12 #include "lib/ieee754.h"
14 #include "lib/getrusage.h"
17 #include "lib/output_buffer.h"
18 #include "helper_thread.h"
21 #define LOG_MSEC_SLACK 10
23 struct fio_mutex *stat_mutex;
25 void clear_rusage_stat(struct thread_data *td)
27 struct thread_stat *ts = &td->ts;
29 fio_getrusage(&td->ru_start);
30 ts->usr_time = ts->sys_time = 0;
32 ts->minf = ts->majf = 0;
35 void update_rusage_stat(struct thread_data *td)
37 struct thread_stat *ts = &td->ts;
39 fio_getrusage(&td->ru_end);
40 ts->usr_time += mtime_since(&td->ru_start.ru_utime,
41 &td->ru_end.ru_utime);
42 ts->sys_time += mtime_since(&td->ru_start.ru_stime,
43 &td->ru_end.ru_stime);
44 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
45 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
46 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
47 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
49 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
53 * Given a latency, return the index of the corresponding bucket in
54 * the structure tracking percentiles.
56 * (1) find the group (and error bits) that the value (latency)
57 * belongs to by looking at its MSB. (2) find the bucket number in the
58 * group by looking at the index bits.
61 static unsigned int plat_val_to_idx(unsigned int val)
63 unsigned int msb, error_bits, base, offset, idx;
65 /* Find MSB starting from bit 0 */
69 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
72 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
73 * all bits of the sample as index
75 if (msb <= FIO_IO_U_PLAT_BITS)
78 /* Compute the number of error bits to discard*/
79 error_bits = msb - FIO_IO_U_PLAT_BITS;
81 /* Compute the number of buckets before the group */
82 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
85 * Discard the error bits and apply the mask to find the
86 * index for the buckets in the group
88 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
90 /* Make sure the index does not exceed (array size - 1) */
91 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
92 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
98 * Convert the given index of the bucket array to the value
99 * represented by the bucket
101 static unsigned int plat_idx_to_val(unsigned int idx)
103 unsigned int error_bits, k, base;
105 assert(idx < FIO_IO_U_PLAT_NR);
107 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
108 * all bits of the sample as index */
109 if (idx < (FIO_IO_U_PLAT_VAL << 1))
112 /* Find the group and compute the minimum value of that group */
113 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
114 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
116 /* Find its bucket number of the group */
117 k = idx % FIO_IO_U_PLAT_VAL;
119 /* Return the mean of the range of the bucket */
120 return base + ((k + 0.5) * (1 << error_bits));
123 static int double_cmp(const void *a, const void *b)
125 const fio_fp64_t fa = *(const fio_fp64_t *) a;
126 const fio_fp64_t fb = *(const fio_fp64_t *) b;
131 else if (fa.u.f < fb.u.f)
137 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
138 fio_fp64_t *plist, unsigned int **output,
139 unsigned int *maxv, unsigned int *minv)
141 unsigned long sum = 0;
142 unsigned int len, i, j = 0;
143 unsigned int oval_len = 0;
144 unsigned int *ovals = NULL;
151 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
158 * Sort the percentile list. Note that it may already be sorted if
159 * we are using the default values, but since it's a short list this
160 * isn't a worry. Also note that this does not work for NaN values.
163 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
166 * Calculate bucket values, note down max and min values
169 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
171 while (sum >= (plist[j].u.f / 100.0 * nr)) {
172 assert(plist[j].u.f <= 100.0);
176 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
179 ovals[j] = plat_idx_to_val(i);
180 if (ovals[j] < *minv)
182 if (ovals[j] > *maxv)
185 is_last = (j == len - 1);
198 * Find and display the p-th percentile of clat
200 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
201 fio_fp64_t *plist, unsigned int precision,
202 struct buf_output *out)
204 unsigned int len, j = 0, minv, maxv;
206 int is_last, per_line, scale_down;
209 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
214 * We default to usecs, but if the value range is such that we
215 * should scale down to msecs, do that.
217 if (minv > 2000 && maxv > 99999) {
219 log_buf(out, " clat percentiles (msec):\n |");
222 log_buf(out, " clat percentiles (usec):\n |");
225 snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
226 per_line = (80 - 7) / (precision + 14);
228 for (j = 0; j < len; j++) {
229 char fbuf[16], *ptr = fbuf;
232 if (j != 0 && (j % per_line) == 0)
235 /* end of the list */
236 is_last = (j == len - 1);
238 if (plist[j].u.f < 10.0)
239 ptr += sprintf(fbuf, " ");
241 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
244 ovals[j] = (ovals[j] + 999) / 1000;
246 log_buf(out, " %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
251 if ((j % per_line) == per_line - 1) /* for formatting */
260 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
261 double *mean, double *dev)
263 double n = (double) is->samples;
270 *mean = is->mean.u.f;
273 *dev = sqrt(is->S.u.f / (n - 1.0));
280 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
282 char *p1, *p2, *p3, *p4;
283 const char *str[] = { " READ", " WRITE" , " TRIM"};
286 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
288 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
289 const int i2p = is_power_of_2(rs->kb_base);
294 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
295 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
296 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
297 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
299 log_buf(out, "%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
300 " mint=%llumsec, maxt=%llumsec\n",
301 rs->unified_rw_rep ? " MIXED" : str[i],
303 (unsigned long long) rs->min_run[i],
304 (unsigned long long) rs->max_run[i]);
313 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
318 * Do depth distribution calculations
320 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
322 io_u_dist[i] = (double) map[i] / (double) total;
323 io_u_dist[i] *= 100.0;
324 if (io_u_dist[i] < 0.1 && map[i])
331 static void stat_calc_lat(struct thread_stat *ts, double *dst,
332 unsigned int *src, int nr)
334 unsigned long total = ddir_rw_sum(ts->total_io_u);
338 * Do latency distribution calculations
340 for (i = 0; i < nr; i++) {
342 dst[i] = (double) src[i] / (double) total;
344 if (dst[i] < 0.01 && src[i])
351 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
353 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
356 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
358 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
361 static void display_lat(const char *name, unsigned long min, unsigned long max,
362 double mean, double dev, struct buf_output *out)
364 const char *base = "(usec)";
367 if (!usec_to_msec(&min, &max, &mean, &dev))
370 minp = num2str(min, 6, 1, 0, 0);
371 maxp = num2str(max, 6, 1, 0, 0);
373 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
374 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
380 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
381 int ddir, struct buf_output *out)
383 const char *str[] = { "read ", "write", "trim" };
384 unsigned long min, max, runt;
385 unsigned long long bw, iops;
387 char *io_p, *bw_p, *iops_p;
390 assert(ddir_rw(ddir));
392 if (!ts->runtime[ddir])
395 i2p = is_power_of_2(rs->kb_base);
396 runt = ts->runtime[ddir];
398 bw = (1000 * ts->io_bytes[ddir]) / runt;
399 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
400 bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
402 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
403 iops_p = num2str(iops, 6, 1, 0, 0);
405 log_buf(out, " %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
406 rs->unified_rw_rep ? "mixed" : str[ddir],
408 (unsigned long long) ts->runtime[ddir]);
414 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
415 display_lat("slat", min, max, mean, dev, out);
416 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
417 display_lat("clat", min, max, mean, dev, out);
418 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
419 display_lat(" lat", min, max, mean, dev, out);
421 if (ts->clat_percentiles) {
422 show_clat_percentiles(ts->io_u_plat[ddir],
423 ts->clat_stat[ddir].samples,
425 ts->percentile_precision, out);
427 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
428 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
429 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
431 if (rs->unit_base == 1) {
439 p_of_agg = mean * 100 / (double) rs->agg[ddir];
440 if (p_of_agg > 100.0)
444 if (mean > fkb_base * fkb_base) {
449 bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
452 log_buf(out, " bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
453 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
454 p_of_agg, mean, dev);
458 static int show_lat(double *io_u_lat, int nr, const char **ranges,
459 const char *msg, struct buf_output *out)
461 int new_line = 1, i, line = 0, shown = 0;
463 for (i = 0; i < nr; i++) {
464 if (io_u_lat[i] <= 0.0)
470 log_buf(out, " lat (%s) : ", msg);
476 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
488 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
490 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
491 "250=", "500=", "750=", "1000=", };
493 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
496 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
498 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
499 "250=", "500=", "750=", "1000=", "2000=",
502 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
505 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
507 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
508 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
510 stat_calc_lat_u(ts, io_u_lat_u);
511 stat_calc_lat_m(ts, io_u_lat_m);
513 show_lat_u(io_u_lat_u, out);
514 show_lat_m(io_u_lat_m, out);
517 static int block_state_category(int block_state)
519 switch (block_state) {
520 case BLOCK_STATE_UNINIT:
522 case BLOCK_STATE_TRIMMED:
523 case BLOCK_STATE_WRITTEN:
525 case BLOCK_STATE_WRITE_FAILURE:
526 case BLOCK_STATE_TRIM_FAILURE:
529 /* Silence compile warning on some BSDs and have a return */
535 static int compare_block_infos(const void *bs1, const void *bs2)
537 uint32_t block1 = *(uint32_t *)bs1;
538 uint32_t block2 = *(uint32_t *)bs2;
539 int state1 = BLOCK_INFO_STATE(block1);
540 int state2 = BLOCK_INFO_STATE(block2);
541 int bscat1 = block_state_category(state1);
542 int bscat2 = block_state_category(state2);
543 int cycles1 = BLOCK_INFO_TRIMS(block1);
544 int cycles2 = BLOCK_INFO_TRIMS(block2);
551 if (cycles1 < cycles2)
553 if (cycles1 > cycles2)
561 assert(block1 == block2);
565 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
566 fio_fp64_t *plist, unsigned int **percentiles,
572 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
574 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
581 * Sort the percentile list. Note that it may already be sorted if
582 * we are using the default values, but since it's a short list this
583 * isn't a worry. Also note that this does not work for NaN values.
586 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
589 /* Start only after the uninit entries end */
591 nr_uninit < nr_block_infos
592 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
596 if (nr_uninit == nr_block_infos)
599 *percentiles = calloc(len, sizeof(**percentiles));
601 for (i = 0; i < len; i++) {
602 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
604 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
607 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
608 for (i = 0; i < nr_block_infos; i++)
609 types[BLOCK_INFO_STATE(block_infos[i])]++;
614 static const char *block_state_names[] = {
615 [BLOCK_STATE_UNINIT] = "unwritten",
616 [BLOCK_STATE_TRIMMED] = "trimmed",
617 [BLOCK_STATE_WRITTEN] = "written",
618 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
619 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
622 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
623 fio_fp64_t *plist, struct buf_output *out)
626 unsigned int *percentiles = NULL;
627 unsigned int block_state_counts[BLOCK_STATE_COUNT];
629 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
630 &percentiles, block_state_counts);
632 log_buf(out, " block lifetime percentiles :\n |");
634 for (i = 0; i < len; i++) {
635 uint32_t block_info = percentiles[i];
636 #define LINE_LENGTH 75
637 char str[LINE_LENGTH];
638 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
639 plist[i].u.f, block_info,
640 i == len - 1 ? '\n' : ',');
641 assert(strln < LINE_LENGTH);
642 if (pos + strln > LINE_LENGTH) {
644 log_buf(out, "\n |");
646 log_buf(out, "%s", str);
653 log_buf(out, " states :");
654 for (i = 0; i < BLOCK_STATE_COUNT; i++)
655 log_buf(out, " %s=%u%c",
656 block_state_names[i], block_state_counts[i],
657 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
660 static void show_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))
674 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
677 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
678 ts->name, ts->groupid, ts->members,
679 ts->error, (int) ts->pid, time_buf);
681 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
682 ts->name, ts->groupid, ts->members,
683 ts->error, ts->verror, (int) ts->pid,
687 if (strlen(ts->description))
688 log_buf(out, " Description : [%s]\n", ts->description);
690 if (ts->io_bytes[DDIR_READ])
691 show_ddir_status(rs, ts, DDIR_READ, out);
692 if (ts->io_bytes[DDIR_WRITE])
693 show_ddir_status(rs, ts, DDIR_WRITE, out);
694 if (ts->io_bytes[DDIR_TRIM])
695 show_ddir_status(rs, ts, DDIR_TRIM, out);
697 show_latencies(ts, out);
699 runtime = ts->total_run_time;
701 double runt = (double) runtime;
703 usr_cpu = (double) ts->usr_time * 100 / runt;
704 sys_cpu = (double) ts->sys_time * 100 / runt;
710 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
711 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
712 (unsigned long long) ts->ctx,
713 (unsigned long long) ts->majf,
714 (unsigned long long) ts->minf);
716 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
717 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
718 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
719 io_u_dist[1], io_u_dist[2],
720 io_u_dist[3], io_u_dist[4],
721 io_u_dist[5], io_u_dist[6]);
723 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
724 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
725 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
726 io_u_dist[1], io_u_dist[2],
727 io_u_dist[3], io_u_dist[4],
728 io_u_dist[5], io_u_dist[6]);
729 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
730 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
731 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
732 io_u_dist[1], io_u_dist[2],
733 io_u_dist[3], io_u_dist[4],
734 io_u_dist[5], io_u_dist[6]);
735 log_buf(out, " issued : total=r=%llu/w=%llu/d=%llu,"
736 " short=r=%llu/w=%llu/d=%llu,"
737 " drop=r=%llu/w=%llu/d=%llu\n",
738 (unsigned long long) ts->total_io_u[0],
739 (unsigned long long) ts->total_io_u[1],
740 (unsigned long long) ts->total_io_u[2],
741 (unsigned long long) ts->short_io_u[0],
742 (unsigned long long) ts->short_io_u[1],
743 (unsigned long long) ts->short_io_u[2],
744 (unsigned long long) ts->drop_io_u[0],
745 (unsigned long long) ts->drop_io_u[1],
746 (unsigned long long) ts->drop_io_u[2]);
747 if (ts->continue_on_error) {
748 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
749 (unsigned long long)ts->total_err_count,
751 strerror(ts->first_error));
753 if (ts->latency_depth) {
754 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
755 (unsigned long long)ts->latency_target,
756 (unsigned long long)ts->latency_window,
757 ts->latency_percentile.u.f,
761 if (ts->nr_block_infos)
762 show_block_infos(ts->nr_block_infos, ts->block_infos,
763 ts->percentile_list, out);
766 static void show_ddir_status_terse(struct thread_stat *ts,
767 struct group_run_stats *rs, int ddir,
768 struct buf_output *out)
770 unsigned long min, max;
771 unsigned long long bw, iops;
772 unsigned int *ovals = NULL;
774 unsigned int len, minv, maxv;
777 assert(ddir_rw(ddir));
780 if (ts->runtime[ddir]) {
781 uint64_t runt = ts->runtime[ddir];
783 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
784 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
787 log_buf(out, ";%llu;%llu;%llu;%llu",
788 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
789 (unsigned long long) ts->runtime[ddir]);
791 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
792 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
794 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
796 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
797 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
799 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
801 if (ts->clat_percentiles) {
802 len = calc_clat_percentiles(ts->io_u_plat[ddir],
803 ts->clat_stat[ddir].samples,
804 ts->percentile_list, &ovals, &maxv,
809 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
811 log_buf(out, ";0%%=0");
814 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
817 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
818 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
820 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
825 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
826 double p_of_agg = 100.0;
829 p_of_agg = mean * 100 / (double) rs->agg[ddir];
830 if (p_of_agg > 100.0)
834 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
836 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
839 static void add_ddir_status_json(struct thread_stat *ts,
840 struct group_run_stats *rs, int ddir, struct json_object *parent)
842 unsigned long min, max;
843 unsigned long long bw;
844 unsigned int *ovals = NULL;
845 double mean, dev, iops;
846 unsigned int len, minv, maxv;
848 const char *ddirname[] = {"read", "write", "trim"};
849 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
851 double p_of_agg = 100.0;
853 assert(ddir_rw(ddir));
855 if (ts->unified_rw_rep && ddir != DDIR_READ)
858 dir_object = json_create_object();
859 json_object_add_value_object(parent,
860 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
864 if (ts->runtime[ddir]) {
865 uint64_t runt = ts->runtime[ddir];
867 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
868 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
871 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
872 json_object_add_value_int(dir_object, "bw", bw);
873 json_object_add_value_float(dir_object, "iops", iops);
874 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
875 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
876 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
877 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
879 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
883 tmp_object = json_create_object();
884 json_object_add_value_object(dir_object, "slat", tmp_object);
885 json_object_add_value_int(tmp_object, "min", min);
886 json_object_add_value_int(tmp_object, "max", max);
887 json_object_add_value_float(tmp_object, "mean", mean);
888 json_object_add_value_float(tmp_object, "stddev", dev);
890 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
894 tmp_object = json_create_object();
895 json_object_add_value_object(dir_object, "clat", tmp_object);
896 json_object_add_value_int(tmp_object, "min", min);
897 json_object_add_value_int(tmp_object, "max", max);
898 json_object_add_value_float(tmp_object, "mean", mean);
899 json_object_add_value_float(tmp_object, "stddev", dev);
901 if (ts->clat_percentiles) {
902 len = calc_clat_percentiles(ts->io_u_plat[ddir],
903 ts->clat_stat[ddir].samples,
904 ts->percentile_list, &ovals, &maxv,
909 percentile_object = json_create_object();
910 json_object_add_value_object(tmp_object, "percentile", percentile_object);
911 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
913 json_object_add_value_int(percentile_object, "0.00", 0);
916 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
917 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
920 if (output_format & FIO_OUTPUT_JSON_PLUS) {
921 clat_bins_object = json_create_object();
922 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
923 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
924 snprintf(buf, sizeof(buf), "%d", i);
925 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
927 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS);
928 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL);
929 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR);
932 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
936 tmp_object = json_create_object();
937 json_object_add_value_object(dir_object, "lat", tmp_object);
938 json_object_add_value_int(tmp_object, "min", min);
939 json_object_add_value_int(tmp_object, "max", max);
940 json_object_add_value_float(tmp_object, "mean", mean);
941 json_object_add_value_float(tmp_object, "stddev", dev);
945 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
947 p_of_agg = mean * 100 / (double) rs->agg[ddir];
948 if (p_of_agg > 100.0)
953 p_of_agg = mean = dev = 0.0;
955 json_object_add_value_int(dir_object, "bw_min", min);
956 json_object_add_value_int(dir_object, "bw_max", max);
957 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
958 json_object_add_value_float(dir_object, "bw_mean", mean);
959 json_object_add_value_float(dir_object, "bw_dev", dev);
962 static void show_thread_status_terse_v2(struct thread_stat *ts,
963 struct group_run_stats *rs,
964 struct buf_output *out)
966 double io_u_dist[FIO_IO_U_MAP_NR];
967 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
968 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
969 double usr_cpu, sys_cpu;
973 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
974 /* Log Read Status */
975 show_ddir_status_terse(ts, rs, DDIR_READ, out);
976 /* Log Write Status */
977 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
978 /* Log Trim Status */
979 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
982 if (ts->total_run_time) {
983 double runt = (double) ts->total_run_time;
985 usr_cpu = (double) ts->usr_time * 100 / runt;
986 sys_cpu = (double) ts->sys_time * 100 / runt;
992 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
993 (unsigned long long) ts->ctx,
994 (unsigned long long) ts->majf,
995 (unsigned long long) ts->minf);
997 /* Calc % distribution of IO depths, usecond, msecond latency */
998 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
999 stat_calc_lat_u(ts, io_u_lat_u);
1000 stat_calc_lat_m(ts, io_u_lat_m);
1002 /* Only show fixed 7 I/O depth levels*/
1003 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1004 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1005 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1007 /* Microsecond latency */
1008 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1009 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1010 /* Millisecond latency */
1011 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1012 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1013 /* Additional output if continue_on_error set - default off*/
1014 if (ts->continue_on_error)
1015 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1018 /* Additional output if description is set */
1019 if (strlen(ts->description))
1020 log_buf(out, ";%s", ts->description);
1025 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1026 struct group_run_stats *rs, int ver,
1027 struct buf_output *out)
1029 double io_u_dist[FIO_IO_U_MAP_NR];
1030 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1031 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1032 double usr_cpu, sys_cpu;
1036 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1037 ts->name, ts->groupid, ts->error);
1038 /* Log Read Status */
1039 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1040 /* Log Write Status */
1041 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1042 /* Log Trim Status */
1044 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1047 if (ts->total_run_time) {
1048 double runt = (double) ts->total_run_time;
1050 usr_cpu = (double) ts->usr_time * 100 / runt;
1051 sys_cpu = (double) ts->sys_time * 100 / runt;
1057 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1058 (unsigned long long) ts->ctx,
1059 (unsigned long long) ts->majf,
1060 (unsigned long long) ts->minf);
1062 /* Calc % distribution of IO depths, usecond, msecond latency */
1063 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1064 stat_calc_lat_u(ts, io_u_lat_u);
1065 stat_calc_lat_m(ts, io_u_lat_m);
1067 /* Only show fixed 7 I/O depth levels*/
1068 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1069 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1070 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1072 /* Microsecond latency */
1073 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1074 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1075 /* Millisecond latency */
1076 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1077 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1079 /* disk util stats, if any */
1080 show_disk_util(1, NULL, out);
1082 /* Additional output if continue_on_error set - default off*/
1083 if (ts->continue_on_error)
1084 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1086 /* Additional output if description is set */
1087 if (strlen(ts->description))
1088 log_buf(out, ";%s", ts->description);
1093 void json_add_job_opts(struct json_object *root, const char *name,
1094 struct flist_head *opt_list, bool num_jobs)
1096 struct json_object *dir_object;
1097 struct flist_head *entry;
1098 struct print_option *p;
1100 if (flist_empty(opt_list))
1103 dir_object = json_create_object();
1104 json_object_add_value_object(root, name, dir_object);
1106 flist_for_each(entry, opt_list) {
1107 const char *pos = "";
1109 p = flist_entry(entry, struct print_option, list);
1110 if (!num_jobs && !strcmp(p->name, "numjobs"))
1114 json_object_add_value_string(dir_object, p->name, pos);
1118 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1119 struct group_run_stats *rs,
1120 struct flist_head *opt_list)
1122 struct json_object *root, *tmp;
1123 struct jobs_eta *je;
1124 double io_u_dist[FIO_IO_U_MAP_NR];
1125 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1126 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1127 double usr_cpu, sys_cpu;
1131 root = json_create_object();
1132 json_object_add_value_string(root, "jobname", ts->name);
1133 json_object_add_value_int(root, "groupid", ts->groupid);
1134 json_object_add_value_int(root, "error", ts->error);
1137 je = get_jobs_eta(true, &size);
1139 json_object_add_value_int(root, "eta", je->eta_sec);
1140 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1144 json_add_job_opts(root, "job options", opt_list, true);
1146 add_ddir_status_json(ts, rs, DDIR_READ, root);
1147 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1148 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1151 if (ts->total_run_time) {
1152 double runt = (double) ts->total_run_time;
1154 usr_cpu = (double) ts->usr_time * 100 / runt;
1155 sys_cpu = (double) ts->sys_time * 100 / runt;
1160 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1161 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1162 json_object_add_value_int(root, "ctx", ts->ctx);
1163 json_object_add_value_int(root, "majf", ts->majf);
1164 json_object_add_value_int(root, "minf", ts->minf);
1167 /* Calc % distribution of IO depths, usecond, msecond latency */
1168 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1169 stat_calc_lat_u(ts, io_u_lat_u);
1170 stat_calc_lat_m(ts, io_u_lat_m);
1172 tmp = json_create_object();
1173 json_object_add_value_object(root, "iodepth_level", tmp);
1174 /* Only show fixed 7 I/O depth levels*/
1175 for (i = 0; i < 7; i++) {
1178 snprintf(name, 20, "%d", 1 << i);
1180 snprintf(name, 20, ">=%d", 1 << i);
1181 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1184 tmp = json_create_object();
1185 json_object_add_value_object(root, "latency_us", tmp);
1186 /* Microsecond latency */
1187 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1188 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1189 "250", "500", "750", "1000", };
1190 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1192 /* Millisecond latency */
1193 tmp = json_create_object();
1194 json_object_add_value_object(root, "latency_ms", tmp);
1195 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1196 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1197 "250", "500", "750", "1000", "2000",
1199 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1202 /* Additional output if continue_on_error set - default off*/
1203 if (ts->continue_on_error) {
1204 json_object_add_value_int(root, "total_err", ts->total_err_count);
1205 json_object_add_value_int(root, "first_error", ts->first_error);
1208 if (ts->latency_depth) {
1209 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1210 json_object_add_value_int(root, "latency_target", ts->latency_target);
1211 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1212 json_object_add_value_int(root, "latency_window", ts->latency_window);
1215 /* Additional output if description is set */
1216 if (strlen(ts->description))
1217 json_object_add_value_string(root, "desc", ts->description);
1219 if (ts->nr_block_infos) {
1220 /* Block error histogram and types */
1222 unsigned int *percentiles = NULL;
1223 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1225 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1226 ts->percentile_list,
1227 &percentiles, block_state_counts);
1230 struct json_object *block, *percentile_object, *states;
1232 block = json_create_object();
1233 json_object_add_value_object(root, "block", block);
1235 percentile_object = json_create_object();
1236 json_object_add_value_object(block, "percentiles",
1238 for (i = 0; i < len; i++) {
1240 snprintf(buf, sizeof(buf), "%f",
1241 ts->percentile_list[i].u.f);
1242 json_object_add_value_int(percentile_object,
1247 states = json_create_object();
1248 json_object_add_value_object(block, "states", states);
1249 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1250 json_object_add_value_int(states,
1251 block_state_names[state],
1252 block_state_counts[state]);
1261 static void show_thread_status_terse(struct thread_stat *ts,
1262 struct group_run_stats *rs,
1263 struct buf_output *out)
1265 if (terse_version == 2)
1266 show_thread_status_terse_v2(ts, rs, out);
1267 else if (terse_version == 3 || terse_version == 4)
1268 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1270 log_err("fio: bad terse version!? %d\n", terse_version);
1273 struct json_object *show_thread_status(struct thread_stat *ts,
1274 struct group_run_stats *rs,
1275 struct flist_head *opt_list,
1276 struct buf_output *out)
1278 struct json_object *ret = NULL;
1280 if (output_format & FIO_OUTPUT_TERSE)
1281 show_thread_status_terse(ts, rs, out);
1282 if (output_format & FIO_OUTPUT_JSON)
1283 ret = show_thread_status_json(ts, rs, opt_list);
1284 if (output_format & FIO_OUTPUT_NORMAL)
1285 show_thread_status_normal(ts, rs, out);
1290 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1294 if (src->samples == 0)
1297 dst->min_val = min(dst->min_val, src->min_val);
1298 dst->max_val = max(dst->max_val, src->max_val);
1301 * Compute new mean and S after the merge
1302 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1303 * #Parallel_algorithm>
1306 mean = src->mean.u.f;
1309 double delta = src->mean.u.f - dst->mean.u.f;
1311 mean = ((src->mean.u.f * src->samples) +
1312 (dst->mean.u.f * dst->samples)) /
1313 (dst->samples + src->samples);
1315 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1316 (dst->samples * src->samples) /
1317 (dst->samples + src->samples);
1320 dst->samples += src->samples;
1321 dst->mean.u.f = mean;
1325 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1329 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1330 if (dst->max_run[i] < src->max_run[i])
1331 dst->max_run[i] = src->max_run[i];
1332 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1333 dst->min_run[i] = src->min_run[i];
1334 if (dst->max_bw[i] < src->max_bw[i])
1335 dst->max_bw[i] = src->max_bw[i];
1336 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1337 dst->min_bw[i] = src->min_bw[i];
1339 dst->io_kb[i] += src->io_kb[i];
1340 dst->agg[i] += src->agg[i];
1344 dst->kb_base = src->kb_base;
1345 if (!dst->unit_base)
1346 dst->unit_base = src->unit_base;
1349 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1354 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1355 if (!dst->unified_rw_rep) {
1356 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1357 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1358 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1359 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1361 dst->io_bytes[l] += src->io_bytes[l];
1363 if (dst->runtime[l] < src->runtime[l])
1364 dst->runtime[l] = src->runtime[l];
1366 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1367 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1368 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1369 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1371 dst->io_bytes[0] += src->io_bytes[l];
1373 if (dst->runtime[0] < src->runtime[l])
1374 dst->runtime[0] = src->runtime[l];
1377 * We're summing to the same destination, so override
1378 * 'first' after the first iteration of the loop
1384 dst->usr_time += src->usr_time;
1385 dst->sys_time += src->sys_time;
1386 dst->ctx += src->ctx;
1387 dst->majf += src->majf;
1388 dst->minf += src->minf;
1390 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1391 dst->io_u_map[k] += src->io_u_map[k];
1392 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1393 dst->io_u_submit[k] += src->io_u_submit[k];
1394 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1395 dst->io_u_complete[k] += src->io_u_complete[k];
1396 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1397 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1398 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1399 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1401 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1402 if (!dst->unified_rw_rep) {
1403 dst->total_io_u[k] += src->total_io_u[k];
1404 dst->short_io_u[k] += src->short_io_u[k];
1405 dst->drop_io_u[k] += src->drop_io_u[k];
1407 dst->total_io_u[0] += src->total_io_u[k];
1408 dst->short_io_u[0] += src->short_io_u[k];
1409 dst->drop_io_u[0] += src->drop_io_u[k];
1413 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1416 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1417 if (!dst->unified_rw_rep)
1418 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1420 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1424 dst->total_run_time += src->total_run_time;
1425 dst->total_submit += src->total_submit;
1426 dst->total_complete += src->total_complete;
1429 void init_group_run_stat(struct group_run_stats *gs)
1432 memset(gs, 0, sizeof(*gs));
1434 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1435 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1438 void init_thread_stat(struct thread_stat *ts)
1442 memset(ts, 0, sizeof(*ts));
1444 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1445 ts->lat_stat[j].min_val = -1UL;
1446 ts->clat_stat[j].min_val = -1UL;
1447 ts->slat_stat[j].min_val = -1UL;
1448 ts->bw_stat[j].min_val = -1UL;
1453 void __show_run_stats(void)
1455 struct group_run_stats *runstats, *rs;
1456 struct thread_data *td;
1457 struct thread_stat *threadstats, *ts;
1458 int i, j, k, nr_ts, last_ts, idx;
1459 int kb_base_warned = 0;
1460 int unit_base_warned = 0;
1461 struct json_object *root = NULL;
1462 struct json_array *array = NULL;
1463 struct buf_output output[FIO_OUTPUT_NR];
1464 struct flist_head **opt_lists;
1466 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1468 for (i = 0; i < groupid + 1; i++)
1469 init_group_run_stat(&runstats[i]);
1472 * find out how many threads stats we need. if group reporting isn't
1473 * enabled, it's one-per-td.
1477 for_each_td(td, i) {
1478 if (!td->o.group_reporting) {
1482 if (last_ts == td->groupid)
1485 last_ts = td->groupid;
1489 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1490 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1492 for (i = 0; i < nr_ts; i++) {
1493 init_thread_stat(&threadstats[i]);
1494 opt_lists[i] = NULL;
1500 for_each_td(td, i) {
1501 if (idx && (!td->o.group_reporting ||
1502 (td->o.group_reporting && last_ts != td->groupid))) {
1507 last_ts = td->groupid;
1509 ts = &threadstats[j];
1511 ts->clat_percentiles = td->o.clat_percentiles;
1512 ts->percentile_precision = td->o.percentile_precision;
1513 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1514 opt_lists[j] = &td->opt_list;
1519 if (ts->groupid == -1) {
1521 * These are per-group shared already
1523 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1524 if (td->o.description)
1525 strncpy(ts->description, td->o.description,
1526 FIO_JOBDESC_SIZE - 1);
1528 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1531 * If multiple entries in this group, this is
1534 ts->thread_number = td->thread_number;
1535 ts->groupid = td->groupid;
1538 * first pid in group, not very useful...
1542 ts->kb_base = td->o.kb_base;
1543 ts->unit_base = td->o.unit_base;
1544 ts->unified_rw_rep = td->o.unified_rw_rep;
1545 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1546 log_info("fio: kb_base differs for jobs in group, using"
1547 " %u as the base\n", ts->kb_base);
1549 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1550 log_info("fio: unit_base differs for jobs in group, using"
1551 " %u as the base\n", ts->unit_base);
1552 unit_base_warned = 1;
1555 ts->continue_on_error = td->o.continue_on_error;
1556 ts->total_err_count += td->total_err_count;
1557 ts->first_error = td->first_error;
1559 if (!td->error && td->o.continue_on_error &&
1561 ts->error = td->first_error;
1562 ts->verror[sizeof(ts->verror) - 1] = '\0';
1563 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1564 } else if (td->error) {
1565 ts->error = td->error;
1566 ts->verror[sizeof(ts->verror) - 1] = '\0';
1567 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1571 ts->latency_depth = td->latency_qd;
1572 ts->latency_target = td->o.latency_target;
1573 ts->latency_percentile = td->o.latency_percentile;
1574 ts->latency_window = td->o.latency_window;
1576 ts->nr_block_infos = td->ts.nr_block_infos;
1577 for (k = 0; k < ts->nr_block_infos; k++)
1578 ts->block_infos[k] = td->ts.block_infos[k];
1580 sum_thread_stats(ts, &td->ts, idx == 1);
1583 for (i = 0; i < nr_ts; i++) {
1584 unsigned long long bw;
1586 ts = &threadstats[i];
1587 if (ts->groupid == -1)
1589 rs = &runstats[ts->groupid];
1590 rs->kb_base = ts->kb_base;
1591 rs->unit_base = ts->unit_base;
1592 rs->unified_rw_rep += ts->unified_rw_rep;
1594 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1595 if (!ts->runtime[j])
1597 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1598 rs->min_run[j] = ts->runtime[j];
1599 if (ts->runtime[j] > rs->max_run[j])
1600 rs->max_run[j] = ts->runtime[j];
1603 if (ts->runtime[j]) {
1604 unsigned long runt = ts->runtime[j];
1605 unsigned long long kb;
1607 kb = ts->io_bytes[j] / rs->kb_base;
1608 bw = kb * 1000 / runt;
1610 if (bw < rs->min_bw[j])
1612 if (bw > rs->max_bw[j])
1615 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1619 for (i = 0; i < groupid + 1; i++) {
1624 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1625 if (rs->max_run[ddir])
1626 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1631 for (i = 0; i < FIO_OUTPUT_NR; i++)
1632 buf_output_init(&output[i]);
1635 * don't overwrite last signal output
1637 if (output_format & FIO_OUTPUT_NORMAL)
1638 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1639 if (output_format & FIO_OUTPUT_JSON) {
1640 struct thread_data *global;
1643 unsigned long long ms_since_epoch;
1645 gettimeofday(&now, NULL);
1646 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1647 (unsigned long long)(now.tv_usec) / 1000;
1649 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1651 time_buf[strlen(time_buf) - 1] = '\0';
1653 root = json_create_object();
1654 json_object_add_value_string(root, "fio version", fio_version_string);
1655 json_object_add_value_int(root, "timestamp", now.tv_sec);
1656 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1657 json_object_add_value_string(root, "time", time_buf);
1658 global = get_global_options();
1659 json_add_job_opts(root, "global options", &global->opt_list, false);
1660 array = json_create_array();
1661 json_object_add_value_array(root, "jobs", array);
1665 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1667 for (i = 0; i < nr_ts; i++) {
1668 ts = &threadstats[i];
1669 rs = &runstats[ts->groupid];
1672 fio_server_send_job_options(opt_lists[i], i);
1673 fio_server_send_ts(ts, rs);
1675 if (output_format & FIO_OUTPUT_TERSE)
1676 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1677 if (output_format & FIO_OUTPUT_JSON) {
1678 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1679 json_array_add_value_object(array, tmp);
1681 if (output_format & FIO_OUTPUT_NORMAL)
1682 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1685 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1686 /* disk util stats, if any */
1687 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1689 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1691 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1692 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1693 json_free_object(root);
1696 for (i = 0; i < groupid + 1; i++) {
1701 fio_server_send_gs(rs);
1702 else if (output_format & FIO_OUTPUT_NORMAL)
1703 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1707 fio_server_send_du();
1708 else if (output_format & FIO_OUTPUT_NORMAL) {
1709 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1710 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1713 for (i = 0; i < FIO_OUTPUT_NR; i++) {
1714 buf_output_flush(&output[i]);
1715 buf_output_free(&output[i]);
1724 void show_run_stats(void)
1726 fio_mutex_down(stat_mutex);
1728 fio_mutex_up(stat_mutex);
1731 void __show_running_run_stats(void)
1733 struct thread_data *td;
1734 unsigned long long *rt;
1738 fio_mutex_down(stat_mutex);
1740 rt = malloc(thread_number * sizeof(unsigned long long));
1741 fio_gettime(&tv, NULL);
1743 for_each_td(td, i) {
1744 td->update_rusage = 1;
1745 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1746 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1747 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1748 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1750 rt[i] = mtime_since(&td->start, &tv);
1751 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1752 td->ts.runtime[DDIR_READ] += rt[i];
1753 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1754 td->ts.runtime[DDIR_WRITE] += rt[i];
1755 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1756 td->ts.runtime[DDIR_TRIM] += rt[i];
1759 for_each_td(td, i) {
1760 if (td->runstate >= TD_EXITED)
1762 if (td->rusage_sem) {
1763 td->update_rusage = 1;
1764 fio_mutex_down(td->rusage_sem);
1766 td->update_rusage = 0;
1771 for_each_td(td, i) {
1772 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1773 td->ts.runtime[DDIR_READ] -= rt[i];
1774 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1775 td->ts.runtime[DDIR_WRITE] -= rt[i];
1776 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1777 td->ts.runtime[DDIR_TRIM] -= rt[i];
1781 fio_mutex_up(stat_mutex);
1784 static int status_interval_init;
1785 static struct timeval status_time;
1786 static int status_file_disabled;
1788 #define FIO_STATUS_FILE "fio-dump-status"
1790 static int check_status_file(void)
1793 const char *temp_dir;
1794 char fio_status_file_path[PATH_MAX];
1796 if (status_file_disabled)
1799 temp_dir = getenv("TMPDIR");
1800 if (temp_dir == NULL) {
1801 temp_dir = getenv("TEMP");
1802 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1805 if (temp_dir == NULL)
1808 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1810 if (stat(fio_status_file_path, &sb))
1813 if (unlink(fio_status_file_path) < 0) {
1814 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1816 log_err("fio: disabling status file updates\n");
1817 status_file_disabled = 1;
1823 void check_for_running_stats(void)
1825 if (status_interval) {
1826 if (!status_interval_init) {
1827 fio_gettime(&status_time, NULL);
1828 status_interval_init = 1;
1829 } else if (mtime_since_now(&status_time) >= status_interval) {
1830 show_running_run_stats();
1831 fio_gettime(&status_time, NULL);
1835 if (check_status_file()) {
1836 show_running_run_stats();
1841 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1846 if (data > is->max_val)
1848 if (data < is->min_val)
1851 delta = val - is->mean.u.f;
1853 is->mean.u.f += delta / (is->samples + 1.0);
1854 is->S.u.f += delta * (val - is->mean.u.f);
1861 * Return a struct io_logs, which is added to the tail of the log
1864 static struct io_logs *get_new_log(struct io_log *iolog)
1866 size_t new_size, new_samples;
1867 struct io_logs *cur_log;
1870 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1873 if (!iolog->cur_log_max)
1874 new_samples = DEF_LOG_ENTRIES;
1876 new_samples = iolog->cur_log_max * 2;
1877 if (new_samples > MAX_LOG_ENTRIES)
1878 new_samples = MAX_LOG_ENTRIES;
1881 new_size = new_samples * log_entry_sz(iolog);
1883 cur_log = smalloc(sizeof(*cur_log));
1885 INIT_FLIST_HEAD(&cur_log->list);
1886 cur_log->log = malloc(new_size);
1888 cur_log->nr_samples = 0;
1889 cur_log->max_samples = new_samples;
1890 flist_add_tail(&cur_log->list, &iolog->io_logs);
1891 iolog->cur_log_max = new_samples;
1901 * Add and return a new log chunk, or return current log if big enough
1903 static struct io_logs *regrow_log(struct io_log *iolog)
1905 struct io_logs *cur_log;
1908 if (!iolog || iolog->disabled)
1911 cur_log = iolog_cur_log(iolog);
1913 cur_log = get_new_log(iolog);
1918 if (cur_log->nr_samples < cur_log->max_samples)
1922 * No room for a new sample. If we're compressing on the fly, flush
1923 * out the current chunk
1925 if (iolog->log_gz) {
1926 if (iolog_cur_flush(iolog, cur_log)) {
1927 log_err("fio: failed flushing iolog! Will stop logging.\n");
1933 * Get a new log array, and add to our list
1935 cur_log = get_new_log(iolog);
1937 log_err("fio: failed extending iolog! Will stop logging.\n");
1941 if (!iolog->pending || !iolog->pending->nr_samples)
1945 * Flush pending items to new log
1947 for (i = 0; i < iolog->pending->nr_samples; i++) {
1948 struct io_sample *src, *dst;
1950 src = get_sample(iolog, iolog->pending, i);
1951 dst = get_sample(iolog, cur_log, i);
1952 memcpy(dst, src, log_entry_sz(iolog));
1954 cur_log->nr_samples = iolog->pending->nr_samples;
1956 iolog->pending->nr_samples = 0;
1960 iolog->disabled = true;
1964 void regrow_logs(struct thread_data *td)
1966 regrow_log(td->slat_log);
1967 regrow_log(td->clat_log);
1968 regrow_log(td->clat_hist_log);
1969 regrow_log(td->lat_log);
1970 regrow_log(td->bw_log);
1971 regrow_log(td->iops_log);
1972 td->flags &= ~TD_F_REGROW_LOGS;
1975 static struct io_logs *get_cur_log(struct io_log *iolog)
1977 struct io_logs *cur_log;
1979 cur_log = iolog_cur_log(iolog);
1981 cur_log = get_new_log(iolog);
1986 if (cur_log->nr_samples < cur_log->max_samples)
1990 * Out of space. If we're in IO offload mode, or we're not doing
1991 * per unit logging (hence logging happens outside of the IO thread
1992 * as well), add a new log chunk inline. If we're doing inline
1993 * submissions, flag 'td' as needing a log regrow and we'll take
1994 * care of it on the submission side.
1996 if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
1997 !per_unit_log(iolog))
1998 return regrow_log(iolog);
2000 iolog->td->flags |= TD_F_REGROW_LOGS;
2001 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2002 return iolog->pending;
2005 static void __add_log_sample(struct io_log *iolog, unsigned long val,
2006 enum fio_ddir ddir, unsigned int bs,
2007 unsigned long t, uint64_t offset)
2009 struct io_logs *cur_log;
2011 if (iolog->disabled)
2013 if (flist_empty(&iolog->io_logs))
2014 iolog->avg_last = t;
2016 cur_log = get_cur_log(iolog);
2018 struct io_sample *s;
2020 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2024 io_sample_set_ddir(iolog, s, ddir);
2027 if (iolog->log_offset) {
2028 struct io_sample_offset *so = (void *) s;
2030 so->offset = offset;
2033 cur_log->nr_samples++;
2037 iolog->disabled = true;
2040 static inline void reset_io_stat(struct io_stat *ios)
2042 ios->max_val = ios->min_val = ios->samples = 0;
2043 ios->mean.u.f = ios->S.u.f = 0;
2046 void reset_io_stats(struct thread_data *td)
2048 struct thread_stat *ts = &td->ts;
2051 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2052 reset_io_stat(&ts->clat_stat[i]);
2053 reset_io_stat(&ts->slat_stat[i]);
2054 reset_io_stat(&ts->lat_stat[i]);
2055 reset_io_stat(&ts->bw_stat[i]);
2056 reset_io_stat(&ts->iops_stat[i]);
2058 ts->io_bytes[i] = 0;
2061 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2062 ts->io_u_plat[i][j] = 0;
2065 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2066 ts->io_u_map[i] = 0;
2067 ts->io_u_submit[i] = 0;
2068 ts->io_u_complete[i] = 0;
2069 ts->io_u_lat_u[i] = 0;
2070 ts->io_u_lat_m[i] = 0;
2071 ts->total_submit = 0;
2072 ts->total_complete = 0;
2075 for (i = 0; i < 3; i++) {
2076 ts->total_io_u[i] = 0;
2077 ts->short_io_u[i] = 0;
2078 ts->drop_io_u[i] = 0;
2082 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2083 unsigned long elapsed, bool log_max)
2086 * Note an entry in the log. Use the mean from the logged samples,
2087 * making sure to properly round up. Only write a log entry if we
2088 * had actual samples done.
2090 if (iolog->avg_window[ddir].samples) {
2094 val = iolog->avg_window[ddir].max_val;
2096 val = iolog->avg_window[ddir].mean.u.f + 0.50;
2098 __add_log_sample(iolog, val, ddir, 0, elapsed, 0);
2101 reset_io_stat(&iolog->avg_window[ddir]);
2104 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2109 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2110 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2113 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2114 unsigned long val, enum fio_ddir ddir,
2115 unsigned int bs, uint64_t offset)
2117 unsigned long elapsed, this_window;
2122 elapsed = mtime_since_now(&td->epoch);
2125 * If no time averaging, just add the log sample.
2127 if (!iolog->avg_msec) {
2128 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
2133 * Add the sample. If the time period has passed, then
2134 * add that entry to the log and clear.
2136 add_stat_sample(&iolog->avg_window[ddir], val);
2139 * If period hasn't passed, adding the above sample is all we
2142 this_window = elapsed - iolog->avg_last;
2143 if (this_window < iolog->avg_msec) {
2144 int diff = iolog->avg_msec - this_window;
2146 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2150 _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2152 iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2153 return iolog->avg_msec;
2156 void finalize_logs(struct thread_data *td, bool unit_logs)
2158 unsigned long elapsed;
2160 elapsed = mtime_since_now(&td->epoch);
2162 if (td->clat_log && unit_logs)
2163 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2164 if (td->slat_log && unit_logs)
2165 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2166 if (td->lat_log && unit_logs)
2167 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2168 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2169 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2170 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2171 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2174 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
2176 struct io_log *iolog;
2181 iolog = agg_io_log[ddir];
2182 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
2185 static void add_clat_percentile_sample(struct thread_stat *ts,
2186 unsigned long usec, enum fio_ddir ddir)
2188 unsigned int idx = plat_val_to_idx(usec);
2189 assert(idx < FIO_IO_U_PLAT_NR);
2191 ts->io_u_plat[ddir][idx]++;
2194 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2195 unsigned long usec, unsigned int bs, uint64_t offset)
2197 unsigned long elapsed, this_window;
2198 struct thread_stat *ts = &td->ts;
2199 struct io_log *iolog = td->clat_hist_log;
2203 add_stat_sample(&ts->clat_stat[ddir], usec);
2206 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
2208 if (ts->clat_percentiles)
2209 add_clat_percentile_sample(ts, usec, ddir);
2211 if (iolog && iolog->hist_msec) {
2212 struct io_hist *hw = &(iolog->hist_window[ddir]);
2214 elapsed = mtime_since_now(&td->epoch);
2215 if (! hw->hist_last)
2216 hw->hist_last = elapsed;
2217 this_window = elapsed - hw->hist_last;
2219 if (this_window >= iolog->hist_msec) {
2221 * Make a byte-for-byte copy of the latency histogram stored in
2222 * td->ts.io_u_plat[ddir], recording it in a log sample. Note that
2223 * the matching call to free() is located in iolog.c after printing
2224 * this sample to the log file.
2226 unsigned int *io_u_plat = (unsigned int *)(td->ts.io_u_plat[ddir]);
2227 unsigned int *dst = malloc(FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2228 memcpy(dst, io_u_plat, FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2229 __add_log_sample(iolog, (uint64_t)dst, ddir, bs, elapsed, offset);
2232 * Update the last time we recorded as being now, minus any drift
2233 * in time we encountered before actually making the record.
2235 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2243 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2244 unsigned long usec, unsigned int bs, uint64_t offset)
2246 struct thread_stat *ts = &td->ts;
2253 add_stat_sample(&ts->slat_stat[ddir], usec);
2256 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
2261 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2262 unsigned long usec, unsigned int bs, uint64_t offset)
2264 struct thread_stat *ts = &td->ts;
2271 add_stat_sample(&ts->lat_stat[ddir], usec);
2274 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2279 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2280 unsigned int bytes, unsigned long spent)
2282 struct thread_stat *ts = &td->ts;
2286 rate = bytes * 1000 / spent;
2292 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2295 add_log_sample(td, td->bw_log, rate, io_u->ddir, bytes, io_u->offset);
2297 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2301 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2303 struct thread_stat *ts = &td->ts;
2304 unsigned long spent, rate;
2306 unsigned int next, next_log;
2308 next_log = td->o.bw_avg_time;
2310 spent = mtime_since(&td->bw_sample_time, t);
2311 if (spent < td->o.bw_avg_time &&
2312 td->o.bw_avg_time - spent >= LOG_MSEC_SLACK)
2313 return td->o.bw_avg_time - spent;
2318 * Compute both read and write rates for the interval.
2320 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2323 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2325 continue; /* No entries for interval */
2328 rate = delta * 1000 / spent / 1024;
2332 add_stat_sample(&ts->bw_stat[ddir], rate);
2335 unsigned int bs = 0;
2337 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2338 bs = td->o.min_bs[ddir];
2340 next = add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2341 next_log = min(next_log, next);
2344 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2347 timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time);
2351 if (spent <= td->o.bw_avg_time)
2352 return min(next_log, td->o.bw_avg_time);
2354 next = td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time);
2355 return min(next, next_log);
2358 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2361 struct thread_stat *ts = &td->ts;
2365 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2368 add_log_sample(td, td->iops_log, 1, io_u->ddir, bytes, io_u->offset);
2370 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2374 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2376 struct thread_stat *ts = &td->ts;
2377 unsigned long spent, iops;
2379 unsigned int next, next_log;
2381 next_log = td->o.iops_avg_time;
2383 spent = mtime_since(&td->iops_sample_time, t);
2384 if (spent < td->o.iops_avg_time &&
2385 td->o.iops_avg_time - spent >= LOG_MSEC_SLACK)
2386 return td->o.iops_avg_time - spent;
2391 * Compute both read and write rates for the interval.
2393 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2396 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2398 continue; /* No entries for interval */
2401 iops = (delta * 1000) / spent;
2405 add_stat_sample(&ts->iops_stat[ddir], iops);
2408 unsigned int bs = 0;
2410 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2411 bs = td->o.min_bs[ddir];
2413 next = add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2414 next_log = min(next_log, next);
2417 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2420 timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time);
2424 if (spent <= td->o.iops_avg_time)
2425 return min(next_log, td->o.iops_avg_time);
2427 next = td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
2428 return min(next, next_log);
2432 * Returns msecs to next event
2434 int calc_log_samples(void)
2436 struct thread_data *td;
2437 unsigned int next = ~0U, tmp;
2441 fio_gettime(&now, NULL);
2443 for_each_td(td, i) {
2444 if (in_ramp_time(td) ||
2445 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2446 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2449 if (!per_unit_log(td->bw_log)) {
2450 tmp = add_bw_samples(td, &now);
2454 if (!per_unit_log(td->iops_log)) {
2455 tmp = add_iops_samples(td, &now);
2461 return next == ~0U ? 0 : next;
2464 void stat_init(void)
2466 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2469 void stat_exit(void)
2472 * When we have the mutex, we know out-of-band access to it
2475 fio_mutex_down(stat_mutex);
2476 fio_mutex_remove(stat_mutex);
2480 * Called from signal handler. Wake up status thread.
2482 void show_running_run_stats(void)
2487 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2489 /* Ignore io_u's which span multiple blocks--they will just get
2490 * inaccurate counts. */
2491 int idx = (io_u->offset - io_u->file->file_offset)
2492 / td->o.bs[DDIR_TRIM];
2493 uint32_t *info = &td->ts.block_infos[idx];
2494 assert(idx < td->ts.nr_block_infos);