_m);
1202 tmp = json_create_object();
1203 json_object_add_value_object(root, "iodepth_level", tmp);
1204 /* Only show fixed 7 I/O depth levels*/
1205 for (i = 0; i < 7; i++) {
1208 snprintf(name, 20, "%d", 1 << i);
1210 snprintf(name, 20, ">=%d", 1 << i);
1211 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1214 tmp = json_create_object();
1215 json_object_add_value_object(root, "latency_us", tmp);
1216 /* Microsecond latency */
1217 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1218 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1219 "250", "500", "750", "1000", };
1220 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1222 /* Millisecond latency */
1223 tmp = json_create_object();
1224 json_object_add_value_object(root, "latency_ms", tmp);
1225 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1226 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1227 "250", "500", "750", "1000", "2000",
1229 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1232 /* Additional output if continue_on_error set - default off*/
1233 if (ts->continue_on_error) {
1234 json_object_add_value_int(root, "total_err", ts->total_err_count);
1235 json_object_add_value_int(root, "first_error", ts->first_error);
1238 if (ts->latency_depth) {
1239 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1240 json_object_add_value_int(root, "latency_target", ts->latency_target);
1241 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1242 json_object_add_value_int(root, "latency_window", ts->latency_window);
1245 /* Additional output if description is set */
1246 if (strlen(ts->description))
1247 json_object_add_value_string(root, "desc", ts->description);
1249 if (ts->nr_block_infos) {
1250 /* Block error histogram and types */
1252 unsigned int *percentiles = NULL;
1253 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1255 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1256 ts->percentile_list,
1257 &percentiles, block_state_counts);
1260 struct json_object *block, *percentile_object, *states;
1262 block = json_create_object();
1263 json_object_add_value_object(root, "block", block);
1265 percentile_object = json_create_object();
1266 json_object_add_value_object(block, "percentiles",
1268 for (i = 0; i < len; i++) {
1270 snprintf(buf, sizeof(buf), "%f",
1271 ts->percentile_list[i].u.f);
1272 json_object_add_value_int(percentile_object,
1277 states = json_create_object();
1278 json_object_add_value_object(block, "states", states);
1279 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1280 json_object_add_value_int(states,
1281 block_state_names[state],
1282 block_state_counts[state]);
1289 struct json_object *data;
1290 struct json_array *iops, *bw;
1294 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1295 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1296 ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1297 (float) ts->ss_limit.u.f,
1298 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1300 tmp = json_create_object();
1301 json_object_add_value_object(root, "steadystate", tmp);
1302 json_object_add_value_string(tmp, "ss", ss_buf);
1303 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1304 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1306 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1307 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1308 json_object_add_value_string(tmp, "criterion", ss_buf);
1309 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1310 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1312 data = json_create_object();
1313 json_object_add_value_object(tmp, "data", data);
1314 bw = json_create_array();
1315 iops = json_create_array();
1318 ** if ss was attained or the buffer is not full,
1319 ** ss->head points to the first element in the list.
1320 ** otherwise it actually points to the second element
1323 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1326 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1327 for (i = 0; i < ts->ss_dur; i++) {
1328 k = (j + i) % ts->ss_dur;
1329 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1330 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1332 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1333 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1334 json_object_add_value_array(data, "iops", iops);
1335 json_object_add_value_array(data, "bw", bw);
1341 static void show_thread_status_terse(struct thread_stat *ts,
1342 struct group_run_stats *rs,
1343 struct buf_output *out)
1345 if (terse_version == 2)
1346 show_thread_status_terse_v2(ts, rs, out);
1347 else if (terse_version == 3 || terse_version == 4)
1348 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1350 log_err("fio: bad terse version!? %d\n", terse_version);
1353 struct json_object *show_thread_status(struct thread_stat *ts,
1354 struct group_run_stats *rs,
1355 struct flist_head *opt_list,
1356 struct buf_output *out)
1358 struct json_object *ret = NULL;
1360 if (output_format & FIO_OUTPUT_TERSE)
1361 show_thread_status_terse(ts, rs, out);
1362 if (output_format & FIO_OUTPUT_JSON)
1363 ret = show_thread_status_json(ts, rs, opt_list);
1364 if (output_format & FIO_OUTPUT_NORMAL)
1365 show_thread_status_normal(ts, rs, out);
1370 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1374 if (src->samples == 0)
1377 dst->min_val = min(dst->min_val, src->min_val);
1378 dst->max_val = max(dst->max_val, src->max_val);
1381 * Compute new mean and S after the merge
1382 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1383 * #Parallel_algorithm>
1386 mean = src->mean.u.f;
1389 double delta = src->mean.u.f - dst->mean.u.f;
1391 mean = ((src->mean.u.f * src->samples) +
1392 (dst->mean.u.f * dst->samples)) /
1393 (dst->samples + src->samples);
1395 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1396 (dst->samples * src->samples) /
1397 (dst->samples + src->samples);
1400 dst->samples += src->samples;
1401 dst->mean.u.f = mean;
1405 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1409 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1410 if (dst->max_run[i] < src->max_run[i])
1411 dst->max_run[i] = src->max_run[i];
1412 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1413 dst->min_run[i] = src->min_run[i];
1414 if (dst->max_bw[i] < src->max_bw[i])
1415 dst->max_bw[i] = src->max_bw[i];
1416 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1417 dst->min_bw[i] = src->min_bw[i];
1419 dst->io_kb[i] += src->io_kb[i];
1420 dst->agg[i] += src->agg[i];
1424 dst->kb_base = src->kb_base;
1425 if (!dst->unit_base)
1426 dst->unit_base = src->unit_base;
1429 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1434 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1435 if (!dst->unified_rw_rep) {
1436 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1437 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1438 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1439 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1441 dst->io_bytes[l] += src->io_bytes[l];
1443 if (dst->runtime[l] < src->runtime[l])
1444 dst->runtime[l] = src->runtime[l];
1446 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1447 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1448 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1449 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1451 dst->io_bytes[0] += src->io_bytes[l];
1453 if (dst->runtime[0] < src->runtime[l])
1454 dst->runtime[0] = src->runtime[l];
1457 * We're summing to the same destination, so override
1458 * 'first' after the first iteration of the loop
1464 dst->usr_time += src->usr_time;
1465 dst->sys_time += src->sys_time;
1466 dst->ctx += src->ctx;
1467 dst->majf += src->majf;
1468 dst->minf += src->minf;
1470 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1471 dst->io_u_map[k] += src->io_u_map[k];
1472 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1473 dst->io_u_submit[k] += src->io_u_submit[k];
1474 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1475 dst->io_u_complete[k] += src->io_u_complete[k];
1476 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1477 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1478 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1479 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1481 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1482 if (!dst->unified_rw_rep) {
1483 dst->total_io_u[k] += src->total_io_u[k];
1484 dst->short_io_u[k] += src->short_io_u[k];
1485 dst->drop_io_u[k] += src->drop_io_u[k];
1487 dst->total_io_u[0] += src->total_io_u[k];
1488 dst->short_io_u[0] += src->short_io_u[k];
1489 dst->drop_io_u[0] += src->drop_io_u[k];
1493 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1496 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1497 if (!dst->unified_rw_rep)
1498 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1500 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1504 dst->total_run_time += src->total_run_time;
1505 dst->total_submit += src->total_submit;
1506 dst->total_complete += src->total_complete;
1509 void init_group_run_stat(struct group_run_stats *gs)
1512 memset(gs, 0, sizeof(*gs));
1514 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1515 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1518 void init_thread_stat(struct thread_stat *ts)
1522 memset(ts, 0, sizeof(*ts));
1524 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1525 ts->lat_stat[j].min_val = -1UL;
1526 ts->clat_stat[j].min_val = -1UL;
1527 ts->slat_stat[j].min_val = -1UL;
1528 ts->bw_stat[j].min_val = -1UL;
1533 void __show_run_stats(void)
1535 struct group_run_stats *runstats, *rs;
1536 struct thread_data *td;
1537 struct thread_stat *threadstats, *ts;
1538 int i, j, k, nr_ts, last_ts, idx;
1539 int kb_base_warned = 0;
1540 int unit_base_warned = 0;
1541 struct json_object *root = NULL;
1542 struct json_array *array = NULL;
1543 struct buf_output output[FIO_OUTPUT_NR];
1544 struct flist_head **opt_lists;
1546 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1548 for (i = 0; i < groupid + 1; i++)
1549 init_group_run_stat(&runstats[i]);
1552 * find out how many threads stats we need. if group reporting isn't
1553 * enabled, it's one-per-td.
1557 for_each_td(td, i) {
1558 if (!td->o.group_reporting) {
1562 if (last_ts == td->groupid)
1565 last_ts = td->groupid;
1569 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1570 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1572 for (i = 0; i < nr_ts; i++) {
1573 init_thread_stat(&threadstats[i]);
1574 opt_lists[i] = NULL;
1580 for_each_td(td, i) {
1581 if (idx && (!td->o.group_reporting ||
1582 (td->o.group_reporting && last_ts != td->groupid))) {
1587 last_ts = td->groupid;
1589 ts = &threadstats[j];
1591 ts->clat_percentiles = td->o.clat_percentiles;
1592 ts->percentile_precision = td->o.percentile_precision;
1593 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1594 opt_lists[j] = &td->opt_list;
1599 if (ts->groupid == -1) {
1601 * These are per-group shared already
1603 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1604 if (td->o.description)
1605 strncpy(ts->description, td->o.description,
1606 FIO_JOBDESC_SIZE - 1);
1608 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1611 * If multiple entries in this group, this is
1614 ts->thread_number = td->thread_number;
1615 ts->groupid = td->groupid;
1618 * first pid in group, not very useful...
1622 ts->kb_base = td->o.kb_base;
1623 ts->unit_base = td->o.unit_base;
1624 ts->unified_rw_rep = td->o.unified_rw_rep;
1625 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1626 log_info("fio: kb_base differs for jobs in group, using"
1627 " %u as the base\n", ts->kb_base);
1629 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1630 log_info("fio: unit_base differs for jobs in group, using"
1631 " %u as the base\n", ts->unit_base);
1632 unit_base_warned = 1;
1635 ts->continue_on_error = td->o.continue_on_error;
1636 ts->total_err_count += td->total_err_count;
1637 ts->first_error = td->first_error;
1639 if (!td->error && td->o.continue_on_error &&
1641 ts->error = td->first_error;
1642 ts->verror[sizeof(ts->verror) - 1] = '\0';
1643 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1644 } else if (td->error) {
1645 ts->error = td->error;
1646 ts->verror[sizeof(ts->verror) - 1] = '\0';
1647 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1651 ts->latency_depth = td->latency_qd;
1652 ts->latency_target = td->o.latency_target;
1653 ts->latency_percentile = td->o.latency_percentile;
1654 ts->latency_window = td->o.latency_window;
1656 ts->nr_block_infos = td->ts.nr_block_infos;
1657 for (k = 0; k < ts->nr_block_infos; k++)
1658 ts->block_infos[k] = td->ts.block_infos[k];
1660 sum_thread_stats(ts, &td->ts, idx == 1);
1663 ts->ss_state = td->ss.state;
1664 ts->ss_dur = td->ss.dur;
1665 ts->ss_head = td->ss.head;
1666 ts->ss_bw_data = td->ss.bw_data;
1667 ts->ss_iops_data = td->ss.iops_data;
1668 ts->ss_limit.u.f = td->ss.limit;
1669 ts->ss_slope.u.f = td->ss.slope;
1670 ts->ss_deviation.u.f = td->ss.deviation;
1671 ts->ss_criterion.u.f = td->ss.criterion;
1674 ts->ss_dur = ts->ss_state = 0;
1677 for (i = 0; i < nr_ts; i++) {
1678 unsigned long long bw;
1680 ts = &threadstats[i];
1681 if (ts->groupid == -1)
1683 rs = &runstats[ts->groupid];
1684 rs->kb_base = ts->kb_base;
1685 rs->unit_base = ts->unit_base;
1686 rs->unified_rw_rep += ts->unified_rw_rep;
1688 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1689 if (!ts->runtime[j])
1691 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1692 rs->min_run[j] = ts->runtime[j];
1693 if (ts->runtime[j] > rs->max_run[j])
1694 rs->max_run[j] = ts->runtime[j];
1697 if (ts->runtime[j]) {
1698 unsigned long runt = ts->runtime[j];
1699 unsigned long long kb;
1701 kb = ts->io_bytes[j] / rs->kb_base;
1702 bw = kb * 1000 / runt;
1704 if (bw < rs->min_bw[j])
1706 if (bw > rs->max_bw[j])
1709 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1713 for (i = 0; i < groupid + 1; i++) {
1718 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1719 if (rs->max_run[ddir])
1720 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1725 for (i = 0; i < FIO_OUTPUT_NR; i++)
1726 buf_output_init(&output[i]);
1729 * don't overwrite last signal output
1731 if (output_format & FIO_OUTPUT_NORMAL)
1732 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1733 if (output_format & FIO_OUTPUT_JSON) {
1734 struct thread_data *global;
1737 unsigned long long ms_since_epoch;
1739 gettimeofday(&now, NULL);
1740 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1741 (unsigned long long)(now.tv_usec) / 1000;
1743 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1745 time_buf[strlen(time_buf) - 1] = '\0';
1747 root = json_create_object();
1748 json_object_add_value_string(root, "fio version", fio_version_string);
1749 json_object_add_value_int(root, "timestamp", now.tv_sec);
1750 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1751 json_object_add_value_string(root, "time", time_buf);
1752 global = get_global_options();
1753 json_add_job_opts(root, "global options", &global->opt_list, false);
1754 array = json_create_array();
1755 json_object_add_value_array(root, "jobs", array);
1759 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1761 for (i = 0; i < nr_ts; i++) {
1762 ts = &threadstats[i];
1763 rs = &runstats[ts->groupid];
1766 fio_server_send_job_options(opt_lists[i], i);
1767 fio_server_send_ts(ts, rs);
1769 if (output_format & FIO_OUTPUT_TERSE)
1770 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1771 if (output_format & FIO_OUTPUT_JSON) {
1772 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1773 json_array_add_value_object(array, tmp);
1775 if (output_format & FIO_OUTPUT_NORMAL)
1776 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1779 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1780 /* disk util stats, if any */
1781 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1783 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1785 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1786 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1787 json_free_object(root);
1790 for (i = 0; i < groupid + 1; i++) {
1795 fio_server_send_gs(rs);
1796 else if (output_format & FIO_OUTPUT_NORMAL)
1797 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1801 fio_server_send_du();
1802 else if (output_format & FIO_OUTPUT_NORMAL) {
1803 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1804 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1807 for (i = 0; i < FIO_OUTPUT_NR; i++) {
1808 buf_output_flush(&output[i]);
1809 buf_output_free(&output[i]);
1818 void show_run_stats(void)
1820 fio_mutex_down(stat_mutex);
1822 fio_mutex_up(stat_mutex);
1825 void __show_running_run_stats(void)
1827 struct thread_data *td;
1828 unsigned long long *rt;
1832 fio_mutex_down(stat_mutex);
1834 rt = malloc(thread_number * sizeof(unsigned long long));
1835 fio_gettime(&tv, NULL);
1837 for_each_td(td, i) {
1838 td->update_rusage = 1;
1839 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1840 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1841 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1842 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1844 rt[i] = mtime_since(&td->start, &tv);
1845 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1846 td->ts.runtime[DDIR_READ] += rt[i];
1847 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1848 td->ts.runtime[DDIR_WRITE] += rt[i];
1849 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1850 td->ts.runtime[DDIR_TRIM] += rt[i];
1853 for_each_td(td, i) {
1854 if (td->runstate >= TD_EXITED)
1856 if (td->rusage_sem) {
1857 td->update_rusage = 1;
1858 fio_mutex_down(td->rusage_sem);
1860 td->update_rusage = 0;
1865 for_each_td(td, i) {
1866 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1867 td->ts.runtime[DDIR_READ] -= rt[i];
1868 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1869 td->ts.runtime[DDIR_WRITE] -= rt[i];
1870 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1871 td->ts.runtime[DDIR_TRIM] -= rt[i];
1875 fio_mutex_up(stat_mutex);
1878 static int status_interval_init;
1879 static struct timeval status_time;
1880 static int status_file_disabled;
1882 #define FIO_STATUS_FILE "fio-dump-status"
1884 static int check_status_file(void)
1887 const char *temp_dir;
1888 char fio_status_file_path[PATH_MAX];
1890 if (status_file_disabled)
1893 temp_dir = getenv("TMPDIR");
1894 if (temp_dir == NULL) {
1895 temp_dir = getenv("TEMP");
1896 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1899 if (temp_dir == NULL)
1902 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1904 if (stat(fio_status_file_path, &sb))
1907 if (unlink(fio_status_file_path) < 0) {
1908 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1910 log_err("fio: disabling status file updates\n");
1911 status_file_disabled = 1;
1917 void check_for_running_stats(void)
1919 if (status_interval) {
1920 if (!status_interval_init) {
1921 fio_gettime(&status_time, NULL);
1922 status_interval_init = 1;
1923 } else if (mtime_since_now(&status_time) >= status_interval) {
1924 show_running_run_stats();
1925 fio_gettime(&status_time, NULL);
1929 if (check_status_file()) {
1930 show_running_run_stats();
1935 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1940 if (data > is->max_val)
1942 if (data < is->min_val)
1945 delta = val - is->mean.u.f;
1947 is->mean.u.f += delta / (is->samples + 1.0);
1948 is->S.u.f += delta * (val - is->mean.u.f);
1955 * Return a struct io_logs, which is added to the tail of the log
1958 static struct io_logs *get_new_log(struct io_log *iolog)
1960 size_t new_size, new_samples;
1961 struct io_logs *cur_log;
1964 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1967 if (!iolog->cur_log_max)
1968 new_samples = DEF_LOG_ENTRIES;
1970 new_samples = iolog->cur_log_max * 2;
1971 if (new_samples > MAX_LOG_ENTRIES)
1972 new_samples = MAX_LOG_ENTRIES;
1975 new_size = new_samples * log_entry_sz(iolog);
1977 cur_log = smalloc(sizeof(*cur_log));
1979 INIT_FLIST_HEAD(&cur_log->list);
1980 cur_log->log = malloc(new_size);
1982 cur_log->nr_samples = 0;
1983 cur_log->max_samples = new_samples;
1984 flist_add_tail(&cur_log->list, &iolog->io_logs);
1985 iolog->cur_log_max = new_samples;
1995 * Add and return a new log chunk, or return current log if big enough
1997 static struct io_logs *regrow_log(struct io_log *iolog)
1999 struct io_logs *cur_log;
2002 if (!iolog || iolog->disabled)
2005 cur_log = iolog_cur_log(iolog);
2007 cur_log = get_new_log(iolog);
2012 if (cur_log->nr_samples < cur_log->max_samples)
2016 * No room for a new sample. If we're compressing on the fly, flush
2017 * out the current chunk
2019 if (iolog->log_gz) {
2020 if (iolog_cur_flush(iolog, cur_log)) {
2021 log_err("fio: failed flushing iolog! Will stop logging.\n");
2027 * Get a new log array, and add to our list
2029 cur_log = get_new_log(iolog);
2031 log_err("fio: failed extending iolog! Will stop logging.\n");
2035 if (!iolog->pending || !iolog->pending->nr_samples)
2039 * Flush pending items to new log
2041 for (i = 0; i < iolog->pending->nr_samples; i++) {
2042 struct io_sample *src, *dst;
2044 src = get_sample(iolog, iolog->pending, i);
2045 dst = get_sample(iolog, cur_log, i);
2046 memcpy(dst, src, log_entry_sz(iolog));
2048 cur_log->nr_samples = iolog->pending->nr_samples;
2050 iolog->pending->nr_samples = 0;
2054 iolog->disabled = true;
2058 void regrow_logs(struct thread_data *td)
2060 regrow_log(td->slat_log);
2061 regrow_log(td->clat_log);
2062 regrow_log(td->clat_hist_log);
2063 regrow_log(td->lat_log);
2064 regrow_log(td->bw_log);
2065 regrow_log(td->iops_log);
2066 td->flags &= ~TD_F_REGROW_LOGS;
2069 static struct io_logs *get_cur_log(struct io_log *iolog)
2071 struct io_logs *cur_log;
2073 cur_log = iolog_cur_log(iolog);
2075 cur_log = get_new_log(iolog);
2080 if (cur_log->nr_samples < cur_log->max_samples)
2084 * Out of space. If we're in IO offload mode, or we're not doing
2085 * per unit logging (hence logging happens outside of the IO thread
2086 * as well), add a new log chunk inline. If we're doing inline
2087 * submissions, flag 'td' as needing a log regrow and we'll take
2088 * care of it on the submission side.
2090 if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2091 !per_unit_log(iolog))
2092 return regrow_log(iolog);
2094 iolog->td->flags |= TD_F_REGROW_LOGS;
2095 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2096 return iolog->pending;
2099 static void __add_log_sample(struct io_log *iolog, unsigned long val,
2100 enum fio_ddir ddir, unsigned int bs,
2101 unsigned long t, uint64_t offset)
2103 struct io_logs *cur_log;
2105 if (iolog->disabled)
2107 if (flist_empty(&iolog->io_logs))
2108 iolog->avg_last = t;
2110 cur_log = get_cur_log(iolog);
2112 struct io_sample *s;
2114 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2118 io_sample_set_ddir(iolog, s, ddir);
2121 if (iolog->log_offset) {
2122 struct io_sample_offset *so = (void *) s;
2124 so->offset = offset;
2127 cur_log->nr_samples++;
2131 iolog->disabled = true;
2134 static inline void reset_io_stat(struct io_stat *ios)
2136 ios->max_val = ios->min_val = ios->samples = 0;
2137 ios->mean.u.f = ios->S.u.f = 0;
2140 void reset_io_stats(struct thread_data *td)
2142 struct thread_stat *ts = &td->ts;
2145 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2146 reset_io_stat(&ts->clat_stat[i]);
2147 reset_io_stat(&ts->slat_stat[i]);
2148 reset_io_stat(&ts->lat_stat[i]);
2149 reset_io_stat(&ts->bw_stat[i]);
2150 reset_io_stat(&ts->iops_stat[i]);
2152 ts->io_bytes[i] = 0;
2155 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2156 ts->io_u_plat[i][j] = 0;
2159 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2160 ts->io_u_map[i] = 0;
2161 ts->io_u_submit[i] = 0;
2162 ts->io_u_complete[i] = 0;
2163 ts->io_u_lat_u[i] = 0;
2164 ts->io_u_lat_m[i] = 0;
2165 ts->total_submit = 0;
2166 ts->total_complete = 0;
2169 for (i = 0; i < 3; i++) {
2170 ts->total_io_u[i] = 0;
2171 ts->short_io_u[i] = 0;
2172 ts->drop_io_u[i] = 0;
2176 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2177 unsigned long elapsed, bool log_max)
2180 * Note an entry in the log. Use the mean from the logged samples,
2181 * making sure to properly round up. Only write a log entry if we
2182 * had actual samples done.
2184 if (iolog->avg_window[ddir].samples) {
2188 val = iolog->avg_window[ddir].max_val;
2190 val = iolog->avg_window[ddir].mean.u.f + 0.50;
2192 __add_log_sample(iolog, val, ddir, 0, elapsed, 0);
2195 reset_io_stat(&iolog->avg_window[ddir]);
2198 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2203 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2204 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2207 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2208 unsigned long val, enum fio_ddir ddir,
2209 unsigned int bs, uint64_t offset)
2211 unsigned long elapsed, this_window;
2216 elapsed = mtime_since_now(&td->epoch);
2219 * If no time averaging, just add the log sample.
2221 if (!iolog->avg_msec) {
2222 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
2227 * Add the sample. If the time period has passed, then
2228 * add that entry to the log and clear.
2230 add_stat_sample(&iolog->avg_window[ddir], val);
2233 * If period hasn't passed, adding the above sample is all we
2236 this_window = elapsed - iolog->avg_last;
2237 if (elapsed < iolog->avg_last)
2238 return iolog->avg_last - elapsed;
2239 else if (this_window < iolog->avg_msec) {
2240 int diff = iolog->avg_msec - this_window;
2242 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2246 _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2248 iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2249 return iolog->avg_msec;
2252 void finalize_logs(struct thread_data *td, bool unit_logs)
2254 unsigned long elapsed;
2256 elapsed = mtime_since_now(&td->epoch);
2258 if (td->clat_log && unit_logs)
2259 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2260 if (td->slat_log && unit_logs)
2261 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2262 if (td->lat_log && unit_logs)
2263 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2264 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2265 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2266 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2267 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2270 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
2272 struct io_log *iolog;
2277 iolog = agg_io_log[ddir];
2278 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
2281 static void add_clat_percentile_sample(struct thread_stat *ts,
2282 unsigned long usec, enum fio_ddir ddir)
2284 unsigned int idx = plat_val_to_idx(usec);
2285 assert(idx < FIO_IO_U_PLAT_NR);
2287 ts->io_u_plat[ddir][idx]++;
2290 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2291 unsigned long usec, unsigned int bs, uint64_t offset)
2293 unsigned long elapsed, this_window;
2294 struct thread_stat *ts = &td->ts;
2295 struct io_log *iolog = td->clat_hist_log;
2299 add_stat_sample(&ts->clat_stat[ddir], usec);
2302 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
2304 if (ts->clat_percentiles)
2305 add_clat_percentile_sample(ts, usec, ddir);
2307 if (iolog && iolog->hist_msec) {
2308 struct io_hist *hw = &iolog->hist_window[ddir];
2311 elapsed = mtime_since_now(&td->epoch);
2313 hw->hist_last = elapsed;
2314 this_window = elapsed - hw->hist_last;
2316 if (this_window >= iolog->hist_msec) {
2317 unsigned int *io_u_plat;
2321 * Make a byte-for-byte copy of the latency histogram
2322 * stored in td->ts.io_u_plat[ddir], recording it in a
2323 * log sample. Note that the matching call to free() is
2324 * located in iolog.c after printing this sample to the
2327 io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2328 dst = malloc(FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2329 memcpy(dst, io_u_plat,
2330 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2331 __add_log_sample(iolog, (unsigned long )dst, ddir, bs,
2335 * Update the last time we recorded as being now, minus
2336 * any drift in time we encountered before actually
2337 * making the record.
2339 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2347 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2348 unsigned long usec, unsigned int bs, uint64_t offset)
2350 struct thread_stat *ts = &td->ts;
2357 add_stat_sample(&ts->slat_stat[ddir], usec);
2360 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
2365 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2366 unsigned long usec, unsigned int bs, uint64_t offset)
2368 struct thread_stat *ts = &td->ts;
2375 add_stat_sample(&ts->lat_stat[ddir], usec);
2378 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2383 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2384 unsigned int bytes, unsigned long spent)
2386 struct thread_stat *ts = &td->ts;
2390 rate = bytes * 1000 / spent;
2396 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2399 add_log_sample(td, td->bw_log, rate, io_u->ddir, bytes, io_u->offset);
2401 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2405 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2407 struct thread_stat *ts = &td->ts;
2408 unsigned long spent, rate;
2410 unsigned int next, next_log;
2412 next_log = td->o.bw_avg_time;
2414 spent = mtime_since(&td->bw_sample_time, t);
2415 if (spent < td->o.bw_avg_time &&
2416 td->o.bw_avg_time - spent >= LOG_MSEC_SLACK)
2417 return td->o.bw_avg_time - spent;
2422 * Compute both read and write rates for the interval.
2424 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2427 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2429 continue; /* No entries for interval */
2432 rate = delta * 1000 / spent / 1024;
2436 add_stat_sample(&ts->bw_stat[ddir], rate);
2439 unsigned int bs = 0;
2441 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2442 bs = td->o.min_bs[ddir];
2444 next = add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2445 next_log = min(next_log, next);
2448 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2451 timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time);
2455 if (spent <= td->o.bw_avg_time)
2456 return min(next_log, td->o.bw_avg_time);
2458 next = td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time);
2459 return min(next, next_log);
2462 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2465 struct thread_stat *ts = &td->ts;
2469 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2472 add_log_sample(td, td->iops_log, 1, io_u->ddir, bytes, io_u->offset);
2474 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2478 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2480 struct thread_stat *ts = &td->ts;
2481 unsigned long spent, iops;
2483 unsigned int next, next_log;
2485 next_log = td->o.iops_avg_time;
2487 spent = mtime_since(&td->iops_sample_time, t);
2488 if (spent < td->o.iops_avg_time &&
2489 td->o.iops_avg_time - spent >= LOG_MSEC_SLACK)
2490 return td->o.iops_avg_time - spent;
2495 * Compute both read and write rates for the interval.
2497 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2500 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2502 continue; /* No entries for interval */
2505 iops = (delta * 1000) / spent;
2509 add_stat_sample(&ts->iops_stat[ddir], iops);
2512 unsigned int bs = 0;
2514 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2515 bs = td->o.min_bs[ddir];
2517 next = add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2518 next_log = min(next_log, next);
2521 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2524 timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time);
2528 if (spent <= td->o.iops_avg_time)
2529 return min(next_log, td->o.iops_avg_time);
2531 next = td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
2532 return min(next, next_log);
2536 * Returns msecs to next event
2538 int calc_log_samples(void)
2540 struct thread_data *td;
2541 unsigned int next = ~0U, tmp;
2545 fio_gettime(&now, NULL);
2547 for_each_td(td, i) {
2548 if (in_ramp_time(td) ||
2549 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2550 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2553 if (!per_unit_log(td->bw_log)) {
2554 tmp = add_bw_samples(td, &now);
2558 if (!per_unit_log(td->iops_log)) {
2559 tmp = add_iops_samples(td, &now);
2565 return next == ~0U ? 0 : next;
2568 void stat_init(void)
2570 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2573 void stat_exit(void)
2576 * When we have the mutex, we know out-of-band access to it
2579 fio_mutex_down(stat_mutex);
2580 fio_mutex_remove(stat_mutex);
2584 * Called from signal handler. Wake up status thread.
2586 void show_running_run_stats(void)
2591 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2593 /* Ignore io_u's which span multiple blocks--they will just get
2594 * inaccurate counts. */
2595 int idx = (io_u->offset - io_u->file->file_offset)
2596 / td->o.bs[DDIR_TRIM];
2597 uint32_t *info = &td->ts.block_infos[idx];
2598 assert(idx < td->ts.nr_block_infos);