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 bool 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_thread_status_normal(struct thread_stat *ts,
661 struct group_run_stats *rs,
662 struct buf_output *out)
664 double usr_cpu, sys_cpu;
665 unsigned long runtime;
666 double io_u_dist[FIO_IO_U_MAP_NR];
670 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
673 memset(time_buf, 0, sizeof(time_buf));
676 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
679 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
680 ts->name, ts->groupid, ts->members,
681 ts->error, (int) ts->pid, time_buf);
683 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
684 ts->name, ts->groupid, ts->members,
685 ts->error, ts->verror, (int) ts->pid,
689 if (strlen(ts->description))
690 log_buf(out, " Description : [%s]\n", ts->description);
692 if (ts->io_bytes[DDIR_READ])
693 show_ddir_status(rs, ts, DDIR_READ, out);
694 if (ts->io_bytes[DDIR_WRITE])
695 show_ddir_status(rs, ts, DDIR_WRITE, out);
696 if (ts->io_bytes[DDIR_TRIM])
697 show_ddir_status(rs, ts, DDIR_TRIM, out);
699 show_latencies(ts, out);
701 runtime = ts->total_run_time;
703 double runt = (double) runtime;
705 usr_cpu = (double) ts->usr_time * 100 / runt;
706 sys_cpu = (double) ts->sys_time * 100 / runt;
712 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
713 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
714 (unsigned long long) ts->ctx,
715 (unsigned long long) ts->majf,
716 (unsigned long long) ts->minf);
718 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
719 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
720 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
721 io_u_dist[1], io_u_dist[2],
722 io_u_dist[3], io_u_dist[4],
723 io_u_dist[5], io_u_dist[6]);
725 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
726 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
727 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
728 io_u_dist[1], io_u_dist[2],
729 io_u_dist[3], io_u_dist[4],
730 io_u_dist[5], io_u_dist[6]);
731 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
732 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
733 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
734 io_u_dist[1], io_u_dist[2],
735 io_u_dist[3], io_u_dist[4],
736 io_u_dist[5], io_u_dist[6]);
737 log_buf(out, " issued : total=r=%llu/w=%llu/d=%llu,"
738 " short=r=%llu/w=%llu/d=%llu,"
739 " drop=r=%llu/w=%llu/d=%llu\n",
740 (unsigned long long) ts->total_io_u[0],
741 (unsigned long long) ts->total_io_u[1],
742 (unsigned long long) ts->total_io_u[2],
743 (unsigned long long) ts->short_io_u[0],
744 (unsigned long long) ts->short_io_u[1],
745 (unsigned long long) ts->short_io_u[2],
746 (unsigned long long) ts->drop_io_u[0],
747 (unsigned long long) ts->drop_io_u[1],
748 (unsigned long long) ts->drop_io_u[2]);
749 if (ts->continue_on_error) {
750 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
751 (unsigned long long)ts->total_err_count,
753 strerror(ts->first_error));
755 if (ts->latency_depth) {
756 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
757 (unsigned long long)ts->latency_target,
758 (unsigned long long)ts->latency_window,
759 ts->latency_percentile.u.f,
763 if (ts->nr_block_infos)
764 show_block_infos(ts->nr_block_infos, ts->block_infos,
765 ts->percentile_list, out);
768 static void show_ddir_status_terse(struct thread_stat *ts,
769 struct group_run_stats *rs, int ddir,
770 struct buf_output *out)
772 unsigned long min, max;
773 unsigned long long bw, iops;
774 unsigned int *ovals = NULL;
776 unsigned int len, minv, maxv;
779 assert(ddir_rw(ddir));
782 if (ts->runtime[ddir]) {
783 uint64_t runt = ts->runtime[ddir];
785 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
786 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
789 log_buf(out, ";%llu;%llu;%llu;%llu",
790 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
791 (unsigned long long) ts->runtime[ddir]);
793 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
794 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
796 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
798 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
799 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
801 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
803 if (ts->clat_percentiles) {
804 len = calc_clat_percentiles(ts->io_u_plat[ddir],
805 ts->clat_stat[ddir].samples,
806 ts->percentile_list, &ovals, &maxv,
811 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
813 log_buf(out, ";0%%=0");
816 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
819 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
820 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
822 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
827 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
828 double p_of_agg = 100.0;
831 p_of_agg = mean * 100 / (double) rs->agg[ddir];
832 if (p_of_agg > 100.0)
836 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
838 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
841 static void add_ddir_status_json(struct thread_stat *ts,
842 struct group_run_stats *rs, int ddir, struct json_object *parent)
844 unsigned long min, max;
845 unsigned long long bw;
846 unsigned int *ovals = NULL;
847 double mean, dev, iops;
848 unsigned int len, minv, maxv;
850 const char *ddirname[] = {"read", "write", "trim"};
851 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
853 double p_of_agg = 100.0;
855 assert(ddir_rw(ddir));
857 if (ts->unified_rw_rep && ddir != DDIR_READ)
860 dir_object = json_create_object();
861 json_object_add_value_object(parent,
862 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
866 if (ts->runtime[ddir]) {
867 uint64_t runt = ts->runtime[ddir];
869 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
870 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
873 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
874 json_object_add_value_int(dir_object, "bw", bw);
875 json_object_add_value_float(dir_object, "iops", iops);
876 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
877 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
878 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
879 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
881 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
885 tmp_object = json_create_object();
886 json_object_add_value_object(dir_object, "slat", tmp_object);
887 json_object_add_value_int(tmp_object, "min", min);
888 json_object_add_value_int(tmp_object, "max", max);
889 json_object_add_value_float(tmp_object, "mean", mean);
890 json_object_add_value_float(tmp_object, "stddev", dev);
892 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
896 tmp_object = json_create_object();
897 json_object_add_value_object(dir_object, "clat", tmp_object);
898 json_object_add_value_int(tmp_object, "min", min);
899 json_object_add_value_int(tmp_object, "max", max);
900 json_object_add_value_float(tmp_object, "mean", mean);
901 json_object_add_value_float(tmp_object, "stddev", dev);
903 if (ts->clat_percentiles) {
904 len = calc_clat_percentiles(ts->io_u_plat[ddir],
905 ts->clat_stat[ddir].samples,
906 ts->percentile_list, &ovals, &maxv,
911 percentile_object = json_create_object();
912 json_object_add_value_object(tmp_object, "percentile", percentile_object);
913 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
915 json_object_add_value_int(percentile_object, "0.00", 0);
918 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
919 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
922 if (output_format & FIO_OUTPUT_JSON_PLUS) {
923 clat_bins_object = json_create_object();
924 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
925 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
926 snprintf(buf, sizeof(buf), "%d", i);
927 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
929 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS);
930 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL);
931 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR);
934 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
938 tmp_object = json_create_object();
939 json_object_add_value_object(dir_object, "lat", tmp_object);
940 json_object_add_value_int(tmp_object, "min", min);
941 json_object_add_value_int(tmp_object, "max", max);
942 json_object_add_value_float(tmp_object, "mean", mean);
943 json_object_add_value_float(tmp_object, "stddev", dev);
947 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
949 p_of_agg = mean * 100 / (double) rs->agg[ddir];
950 if (p_of_agg > 100.0)
955 p_of_agg = mean = dev = 0.0;
957 json_object_add_value_int(dir_object, "bw_min", min);
958 json_object_add_value_int(dir_object, "bw_max", max);
959 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
960 json_object_add_value_float(dir_object, "bw_mean", mean);
961 json_object_add_value_float(dir_object, "bw_dev", dev);
964 static void show_thread_status_terse_v2(struct thread_stat *ts,
965 struct group_run_stats *rs,
966 struct buf_output *out)
968 double io_u_dist[FIO_IO_U_MAP_NR];
969 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
970 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
971 double usr_cpu, sys_cpu;
975 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
976 /* Log Read Status */
977 show_ddir_status_terse(ts, rs, DDIR_READ, out);
978 /* Log Write Status */
979 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
980 /* Log Trim Status */
981 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
984 if (ts->total_run_time) {
985 double runt = (double) ts->total_run_time;
987 usr_cpu = (double) ts->usr_time * 100 / runt;
988 sys_cpu = (double) ts->sys_time * 100 / runt;
994 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
995 (unsigned long long) ts->ctx,
996 (unsigned long long) ts->majf,
997 (unsigned long long) ts->minf);
999 /* Calc % distribution of IO depths, usecond, msecond latency */
1000 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1001 stat_calc_lat_u(ts, io_u_lat_u);
1002 stat_calc_lat_m(ts, io_u_lat_m);
1004 /* Only show fixed 7 I/O depth levels*/
1005 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1006 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1007 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1009 /* Microsecond latency */
1010 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1011 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1012 /* Millisecond latency */
1013 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1014 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1015 /* Additional output if continue_on_error set - default off*/
1016 if (ts->continue_on_error)
1017 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1020 /* Additional output if description is set */
1021 if (strlen(ts->description))
1022 log_buf(out, ";%s", ts->description);
1027 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1028 struct group_run_stats *rs, int ver,
1029 struct buf_output *out)
1031 double io_u_dist[FIO_IO_U_MAP_NR];
1032 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1033 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1034 double usr_cpu, sys_cpu;
1038 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1039 ts->name, ts->groupid, ts->error);
1040 /* Log Read Status */
1041 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1042 /* Log Write Status */
1043 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1044 /* Log Trim Status */
1046 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1049 if (ts->total_run_time) {
1050 double runt = (double) ts->total_run_time;
1052 usr_cpu = (double) ts->usr_time * 100 / runt;
1053 sys_cpu = (double) ts->sys_time * 100 / runt;
1059 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1060 (unsigned long long) ts->ctx,
1061 (unsigned long long) ts->majf,
1062 (unsigned long long) ts->minf);
1064 /* Calc % distribution of IO depths, usecond, msecond latency */
1065 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1066 stat_calc_lat_u(ts, io_u_lat_u);
1067 stat_calc_lat_m(ts, io_u_lat_m);
1069 /* Only show fixed 7 I/O depth levels*/
1070 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1071 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1072 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1074 /* Microsecond latency */
1075 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1076 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1077 /* Millisecond latency */
1078 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1079 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1081 /* disk util stats, if any */
1082 show_disk_util(1, NULL, out);
1084 /* Additional output if continue_on_error set - default off*/
1085 if (ts->continue_on_error)
1086 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1088 /* Additional output if description is set */
1089 if (strlen(ts->description))
1090 log_buf(out, ";%s", ts->description);
1095 static void json_add_job_opts(struct json_object *root, const char *name,
1096 struct flist_head *opt_list, bool num_jobs)
1098 struct json_object *dir_object;
1099 struct flist_head *entry;
1100 struct print_option *p;
1102 if (flist_empty(opt_list))
1105 dir_object = json_create_object();
1106 json_object_add_value_object(root, name, dir_object);
1108 flist_for_each(entry, opt_list) {
1109 const char *pos = "";
1111 p = flist_entry(entry, struct print_option, list);
1112 if (!num_jobs && !strcmp(p->name, "numjobs"))
1116 json_object_add_value_string(dir_object, p->name, pos);
1120 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1121 struct group_run_stats *rs,
1122 struct flist_head *opt_list)
1124 struct json_object *root, *tmp;
1125 struct jobs_eta *je;
1126 double io_u_dist[FIO_IO_U_MAP_NR];
1127 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1128 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1129 double usr_cpu, sys_cpu;
1133 root = json_create_object();
1134 json_object_add_value_string(root, "jobname", ts->name);
1135 json_object_add_value_int(root, "groupid", ts->groupid);
1136 json_object_add_value_int(root, "error", ts->error);
1139 je = get_jobs_eta(true, &size);
1141 json_object_add_value_int(root, "eta", je->eta_sec);
1142 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1146 json_add_job_opts(root, "job options", opt_list, true);
1148 add_ddir_status_json(ts, rs, DDIR_READ, root);
1149 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1150 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1153 if (ts->total_run_time) {
1154 double runt = (double) ts->total_run_time;
1156 usr_cpu = (double) ts->usr_time * 100 / runt;
1157 sys_cpu = (double) ts->sys_time * 100 / runt;
1162 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1163 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1164 json_object_add_value_int(root, "ctx", ts->ctx);
1165 json_object_add_value_int(root, "majf", ts->majf);
1166 json_object_add_value_int(root, "minf", ts->minf);
1169 /* Calc % distribution of IO depths, usecond, msecond latency */
1170 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1171 stat_calc_lat_u(ts, io_u_lat_u);
1172 stat_calc_lat_m(ts, io_u_lat_m);
1174 tmp = json_create_object();
1175 json_object_add_value_object(root, "iodepth_level", tmp);
1176 /* Only show fixed 7 I/O depth levels*/
1177 for (i = 0; i < 7; i++) {
1180 snprintf(name, 20, "%d", 1 << i);
1182 snprintf(name, 20, ">=%d", 1 << i);
1183 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1186 tmp = json_create_object();
1187 json_object_add_value_object(root, "latency_us", tmp);
1188 /* Microsecond latency */
1189 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1190 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1191 "250", "500", "750", "1000", };
1192 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1194 /* Millisecond latency */
1195 tmp = json_create_object();
1196 json_object_add_value_object(root, "latency_ms", tmp);
1197 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1198 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1199 "250", "500", "750", "1000", "2000",
1201 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1204 /* Additional output if continue_on_error set - default off*/
1205 if (ts->continue_on_error) {
1206 json_object_add_value_int(root, "total_err", ts->total_err_count);
1207 json_object_add_value_int(root, "first_error", ts->first_error);
1210 if (ts->latency_depth) {
1211 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1212 json_object_add_value_int(root, "latency_target", ts->latency_target);
1213 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1214 json_object_add_value_int(root, "latency_window", ts->latency_window);
1217 /* Additional output if description is set */
1218 if (strlen(ts->description))
1219 json_object_add_value_string(root, "desc", ts->description);
1221 if (ts->nr_block_infos) {
1222 /* Block error histogram and types */
1224 unsigned int *percentiles = NULL;
1225 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1227 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1228 ts->percentile_list,
1229 &percentiles, block_state_counts);
1232 struct json_object *block, *percentile_object, *states;
1234 block = json_create_object();
1235 json_object_add_value_object(root, "block", block);
1237 percentile_object = json_create_object();
1238 json_object_add_value_object(block, "percentiles",
1240 for (i = 0; i < len; i++) {
1242 snprintf(buf, sizeof(buf), "%f",
1243 ts->percentile_list[i].u.f);
1244 json_object_add_value_int(percentile_object,
1249 states = json_create_object();
1250 json_object_add_value_object(block, "states", states);
1251 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1252 json_object_add_value_int(states,
1253 block_state_names[state],
1254 block_state_counts[state]);
1263 static void show_thread_status_terse(struct thread_stat *ts,
1264 struct group_run_stats *rs,
1265 struct buf_output *out)
1267 if (terse_version == 2)
1268 show_thread_status_terse_v2(ts, rs, out);
1269 else if (terse_version == 3 || terse_version == 4)
1270 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1272 log_err("fio: bad terse version!? %d\n", terse_version);
1275 struct json_object *show_thread_status(struct thread_stat *ts,
1276 struct group_run_stats *rs,
1277 struct flist_head *opt_list,
1278 struct buf_output *out)
1280 struct json_object *ret = NULL;
1282 if (output_format & FIO_OUTPUT_TERSE)
1283 show_thread_status_terse(ts, rs, out);
1284 if (output_format & FIO_OUTPUT_JSON)
1285 ret = show_thread_status_json(ts, rs, opt_list);
1286 if (output_format & FIO_OUTPUT_NORMAL)
1287 show_thread_status_normal(ts, rs, out);
1292 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1296 if (src->samples == 0)
1299 dst->min_val = min(dst->min_val, src->min_val);
1300 dst->max_val = max(dst->max_val, src->max_val);
1303 * Compute new mean and S after the merge
1304 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1305 * #Parallel_algorithm>
1308 mean = src->mean.u.f;
1311 double delta = src->mean.u.f - dst->mean.u.f;
1313 mean = ((src->mean.u.f * src->samples) +
1314 (dst->mean.u.f * dst->samples)) /
1315 (dst->samples + src->samples);
1317 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1318 (dst->samples * src->samples) /
1319 (dst->samples + src->samples);
1322 dst->samples += src->samples;
1323 dst->mean.u.f = mean;
1327 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1331 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1332 if (dst->max_run[i] < src->max_run[i])
1333 dst->max_run[i] = src->max_run[i];
1334 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1335 dst->min_run[i] = src->min_run[i];
1336 if (dst->max_bw[i] < src->max_bw[i])
1337 dst->max_bw[i] = src->max_bw[i];
1338 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1339 dst->min_bw[i] = src->min_bw[i];
1341 dst->io_kb[i] += src->io_kb[i];
1342 dst->agg[i] += src->agg[i];
1346 dst->kb_base = src->kb_base;
1347 if (!dst->unit_base)
1348 dst->unit_base = src->unit_base;
1351 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1356 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1357 if (!dst->unified_rw_rep) {
1358 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1359 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1360 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1361 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1363 dst->io_bytes[l] += src->io_bytes[l];
1365 if (dst->runtime[l] < src->runtime[l])
1366 dst->runtime[l] = src->runtime[l];
1368 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1369 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1370 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1371 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1373 dst->io_bytes[0] += src->io_bytes[l];
1375 if (dst->runtime[0] < src->runtime[l])
1376 dst->runtime[0] = src->runtime[l];
1379 * We're summing to the same destination, so override
1380 * 'first' after the first iteration of the loop
1386 dst->usr_time += src->usr_time;
1387 dst->sys_time += src->sys_time;
1388 dst->ctx += src->ctx;
1389 dst->majf += src->majf;
1390 dst->minf += src->minf;
1392 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1393 dst->io_u_map[k] += src->io_u_map[k];
1394 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1395 dst->io_u_submit[k] += src->io_u_submit[k];
1396 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1397 dst->io_u_complete[k] += src->io_u_complete[k];
1398 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1399 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1400 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1401 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1403 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1404 if (!dst->unified_rw_rep) {
1405 dst->total_io_u[k] += src->total_io_u[k];
1406 dst->short_io_u[k] += src->short_io_u[k];
1407 dst->drop_io_u[k] += src->drop_io_u[k];
1409 dst->total_io_u[0] += src->total_io_u[k];
1410 dst->short_io_u[0] += src->short_io_u[k];
1411 dst->drop_io_u[0] += src->drop_io_u[k];
1415 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1418 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1419 if (!dst->unified_rw_rep)
1420 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1422 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1426 dst->total_run_time += src->total_run_time;
1427 dst->total_submit += src->total_submit;
1428 dst->total_complete += src->total_complete;
1431 void init_group_run_stat(struct group_run_stats *gs)
1434 memset(gs, 0, sizeof(*gs));
1436 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1437 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1440 void init_thread_stat(struct thread_stat *ts)
1444 memset(ts, 0, sizeof(*ts));
1446 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1447 ts->lat_stat[j].min_val = -1UL;
1448 ts->clat_stat[j].min_val = -1UL;
1449 ts->slat_stat[j].min_val = -1UL;
1450 ts->bw_stat[j].min_val = -1UL;
1455 void __show_run_stats(void)
1457 struct group_run_stats *runstats, *rs;
1458 struct thread_data *td;
1459 struct thread_stat *threadstats, *ts;
1460 int i, j, k, nr_ts, last_ts, idx;
1461 int kb_base_warned = 0;
1462 int unit_base_warned = 0;
1463 struct json_object *root = NULL;
1464 struct json_array *array = NULL;
1465 struct buf_output output[FIO_OUTPUT_NR];
1466 struct flist_head **opt_lists;
1468 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1470 for (i = 0; i < groupid + 1; i++)
1471 init_group_run_stat(&runstats[i]);
1474 * find out how many threads stats we need. if group reporting isn't
1475 * enabled, it's one-per-td.
1479 for_each_td(td, i) {
1480 if (!td->o.group_reporting) {
1484 if (last_ts == td->groupid)
1487 last_ts = td->groupid;
1491 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1492 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1494 for (i = 0; i < nr_ts; i++) {
1495 init_thread_stat(&threadstats[i]);
1496 opt_lists[i] = NULL;
1502 for_each_td(td, i) {
1503 if (idx && (!td->o.group_reporting ||
1504 (td->o.group_reporting && last_ts != td->groupid))) {
1509 last_ts = td->groupid;
1511 ts = &threadstats[j];
1513 ts->clat_percentiles = td->o.clat_percentiles;
1514 ts->percentile_precision = td->o.percentile_precision;
1515 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1516 opt_lists[j] = &td->opt_list;
1521 if (ts->groupid == -1) {
1523 * These are per-group shared already
1525 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1526 if (td->o.description)
1527 strncpy(ts->description, td->o.description,
1528 FIO_JOBDESC_SIZE - 1);
1530 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1533 * If multiple entries in this group, this is
1536 ts->thread_number = td->thread_number;
1537 ts->groupid = td->groupid;
1540 * first pid in group, not very useful...
1544 ts->kb_base = td->o.kb_base;
1545 ts->unit_base = td->o.unit_base;
1546 ts->unified_rw_rep = td->o.unified_rw_rep;
1547 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1548 log_info("fio: kb_base differs for jobs in group, using"
1549 " %u as the base\n", ts->kb_base);
1551 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1552 log_info("fio: unit_base differs for jobs in group, using"
1553 " %u as the base\n", ts->unit_base);
1554 unit_base_warned = 1;
1557 ts->continue_on_error = td->o.continue_on_error;
1558 ts->total_err_count += td->total_err_count;
1559 ts->first_error = td->first_error;
1561 if (!td->error && td->o.continue_on_error &&
1563 ts->error = td->first_error;
1564 ts->verror[sizeof(ts->verror) - 1] = '\0';
1565 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1566 } else if (td->error) {
1567 ts->error = td->error;
1568 ts->verror[sizeof(ts->verror) - 1] = '\0';
1569 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1573 ts->latency_depth = td->latency_qd;
1574 ts->latency_target = td->o.latency_target;
1575 ts->latency_percentile = td->o.latency_percentile;
1576 ts->latency_window = td->o.latency_window;
1578 ts->nr_block_infos = td->ts.nr_block_infos;
1579 for (k = 0; k < ts->nr_block_infos; k++)
1580 ts->block_infos[k] = td->ts.block_infos[k];
1582 sum_thread_stats(ts, &td->ts, idx == 1);
1585 for (i = 0; i < nr_ts; i++) {
1586 unsigned long long bw;
1588 ts = &threadstats[i];
1589 if (ts->groupid == -1)
1591 rs = &runstats[ts->groupid];
1592 rs->kb_base = ts->kb_base;
1593 rs->unit_base = ts->unit_base;
1594 rs->unified_rw_rep += ts->unified_rw_rep;
1596 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1597 if (!ts->runtime[j])
1599 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1600 rs->min_run[j] = ts->runtime[j];
1601 if (ts->runtime[j] > rs->max_run[j])
1602 rs->max_run[j] = ts->runtime[j];
1605 if (ts->runtime[j]) {
1606 unsigned long runt = ts->runtime[j];
1607 unsigned long long kb;
1609 kb = ts->io_bytes[j] / rs->kb_base;
1610 bw = kb * 1000 / runt;
1612 if (bw < rs->min_bw[j])
1614 if (bw > rs->max_bw[j])
1617 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1621 for (i = 0; i < groupid + 1; i++) {
1626 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1627 if (rs->max_run[ddir])
1628 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1633 for (i = 0; i < FIO_OUTPUT_NR; i++)
1634 buf_output_init(&output[i]);
1637 * don't overwrite last signal output
1639 if (output_format & FIO_OUTPUT_NORMAL)
1640 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1641 if (output_format & FIO_OUTPUT_JSON) {
1642 struct thread_data *global;
1645 unsigned long long ms_since_epoch;
1647 gettimeofday(&now, NULL);
1648 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1649 (unsigned long long)(now.tv_usec) / 1000;
1651 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1653 if (time_buf[strlen(time_buf) - 1] == '\n')
1654 time_buf[strlen(time_buf) - 1] = '\0';
1656 root = json_create_object();
1657 json_object_add_value_string(root, "fio version", fio_version_string);
1658 json_object_add_value_int(root, "timestamp", now.tv_sec);
1659 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1660 json_object_add_value_string(root, "time", time_buf);
1661 global = get_global_options();
1662 json_add_job_opts(root, "global options", &global->opt_list, false);
1663 array = json_create_array();
1664 json_object_add_value_array(root, "jobs", array);
1668 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1670 for (i = 0; i < nr_ts; i++) {
1671 ts = &threadstats[i];
1672 rs = &runstats[ts->groupid];
1675 fio_server_send_job_options(opt_lists[i], i);
1676 fio_server_send_ts(ts, rs);
1678 if (output_format & FIO_OUTPUT_TERSE)
1679 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1680 if (output_format & FIO_OUTPUT_JSON) {
1681 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1682 json_array_add_value_object(array, tmp);
1684 if (output_format & FIO_OUTPUT_NORMAL)
1685 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1688 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1689 /* disk util stats, if any */
1690 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1692 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1694 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1695 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1696 json_free_object(root);
1699 for (i = 0; i < groupid + 1; i++) {
1704 fio_server_send_gs(rs);
1705 else if (output_format & FIO_OUTPUT_NORMAL)
1706 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1710 fio_server_send_du();
1711 else if (output_format & FIO_OUTPUT_NORMAL) {
1712 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1713 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1716 for (i = 0; i < FIO_OUTPUT_NR; i++) {
1717 buf_output_flush(&output[i]);
1718 buf_output_free(&output[i]);
1727 void show_run_stats(void)
1729 fio_mutex_down(stat_mutex);
1731 fio_mutex_up(stat_mutex);
1734 void __show_running_run_stats(void)
1736 struct thread_data *td;
1737 unsigned long long *rt;
1741 fio_mutex_down(stat_mutex);
1743 rt = malloc(thread_number * sizeof(unsigned long long));
1744 fio_gettime(&tv, NULL);
1746 for_each_td(td, i) {
1747 td->update_rusage = 1;
1748 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1749 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1750 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1751 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1753 rt[i] = mtime_since(&td->start, &tv);
1754 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1755 td->ts.runtime[DDIR_READ] += rt[i];
1756 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1757 td->ts.runtime[DDIR_WRITE] += rt[i];
1758 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1759 td->ts.runtime[DDIR_TRIM] += rt[i];
1762 for_each_td(td, i) {
1763 if (td->runstate >= TD_EXITED)
1765 if (td->rusage_sem) {
1766 td->update_rusage = 1;
1767 fio_mutex_down(td->rusage_sem);
1769 td->update_rusage = 0;
1774 for_each_td(td, i) {
1775 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1776 td->ts.runtime[DDIR_READ] -= rt[i];
1777 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1778 td->ts.runtime[DDIR_WRITE] -= rt[i];
1779 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1780 td->ts.runtime[DDIR_TRIM] -= rt[i];
1784 fio_mutex_up(stat_mutex);
1787 static int status_interval_init;
1788 static struct timeval status_time;
1789 static int status_file_disabled;
1791 #define FIO_STATUS_FILE "fio-dump-status"
1793 static int check_status_file(void)
1796 const char *temp_dir;
1797 char fio_status_file_path[PATH_MAX];
1799 if (status_file_disabled)
1802 temp_dir = getenv("TMPDIR");
1803 if (temp_dir == NULL) {
1804 temp_dir = getenv("TEMP");
1805 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1808 if (temp_dir == NULL)
1811 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1813 if (stat(fio_status_file_path, &sb))
1816 if (unlink(fio_status_file_path) < 0) {
1817 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1819 log_err("fio: disabling status file updates\n");
1820 status_file_disabled = 1;
1826 void check_for_running_stats(void)
1828 if (status_interval) {
1829 if (!status_interval_init) {
1830 fio_gettime(&status_time, NULL);
1831 status_interval_init = 1;
1832 } else if (mtime_since_now(&status_time) >= status_interval) {
1833 show_running_run_stats();
1834 fio_gettime(&status_time, NULL);
1838 if (check_status_file()) {
1839 show_running_run_stats();
1844 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1849 if (data > is->max_val)
1851 if (data < is->min_val)
1854 delta = val - is->mean.u.f;
1856 is->mean.u.f += delta / (is->samples + 1.0);
1857 is->S.u.f += delta * (val - is->mean.u.f);
1864 * Return a struct io_logs, which is added to the tail of the log
1867 static struct io_logs *get_new_log(struct io_log *iolog)
1869 size_t new_size, new_samples;
1870 struct io_logs *cur_log;
1873 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1876 if (!iolog->cur_log_max)
1877 new_samples = DEF_LOG_ENTRIES;
1879 new_samples = iolog->cur_log_max * 2;
1880 if (new_samples > MAX_LOG_ENTRIES)
1881 new_samples = MAX_LOG_ENTRIES;
1884 new_size = new_samples * log_entry_sz(iolog);
1886 cur_log = smalloc(sizeof(*cur_log));
1888 INIT_FLIST_HEAD(&cur_log->list);
1889 cur_log->log = malloc(new_size);
1891 cur_log->nr_samples = 0;
1892 cur_log->max_samples = new_samples;
1893 flist_add_tail(&cur_log->list, &iolog->io_logs);
1894 iolog->cur_log_max = new_samples;
1904 * Add and return a new log chunk, or return current log if big enough
1906 static struct io_logs *regrow_log(struct io_log *iolog)
1908 struct io_logs *cur_log;
1911 if (!iolog || iolog->disabled)
1914 cur_log = iolog_cur_log(iolog);
1916 cur_log = get_new_log(iolog);
1921 if (cur_log->nr_samples < cur_log->max_samples)
1925 * No room for a new sample. If we're compressing on the fly, flush
1926 * out the current chunk
1928 if (iolog->log_gz) {
1929 if (iolog_cur_flush(iolog, cur_log)) {
1930 log_err("fio: failed flushing iolog! Will stop logging.\n");
1936 * Get a new log array, and add to our list
1938 cur_log = get_new_log(iolog);
1940 log_err("fio: failed extending iolog! Will stop logging.\n");
1944 if (!iolog->pending || !iolog->pending->nr_samples)
1948 * Flush pending items to new log
1950 for (i = 0; i < iolog->pending->nr_samples; i++) {
1951 struct io_sample *src, *dst;
1953 src = get_sample(iolog, iolog->pending, i);
1954 dst = get_sample(iolog, cur_log, i);
1955 memcpy(dst, src, log_entry_sz(iolog));
1957 cur_log->nr_samples = iolog->pending->nr_samples;
1959 iolog->pending->nr_samples = 0;
1963 iolog->disabled = true;
1967 void regrow_logs(struct thread_data *td)
1969 regrow_log(td->slat_log);
1970 regrow_log(td->clat_log);
1971 regrow_log(td->clat_hist_log);
1972 regrow_log(td->lat_log);
1973 regrow_log(td->bw_log);
1974 regrow_log(td->iops_log);
1975 td->flags &= ~TD_F_REGROW_LOGS;
1978 static struct io_logs *get_cur_log(struct io_log *iolog)
1980 struct io_logs *cur_log;
1982 cur_log = iolog_cur_log(iolog);
1984 cur_log = get_new_log(iolog);
1989 if (cur_log->nr_samples < cur_log->max_samples)
1993 * Out of space. If we're in IO offload mode, or we're not doing
1994 * per unit logging (hence logging happens outside of the IO thread
1995 * as well), add a new log chunk inline. If we're doing inline
1996 * submissions, flag 'td' as needing a log regrow and we'll take
1997 * care of it on the submission side.
1999 if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2000 !per_unit_log(iolog))
2001 return regrow_log(iolog);
2003 iolog->td->flags |= TD_F_REGROW_LOGS;
2004 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2005 return iolog->pending;
2008 static void __add_log_sample(struct io_log *iolog, unsigned long val,
2009 enum fio_ddir ddir, unsigned int bs,
2010 unsigned long t, uint64_t offset)
2012 struct io_logs *cur_log;
2014 if (iolog->disabled)
2016 if (flist_empty(&iolog->io_logs))
2017 iolog->avg_last = t;
2019 cur_log = get_cur_log(iolog);
2021 struct io_sample *s;
2023 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2026 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2027 io_sample_set_ddir(iolog, s, ddir);
2030 if (iolog->log_offset) {
2031 struct io_sample_offset *so = (void *) s;
2033 so->offset = offset;
2036 cur_log->nr_samples++;
2040 iolog->disabled = true;
2043 static inline void reset_io_stat(struct io_stat *ios)
2045 ios->max_val = ios->min_val = ios->samples = 0;
2046 ios->mean.u.f = ios->S.u.f = 0;
2049 void reset_io_stats(struct thread_data *td)
2051 struct thread_stat *ts = &td->ts;
2054 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2055 reset_io_stat(&ts->clat_stat[i]);
2056 reset_io_stat(&ts->slat_stat[i]);
2057 reset_io_stat(&ts->lat_stat[i]);
2058 reset_io_stat(&ts->bw_stat[i]);
2059 reset_io_stat(&ts->iops_stat[i]);
2061 ts->io_bytes[i] = 0;
2064 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2065 ts->io_u_plat[i][j] = 0;
2068 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2069 ts->io_u_map[i] = 0;
2070 ts->io_u_submit[i] = 0;
2071 ts->io_u_complete[i] = 0;
2072 ts->io_u_lat_u[i] = 0;
2073 ts->io_u_lat_m[i] = 0;
2074 ts->total_submit = 0;
2075 ts->total_complete = 0;
2078 for (i = 0; i < 3; i++) {
2079 ts->total_io_u[i] = 0;
2080 ts->short_io_u[i] = 0;
2081 ts->drop_io_u[i] = 0;
2085 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2086 unsigned long elapsed, bool log_max)
2089 * Note an entry in the log. Use the mean from the logged samples,
2090 * making sure to properly round up. Only write a log entry if we
2091 * had actual samples done.
2093 if (iolog->avg_window[ddir].samples) {
2097 val = iolog->avg_window[ddir].max_val;
2099 val = iolog->avg_window[ddir].mean.u.f + 0.50;
2101 __add_log_sample(iolog, val, ddir, 0, elapsed, 0);
2104 reset_io_stat(&iolog->avg_window[ddir]);
2107 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2112 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2113 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2116 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2117 unsigned long val, enum fio_ddir ddir,
2118 unsigned int bs, uint64_t offset)
2120 unsigned long elapsed, this_window;
2125 elapsed = mtime_since_now(&td->epoch);
2128 * If no time averaging, just add the log sample.
2130 if (!iolog->avg_msec) {
2131 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
2136 * Add the sample. If the time period has passed, then
2137 * add that entry to the log and clear.
2139 add_stat_sample(&iolog->avg_window[ddir], val);
2142 * If period hasn't passed, adding the above sample is all we
2145 this_window = elapsed - iolog->avg_last;
2146 if (elapsed < iolog->avg_last)
2147 return iolog->avg_last - elapsed;
2148 else if (this_window < iolog->avg_msec) {
2149 int diff = iolog->avg_msec - this_window;
2151 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2155 _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2157 iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2158 return iolog->avg_msec;
2161 void finalize_logs(struct thread_data *td, bool unit_logs)
2163 unsigned long elapsed;
2165 elapsed = mtime_since_now(&td->epoch);
2167 if (td->clat_log && unit_logs)
2168 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2169 if (td->slat_log && unit_logs)
2170 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2171 if (td->lat_log && unit_logs)
2172 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2173 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2174 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2175 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2176 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2179 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
2181 struct io_log *iolog;
2186 iolog = agg_io_log[ddir];
2187 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
2190 static void add_clat_percentile_sample(struct thread_stat *ts,
2191 unsigned long usec, enum fio_ddir ddir)
2193 unsigned int idx = plat_val_to_idx(usec);
2194 assert(idx < FIO_IO_U_PLAT_NR);
2196 ts->io_u_plat[ddir][idx]++;
2199 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2200 unsigned long usec, unsigned int bs, uint64_t offset)
2202 unsigned long elapsed, this_window;
2203 struct thread_stat *ts = &td->ts;
2204 struct io_log *iolog = td->clat_hist_log;
2208 add_stat_sample(&ts->clat_stat[ddir], usec);
2211 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
2213 if (ts->clat_percentiles)
2214 add_clat_percentile_sample(ts, usec, ddir);
2216 if (iolog && iolog->hist_msec) {
2217 struct io_hist *hw = &iolog->hist_window[ddir];
2220 elapsed = mtime_since_now(&td->epoch);
2222 hw->hist_last = elapsed;
2223 this_window = elapsed - hw->hist_last;
2225 if (this_window >= iolog->hist_msec) {
2226 unsigned int *io_u_plat;
2227 struct io_u_plat_entry *dst;
2230 * Make a byte-for-byte copy of the latency histogram
2231 * stored in td->ts.io_u_plat[ddir], recording it in a
2232 * log sample. Note that the matching call to free() is
2233 * located in iolog.c after printing this sample to the
2236 io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2237 dst = malloc(sizeof(struct io_u_plat_entry));
2238 memcpy(&(dst->io_u_plat), io_u_plat,
2239 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2240 flist_add(&dst->list, &hw->list);
2241 __add_log_sample(iolog, (unsigned long)dst, ddir, bs,
2245 * Update the last time we recorded as being now, minus
2246 * any drift in time we encountered before actually
2247 * making the record.
2249 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2257 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2258 unsigned long usec, unsigned int bs, uint64_t offset)
2260 struct thread_stat *ts = &td->ts;
2267 add_stat_sample(&ts->slat_stat[ddir], usec);
2270 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
2275 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2276 unsigned long usec, unsigned int bs, uint64_t offset)
2278 struct thread_stat *ts = &td->ts;
2285 add_stat_sample(&ts->lat_stat[ddir], usec);
2288 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2293 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2294 unsigned int bytes, unsigned long spent)
2296 struct thread_stat *ts = &td->ts;
2300 rate = bytes * 1000 / spent;
2306 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2309 add_log_sample(td, td->bw_log, rate, io_u->ddir, bytes, io_u->offset);
2311 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2315 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2317 struct thread_stat *ts = &td->ts;
2318 unsigned long spent, rate;
2320 unsigned int next, next_log;
2322 next_log = td->o.bw_avg_time;
2324 spent = mtime_since(&td->bw_sample_time, t);
2325 if (spent < td->o.bw_avg_time &&
2326 td->o.bw_avg_time - spent >= LOG_MSEC_SLACK)
2327 return td->o.bw_avg_time - spent;
2332 * Compute both read and write rates for the interval.
2334 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2337 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2339 continue; /* No entries for interval */
2342 rate = delta * 1000 / spent / 1024;
2346 add_stat_sample(&ts->bw_stat[ddir], rate);
2349 unsigned int bs = 0;
2351 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2352 bs = td->o.min_bs[ddir];
2354 next = add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2355 next_log = min(next_log, next);
2358 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2361 timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time);
2365 if (spent <= td->o.bw_avg_time)
2366 return min(next_log, td->o.bw_avg_time);
2368 next = td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time);
2369 return min(next, next_log);
2372 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2375 struct thread_stat *ts = &td->ts;
2379 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2382 add_log_sample(td, td->iops_log, 1, io_u->ddir, bytes, io_u->offset);
2384 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2388 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2390 struct thread_stat *ts = &td->ts;
2391 unsigned long spent, iops;
2393 unsigned int next, next_log;
2395 next_log = td->o.iops_avg_time;
2397 spent = mtime_since(&td->iops_sample_time, t);
2398 if (spent < td->o.iops_avg_time &&
2399 td->o.iops_avg_time - spent >= LOG_MSEC_SLACK)
2400 return td->o.iops_avg_time - spent;
2405 * Compute both read and write rates for the interval.
2407 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2410 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2412 continue; /* No entries for interval */
2415 iops = (delta * 1000) / spent;
2419 add_stat_sample(&ts->iops_stat[ddir], iops);
2422 unsigned int bs = 0;
2424 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2425 bs = td->o.min_bs[ddir];
2427 next = add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2428 next_log = min(next_log, next);
2431 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2434 timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time);
2438 if (spent <= td->o.iops_avg_time)
2439 return min(next_log, td->o.iops_avg_time);
2441 next = td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
2442 return min(next, next_log);
2446 * Returns msecs to next event
2448 int calc_log_samples(void)
2450 struct thread_data *td;
2451 unsigned int next = ~0U, tmp;
2455 fio_gettime(&now, NULL);
2457 for_each_td(td, i) {
2458 if (in_ramp_time(td) ||
2459 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2460 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2463 if (td->bw_log && !per_unit_log(td->bw_log)) {
2464 tmp = add_bw_samples(td, &now);
2468 if (td->iops_log && !per_unit_log(td->iops_log)) {
2469 tmp = add_iops_samples(td, &now);
2475 return next == ~0U ? 0 : next;
2478 void stat_init(void)
2480 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2483 void stat_exit(void)
2486 * When we have the mutex, we know out-of-band access to it
2489 fio_mutex_down(stat_mutex);
2490 fio_mutex_remove(stat_mutex);
2494 * Called from signal handler. Wake up status thread.
2496 void show_running_run_stats(void)
2501 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2503 /* Ignore io_u's which span multiple blocks--they will just get
2504 * inaccurate counts. */
2505 int idx = (io_u->offset - io_u->file->file_offset)
2506 / td->o.bs[DDIR_TRIM];
2507 uint32_t *info = &td->ts.block_infos[idx];
2508 assert(idx < td->ts.nr_block_infos);