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 long long 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 *io, *agg, *min, *max;
283 char *ioalt, *aggalt, *minalt, *maxalt;
284 const char *str[] = { " READ", " WRITE" , " TRIM"};
287 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
289 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
290 const int i2p = is_power_of_2(rs->kb_base);
295 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
296 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
297 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
298 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
299 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
300 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
301 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
302 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
303 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
304 rs->unified_rw_rep ? " MIXED" : str[i],
305 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
306 (unsigned long long) rs->min_run[i],
307 (unsigned long long) rs->max_run[i]);
320 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
325 * Do depth distribution calculations
327 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
329 io_u_dist[i] = (double) map[i] / (double) total;
330 io_u_dist[i] *= 100.0;
331 if (io_u_dist[i] < 0.1 && map[i])
338 static void stat_calc_lat(struct thread_stat *ts, double *dst,
339 unsigned int *src, int nr)
341 unsigned long total = ddir_rw_sum(ts->total_io_u);
345 * Do latency distribution calculations
347 for (i = 0; i < nr; i++) {
349 dst[i] = (double) src[i] / (double) total;
351 if (dst[i] < 0.01 && src[i])
358 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
360 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
363 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
365 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
368 static void display_lat(const char *name, unsigned long min, unsigned long max,
369 double mean, double dev, struct buf_output *out)
371 const char *base = "(usec)";
374 if (usec_to_msec(&min, &max, &mean, &dev))
377 minp = num2str(min, 6, 1, 0, N2S_NONE);
378 maxp = num2str(max, 6, 1, 0, N2S_NONE);
380 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
381 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
387 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
388 int ddir, struct buf_output *out)
390 const char *str[] = { " read", "write", " trim" };
391 unsigned long min, max, runt;
392 unsigned long long bw, iops;
394 char *io_p, *bw_p, *bw_p_alt, *iops_p;
397 assert(ddir_rw(ddir));
399 if (!ts->runtime[ddir])
402 i2p = is_power_of_2(rs->kb_base);
403 runt = ts->runtime[ddir];
405 bw = (1000 * ts->io_bytes[ddir]) / runt;
406 io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
407 bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
408 bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
410 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
411 iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
413 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
414 rs->unified_rw_rep ? "mixed" : str[ddir],
415 iops_p, bw_p, bw_p_alt, io_p,
416 (unsigned long long) ts->runtime[ddir]);
423 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
424 display_lat("slat", min, max, mean, dev, out);
425 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
426 display_lat("clat", min, max, mean, dev, out);
427 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
428 display_lat(" lat", min, max, mean, dev, out);
430 if (ts->clat_percentiles) {
431 show_clat_percentiles(ts->io_u_plat[ddir],
432 ts->clat_stat[ddir].samples,
434 ts->percentile_precision, out);
436 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
437 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
440 if ((rs->unit_base == 1) && i2p)
442 else if (rs->unit_base == 1)
449 if (rs->unit_base == 1) {
457 p_of_agg = mean * 100 / (double) rs->agg[ddir];
458 if (p_of_agg > 100.0)
462 if (mean > fkb_base * fkb_base) {
467 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
470 log_buf(out, " bw (%5s/s): min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, stdev=%5.02f\n",
471 bw_str, min, max, p_of_agg, mean, dev);
475 static int show_lat(double *io_u_lat, int nr, const char **ranges,
476 const char *msg, struct buf_output *out)
478 int new_line = 1, i, line = 0, shown = 0;
480 for (i = 0; i < nr; i++) {
481 if (io_u_lat[i] <= 0.0)
487 log_buf(out, " lat (%s) : ", msg);
493 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
505 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
507 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
508 "250=", "500=", "750=", "1000=", };
510 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
513 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
515 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
516 "250=", "500=", "750=", "1000=", "2000=",
519 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
522 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
524 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
525 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
527 stat_calc_lat_u(ts, io_u_lat_u);
528 stat_calc_lat_m(ts, io_u_lat_m);
530 show_lat_u(io_u_lat_u, out);
531 show_lat_m(io_u_lat_m, out);
534 static int block_state_category(int block_state)
536 switch (block_state) {
537 case BLOCK_STATE_UNINIT:
539 case BLOCK_STATE_TRIMMED:
540 case BLOCK_STATE_WRITTEN:
542 case BLOCK_STATE_WRITE_FAILURE:
543 case BLOCK_STATE_TRIM_FAILURE:
546 /* Silence compile warning on some BSDs and have a return */
552 static int compare_block_infos(const void *bs1, const void *bs2)
554 uint32_t block1 = *(uint32_t *)bs1;
555 uint32_t block2 = *(uint32_t *)bs2;
556 int state1 = BLOCK_INFO_STATE(block1);
557 int state2 = BLOCK_INFO_STATE(block2);
558 int bscat1 = block_state_category(state1);
559 int bscat2 = block_state_category(state2);
560 int cycles1 = BLOCK_INFO_TRIMS(block1);
561 int cycles2 = BLOCK_INFO_TRIMS(block2);
568 if (cycles1 < cycles2)
570 if (cycles1 > cycles2)
578 assert(block1 == block2);
582 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
583 fio_fp64_t *plist, unsigned int **percentiles,
589 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
591 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
598 * Sort the percentile list. Note that it may already be sorted if
599 * we are using the default values, but since it's a short list this
600 * isn't a worry. Also note that this does not work for NaN values.
603 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
606 /* Start only after the uninit entries end */
608 nr_uninit < nr_block_infos
609 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
613 if (nr_uninit == nr_block_infos)
616 *percentiles = calloc(len, sizeof(**percentiles));
618 for (i = 0; i < len; i++) {
619 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
621 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
624 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
625 for (i = 0; i < nr_block_infos; i++)
626 types[BLOCK_INFO_STATE(block_infos[i])]++;
631 static const char *block_state_names[] = {
632 [BLOCK_STATE_UNINIT] = "unwritten",
633 [BLOCK_STATE_TRIMMED] = "trimmed",
634 [BLOCK_STATE_WRITTEN] = "written",
635 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
636 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
639 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
640 fio_fp64_t *plist, struct buf_output *out)
643 unsigned int *percentiles = NULL;
644 unsigned int block_state_counts[BLOCK_STATE_COUNT];
646 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
647 &percentiles, block_state_counts);
649 log_buf(out, " block lifetime percentiles :\n |");
651 for (i = 0; i < len; i++) {
652 uint32_t block_info = percentiles[i];
653 #define LINE_LENGTH 75
654 char str[LINE_LENGTH];
655 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
656 plist[i].u.f, block_info,
657 i == len - 1 ? '\n' : ',');
658 assert(strln < LINE_LENGTH);
659 if (pos + strln > LINE_LENGTH) {
661 log_buf(out, "\n |");
663 log_buf(out, "%s", str);
670 log_buf(out, " states :");
671 for (i = 0; i < BLOCK_STATE_COUNT; i++)
672 log_buf(out, " %s=%u%c",
673 block_state_names[i], block_state_counts[i],
674 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
677 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
679 char *p1, *p1alt, *p2;
680 unsigned long long bw_mean, iops_mean;
681 const int i2p = is_power_of_2(ts->kb_base);
686 bw_mean = steadystate_bw_mean(ts);
687 iops_mean = steadystate_iops_mean(ts);
689 p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
690 p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
691 p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
693 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
694 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
696 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
697 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
698 ts->ss_criterion.u.f,
699 ts->ss_state & __FIO_SS_PCT ? "%" : "");
706 static void show_thread_status_normal(struct thread_stat *ts,
707 struct group_run_stats *rs,
708 struct buf_output *out)
710 double usr_cpu, sys_cpu;
711 unsigned long runtime;
712 double io_u_dist[FIO_IO_U_MAP_NR];
716 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
719 memset(time_buf, 0, sizeof(time_buf));
722 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
725 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
726 ts->name, ts->groupid, ts->members,
727 ts->error, (int) ts->pid, time_buf);
729 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
730 ts->name, ts->groupid, ts->members,
731 ts->error, ts->verror, (int) ts->pid,
735 if (strlen(ts->description))
736 log_buf(out, " Description : [%s]\n", ts->description);
738 if (ts->io_bytes[DDIR_READ])
739 show_ddir_status(rs, ts, DDIR_READ, out);
740 if (ts->io_bytes[DDIR_WRITE])
741 show_ddir_status(rs, ts, DDIR_WRITE, out);
742 if (ts->io_bytes[DDIR_TRIM])
743 show_ddir_status(rs, ts, DDIR_TRIM, out);
745 show_latencies(ts, out);
747 runtime = ts->total_run_time;
749 double runt = (double) runtime;
751 usr_cpu = (double) ts->usr_time * 100 / runt;
752 sys_cpu = (double) ts->sys_time * 100 / runt;
758 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
759 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
760 (unsigned long long) ts->ctx,
761 (unsigned long long) ts->majf,
762 (unsigned long long) ts->minf);
764 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
765 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
766 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
767 io_u_dist[1], io_u_dist[2],
768 io_u_dist[3], io_u_dist[4],
769 io_u_dist[5], io_u_dist[6]);
771 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
772 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
773 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
774 io_u_dist[1], io_u_dist[2],
775 io_u_dist[3], io_u_dist[4],
776 io_u_dist[5], io_u_dist[6]);
777 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
778 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
779 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
780 io_u_dist[1], io_u_dist[2],
781 io_u_dist[3], io_u_dist[4],
782 io_u_dist[5], io_u_dist[6]);
783 log_buf(out, " issued rwt: total=%llu,%llu,%llu,"
784 " short=%llu,%llu,%llu,"
785 " dropped=%llu,%llu,%llu\n",
786 (unsigned long long) ts->total_io_u[0],
787 (unsigned long long) ts->total_io_u[1],
788 (unsigned long long) ts->total_io_u[2],
789 (unsigned long long) ts->short_io_u[0],
790 (unsigned long long) ts->short_io_u[1],
791 (unsigned long long) ts->short_io_u[2],
792 (unsigned long long) ts->drop_io_u[0],
793 (unsigned long long) ts->drop_io_u[1],
794 (unsigned long long) ts->drop_io_u[2]);
795 if (ts->continue_on_error) {
796 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
797 (unsigned long long)ts->total_err_count,
799 strerror(ts->first_error));
801 if (ts->latency_depth) {
802 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
803 (unsigned long long)ts->latency_target,
804 (unsigned long long)ts->latency_window,
805 ts->latency_percentile.u.f,
809 if (ts->nr_block_infos)
810 show_block_infos(ts->nr_block_infos, ts->block_infos,
811 ts->percentile_list, out);
814 show_ss_normal(ts, out);
817 static void show_ddir_status_terse(struct thread_stat *ts,
818 struct group_run_stats *rs, int ddir,
819 struct buf_output *out)
821 unsigned long min, max;
822 unsigned long long bw, iops;
823 unsigned int *ovals = NULL;
825 unsigned int len, minv, maxv;
828 assert(ddir_rw(ddir));
831 if (ts->runtime[ddir]) {
832 uint64_t runt = ts->runtime[ddir];
834 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
835 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
838 log_buf(out, ";%llu;%llu;%llu;%llu",
839 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
840 (unsigned long long) ts->runtime[ddir]);
842 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
843 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
845 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
847 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
848 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
850 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
852 if (ts->clat_percentiles) {
853 len = calc_clat_percentiles(ts->io_u_plat[ddir],
854 ts->clat_stat[ddir].samples,
855 ts->percentile_list, &ovals, &maxv,
860 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
862 log_buf(out, ";0%%=0");
865 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
868 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
869 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
871 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
876 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
877 double p_of_agg = 100.0;
880 p_of_agg = mean * 100 / (double) rs->agg[ddir];
881 if (p_of_agg > 100.0)
885 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
887 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
890 static void add_ddir_status_json(struct thread_stat *ts,
891 struct group_run_stats *rs, int ddir, struct json_object *parent)
893 unsigned long min, max;
894 unsigned long long bw;
895 unsigned int *ovals = NULL;
896 double mean, dev, iops;
897 unsigned int len, minv, maxv;
899 const char *ddirname[] = {"read", "write", "trim"};
900 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
902 double p_of_agg = 100.0;
904 assert(ddir_rw(ddir));
906 if (ts->unified_rw_rep && ddir != DDIR_READ)
909 dir_object = json_create_object();
910 json_object_add_value_object(parent,
911 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
915 if (ts->runtime[ddir]) {
916 uint64_t runt = ts->runtime[ddir];
918 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
919 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
922 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
923 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
924 json_object_add_value_int(dir_object, "bw", bw);
925 json_object_add_value_float(dir_object, "iops", iops);
926 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
927 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
928 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
929 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
931 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
935 tmp_object = json_create_object();
936 json_object_add_value_object(dir_object, "slat", tmp_object);
937 json_object_add_value_int(tmp_object, "min", min);
938 json_object_add_value_int(tmp_object, "max", max);
939 json_object_add_value_float(tmp_object, "mean", mean);
940 json_object_add_value_float(tmp_object, "stddev", dev);
942 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
946 tmp_object = json_create_object();
947 json_object_add_value_object(dir_object, "clat", tmp_object);
948 json_object_add_value_int(tmp_object, "min", min);
949 json_object_add_value_int(tmp_object, "max", max);
950 json_object_add_value_float(tmp_object, "mean", mean);
951 json_object_add_value_float(tmp_object, "stddev", dev);
953 if (ts->clat_percentiles) {
954 len = calc_clat_percentiles(ts->io_u_plat[ddir],
955 ts->clat_stat[ddir].samples,
956 ts->percentile_list, &ovals, &maxv,
961 percentile_object = json_create_object();
962 json_object_add_value_object(tmp_object, "percentile", percentile_object);
963 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
965 json_object_add_value_int(percentile_object, "0.00", 0);
968 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
969 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
972 if (output_format & FIO_OUTPUT_JSON_PLUS) {
973 clat_bins_object = json_create_object();
974 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
975 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
976 if (ts->io_u_plat[ddir][i]) {
977 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
978 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
983 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
987 tmp_object = json_create_object();
988 json_object_add_value_object(dir_object, "lat", tmp_object);
989 json_object_add_value_int(tmp_object, "min", min);
990 json_object_add_value_int(tmp_object, "max", max);
991 json_object_add_value_float(tmp_object, "mean", mean);
992 json_object_add_value_float(tmp_object, "stddev", dev);
996 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
998 p_of_agg = mean * 100 / (double) rs->agg[ddir];
999 if (p_of_agg > 100.0)
1004 p_of_agg = mean = dev = 0.0;
1006 json_object_add_value_int(dir_object, "bw_min", min);
1007 json_object_add_value_int(dir_object, "bw_max", max);
1008 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1009 json_object_add_value_float(dir_object, "bw_mean", mean);
1010 json_object_add_value_float(dir_object, "bw_dev", dev);
1013 static void show_thread_status_terse_v2(struct thread_stat *ts,
1014 struct group_run_stats *rs,
1015 struct buf_output *out)
1017 double io_u_dist[FIO_IO_U_MAP_NR];
1018 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1019 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1020 double usr_cpu, sys_cpu;
1024 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1025 /* Log Read Status */
1026 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1027 /* Log Write Status */
1028 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1029 /* Log Trim Status */
1030 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1033 if (ts->total_run_time) {
1034 double runt = (double) ts->total_run_time;
1036 usr_cpu = (double) ts->usr_time * 100 / runt;
1037 sys_cpu = (double) ts->sys_time * 100 / runt;
1043 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1044 (unsigned long long) ts->ctx,
1045 (unsigned long long) ts->majf,
1046 (unsigned long long) ts->minf);
1048 /* Calc % distribution of IO depths, usecond, msecond latency */
1049 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1050 stat_calc_lat_u(ts, io_u_lat_u);
1051 stat_calc_lat_m(ts, io_u_lat_m);
1053 /* Only show fixed 7 I/O depth levels*/
1054 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1055 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1056 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1058 /* Microsecond latency */
1059 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1060 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1061 /* Millisecond latency */
1062 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1063 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1064 /* Additional output if continue_on_error set - default off*/
1065 if (ts->continue_on_error)
1066 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1069 /* Additional output if description is set */
1070 if (strlen(ts->description))
1071 log_buf(out, ";%s", ts->description);
1076 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1077 struct group_run_stats *rs, int ver,
1078 struct buf_output *out)
1080 double io_u_dist[FIO_IO_U_MAP_NR];
1081 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1082 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1083 double usr_cpu, sys_cpu;
1087 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1088 ts->name, ts->groupid, ts->error);
1089 /* Log Read Status */
1090 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1091 /* Log Write Status */
1092 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1093 /* Log Trim Status */
1095 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1098 if (ts->total_run_time) {
1099 double runt = (double) ts->total_run_time;
1101 usr_cpu = (double) ts->usr_time * 100 / runt;
1102 sys_cpu = (double) ts->sys_time * 100 / runt;
1108 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1109 (unsigned long long) ts->ctx,
1110 (unsigned long long) ts->majf,
1111 (unsigned long long) ts->minf);
1113 /* Calc % distribution of IO depths, usecond, msecond latency */
1114 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1115 stat_calc_lat_u(ts, io_u_lat_u);
1116 stat_calc_lat_m(ts, io_u_lat_m);
1118 /* Only show fixed 7 I/O depth levels*/
1119 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1120 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1121 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1123 /* Microsecond latency */
1124 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1125 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1126 /* Millisecond latency */
1127 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1128 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1130 /* disk util stats, if any */
1131 show_disk_util(1, NULL, out);
1133 /* Additional output if continue_on_error set - default off*/
1134 if (ts->continue_on_error)
1135 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1137 /* Additional output if description is set */
1138 if (strlen(ts->description))
1139 log_buf(out, ";%s", ts->description);
1144 static void json_add_job_opts(struct json_object *root, const char *name,
1145 struct flist_head *opt_list, bool num_jobs)
1147 struct json_object *dir_object;
1148 struct flist_head *entry;
1149 struct print_option *p;
1151 if (flist_empty(opt_list))
1154 dir_object = json_create_object();
1155 json_object_add_value_object(root, name, dir_object);
1157 flist_for_each(entry, opt_list) {
1158 const char *pos = "";
1160 p = flist_entry(entry, struct print_option, list);
1161 if (!num_jobs && !strcmp(p->name, "numjobs"))
1165 json_object_add_value_string(dir_object, p->name, pos);
1169 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1170 struct group_run_stats *rs,
1171 struct flist_head *opt_list)
1173 struct json_object *root, *tmp;
1174 struct jobs_eta *je;
1175 double io_u_dist[FIO_IO_U_MAP_NR];
1176 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1177 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1178 double usr_cpu, sys_cpu;
1182 root = json_create_object();
1183 json_object_add_value_string(root, "jobname", ts->name);
1184 json_object_add_value_int(root, "groupid", ts->groupid);
1185 json_object_add_value_int(root, "error", ts->error);
1188 je = get_jobs_eta(true, &size);
1190 json_object_add_value_int(root, "eta", je->eta_sec);
1191 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1195 json_add_job_opts(root, "job options", opt_list, true);
1197 add_ddir_status_json(ts, rs, DDIR_READ, root);
1198 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1199 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1202 if (ts->total_run_time) {
1203 double runt = (double) ts->total_run_time;
1205 usr_cpu = (double) ts->usr_time * 100 / runt;
1206 sys_cpu = (double) ts->sys_time * 100 / runt;
1211 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1212 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1213 json_object_add_value_int(root, "ctx", ts->ctx);
1214 json_object_add_value_int(root, "majf", ts->majf);
1215 json_object_add_value_int(root, "minf", ts->minf);
1218 /* Calc % distribution of IO depths, usecond, msecond latency */
1219 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1220 stat_calc_lat_u(ts, io_u_lat_u);
1221 stat_calc_lat_m(ts, io_u_lat_m);
1223 tmp = json_create_object();
1224 json_object_add_value_object(root, "iodepth_level", tmp);
1225 /* Only show fixed 7 I/O depth levels*/
1226 for (i = 0; i < 7; i++) {
1229 snprintf(name, 20, "%d", 1 << i);
1231 snprintf(name, 20, ">=%d", 1 << i);
1232 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1235 tmp = json_create_object();
1236 json_object_add_value_object(root, "latency_us", tmp);
1237 /* Microsecond latency */
1238 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1239 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1240 "250", "500", "750", "1000", };
1241 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1243 /* Millisecond latency */
1244 tmp = json_create_object();
1245 json_object_add_value_object(root, "latency_ms", tmp);
1246 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1247 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1248 "250", "500", "750", "1000", "2000",
1250 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1253 /* Additional output if continue_on_error set - default off*/
1254 if (ts->continue_on_error) {
1255 json_object_add_value_int(root, "total_err", ts->total_err_count);
1256 json_object_add_value_int(root, "first_error", ts->first_error);
1259 if (ts->latency_depth) {
1260 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1261 json_object_add_value_int(root, "latency_target", ts->latency_target);
1262 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1263 json_object_add_value_int(root, "latency_window", ts->latency_window);
1266 /* Additional output if description is set */
1267 if (strlen(ts->description))
1268 json_object_add_value_string(root, "desc", ts->description);
1270 if (ts->nr_block_infos) {
1271 /* Block error histogram and types */
1273 unsigned int *percentiles = NULL;
1274 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1276 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1277 ts->percentile_list,
1278 &percentiles, block_state_counts);
1281 struct json_object *block, *percentile_object, *states;
1283 block = json_create_object();
1284 json_object_add_value_object(root, "block", block);
1286 percentile_object = json_create_object();
1287 json_object_add_value_object(block, "percentiles",
1289 for (i = 0; i < len; i++) {
1291 snprintf(buf, sizeof(buf), "%f",
1292 ts->percentile_list[i].u.f);
1293 json_object_add_value_int(percentile_object,
1298 states = json_create_object();
1299 json_object_add_value_object(block, "states", states);
1300 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1301 json_object_add_value_int(states,
1302 block_state_names[state],
1303 block_state_counts[state]);
1310 struct json_object *data;
1311 struct json_array *iops, *bw;
1315 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1316 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1317 ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1318 (float) ts->ss_limit.u.f,
1319 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1321 tmp = json_create_object();
1322 json_object_add_value_object(root, "steadystate", tmp);
1323 json_object_add_value_string(tmp, "ss", ss_buf);
1324 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1325 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1327 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1328 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1329 json_object_add_value_string(tmp, "criterion", ss_buf);
1330 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1331 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1333 data = json_create_object();
1334 json_object_add_value_object(tmp, "data", data);
1335 bw = json_create_array();
1336 iops = json_create_array();
1339 ** if ss was attained or the buffer is not full,
1340 ** ss->head points to the first element in the list.
1341 ** otherwise it actually points to the second element
1344 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1347 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1348 for (i = 0; i < ts->ss_dur; i++) {
1349 k = (j + i) % ts->ss_dur;
1350 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1351 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1353 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1354 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1355 json_object_add_value_array(data, "iops", iops);
1356 json_object_add_value_array(data, "bw", bw);
1362 static void show_thread_status_terse(struct thread_stat *ts,
1363 struct group_run_stats *rs,
1364 struct buf_output *out)
1366 if (terse_version == 2)
1367 show_thread_status_terse_v2(ts, rs, out);
1368 else if (terse_version == 3 || terse_version == 4)
1369 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1371 log_err("fio: bad terse version!? %d\n", terse_version);
1374 struct json_object *show_thread_status(struct thread_stat *ts,
1375 struct group_run_stats *rs,
1376 struct flist_head *opt_list,
1377 struct buf_output *out)
1379 struct json_object *ret = NULL;
1381 if (output_format & FIO_OUTPUT_TERSE)
1382 show_thread_status_terse(ts, rs, out);
1383 if (output_format & FIO_OUTPUT_JSON)
1384 ret = show_thread_status_json(ts, rs, opt_list);
1385 if (output_format & FIO_OUTPUT_NORMAL)
1386 show_thread_status_normal(ts, rs, out);
1391 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1395 if (src->samples == 0)
1398 dst->min_val = min(dst->min_val, src->min_val);
1399 dst->max_val = max(dst->max_val, src->max_val);
1402 * Compute new mean and S after the merge
1403 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1404 * #Parallel_algorithm>
1407 mean = src->mean.u.f;
1410 double delta = src->mean.u.f - dst->mean.u.f;
1412 mean = ((src->mean.u.f * src->samples) +
1413 (dst->mean.u.f * dst->samples)) /
1414 (dst->samples + src->samples);
1416 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1417 (dst->samples * src->samples) /
1418 (dst->samples + src->samples);
1421 dst->samples += src->samples;
1422 dst->mean.u.f = mean;
1426 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1430 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1431 if (dst->max_run[i] < src->max_run[i])
1432 dst->max_run[i] = src->max_run[i];
1433 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1434 dst->min_run[i] = src->min_run[i];
1435 if (dst->max_bw[i] < src->max_bw[i])
1436 dst->max_bw[i] = src->max_bw[i];
1437 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1438 dst->min_bw[i] = src->min_bw[i];
1440 dst->iobytes[i] += src->iobytes[i];
1441 dst->agg[i] += src->agg[i];
1445 dst->kb_base = src->kb_base;
1446 if (!dst->unit_base)
1447 dst->unit_base = src->unit_base;
1450 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1455 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1456 if (!dst->unified_rw_rep) {
1457 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1458 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1459 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1460 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1462 dst->io_bytes[l] += src->io_bytes[l];
1464 if (dst->runtime[l] < src->runtime[l])
1465 dst->runtime[l] = src->runtime[l];
1467 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1468 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1469 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1470 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1472 dst->io_bytes[0] += src->io_bytes[l];
1474 if (dst->runtime[0] < src->runtime[l])
1475 dst->runtime[0] = src->runtime[l];
1478 * We're summing to the same destination, so override
1479 * 'first' after the first iteration of the loop
1485 dst->usr_time += src->usr_time;
1486 dst->sys_time += src->sys_time;
1487 dst->ctx += src->ctx;
1488 dst->majf += src->majf;
1489 dst->minf += src->minf;
1491 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1492 dst->io_u_map[k] += src->io_u_map[k];
1493 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1494 dst->io_u_submit[k] += src->io_u_submit[k];
1495 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1496 dst->io_u_complete[k] += src->io_u_complete[k];
1497 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1498 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1499 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1500 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1502 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1503 if (!dst->unified_rw_rep) {
1504 dst->total_io_u[k] += src->total_io_u[k];
1505 dst->short_io_u[k] += src->short_io_u[k];
1506 dst->drop_io_u[k] += src->drop_io_u[k];
1508 dst->total_io_u[0] += src->total_io_u[k];
1509 dst->short_io_u[0] += src->short_io_u[k];
1510 dst->drop_io_u[0] += src->drop_io_u[k];
1514 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1517 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1518 if (!dst->unified_rw_rep)
1519 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1521 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1525 dst->total_run_time += src->total_run_time;
1526 dst->total_submit += src->total_submit;
1527 dst->total_complete += src->total_complete;
1530 void init_group_run_stat(struct group_run_stats *gs)
1533 memset(gs, 0, sizeof(*gs));
1535 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1536 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1539 void init_thread_stat(struct thread_stat *ts)
1543 memset(ts, 0, sizeof(*ts));
1545 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1546 ts->lat_stat[j].min_val = -1UL;
1547 ts->clat_stat[j].min_val = -1UL;
1548 ts->slat_stat[j].min_val = -1UL;
1549 ts->bw_stat[j].min_val = -1UL;
1554 void __show_run_stats(void)
1556 struct group_run_stats *runstats, *rs;
1557 struct thread_data *td;
1558 struct thread_stat *threadstats, *ts;
1559 int i, j, k, nr_ts, last_ts, idx;
1560 int kb_base_warned = 0;
1561 int unit_base_warned = 0;
1562 struct json_object *root = NULL;
1563 struct json_array *array = NULL;
1564 struct buf_output output[FIO_OUTPUT_NR];
1565 struct flist_head **opt_lists;
1567 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1569 for (i = 0; i < groupid + 1; i++)
1570 init_group_run_stat(&runstats[i]);
1573 * find out how many threads stats we need. if group reporting isn't
1574 * enabled, it's one-per-td.
1578 for_each_td(td, i) {
1579 if (!td->o.group_reporting) {
1583 if (last_ts == td->groupid)
1588 last_ts = td->groupid;
1592 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1593 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1595 for (i = 0; i < nr_ts; i++) {
1596 init_thread_stat(&threadstats[i]);
1597 opt_lists[i] = NULL;
1603 for_each_td(td, i) {
1606 if (idx && (!td->o.group_reporting ||
1607 (td->o.group_reporting && last_ts != td->groupid))) {
1612 last_ts = td->groupid;
1614 ts = &threadstats[j];
1616 ts->clat_percentiles = td->o.clat_percentiles;
1617 ts->percentile_precision = td->o.percentile_precision;
1618 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1619 opt_lists[j] = &td->opt_list;
1624 if (ts->groupid == -1) {
1626 * These are per-group shared already
1628 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1629 if (td->o.description)
1630 strncpy(ts->description, td->o.description,
1631 FIO_JOBDESC_SIZE - 1);
1633 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1636 * If multiple entries in this group, this is
1639 ts->thread_number = td->thread_number;
1640 ts->groupid = td->groupid;
1643 * first pid in group, not very useful...
1647 ts->kb_base = td->o.kb_base;
1648 ts->unit_base = td->o.unit_base;
1649 ts->unified_rw_rep = td->o.unified_rw_rep;
1650 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1651 log_info("fio: kb_base differs for jobs in group, using"
1652 " %u as the base\n", ts->kb_base);
1654 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1655 log_info("fio: unit_base differs for jobs in group, using"
1656 " %u as the base\n", ts->unit_base);
1657 unit_base_warned = 1;
1660 ts->continue_on_error = td->o.continue_on_error;
1661 ts->total_err_count += td->total_err_count;
1662 ts->first_error = td->first_error;
1664 if (!td->error && td->o.continue_on_error &&
1666 ts->error = td->first_error;
1667 ts->verror[sizeof(ts->verror) - 1] = '\0';
1668 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1669 } else if (td->error) {
1670 ts->error = td->error;
1671 ts->verror[sizeof(ts->verror) - 1] = '\0';
1672 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1676 ts->latency_depth = td->latency_qd;
1677 ts->latency_target = td->o.latency_target;
1678 ts->latency_percentile = td->o.latency_percentile;
1679 ts->latency_window = td->o.latency_window;
1681 ts->nr_block_infos = td->ts.nr_block_infos;
1682 for (k = 0; k < ts->nr_block_infos; k++)
1683 ts->block_infos[k] = td->ts.block_infos[k];
1685 sum_thread_stats(ts, &td->ts, idx == 1);
1688 ts->ss_state = td->ss.state;
1689 ts->ss_dur = td->ss.dur;
1690 ts->ss_head = td->ss.head;
1691 ts->ss_bw_data = td->ss.bw_data;
1692 ts->ss_iops_data = td->ss.iops_data;
1693 ts->ss_limit.u.f = td->ss.limit;
1694 ts->ss_slope.u.f = td->ss.slope;
1695 ts->ss_deviation.u.f = td->ss.deviation;
1696 ts->ss_criterion.u.f = td->ss.criterion;
1699 ts->ss_dur = ts->ss_state = 0;
1702 for (i = 0; i < nr_ts; i++) {
1703 unsigned long long bw;
1705 ts = &threadstats[i];
1706 if (ts->groupid == -1)
1708 rs = &runstats[ts->groupid];
1709 rs->kb_base = ts->kb_base;
1710 rs->unit_base = ts->unit_base;
1711 rs->unified_rw_rep += ts->unified_rw_rep;
1713 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1714 if (!ts->runtime[j])
1716 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1717 rs->min_run[j] = ts->runtime[j];
1718 if (ts->runtime[j] > rs->max_run[j])
1719 rs->max_run[j] = ts->runtime[j];
1723 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1724 if (bw < rs->min_bw[j])
1726 if (bw > rs->max_bw[j])
1729 rs->iobytes[j] += ts->io_bytes[j];
1733 for (i = 0; i < groupid + 1; i++) {
1738 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1739 if (rs->max_run[ddir])
1740 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1745 for (i = 0; i < FIO_OUTPUT_NR; i++)
1746 buf_output_init(&output[i]);
1749 * don't overwrite last signal output
1751 if (output_format & FIO_OUTPUT_NORMAL)
1752 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1753 if (output_format & FIO_OUTPUT_JSON) {
1754 struct thread_data *global;
1757 unsigned long long ms_since_epoch;
1759 gettimeofday(&now, NULL);
1760 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1761 (unsigned long long)(now.tv_usec) / 1000;
1763 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1765 if (time_buf[strlen(time_buf) - 1] == '\n')
1766 time_buf[strlen(time_buf) - 1] = '\0';
1768 root = json_create_object();
1769 json_object_add_value_string(root, "fio version", fio_version_string);
1770 json_object_add_value_int(root, "timestamp", now.tv_sec);
1771 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1772 json_object_add_value_string(root, "time", time_buf);
1773 global = get_global_options();
1774 json_add_job_opts(root, "global options", &global->opt_list, false);
1775 array = json_create_array();
1776 json_object_add_value_array(root, "jobs", array);
1780 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1782 for (i = 0; i < nr_ts; i++) {
1783 ts = &threadstats[i];
1784 rs = &runstats[ts->groupid];
1787 fio_server_send_job_options(opt_lists[i], i);
1788 fio_server_send_ts(ts, rs);
1790 if (output_format & FIO_OUTPUT_TERSE)
1791 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1792 if (output_format & FIO_OUTPUT_JSON) {
1793 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1794 json_array_add_value_object(array, tmp);
1796 if (output_format & FIO_OUTPUT_NORMAL)
1797 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1800 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1801 /* disk util stats, if any */
1802 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1804 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1806 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1807 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1808 json_free_object(root);
1811 for (i = 0; i < groupid + 1; i++) {
1816 fio_server_send_gs(rs);
1817 else if (output_format & FIO_OUTPUT_NORMAL)
1818 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1822 fio_server_send_du();
1823 else if (output_format & FIO_OUTPUT_NORMAL) {
1824 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1825 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1828 for (i = 0; i < FIO_OUTPUT_NR; i++) {
1829 struct buf_output *out = &output[i];
1831 log_info_buf(out->buf, out->buflen);
1832 buf_output_free(out);
1841 void show_run_stats(void)
1843 fio_mutex_down(stat_mutex);
1845 fio_mutex_up(stat_mutex);
1848 void __show_running_run_stats(void)
1850 struct thread_data *td;
1851 unsigned long long *rt;
1855 fio_mutex_down(stat_mutex);
1857 rt = malloc(thread_number * sizeof(unsigned long long));
1858 fio_gettime(&tv, NULL);
1860 for_each_td(td, i) {
1861 td->update_rusage = 1;
1862 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1863 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1864 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1865 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1867 rt[i] = mtime_since(&td->start, &tv);
1868 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1869 td->ts.runtime[DDIR_READ] += rt[i];
1870 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1871 td->ts.runtime[DDIR_WRITE] += rt[i];
1872 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1873 td->ts.runtime[DDIR_TRIM] += rt[i];
1876 for_each_td(td, i) {
1877 if (td->runstate >= TD_EXITED)
1879 if (td->rusage_sem) {
1880 td->update_rusage = 1;
1881 fio_mutex_down(td->rusage_sem);
1883 td->update_rusage = 0;
1888 for_each_td(td, i) {
1889 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1890 td->ts.runtime[DDIR_READ] -= rt[i];
1891 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1892 td->ts.runtime[DDIR_WRITE] -= rt[i];
1893 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1894 td->ts.runtime[DDIR_TRIM] -= rt[i];
1898 fio_mutex_up(stat_mutex);
1901 static int status_interval_init;
1902 static struct timeval status_time;
1903 static int status_file_disabled;
1905 #define FIO_STATUS_FILE "fio-dump-status"
1907 static int check_status_file(void)
1910 const char *temp_dir;
1911 char fio_status_file_path[PATH_MAX];
1913 if (status_file_disabled)
1916 temp_dir = getenv("TMPDIR");
1917 if (temp_dir == NULL) {
1918 temp_dir = getenv("TEMP");
1919 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1922 if (temp_dir == NULL)
1925 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1927 if (stat(fio_status_file_path, &sb))
1930 if (unlink(fio_status_file_path) < 0) {
1931 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1933 log_err("fio: disabling status file updates\n");
1934 status_file_disabled = 1;
1940 void check_for_running_stats(void)
1942 if (status_interval) {
1943 if (!status_interval_init) {
1944 fio_gettime(&status_time, NULL);
1945 status_interval_init = 1;
1946 } else if (mtime_since_now(&status_time) >= status_interval) {
1947 show_running_run_stats();
1948 fio_gettime(&status_time, NULL);
1952 if (check_status_file()) {
1953 show_running_run_stats();
1958 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1963 if (data > is->max_val)
1965 if (data < is->min_val)
1968 delta = val - is->mean.u.f;
1970 is->mean.u.f += delta / (is->samples + 1.0);
1971 is->S.u.f += delta * (val - is->mean.u.f);
1978 * Return a struct io_logs, which is added to the tail of the log
1981 static struct io_logs *get_new_log(struct io_log *iolog)
1983 size_t new_size, new_samples;
1984 struct io_logs *cur_log;
1987 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1990 if (!iolog->cur_log_max)
1991 new_samples = DEF_LOG_ENTRIES;
1993 new_samples = iolog->cur_log_max * 2;
1994 if (new_samples > MAX_LOG_ENTRIES)
1995 new_samples = MAX_LOG_ENTRIES;
1998 new_size = new_samples * log_entry_sz(iolog);
2000 cur_log = smalloc(sizeof(*cur_log));
2002 INIT_FLIST_HEAD(&cur_log->list);
2003 cur_log->log = malloc(new_size);
2005 cur_log->nr_samples = 0;
2006 cur_log->max_samples = new_samples;
2007 flist_add_tail(&cur_log->list, &iolog->io_logs);
2008 iolog->cur_log_max = new_samples;
2018 * Add and return a new log chunk, or return current log if big enough
2020 static struct io_logs *regrow_log(struct io_log *iolog)
2022 struct io_logs *cur_log;
2025 if (!iolog || iolog->disabled)
2028 cur_log = iolog_cur_log(iolog);
2030 cur_log = get_new_log(iolog);
2035 if (cur_log->nr_samples < cur_log->max_samples)
2039 * No room for a new sample. If we're compressing on the fly, flush
2040 * out the current chunk
2042 if (iolog->log_gz) {
2043 if (iolog_cur_flush(iolog, cur_log)) {
2044 log_err("fio: failed flushing iolog! Will stop logging.\n");
2050 * Get a new log array, and add to our list
2052 cur_log = get_new_log(iolog);
2054 log_err("fio: failed extending iolog! Will stop logging.\n");
2058 if (!iolog->pending || !iolog->pending->nr_samples)
2062 * Flush pending items to new log
2064 for (i = 0; i < iolog->pending->nr_samples; i++) {
2065 struct io_sample *src, *dst;
2067 src = get_sample(iolog, iolog->pending, i);
2068 dst = get_sample(iolog, cur_log, i);
2069 memcpy(dst, src, log_entry_sz(iolog));
2071 cur_log->nr_samples = iolog->pending->nr_samples;
2073 iolog->pending->nr_samples = 0;
2077 iolog->disabled = true;
2081 void regrow_logs(struct thread_data *td)
2083 regrow_log(td->slat_log);
2084 regrow_log(td->clat_log);
2085 regrow_log(td->clat_hist_log);
2086 regrow_log(td->lat_log);
2087 regrow_log(td->bw_log);
2088 regrow_log(td->iops_log);
2089 td->flags &= ~TD_F_REGROW_LOGS;
2092 static struct io_logs *get_cur_log(struct io_log *iolog)
2094 struct io_logs *cur_log;
2096 cur_log = iolog_cur_log(iolog);
2098 cur_log = get_new_log(iolog);
2103 if (cur_log->nr_samples < cur_log->max_samples)
2107 * Out of space. If we're in IO offload mode, or we're not doing
2108 * per unit logging (hence logging happens outside of the IO thread
2109 * as well), add a new log chunk inline. If we're doing inline
2110 * submissions, flag 'td' as needing a log regrow and we'll take
2111 * care of it on the submission side.
2113 if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2114 !per_unit_log(iolog))
2115 return regrow_log(iolog);
2117 iolog->td->flags |= TD_F_REGROW_LOGS;
2118 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2119 return iolog->pending;
2122 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2123 enum fio_ddir ddir, unsigned int bs,
2124 unsigned long t, uint64_t offset)
2126 struct io_logs *cur_log;
2128 if (iolog->disabled)
2130 if (flist_empty(&iolog->io_logs))
2131 iolog->avg_last = t;
2133 cur_log = get_cur_log(iolog);
2135 struct io_sample *s;
2137 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2140 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2141 io_sample_set_ddir(iolog, s, ddir);
2144 if (iolog->log_offset) {
2145 struct io_sample_offset *so = (void *) s;
2147 so->offset = offset;
2150 cur_log->nr_samples++;
2154 iolog->disabled = true;
2157 static inline void reset_io_stat(struct io_stat *ios)
2159 ios->max_val = ios->min_val = ios->samples = 0;
2160 ios->mean.u.f = ios->S.u.f = 0;
2163 void reset_io_stats(struct thread_data *td)
2165 struct thread_stat *ts = &td->ts;
2168 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2169 reset_io_stat(&ts->clat_stat[i]);
2170 reset_io_stat(&ts->slat_stat[i]);
2171 reset_io_stat(&ts->lat_stat[i]);
2172 reset_io_stat(&ts->bw_stat[i]);
2173 reset_io_stat(&ts->iops_stat[i]);
2175 ts->io_bytes[i] = 0;
2177 ts->total_io_u[i] = 0;
2178 ts->short_io_u[i] = 0;
2179 ts->drop_io_u[i] = 0;
2181 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2182 ts->io_u_plat[i][j] = 0;
2185 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2186 ts->io_u_map[i] = 0;
2187 ts->io_u_submit[i] = 0;
2188 ts->io_u_complete[i] = 0;
2191 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2192 ts->io_u_lat_u[i] = 0;
2193 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2194 ts->io_u_lat_m[i] = 0;
2196 ts->total_submit = 0;
2197 ts->total_complete = 0;
2200 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2201 unsigned long elapsed, bool log_max)
2204 * Note an entry in the log. Use the mean from the logged samples,
2205 * making sure to properly round up. Only write a log entry if we
2206 * had actual samples done.
2208 if (iolog->avg_window[ddir].samples) {
2209 union io_sample_data data;
2212 data.val = iolog->avg_window[ddir].max_val;
2214 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2216 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2219 reset_io_stat(&iolog->avg_window[ddir]);
2222 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2227 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2228 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2231 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2232 union io_sample_data data, enum fio_ddir ddir,
2233 unsigned int bs, uint64_t offset)
2235 unsigned long elapsed, this_window;
2240 elapsed = mtime_since_now(&td->epoch);
2243 * If no time averaging, just add the log sample.
2245 if (!iolog->avg_msec) {
2246 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2251 * Add the sample. If the time period has passed, then
2252 * add that entry to the log and clear.
2254 add_stat_sample(&iolog->avg_window[ddir], data.val);
2257 * If period hasn't passed, adding the above sample is all we
2260 this_window = elapsed - iolog->avg_last;
2261 if (elapsed < iolog->avg_last)
2262 return iolog->avg_last - elapsed;
2263 else if (this_window < iolog->avg_msec) {
2264 int diff = iolog->avg_msec - this_window;
2266 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2270 _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2272 iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2273 return iolog->avg_msec;
2276 void finalize_logs(struct thread_data *td, bool unit_logs)
2278 unsigned long elapsed;
2280 elapsed = mtime_since_now(&td->epoch);
2282 if (td->clat_log && unit_logs)
2283 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2284 if (td->slat_log && unit_logs)
2285 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2286 if (td->lat_log && unit_logs)
2287 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2288 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2289 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2290 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2291 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2294 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2296 struct io_log *iolog;
2301 iolog = agg_io_log[ddir];
2302 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2305 static void add_clat_percentile_sample(struct thread_stat *ts,
2306 unsigned long usec, enum fio_ddir ddir)
2308 unsigned int idx = plat_val_to_idx(usec);
2309 assert(idx < FIO_IO_U_PLAT_NR);
2311 ts->io_u_plat[ddir][idx]++;
2314 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2315 unsigned long usec, unsigned int bs, uint64_t offset)
2317 unsigned long elapsed, this_window;
2318 struct thread_stat *ts = &td->ts;
2319 struct io_log *iolog = td->clat_hist_log;
2323 add_stat_sample(&ts->clat_stat[ddir], usec);
2326 add_log_sample(td, td->clat_log, sample_val(usec), ddir, bs,
2329 if (ts->clat_percentiles)
2330 add_clat_percentile_sample(ts, usec, ddir);
2332 if (iolog && iolog->hist_msec) {
2333 struct io_hist *hw = &iolog->hist_window[ddir];
2336 elapsed = mtime_since_now(&td->epoch);
2338 hw->hist_last = elapsed;
2339 this_window = elapsed - hw->hist_last;
2341 if (this_window >= iolog->hist_msec) {
2342 unsigned int *io_u_plat;
2343 struct io_u_plat_entry *dst;
2346 * Make a byte-for-byte copy of the latency histogram
2347 * stored in td->ts.io_u_plat[ddir], recording it in a
2348 * log sample. Note that the matching call to free() is
2349 * located in iolog.c after printing this sample to the
2352 io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2353 dst = malloc(sizeof(struct io_u_plat_entry));
2354 memcpy(&(dst->io_u_plat), io_u_plat,
2355 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2356 flist_add(&dst->list, &hw->list);
2357 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2361 * Update the last time we recorded as being now, minus
2362 * any drift in time we encountered before actually
2363 * making the record.
2365 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2373 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2374 unsigned long usec, unsigned int bs, uint64_t offset)
2376 struct thread_stat *ts = &td->ts;
2383 add_stat_sample(&ts->slat_stat[ddir], usec);
2386 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2391 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2392 unsigned long usec, unsigned int bs, uint64_t offset)
2394 struct thread_stat *ts = &td->ts;
2401 add_stat_sample(&ts->lat_stat[ddir], usec);
2404 add_log_sample(td, td->lat_log, sample_val(usec), ddir, bs,
2410 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2411 unsigned int bytes, unsigned long spent)
2413 struct thread_stat *ts = &td->ts;
2417 rate = bytes * 1000 / spent;
2423 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2426 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2427 bytes, io_u->offset);
2429 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2433 static int __add_samples(struct thread_data *td, struct timeval *parent_tv,
2434 struct timeval *t, unsigned int avg_time,
2435 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2436 struct io_stat *stat, struct io_log *log,
2439 unsigned long spent, rate;
2441 unsigned int next, next_log;
2443 next_log = avg_time;
2445 spent = mtime_since(parent_tv, t);
2446 if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2447 return avg_time - spent;
2452 * Compute both read and write rates for the interval.
2454 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2457 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2459 continue; /* No entries for interval */
2463 rate = delta * 1000 / spent / 1024; /* KiB/s */
2465 rate = (delta * 1000) / spent;
2469 add_stat_sample(&stat[ddir], rate);
2472 unsigned int bs = 0;
2474 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2475 bs = td->o.min_bs[ddir];
2477 next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2478 next_log = min(next_log, next);
2481 stat_io_bytes[ddir] = this_io_bytes[ddir];
2484 timeval_add_msec(parent_tv, avg_time);
2488 if (spent <= avg_time)
2491 next = avg_time - (1 + spent - avg_time);
2493 return min(next, next_log);
2496 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2498 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2499 td->this_io_bytes, td->stat_io_bytes,
2500 td->ts.bw_stat, td->bw_log, true);
2503 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2506 struct thread_stat *ts = &td->ts;
2510 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2513 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2514 bytes, io_u->offset);
2516 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2520 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2522 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2523 td->this_io_blocks, td->stat_io_blocks,
2524 td->ts.iops_stat, td->iops_log, false);
2528 * Returns msecs to next event
2530 int calc_log_samples(void)
2532 struct thread_data *td;
2533 unsigned int next = ~0U, tmp;
2537 fio_gettime(&now, NULL);
2539 for_each_td(td, i) {
2542 if (in_ramp_time(td) ||
2543 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2544 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2548 (td->bw_log && !per_unit_log(td->bw_log))) {
2549 tmp = add_bw_samples(td, &now);
2553 if (!td->iops_log ||
2554 (td->iops_log && !per_unit_log(td->iops_log))) {
2555 tmp = add_iops_samples(td, &now);
2561 return next == ~0U ? 0 : next;
2564 void stat_init(void)
2566 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2569 void stat_exit(void)
2572 * When we have the mutex, we know out-of-band access to it
2575 fio_mutex_down(stat_mutex);
2576 fio_mutex_remove(stat_mutex);
2580 * Called from signal handler. Wake up status thread.
2582 void show_running_run_stats(void)
2587 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2589 /* Ignore io_u's which span multiple blocks--they will just get
2590 * inaccurate counts. */
2591 int idx = (io_u->offset - io_u->file->file_offset)
2592 / td->o.bs[DDIR_TRIM];
2593 uint32_t *info = &td->ts.block_infos[idx];
2594 assert(idx < td->ts.nr_block_infos);