12 #include "lib/ieee754.h"
14 #include "lib/getrusage.h"
18 struct fio_mutex *stat_mutex;
20 void update_rusage_stat(struct thread_data *td)
22 struct thread_stat *ts = &td->ts;
24 fio_getrusage(&td->ru_end);
25 ts->usr_time += mtime_since(&td->ru_start.ru_utime,
26 &td->ru_end.ru_utime);
27 ts->sys_time += mtime_since(&td->ru_start.ru_stime,
28 &td->ru_end.ru_stime);
29 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
30 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
31 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
32 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
34 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
38 * Given a latency, return the index of the corresponding bucket in
39 * the structure tracking percentiles.
41 * (1) find the group (and error bits) that the value (latency)
42 * belongs to by looking at its MSB. (2) find the bucket number in the
43 * group by looking at the index bits.
46 static unsigned int plat_val_to_idx(unsigned int val)
48 unsigned int msb, error_bits, base, offset, idx;
50 /* Find MSB starting from bit 0 */
54 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
57 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
58 * all bits of the sample as index
60 if (msb <= FIO_IO_U_PLAT_BITS)
63 /* Compute the number of error bits to discard*/
64 error_bits = msb - FIO_IO_U_PLAT_BITS;
66 /* Compute the number of buckets before the group */
67 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
70 * Discard the error bits and apply the mask to find the
71 * index for the buckets in the group
73 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
75 /* Make sure the index does not exceed (array size - 1) */
76 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
77 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
83 * Convert the given index of the bucket array to the value
84 * represented by the bucket
86 static unsigned int plat_idx_to_val(unsigned int idx)
88 unsigned int error_bits, k, base;
90 assert(idx < FIO_IO_U_PLAT_NR);
92 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
93 * all bits of the sample as index */
94 if (idx < (FIO_IO_U_PLAT_VAL << 1))
97 /* Find the group and compute the minimum value of that group */
98 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
99 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
101 /* Find its bucket number of the group */
102 k = idx % FIO_IO_U_PLAT_VAL;
104 /* Return the mean of the range of the bucket */
105 return base + ((k + 0.5) * (1 << error_bits));
108 static int double_cmp(const void *a, const void *b)
110 const fio_fp64_t fa = *(const fio_fp64_t *) a;
111 const fio_fp64_t fb = *(const fio_fp64_t *) b;
116 else if (fa.u.f < fb.u.f)
122 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
123 fio_fp64_t *plist, unsigned int **output,
124 unsigned int *maxv, unsigned int *minv)
126 unsigned long sum = 0;
127 unsigned int len, i, j = 0;
128 unsigned int oval_len = 0;
129 unsigned int *ovals = NULL;
136 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
143 * Sort the percentile list. Note that it may already be sorted if
144 * we are using the default values, but since it's a short list this
145 * isn't a worry. Also note that this does not work for NaN values.
148 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
151 * Calculate bucket values, note down max and min values
154 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
156 while (sum >= (plist[j].u.f / 100.0 * nr)) {
157 assert(plist[j].u.f <= 100.0);
161 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
164 ovals[j] = plat_idx_to_val(i);
165 if (ovals[j] < *minv)
167 if (ovals[j] > *maxv)
170 is_last = (j == len - 1);
183 * Find and display the p-th percentile of clat
185 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
186 fio_fp64_t *plist, unsigned int precision)
188 unsigned int len, j = 0, minv, maxv;
190 int is_last, per_line, scale_down;
193 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
198 * We default to usecs, but if the value range is such that we
199 * should scale down to msecs, do that.
201 if (minv > 2000 && maxv > 99999) {
203 log_info(" clat percentiles (msec):\n |");
206 log_info(" clat percentiles (usec):\n |");
209 snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
210 per_line = (80 - 7) / (precision + 14);
212 for (j = 0; j < len; j++) {
213 char fbuf[16], *ptr = fbuf;
216 if (j != 0 && (j % per_line) == 0)
219 /* end of the list */
220 is_last = (j == len - 1);
222 if (plist[j].u.f < 10.0)
223 ptr += sprintf(fbuf, " ");
225 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
228 ovals[j] = (ovals[j] + 999) / 1000;
230 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
235 if ((j % per_line) == per_line - 1) /* for formatting */
244 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
245 double *mean, double *dev)
247 double n = (double) is->samples;
254 *mean = is->mean.u.f;
257 *dev = sqrt(is->S.u.f / (n - 1.0));
264 void show_group_stats(struct group_run_stats *rs)
266 char *p1, *p2, *p3, *p4;
267 const char *str[] = { " READ", " WRITE" , " TRIM"};
270 log_info("\nRun status group %d (all jobs):\n", rs->groupid);
272 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
273 const int i2p = is_power_of_2(rs->kb_base);
278 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
279 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
280 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
281 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
283 log_info("%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
284 " mint=%llumsec, maxt=%llumsec\n",
285 rs->unified_rw_rep ? " MIXED" : str[i],
287 (unsigned long long) rs->min_run[i],
288 (unsigned long long) rs->max_run[i]);
297 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
302 * Do depth distribution calculations
304 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
306 io_u_dist[i] = (double) map[i] / (double) total;
307 io_u_dist[i] *= 100.0;
308 if (io_u_dist[i] < 0.1 && map[i])
315 static void stat_calc_lat(struct thread_stat *ts, double *dst,
316 unsigned int *src, int nr)
318 unsigned long total = ddir_rw_sum(ts->total_io_u);
322 * Do latency distribution calculations
324 for (i = 0; i < nr; i++) {
326 dst[i] = (double) src[i] / (double) total;
328 if (dst[i] < 0.01 && src[i])
335 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
337 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
340 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
342 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
345 static void display_lat(const char *name, unsigned long min, unsigned long max,
346 double mean, double dev)
348 const char *base = "(usec)";
351 if (!usec_to_msec(&min, &max, &mean, &dev))
354 minp = num2str(min, 6, 1, 0, 0);
355 maxp = num2str(max, 6, 1, 0, 0);
357 log_info(" %s %s: min=%s, max=%s, avg=%5.02f,"
358 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
364 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
367 const char *str[] = { "read ", "write", "trim" };
368 unsigned long min, max, runt;
369 unsigned long long bw, iops;
371 char *io_p, *bw_p, *iops_p;
374 assert(ddir_rw(ddir));
376 if (!ts->runtime[ddir])
379 i2p = is_power_of_2(rs->kb_base);
380 runt = ts->runtime[ddir];
382 bw = (1000 * ts->io_bytes[ddir]) / runt;
383 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
384 bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
386 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
387 iops_p = num2str(iops, 6, 1, 0, 0);
389 log_info(" %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
390 rs->unified_rw_rep ? "mixed" : str[ddir],
392 (unsigned long long) ts->runtime[ddir]);
398 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
399 display_lat("slat", min, max, mean, dev);
400 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
401 display_lat("clat", min, max, mean, dev);
402 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
403 display_lat(" lat", min, max, mean, dev);
405 if (ts->clat_percentiles) {
406 show_clat_percentiles(ts->io_u_plat[ddir],
407 ts->clat_stat[ddir].samples,
409 ts->percentile_precision);
411 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
412 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
413 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
415 if (rs->unit_base == 1) {
423 p_of_agg = mean * 100 / (double) rs->agg[ddir];
424 if (p_of_agg > 100.0)
428 if (mean > fkb_base * fkb_base) {
433 bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
436 log_info(" bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
437 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
438 p_of_agg, mean, dev);
442 static int show_lat(double *io_u_lat, int nr, const char **ranges,
445 int new_line = 1, i, line = 0, shown = 0;
447 for (i = 0; i < nr; i++) {
448 if (io_u_lat[i] <= 0.0)
454 log_info(" lat (%s) : ", msg);
460 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
472 static void show_lat_u(double *io_u_lat_u)
474 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
475 "250=", "500=", "750=", "1000=", };
477 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
480 static void show_lat_m(double *io_u_lat_m)
482 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
483 "250=", "500=", "750=", "1000=", "2000=",
486 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
489 static void show_latencies(struct thread_stat *ts)
491 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
492 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
494 stat_calc_lat_u(ts, io_u_lat_u);
495 stat_calc_lat_m(ts, io_u_lat_m);
497 show_lat_u(io_u_lat_u);
498 show_lat_m(io_u_lat_m);
501 static int block_state_category(int block_state)
503 switch (block_state) {
504 case BLOCK_STATE_UNINIT:
506 case BLOCK_STATE_TRIMMED:
507 case BLOCK_STATE_WRITTEN:
509 case BLOCK_STATE_WRITE_FAILURE:
510 case BLOCK_STATE_TRIM_FAILURE:
513 /* Silence compile warning on some BSDs and have a return */
519 static int compare_block_infos(const void *bs1, const void *bs2)
521 uint32_t block1 = *(uint32_t *)bs1;
522 uint32_t block2 = *(uint32_t *)bs2;
523 int state1 = BLOCK_INFO_STATE(block1);
524 int state2 = BLOCK_INFO_STATE(block2);
525 int bscat1 = block_state_category(state1);
526 int bscat2 = block_state_category(state2);
527 int cycles1 = BLOCK_INFO_TRIMS(block1);
528 int cycles2 = BLOCK_INFO_TRIMS(block2);
535 if (cycles1 < cycles2)
537 if (cycles1 > cycles2)
545 assert(block1 == block2);
549 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
550 fio_fp64_t *plist, unsigned int **percentiles,
556 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
558 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
565 * Sort the percentile list. Note that it may already be sorted if
566 * we are using the default values, but since it's a short list this
567 * isn't a worry. Also note that this does not work for NaN values.
570 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
573 /* Start only after the uninit entries end */
575 nr_uninit < nr_block_infos
576 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
580 if (nr_uninit == nr_block_infos)
583 *percentiles = calloc(len, sizeof(**percentiles));
585 for (i = 0; i < len; i++) {
586 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
588 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
591 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
592 for (i = 0; i < nr_block_infos; i++)
593 types[BLOCK_INFO_STATE(block_infos[i])]++;
598 static const char *block_state_names[] = {
599 [BLOCK_STATE_UNINIT] = "unwritten",
600 [BLOCK_STATE_TRIMMED] = "trimmed",
601 [BLOCK_STATE_WRITTEN] = "written",
602 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
603 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
606 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
610 unsigned int *percentiles = NULL;
611 unsigned int block_state_counts[BLOCK_STATE_COUNT];
613 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
614 &percentiles, block_state_counts);
616 log_info(" block lifetime percentiles :\n |");
618 for (i = 0; i < len; i++) {
619 uint32_t block_info = percentiles[i];
620 #define LINE_LENGTH 75
621 char str[LINE_LENGTH];
622 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
623 plist[i].u.f, block_info,
624 i == len - 1 ? '\n' : ',');
625 assert(strln < LINE_LENGTH);
626 if (pos + strln > LINE_LENGTH) {
637 log_info(" states :");
638 for (i = 0; i < BLOCK_STATE_COUNT; i++)
640 block_state_names[i], block_state_counts[i],
641 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
644 static void show_thread_status_normal(struct thread_stat *ts,
645 struct group_run_stats *rs)
647 double usr_cpu, sys_cpu;
648 unsigned long runtime;
649 double io_u_dist[FIO_IO_U_MAP_NR];
653 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
657 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
660 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
661 ts->name, ts->groupid, ts->members,
662 ts->error, (int) ts->pid, time_buf);
664 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
665 ts->name, ts->groupid, ts->members,
666 ts->error, ts->verror, (int) ts->pid,
670 if (strlen(ts->description))
671 log_info(" Description : [%s]\n", ts->description);
673 if (ts->io_bytes[DDIR_READ])
674 show_ddir_status(rs, ts, DDIR_READ);
675 if (ts->io_bytes[DDIR_WRITE])
676 show_ddir_status(rs, ts, DDIR_WRITE);
677 if (ts->io_bytes[DDIR_TRIM])
678 show_ddir_status(rs, ts, DDIR_TRIM);
682 runtime = ts->total_run_time;
684 double runt = (double) runtime;
686 usr_cpu = (double) ts->usr_time * 100 / runt;
687 sys_cpu = (double) ts->sys_time * 100 / runt;
693 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
694 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
695 (unsigned long long) ts->ctx,
696 (unsigned long long) ts->majf,
697 (unsigned long long) ts->minf);
699 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
700 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
701 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
702 io_u_dist[1], io_u_dist[2],
703 io_u_dist[3], io_u_dist[4],
704 io_u_dist[5], io_u_dist[6]);
706 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
707 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
708 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
709 io_u_dist[1], io_u_dist[2],
710 io_u_dist[3], io_u_dist[4],
711 io_u_dist[5], io_u_dist[6]);
712 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
713 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
714 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
715 io_u_dist[1], io_u_dist[2],
716 io_u_dist[3], io_u_dist[4],
717 io_u_dist[5], io_u_dist[6]);
718 log_info(" issued : total=r=%llu/w=%llu/d=%llu,"
719 " short=r=%llu/w=%llu/d=%llu,"
720 " drop=r=%llu/w=%llu/d=%llu\n",
721 (unsigned long long) ts->total_io_u[0],
722 (unsigned long long) ts->total_io_u[1],
723 (unsigned long long) ts->total_io_u[2],
724 (unsigned long long) ts->short_io_u[0],
725 (unsigned long long) ts->short_io_u[1],
726 (unsigned long long) ts->short_io_u[2],
727 (unsigned long long) ts->drop_io_u[0],
728 (unsigned long long) ts->drop_io_u[1],
729 (unsigned long long) ts->drop_io_u[2]);
730 if (ts->continue_on_error) {
731 log_info(" errors : total=%llu, first_error=%d/<%s>\n",
732 (unsigned long long)ts->total_err_count,
734 strerror(ts->first_error));
736 if (ts->latency_depth) {
737 log_info(" latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
738 (unsigned long long)ts->latency_target,
739 (unsigned long long)ts->latency_window,
740 ts->latency_percentile.u.f,
744 if (ts->nr_block_infos)
745 show_block_infos(ts->nr_block_infos, ts->block_infos,
746 ts->percentile_list);
749 static void show_ddir_status_terse(struct thread_stat *ts,
750 struct group_run_stats *rs, int ddir)
752 unsigned long min, max;
753 unsigned long long bw, iops;
754 unsigned int *ovals = NULL;
756 unsigned int len, minv, maxv;
759 assert(ddir_rw(ddir));
762 if (ts->runtime[ddir]) {
763 uint64_t runt = ts->runtime[ddir];
765 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
766 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
769 log_info(";%llu;%llu;%llu;%llu",
770 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
771 (unsigned long long) ts->runtime[ddir]);
773 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
774 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
776 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
778 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
779 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
781 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
783 if (ts->clat_percentiles) {
784 len = calc_clat_percentiles(ts->io_u_plat[ddir],
785 ts->clat_stat[ddir].samples,
786 ts->percentile_list, &ovals, &maxv,
791 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
796 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
799 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
800 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
802 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
807 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
808 double p_of_agg = 100.0;
811 p_of_agg = mean * 100 / (double) rs->agg[ddir];
812 if (p_of_agg > 100.0)
816 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
818 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
821 static void add_ddir_status_json(struct thread_stat *ts,
822 struct group_run_stats *rs, int ddir, struct json_object *parent)
824 unsigned long min, max;
825 unsigned long long bw;
826 unsigned int *ovals = NULL;
827 double mean, dev, iops;
828 unsigned int len, minv, maxv;
830 const char *ddirname[] = {"read", "write", "trim"};
831 struct json_object *dir_object, *tmp_object, *percentile_object;
833 double p_of_agg = 100.0;
835 assert(ddir_rw(ddir));
837 if (ts->unified_rw_rep && ddir != DDIR_READ)
840 dir_object = json_create_object();
841 json_object_add_value_object(parent,
842 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
846 if (ts->runtime[ddir]) {
847 uint64_t runt = ts->runtime[ddir];
849 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
850 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
853 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
854 json_object_add_value_int(dir_object, "bw", bw);
855 json_object_add_value_float(dir_object, "iops", iops);
856 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
857 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
858 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
859 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
861 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
865 tmp_object = json_create_object();
866 json_object_add_value_object(dir_object, "slat", tmp_object);
867 json_object_add_value_int(tmp_object, "min", min);
868 json_object_add_value_int(tmp_object, "max", max);
869 json_object_add_value_float(tmp_object, "mean", mean);
870 json_object_add_value_float(tmp_object, "stddev", dev);
872 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
876 tmp_object = json_create_object();
877 json_object_add_value_object(dir_object, "clat", tmp_object);
878 json_object_add_value_int(tmp_object, "min", min);
879 json_object_add_value_int(tmp_object, "max", max);
880 json_object_add_value_float(tmp_object, "mean", mean);
881 json_object_add_value_float(tmp_object, "stddev", dev);
883 if (ts->clat_percentiles) {
884 len = calc_clat_percentiles(ts->io_u_plat[ddir],
885 ts->clat_stat[ddir].samples,
886 ts->percentile_list, &ovals, &maxv,
891 percentile_object = json_create_object();
892 json_object_add_value_object(tmp_object, "percentile", percentile_object);
893 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
895 json_object_add_value_int(percentile_object, "0.00", 0);
898 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
899 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
902 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
906 tmp_object = json_create_object();
907 json_object_add_value_object(dir_object, "lat", tmp_object);
908 json_object_add_value_int(tmp_object, "min", min);
909 json_object_add_value_int(tmp_object, "max", max);
910 json_object_add_value_float(tmp_object, "mean", mean);
911 json_object_add_value_float(tmp_object, "stddev", dev);
915 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
917 p_of_agg = mean * 100 / (double) rs->agg[ddir];
918 if (p_of_agg > 100.0)
923 p_of_agg = mean = dev = 0.0;
925 json_object_add_value_int(dir_object, "bw_min", min);
926 json_object_add_value_int(dir_object, "bw_max", max);
927 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
928 json_object_add_value_float(dir_object, "bw_mean", mean);
929 json_object_add_value_float(dir_object, "bw_dev", dev);
932 static void show_thread_status_terse_v2(struct thread_stat *ts,
933 struct group_run_stats *rs)
935 double io_u_dist[FIO_IO_U_MAP_NR];
936 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
937 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
938 double usr_cpu, sys_cpu;
942 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
943 /* Log Read Status */
944 show_ddir_status_terse(ts, rs, DDIR_READ);
945 /* Log Write Status */
946 show_ddir_status_terse(ts, rs, DDIR_WRITE);
947 /* Log Trim Status */
948 show_ddir_status_terse(ts, rs, DDIR_TRIM);
951 if (ts->total_run_time) {
952 double runt = (double) ts->total_run_time;
954 usr_cpu = (double) ts->usr_time * 100 / runt;
955 sys_cpu = (double) ts->sys_time * 100 / runt;
961 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
962 (unsigned long long) ts->ctx,
963 (unsigned long long) ts->majf,
964 (unsigned long long) ts->minf);
966 /* Calc % distribution of IO depths, usecond, msecond latency */
967 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
968 stat_calc_lat_u(ts, io_u_lat_u);
969 stat_calc_lat_m(ts, io_u_lat_m);
971 /* Only show fixed 7 I/O depth levels*/
972 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
973 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
974 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
976 /* Microsecond latency */
977 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
978 log_info(";%3.2f%%", io_u_lat_u[i]);
979 /* Millisecond latency */
980 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
981 log_info(";%3.2f%%", io_u_lat_m[i]);
982 /* Additional output if continue_on_error set - default off*/
983 if (ts->continue_on_error)
984 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
987 /* Additional output if description is set */
988 if (strlen(ts->description))
989 log_info(";%s", ts->description);
994 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
995 struct group_run_stats *rs, int ver)
997 double io_u_dist[FIO_IO_U_MAP_NR];
998 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
999 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1000 double usr_cpu, sys_cpu;
1004 log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
1005 ts->name, ts->groupid, ts->error);
1006 /* Log Read Status */
1007 show_ddir_status_terse(ts, rs, DDIR_READ);
1008 /* Log Write Status */
1009 show_ddir_status_terse(ts, rs, DDIR_WRITE);
1010 /* Log Trim Status */
1012 show_ddir_status_terse(ts, rs, DDIR_TRIM);
1015 if (ts->total_run_time) {
1016 double runt = (double) ts->total_run_time;
1018 usr_cpu = (double) ts->usr_time * 100 / runt;
1019 sys_cpu = (double) ts->sys_time * 100 / runt;
1025 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1026 (unsigned long long) ts->ctx,
1027 (unsigned long long) ts->majf,
1028 (unsigned long long) ts->minf);
1030 /* Calc % distribution of IO depths, usecond, msecond latency */
1031 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1032 stat_calc_lat_u(ts, io_u_lat_u);
1033 stat_calc_lat_m(ts, io_u_lat_m);
1035 /* Only show fixed 7 I/O depth levels*/
1036 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1037 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1038 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1040 /* Microsecond latency */
1041 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1042 log_info(";%3.2f%%", io_u_lat_u[i]);
1043 /* Millisecond latency */
1044 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1045 log_info(";%3.2f%%", io_u_lat_m[i]);
1047 /* disk util stats, if any */
1048 show_disk_util(1, NULL);
1050 /* Additional output if continue_on_error set - default off*/
1051 if (ts->continue_on_error)
1052 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1054 /* Additional output if description is set */
1055 if (strlen(ts->description))
1056 log_info(";%s", ts->description);
1061 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1062 struct group_run_stats *rs)
1064 struct json_object *root, *tmp;
1065 struct jobs_eta *je;
1066 double io_u_dist[FIO_IO_U_MAP_NR];
1067 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1068 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1069 double usr_cpu, sys_cpu;
1074 root = json_create_object();
1075 json_object_add_value_string(root, "jobname", ts->name);
1076 json_object_add_value_int(root, "groupid", ts->groupid);
1077 json_object_add_value_int(root, "error", ts->error);
1080 je = get_jobs_eta(1, &size);
1081 json_object_add_value_int(root, "eta", je->eta_sec);
1082 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1085 add_ddir_status_json(ts, rs, DDIR_READ, root);
1086 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1087 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1090 if (ts->total_run_time) {
1091 double runt = (double) ts->total_run_time;
1093 usr_cpu = (double) ts->usr_time * 100 / runt;
1094 sys_cpu = (double) ts->sys_time * 100 / runt;
1099 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1100 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1101 json_object_add_value_int(root, "ctx", ts->ctx);
1102 json_object_add_value_int(root, "majf", ts->majf);
1103 json_object_add_value_int(root, "minf", ts->minf);
1106 /* Calc % distribution of IO depths, usecond, msecond latency */
1107 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1108 stat_calc_lat_u(ts, io_u_lat_u);
1109 stat_calc_lat_m(ts, io_u_lat_m);
1111 tmp = json_create_object();
1112 json_object_add_value_object(root, "iodepth_level", tmp);
1113 /* Only show fixed 7 I/O depth levels*/
1114 for (i = 0; i < 7; i++) {
1117 snprintf(name, 20, "%d", 1 << i);
1119 snprintf(name, 20, ">=%d", 1 << i);
1120 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1123 tmp = json_create_object();
1124 json_object_add_value_object(root, "latency_us", tmp);
1125 /* Microsecond latency */
1126 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1127 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1128 "250", "500", "750", "1000", };
1129 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1131 /* Millisecond latency */
1132 tmp = json_create_object();
1133 json_object_add_value_object(root, "latency_ms", tmp);
1134 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1135 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1136 "250", "500", "750", "1000", "2000",
1138 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1141 /* Additional output if continue_on_error set - default off*/
1142 if (ts->continue_on_error) {
1143 json_object_add_value_int(root, "total_err", ts->total_err_count);
1144 json_object_add_value_int(root, "first_error", ts->first_error);
1147 if (ts->latency_depth) {
1148 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1149 json_object_add_value_int(root, "latency_target", ts->latency_target);
1150 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1151 json_object_add_value_int(root, "latency_window", ts->latency_window);
1154 /* Additional output if description is set */
1155 if (strlen(ts->description))
1156 json_object_add_value_string(root, "desc", ts->description);
1158 if (ts->nr_block_infos) {
1159 /* Block error histogram and types */
1161 unsigned int *percentiles = NULL;
1162 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1164 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1165 ts->percentile_list,
1166 &percentiles, block_state_counts);
1169 struct json_object *block, *percentile_object, *states;
1171 block = json_create_object();
1172 json_object_add_value_object(root, "block", block);
1174 percentile_object = json_create_object();
1175 json_object_add_value_object(block, "percentiles",
1177 for (i = 0; i < len; i++) {
1179 snprintf(buf, sizeof(buf), "%f",
1180 ts->percentile_list[i].u.f);
1181 json_object_add_value_int(percentile_object,
1186 states = json_create_object();
1187 json_object_add_value_object(block, "states", states);
1188 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1189 json_object_add_value_int(states,
1190 block_state_names[state],
1191 block_state_counts[state]);
1200 static void show_thread_status_terse(struct thread_stat *ts,
1201 struct group_run_stats *rs)
1203 if (terse_version == 2)
1204 show_thread_status_terse_v2(ts, rs);
1205 else if (terse_version == 3 || terse_version == 4)
1206 show_thread_status_terse_v3_v4(ts, rs, terse_version);
1208 log_err("fio: bad terse version!? %d\n", terse_version);
1211 struct json_object *show_thread_status(struct thread_stat *ts,
1212 struct group_run_stats *rs)
1214 if (output_format == FIO_OUTPUT_TERSE)
1215 show_thread_status_terse(ts, rs);
1216 else if (output_format == FIO_OUTPUT_JSON)
1217 return show_thread_status_json(ts, rs);
1219 show_thread_status_normal(ts, rs);
1223 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1227 if (src->samples == 0)
1230 dst->min_val = min(dst->min_val, src->min_val);
1231 dst->max_val = max(dst->max_val, src->max_val);
1234 * Compute new mean and S after the merge
1235 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1236 * #Parallel_algorithm>
1239 mean = src->mean.u.f;
1242 double delta = src->mean.u.f - dst->mean.u.f;
1244 mean = ((src->mean.u.f * src->samples) +
1245 (dst->mean.u.f * dst->samples)) /
1246 (dst->samples + src->samples);
1248 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1249 (dst->samples * src->samples) /
1250 (dst->samples + src->samples);
1253 dst->samples += src->samples;
1254 dst->mean.u.f = mean;
1258 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1262 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1263 if (dst->max_run[i] < src->max_run[i])
1264 dst->max_run[i] = src->max_run[i];
1265 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1266 dst->min_run[i] = src->min_run[i];
1267 if (dst->max_bw[i] < src->max_bw[i])
1268 dst->max_bw[i] = src->max_bw[i];
1269 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1270 dst->min_bw[i] = src->min_bw[i];
1272 dst->io_kb[i] += src->io_kb[i];
1273 dst->agg[i] += src->agg[i];
1277 dst->kb_base = src->kb_base;
1278 if (!dst->unit_base)
1279 dst->unit_base = src->unit_base;
1282 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1286 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1287 if (!dst->unified_rw_rep) {
1288 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1289 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1290 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1291 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1293 dst->io_bytes[l] += src->io_bytes[l];
1295 if (dst->runtime[l] < src->runtime[l])
1296 dst->runtime[l] = src->runtime[l];
1298 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1299 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1300 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1301 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1303 dst->io_bytes[0] += src->io_bytes[l];
1305 if (dst->runtime[0] < src->runtime[l])
1306 dst->runtime[0] = src->runtime[l];
1310 dst->usr_time += src->usr_time;
1311 dst->sys_time += src->sys_time;
1312 dst->ctx += src->ctx;
1313 dst->majf += src->majf;
1314 dst->minf += src->minf;
1316 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1317 dst->io_u_map[k] += src->io_u_map[k];
1318 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1319 dst->io_u_submit[k] += src->io_u_submit[k];
1320 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1321 dst->io_u_complete[k] += src->io_u_complete[k];
1322 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1323 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1324 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1325 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1327 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1328 if (!dst->unified_rw_rep) {
1329 dst->total_io_u[k] += src->total_io_u[k];
1330 dst->short_io_u[k] += src->short_io_u[k];
1331 dst->drop_io_u[k] += src->drop_io_u[k];
1333 dst->total_io_u[0] += src->total_io_u[k];
1334 dst->short_io_u[0] += src->short_io_u[k];
1335 dst->drop_io_u[0] += src->drop_io_u[k];
1339 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1342 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1343 if (!dst->unified_rw_rep)
1344 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1346 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1350 dst->total_run_time += src->total_run_time;
1351 dst->total_submit += src->total_submit;
1352 dst->total_complete += src->total_complete;
1355 void init_group_run_stat(struct group_run_stats *gs)
1358 memset(gs, 0, sizeof(*gs));
1360 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1361 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1364 void init_thread_stat(struct thread_stat *ts)
1368 memset(ts, 0, sizeof(*ts));
1370 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1371 ts->lat_stat[j].min_val = -1UL;
1372 ts->clat_stat[j].min_val = -1UL;
1373 ts->slat_stat[j].min_val = -1UL;
1374 ts->bw_stat[j].min_val = -1UL;
1379 void __show_run_stats(void)
1381 struct group_run_stats *runstats, *rs;
1382 struct thread_data *td;
1383 struct thread_stat *threadstats, *ts;
1384 int i, j, k, nr_ts, last_ts, idx;
1385 int kb_base_warned = 0;
1386 int unit_base_warned = 0;
1387 struct json_object *root = NULL;
1388 struct json_array *array = NULL;
1389 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1391 for (i = 0; i < groupid + 1; i++)
1392 init_group_run_stat(&runstats[i]);
1395 * find out how many threads stats we need. if group reporting isn't
1396 * enabled, it's one-per-td.
1400 for_each_td(td, i) {
1401 if (!td->o.group_reporting) {
1405 if (last_ts == td->groupid)
1408 last_ts = td->groupid;
1412 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1414 for (i = 0; i < nr_ts; i++)
1415 init_thread_stat(&threadstats[i]);
1420 for_each_td(td, i) {
1421 if (idx && (!td->o.group_reporting ||
1422 (td->o.group_reporting && last_ts != td->groupid))) {
1427 last_ts = td->groupid;
1429 ts = &threadstats[j];
1431 ts->clat_percentiles = td->o.clat_percentiles;
1432 ts->percentile_precision = td->o.percentile_precision;
1433 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1438 if (ts->groupid == -1) {
1440 * These are per-group shared already
1442 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1443 if (td->o.description)
1444 strncpy(ts->description, td->o.description,
1445 FIO_JOBDESC_SIZE - 1);
1447 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1450 * If multiple entries in this group, this is
1453 ts->thread_number = td->thread_number;
1454 ts->groupid = td->groupid;
1457 * first pid in group, not very useful...
1461 ts->kb_base = td->o.kb_base;
1462 ts->unit_base = td->o.unit_base;
1463 ts->unified_rw_rep = td->o.unified_rw_rep;
1464 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1465 log_info("fio: kb_base differs for jobs in group, using"
1466 " %u as the base\n", ts->kb_base);
1468 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1469 log_info("fio: unit_base differs for jobs in group, using"
1470 " %u as the base\n", ts->unit_base);
1471 unit_base_warned = 1;
1474 ts->continue_on_error = td->o.continue_on_error;
1475 ts->total_err_count += td->total_err_count;
1476 ts->first_error = td->first_error;
1478 if (!td->error && td->o.continue_on_error &&
1480 ts->error = td->first_error;
1481 ts->verror[sizeof(ts->verror) - 1] = '\0';
1482 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1483 } else if (td->error) {
1484 ts->error = td->error;
1485 ts->verror[sizeof(ts->verror) - 1] = '\0';
1486 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1490 ts->latency_depth = td->latency_qd;
1491 ts->latency_target = td->o.latency_target;
1492 ts->latency_percentile = td->o.latency_percentile;
1493 ts->latency_window = td->o.latency_window;
1495 ts->nr_block_infos = td->ts.nr_block_infos;
1496 for (k = 0; k < ts->nr_block_infos; k++)
1497 ts->block_infos[k] = td->ts.block_infos[k];
1499 sum_thread_stats(ts, &td->ts, idx);
1502 for (i = 0; i < nr_ts; i++) {
1503 unsigned long long bw;
1505 ts = &threadstats[i];
1506 rs = &runstats[ts->groupid];
1507 rs->kb_base = ts->kb_base;
1508 rs->unit_base = ts->unit_base;
1509 rs->unified_rw_rep += ts->unified_rw_rep;
1511 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1512 if (!ts->runtime[j])
1514 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1515 rs->min_run[j] = ts->runtime[j];
1516 if (ts->runtime[j] > rs->max_run[j])
1517 rs->max_run[j] = ts->runtime[j];
1520 if (ts->runtime[j]) {
1521 unsigned long runt = ts->runtime[j];
1522 unsigned long long kb;
1524 kb = ts->io_bytes[j] / rs->kb_base;
1525 bw = kb * 1000 / runt;
1527 if (bw < rs->min_bw[j])
1529 if (bw > rs->max_bw[j])
1532 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1536 for (i = 0; i < groupid + 1; i++) {
1541 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1542 if (rs->max_run[ddir])
1543 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1549 * don't overwrite last signal output
1551 if (output_format == FIO_OUTPUT_NORMAL)
1553 else if (output_format == FIO_OUTPUT_JSON) {
1558 os_ctime_r((const time_t *) &time_p, time_buf,
1560 time_buf[strlen(time_buf) - 1] = '\0';
1562 root = json_create_object();
1563 json_object_add_value_string(root, "fio version", fio_version_string);
1564 json_object_add_value_int(root, "timestamp", time_p);
1565 json_object_add_value_string(root, "time", time_buf);
1566 array = json_create_array();
1567 json_object_add_value_array(root, "jobs", array);
1570 for (i = 0; i < nr_ts; i++) {
1571 ts = &threadstats[i];
1572 rs = &runstats[ts->groupid];
1575 fio_server_send_ts(ts, rs);
1576 else if (output_format == FIO_OUTPUT_TERSE)
1577 show_thread_status_terse(ts, rs);
1578 else if (output_format == FIO_OUTPUT_JSON) {
1579 struct json_object *tmp = show_thread_status_json(ts, rs);
1580 json_array_add_value_object(array, tmp);
1582 show_thread_status_normal(ts, rs);
1584 if (output_format == FIO_OUTPUT_JSON) {
1585 /* disk util stats, if any */
1586 show_disk_util(1, root);
1588 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1590 json_print_object(root);
1592 json_free_object(root);
1595 for (i = 0; i < groupid + 1; i++) {
1600 fio_server_send_gs(rs);
1601 else if (output_format == FIO_OUTPUT_NORMAL)
1602 show_group_stats(rs);
1606 fio_server_send_du();
1607 else if (output_format == FIO_OUTPUT_NORMAL) {
1608 show_disk_util(0, NULL);
1609 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1612 if ( !(output_format == FIO_OUTPUT_TERSE) && append_terse_output) {
1613 log_info("\nAdditional Terse Output:\n");
1615 for (i = 0; i < nr_ts; i++) {
1616 ts = &threadstats[i];
1617 rs = &runstats[ts->groupid];
1618 show_thread_status_terse(ts, rs);
1627 void show_run_stats(void)
1629 fio_mutex_down(stat_mutex);
1631 fio_mutex_up(stat_mutex);
1634 void __show_running_run_stats(void)
1636 struct thread_data *td;
1637 unsigned long long *rt;
1641 fio_mutex_down(stat_mutex);
1643 rt = malloc(thread_number * sizeof(unsigned long long));
1644 fio_gettime(&tv, NULL);
1646 for_each_td(td, i) {
1647 rt[i] = mtime_since(&td->start, &tv);
1648 if (td_read(td) && td->io_bytes[DDIR_READ])
1649 td->ts.runtime[DDIR_READ] += rt[i];
1650 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1651 td->ts.runtime[DDIR_WRITE] += rt[i];
1652 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1653 td->ts.runtime[DDIR_TRIM] += rt[i];
1655 td->update_rusage = 1;
1656 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1657 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1658 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1659 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1662 for_each_td(td, i) {
1663 if (td->runstate >= TD_EXITED)
1665 if (td->rusage_sem) {
1666 td->update_rusage = 1;
1667 fio_mutex_down(td->rusage_sem);
1669 td->update_rusage = 0;
1674 for_each_td(td, i) {
1675 if (td_read(td) && td->io_bytes[DDIR_READ])
1676 td->ts.runtime[DDIR_READ] -= rt[i];
1677 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1678 td->ts.runtime[DDIR_WRITE] -= rt[i];
1679 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1680 td->ts.runtime[DDIR_TRIM] -= rt[i];
1684 fio_mutex_up(stat_mutex);
1687 static int status_interval_init;
1688 static struct timeval status_time;
1689 static int status_file_disabled;
1691 #define FIO_STATUS_FILE "fio-dump-status"
1693 static int check_status_file(void)
1696 const char *temp_dir;
1697 char fio_status_file_path[PATH_MAX];
1699 if (status_file_disabled)
1702 temp_dir = getenv("TMPDIR");
1703 if (temp_dir == NULL) {
1704 temp_dir = getenv("TEMP");
1705 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1708 if (temp_dir == NULL)
1711 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1713 if (stat(fio_status_file_path, &sb))
1716 if (unlink(fio_status_file_path) < 0) {
1717 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1719 log_err("fio: disabling status file updates\n");
1720 status_file_disabled = 1;
1726 void check_for_running_stats(void)
1728 if (status_interval) {
1729 if (!status_interval_init) {
1730 fio_gettime(&status_time, NULL);
1731 status_interval_init = 1;
1732 } else if (mtime_since_now(&status_time) >= status_interval) {
1733 show_running_run_stats();
1734 fio_gettime(&status_time, NULL);
1738 if (check_status_file()) {
1739 show_running_run_stats();
1744 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1749 if (data > is->max_val)
1751 if (data < is->min_val)
1754 delta = val - is->mean.u.f;
1756 is->mean.u.f += delta / (is->samples + 1.0);
1757 is->S.u.f += delta * (val - is->mean.u.f);
1763 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1764 enum fio_ddir ddir, unsigned int bs,
1765 unsigned long t, uint64_t offset)
1767 uint64_t nr_samples = iolog->nr_samples;
1768 struct io_sample *s;
1770 if (iolog->disabled)
1773 if (!iolog->nr_samples)
1774 iolog->avg_last = t;
1776 if (iolog->nr_samples == iolog->max_samples) {
1780 new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
1782 if (iolog->log_gz && (new_size > iolog->log_gz)) {
1783 if (iolog_flush(iolog, 0)) {
1784 log_err("fio: failed flushing iolog! Will stop logging.\n");
1785 iolog->disabled = 1;
1788 nr_samples = iolog->nr_samples;
1790 new_log = realloc(iolog->log, new_size);
1792 log_err("fio: failed extending iolog! Will stop logging.\n");
1793 iolog->disabled = 1;
1796 iolog->log = new_log;
1797 iolog->max_samples <<= 1;
1801 s = get_sample(iolog, nr_samples);
1805 io_sample_set_ddir(iolog, s, ddir);
1808 if (iolog->log_offset) {
1809 struct io_sample_offset *so = (void *) s;
1811 so->offset = offset;
1814 iolog->nr_samples++;
1817 static inline void reset_io_stat(struct io_stat *ios)
1819 ios->max_val = ios->min_val = ios->samples = 0;
1820 ios->mean.u.f = ios->S.u.f = 0;
1823 void reset_io_stats(struct thread_data *td)
1825 struct thread_stat *ts = &td->ts;
1828 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1829 reset_io_stat(&ts->clat_stat[i]);
1830 reset_io_stat(&ts->slat_stat[i]);
1831 reset_io_stat(&ts->lat_stat[i]);
1832 reset_io_stat(&ts->bw_stat[i]);
1833 reset_io_stat(&ts->iops_stat[i]);
1835 ts->io_bytes[i] = 0;
1838 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1839 ts->io_u_plat[i][j] = 0;
1842 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1843 ts->io_u_map[i] = 0;
1844 ts->io_u_submit[i] = 0;
1845 ts->io_u_complete[i] = 0;
1846 ts->io_u_lat_u[i] = 0;
1847 ts->io_u_lat_m[i] = 0;
1848 ts->total_submit = 0;
1849 ts->total_complete = 0;
1852 for (i = 0; i < 3; i++) {
1853 ts->total_io_u[i] = 0;
1854 ts->short_io_u[i] = 0;
1855 ts->drop_io_u[i] = 0;
1859 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
1862 * Note an entry in the log. Use the mean from the logged samples,
1863 * making sure to properly round up. Only write a log entry if we
1864 * had actual samples done.
1866 if (iolog->avg_window[DDIR_READ].samples) {
1869 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1870 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed, 0);
1872 if (iolog->avg_window[DDIR_WRITE].samples) {
1875 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1876 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
1878 if (iolog->avg_window[DDIR_TRIM].samples) {
1881 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1882 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed, 0);
1885 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1886 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1887 reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1890 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1891 unsigned long val, enum fio_ddir ddir,
1892 unsigned int bs, uint64_t offset)
1894 unsigned long elapsed, this_window;
1899 elapsed = mtime_since_now(&td->epoch);
1902 * If no time averaging, just add the log sample.
1904 if (!iolog->avg_msec) {
1905 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
1910 * Add the sample. If the time period has passed, then
1911 * add that entry to the log and clear.
1913 add_stat_sample(&iolog->avg_window[ddir], val);
1916 * If period hasn't passed, adding the above sample is all we
1919 this_window = elapsed - iolog->avg_last;
1920 if (this_window < iolog->avg_msec)
1923 _add_stat_to_log(iolog, elapsed);
1925 iolog->avg_last = elapsed;
1928 void finalize_logs(struct thread_data *td)
1930 unsigned long elapsed;
1932 elapsed = mtime_since_now(&td->epoch);
1935 _add_stat_to_log(td->clat_log, elapsed);
1937 _add_stat_to_log(td->slat_log, elapsed);
1939 _add_stat_to_log(td->lat_log, elapsed);
1941 _add_stat_to_log(td->bw_log, elapsed);
1943 _add_stat_to_log(td->iops_log, elapsed);
1946 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1948 struct io_log *iolog;
1953 iolog = agg_io_log[ddir];
1954 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
1957 static void add_clat_percentile_sample(struct thread_stat *ts,
1958 unsigned long usec, enum fio_ddir ddir)
1960 unsigned int idx = plat_val_to_idx(usec);
1961 assert(idx < FIO_IO_U_PLAT_NR);
1963 ts->io_u_plat[ddir][idx]++;
1966 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1967 unsigned long usec, unsigned int bs, uint64_t offset)
1969 struct thread_stat *ts = &td->ts;
1976 add_stat_sample(&ts->clat_stat[ddir], usec);
1979 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
1981 if (ts->clat_percentiles)
1982 add_clat_percentile_sample(ts, usec, ddir);
1987 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1988 unsigned long usec, unsigned int bs, uint64_t offset)
1990 struct thread_stat *ts = &td->ts;
1997 add_stat_sample(&ts->slat_stat[ddir], usec);
2000 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
2005 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2006 unsigned long usec, unsigned int bs, uint64_t offset)
2008 struct thread_stat *ts = &td->ts;
2015 add_stat_sample(&ts->lat_stat[ddir], usec);
2018 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2023 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
2026 struct thread_stat *ts = &td->ts;
2027 unsigned long spent, rate;
2032 spent = mtime_since(&td->bw_sample_time, t);
2033 if (spent < td->o.bw_avg_time)
2039 * Compute both read and write rates for the interval.
2041 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2044 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2046 continue; /* No entries for interval */
2049 rate = delta * 1000 / spent / 1024;
2053 add_stat_sample(&ts->bw_stat[ddir], rate);
2056 add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2058 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2061 fio_gettime(&td->bw_sample_time, NULL);
2065 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
2068 struct thread_stat *ts = &td->ts;
2069 unsigned long spent, iops;
2074 spent = mtime_since(&td->iops_sample_time, t);
2075 if (spent < td->o.iops_avg_time)
2081 * Compute both read and write rates for the interval.
2083 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2086 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2088 continue; /* No entries for interval */
2091 iops = (delta * 1000) / spent;
2095 add_stat_sample(&ts->iops_stat[ddir], iops);
2098 add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2100 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2103 fio_gettime(&td->iops_sample_time, NULL);
2107 void stat_init(void)
2109 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2112 void stat_exit(void)
2115 * When we have the mutex, we know out-of-band access to it
2118 fio_mutex_down(stat_mutex);
2119 fio_mutex_remove(stat_mutex);
2123 * Called from signal handler. Wake up status thread.
2125 void show_running_run_stats(void)
2128 pthread_cond_signal(&helper_cond);
2131 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2133 /* Ignore io_u's which span multiple blocks--they will just get
2134 * inaccurate counts. */
2135 int idx = (io_u->offset - io_u->file->file_offset)
2136 / td->o.bs[DDIR_TRIM];
2137 uint32_t *info = &td->ts.block_infos[idx];
2138 assert(idx < td->ts.nr_block_infos);