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_tv(&td->ru_start.ru_utime,
41 &td->ru_end.ru_utime);
42 ts->sys_time += mtime_since_tv(&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 long long val)
63 unsigned int msb, error_bits, base, offset, idx;
65 /* Find MSB starting from bit 0 */
69 msb = (sizeof(val)*8) - __builtin_clzll(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 long long **output,
139 unsigned long long *maxv, unsigned long long *minv)
141 unsigned long sum = 0;
142 unsigned int len, i, j = 0;
143 unsigned int oval_len = 0;
144 unsigned long long *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(*ovals));
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 divisor, len, i, j = 0;
205 unsigned long long minv, maxv;
206 unsigned long long *ovals;
207 int is_last, per_line, scale_down, time_width;
210 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
215 * We default to nsecs, but if the value range is such that we
216 * should scale down to usecs or msecs, do that.
218 if (minv > 2000000 && maxv > 99999999ULL) {
221 log_buf(out, " clat percentiles (msec):\n |");
222 } else if (minv > 2000 && maxv > 99999) {
225 log_buf(out, " clat percentiles (usec):\n |");
229 log_buf(out, " clat percentiles (nsec):\n |");
233 time_width = max(5, (int) (log10(maxv / divisor) + 1));
234 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
235 precision, time_width);
236 /* fmt will be something like " %5.2fth=[%4llu]%c" */
237 per_line = (80 - 7) / (precision + 10 + time_width);
239 for (j = 0; j < len; j++) {
241 if (j != 0 && (j % per_line) == 0)
244 /* end of the list */
245 is_last = (j == len - 1);
247 for (i = 0; i < scale_down; i++)
248 ovals[j] = (ovals[j] + 999) / 1000;
250 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
255 if ((j % per_line) == per_line - 1) /* for formatting */
264 bool calc_lat(struct io_stat *is, unsigned long long *min,
265 unsigned long long *max, double *mean, double *dev)
267 double n = (double) is->samples;
274 *mean = is->mean.u.f;
277 *dev = sqrt(is->S.u.f / (n - 1.0));
284 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
286 char *io, *agg, *min, *max;
287 char *ioalt, *aggalt, *minalt, *maxalt;
288 const char *str[] = { " READ", " WRITE" , " TRIM"};
291 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
293 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
294 const int i2p = is_power_of_2(rs->kb_base);
299 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
300 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
301 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
302 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
303 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
304 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
305 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
306 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
307 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
308 rs->unified_rw_rep ? " MIXED" : str[i],
309 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
310 (unsigned long long) rs->min_run[i],
311 (unsigned long long) rs->max_run[i]);
324 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
329 * Do depth distribution calculations
331 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
333 io_u_dist[i] = (double) map[i] / (double) total;
334 io_u_dist[i] *= 100.0;
335 if (io_u_dist[i] < 0.1 && map[i])
342 static void stat_calc_lat(struct thread_stat *ts, double *dst,
343 unsigned int *src, int nr)
345 unsigned long total = ddir_rw_sum(ts->total_io_u);
349 * Do latency distribution calculations
351 for (i = 0; i < nr; i++) {
353 dst[i] = (double) src[i] / (double) total;
355 if (dst[i] < 0.01 && src[i])
363 * To keep the terse format unaltered, add all of the ns latency
364 * buckets to the first us latency bucket
366 void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
368 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
371 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
373 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
374 ntotal += ts->io_u_lat_n[i];
376 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
379 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
381 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
384 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
386 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
389 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
391 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
394 static void display_lat(const char *name, unsigned long long min,
395 unsigned long long max, double mean, double dev,
396 struct buf_output *out)
398 const char *base = "(nsec)";
401 if (nsec_to_msec(&min, &max, &mean, &dev))
403 else if (nsec_to_usec(&min, &max, &mean, &dev))
406 minp = num2str(min, 6, 1, 0, N2S_NONE);
407 maxp = num2str(max, 6, 1, 0, N2S_NONE);
409 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
410 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
416 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
417 int ddir, struct buf_output *out)
419 const char *str[] = { " read", "write", " trim" };
421 unsigned long long min, max, bw, iops;
423 char *io_p, *bw_p, *bw_p_alt, *iops_p;
426 assert(ddir_rw(ddir));
428 if (!ts->runtime[ddir])
431 i2p = is_power_of_2(rs->kb_base);
432 runt = ts->runtime[ddir];
434 bw = (1000 * ts->io_bytes[ddir]) / runt;
435 io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
436 bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
437 bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
439 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
440 iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
442 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
443 rs->unified_rw_rep ? "mixed" : str[ddir],
444 iops_p, bw_p, bw_p_alt, io_p,
445 (unsigned long long) ts->runtime[ddir]);
452 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
453 display_lat("slat", min, max, mean, dev, out);
454 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
455 display_lat("clat", min, max, mean, dev, out);
456 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
457 display_lat(" lat", min, max, mean, dev, out);
459 if (ts->clat_percentiles) {
460 show_clat_percentiles(ts->io_u_plat[ddir],
461 ts->clat_stat[ddir].samples,
463 ts->percentile_precision, out);
465 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
466 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
469 if ((rs->unit_base == 1) && i2p)
471 else if (rs->unit_base == 1)
478 if (rs->unit_base == 1) {
486 p_of_agg = mean * 100 / (double) rs->agg[ddir];
487 if (p_of_agg > 100.0)
491 if (mean > fkb_base * fkb_base) {
496 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
499 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, avg=%5.02f, stdev=%5.02f\n",
500 bw_str, min, max, p_of_agg, mean, dev);
504 static int show_lat(double *io_u_lat, int nr, const char **ranges,
505 const char *msg, struct buf_output *out)
507 int new_line = 1, i, line = 0, shown = 0;
509 for (i = 0; i < nr; i++) {
510 if (io_u_lat[i] <= 0.0)
516 log_buf(out, " lat (%s) : ", msg);
522 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
534 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
536 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
537 "250=", "500=", "750=", "1000=", };
539 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
542 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
544 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
545 "250=", "500=", "750=", "1000=", };
547 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
550 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
552 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
553 "250=", "500=", "750=", "1000=", "2000=",
556 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
559 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
561 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
562 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
563 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
565 stat_calc_lat_n(ts, io_u_lat_n);
566 stat_calc_lat_u(ts, io_u_lat_u);
567 stat_calc_lat_m(ts, io_u_lat_m);
569 show_lat_n(io_u_lat_n, out);
570 show_lat_u(io_u_lat_u, out);
571 show_lat_m(io_u_lat_m, out);
574 static int block_state_category(int block_state)
576 switch (block_state) {
577 case BLOCK_STATE_UNINIT:
579 case BLOCK_STATE_TRIMMED:
580 case BLOCK_STATE_WRITTEN:
582 case BLOCK_STATE_WRITE_FAILURE:
583 case BLOCK_STATE_TRIM_FAILURE:
586 /* Silence compile warning on some BSDs and have a return */
592 static int compare_block_infos(const void *bs1, const void *bs2)
594 uint32_t block1 = *(uint32_t *)bs1;
595 uint32_t block2 = *(uint32_t *)bs2;
596 int state1 = BLOCK_INFO_STATE(block1);
597 int state2 = BLOCK_INFO_STATE(block2);
598 int bscat1 = block_state_category(state1);
599 int bscat2 = block_state_category(state2);
600 int cycles1 = BLOCK_INFO_TRIMS(block1);
601 int cycles2 = BLOCK_INFO_TRIMS(block2);
608 if (cycles1 < cycles2)
610 if (cycles1 > cycles2)
618 assert(block1 == block2);
622 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
623 fio_fp64_t *plist, unsigned int **percentiles,
629 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
631 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
638 * Sort the percentile list. Note that it may already be sorted if
639 * we are using the default values, but since it's a short list this
640 * isn't a worry. Also note that this does not work for NaN values.
643 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
646 /* Start only after the uninit entries end */
648 nr_uninit < nr_block_infos
649 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
653 if (nr_uninit == nr_block_infos)
656 *percentiles = calloc(len, sizeof(**percentiles));
658 for (i = 0; i < len; i++) {
659 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
661 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
664 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
665 for (i = 0; i < nr_block_infos; i++)
666 types[BLOCK_INFO_STATE(block_infos[i])]++;
671 static const char *block_state_names[] = {
672 [BLOCK_STATE_UNINIT] = "unwritten",
673 [BLOCK_STATE_TRIMMED] = "trimmed",
674 [BLOCK_STATE_WRITTEN] = "written",
675 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
676 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
679 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
680 fio_fp64_t *plist, struct buf_output *out)
683 unsigned int *percentiles = NULL;
684 unsigned int block_state_counts[BLOCK_STATE_COUNT];
686 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
687 &percentiles, block_state_counts);
689 log_buf(out, " block lifetime percentiles :\n |");
691 for (i = 0; i < len; i++) {
692 uint32_t block_info = percentiles[i];
693 #define LINE_LENGTH 75
694 char str[LINE_LENGTH];
695 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
696 plist[i].u.f, block_info,
697 i == len - 1 ? '\n' : ',');
698 assert(strln < LINE_LENGTH);
699 if (pos + strln > LINE_LENGTH) {
701 log_buf(out, "\n |");
703 log_buf(out, "%s", str);
710 log_buf(out, " states :");
711 for (i = 0; i < BLOCK_STATE_COUNT; i++)
712 log_buf(out, " %s=%u%c",
713 block_state_names[i], block_state_counts[i],
714 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
717 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
719 char *p1, *p1alt, *p2;
720 unsigned long long bw_mean, iops_mean;
721 const int i2p = is_power_of_2(ts->kb_base);
726 bw_mean = steadystate_bw_mean(ts);
727 iops_mean = steadystate_iops_mean(ts);
729 p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
730 p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
731 p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
733 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
734 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
736 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
737 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
738 ts->ss_criterion.u.f,
739 ts->ss_state & __FIO_SS_PCT ? "%" : "");
746 static void show_thread_status_normal(struct thread_stat *ts,
747 struct group_run_stats *rs,
748 struct buf_output *out)
750 double usr_cpu, sys_cpu;
751 unsigned long runtime;
752 double io_u_dist[FIO_IO_U_MAP_NR];
756 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
759 memset(time_buf, 0, sizeof(time_buf));
762 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
765 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
766 ts->name, ts->groupid, ts->members,
767 ts->error, (int) ts->pid, time_buf);
769 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
770 ts->name, ts->groupid, ts->members,
771 ts->error, ts->verror, (int) ts->pid,
775 if (strlen(ts->description))
776 log_buf(out, " Description : [%s]\n", ts->description);
778 if (ts->io_bytes[DDIR_READ])
779 show_ddir_status(rs, ts, DDIR_READ, out);
780 if (ts->io_bytes[DDIR_WRITE])
781 show_ddir_status(rs, ts, DDIR_WRITE, out);
782 if (ts->io_bytes[DDIR_TRIM])
783 show_ddir_status(rs, ts, DDIR_TRIM, out);
785 show_latencies(ts, out);
787 runtime = ts->total_run_time;
789 double runt = (double) runtime;
791 usr_cpu = (double) ts->usr_time * 100 / runt;
792 sys_cpu = (double) ts->sys_time * 100 / runt;
798 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
799 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
800 (unsigned long long) ts->ctx,
801 (unsigned long long) ts->majf,
802 (unsigned long long) ts->minf);
804 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
805 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
806 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
807 io_u_dist[1], io_u_dist[2],
808 io_u_dist[3], io_u_dist[4],
809 io_u_dist[5], io_u_dist[6]);
811 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
812 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
813 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
814 io_u_dist[1], io_u_dist[2],
815 io_u_dist[3], io_u_dist[4],
816 io_u_dist[5], io_u_dist[6]);
817 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
818 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
819 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
820 io_u_dist[1], io_u_dist[2],
821 io_u_dist[3], io_u_dist[4],
822 io_u_dist[5], io_u_dist[6]);
823 log_buf(out, " issued rwt: total=%llu,%llu,%llu,"
824 " short=%llu,%llu,%llu,"
825 " dropped=%llu,%llu,%llu\n",
826 (unsigned long long) ts->total_io_u[0],
827 (unsigned long long) ts->total_io_u[1],
828 (unsigned long long) ts->total_io_u[2],
829 (unsigned long long) ts->short_io_u[0],
830 (unsigned long long) ts->short_io_u[1],
831 (unsigned long long) ts->short_io_u[2],
832 (unsigned long long) ts->drop_io_u[0],
833 (unsigned long long) ts->drop_io_u[1],
834 (unsigned long long) ts->drop_io_u[2]);
835 if (ts->continue_on_error) {
836 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
837 (unsigned long long)ts->total_err_count,
839 strerror(ts->first_error));
841 if (ts->latency_depth) {
842 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
843 (unsigned long long)ts->latency_target,
844 (unsigned long long)ts->latency_window,
845 ts->latency_percentile.u.f,
849 if (ts->nr_block_infos)
850 show_block_infos(ts->nr_block_infos, ts->block_infos,
851 ts->percentile_list, out);
854 show_ss_normal(ts, out);
857 static void show_ddir_status_terse(struct thread_stat *ts,
858 struct group_run_stats *rs, int ddir,
859 struct buf_output *out)
861 unsigned long long min, max, minv, maxv, bw, iops;
862 unsigned long long *ovals = NULL;
867 assert(ddir_rw(ddir));
870 if (ts->runtime[ddir]) {
871 uint64_t runt = ts->runtime[ddir];
873 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
874 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
877 log_buf(out, ";%llu;%llu;%llu;%llu",
878 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
879 (unsigned long long) ts->runtime[ddir]);
881 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
882 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
884 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
886 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
887 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
889 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
891 if (ts->clat_percentiles) {
892 len = calc_clat_percentiles(ts->io_u_plat[ddir],
893 ts->clat_stat[ddir].samples,
894 ts->percentile_list, &ovals, &maxv,
899 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
901 log_buf(out, ";0%%=0");
904 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
907 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
908 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
910 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
915 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
916 double p_of_agg = 100.0;
919 p_of_agg = mean * 100 / (double) rs->agg[ddir];
920 if (p_of_agg > 100.0)
924 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
926 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
929 static void add_ddir_status_json(struct thread_stat *ts,
930 struct group_run_stats *rs, int ddir, struct json_object *parent)
932 unsigned long long min, max, minv, maxv;
933 unsigned long long bw;
934 unsigned long long *ovals = NULL;
935 double mean, dev, iops;
938 const char *ddirname[] = {"read", "write", "trim"};
939 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
941 double p_of_agg = 100.0;
943 assert(ddir_rw(ddir));
945 if (ts->unified_rw_rep && ddir != DDIR_READ)
948 dir_object = json_create_object();
949 json_object_add_value_object(parent,
950 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
954 if (ts->runtime[ddir]) {
955 uint64_t runt = ts->runtime[ddir];
957 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
958 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
961 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
962 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
963 json_object_add_value_int(dir_object, "bw", bw);
964 json_object_add_value_float(dir_object, "iops", iops);
965 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
966 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
967 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
968 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
970 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
974 tmp_object = json_create_object();
975 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
976 json_object_add_value_int(tmp_object, "min", min);
977 json_object_add_value_int(tmp_object, "max", max);
978 json_object_add_value_float(tmp_object, "mean", mean);
979 json_object_add_value_float(tmp_object, "stddev", dev);
981 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
985 tmp_object = json_create_object();
986 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
987 json_object_add_value_int(tmp_object, "min", min);
988 json_object_add_value_int(tmp_object, "max", max);
989 json_object_add_value_float(tmp_object, "mean", mean);
990 json_object_add_value_float(tmp_object, "stddev", dev);
992 if (ts->clat_percentiles) {
993 len = calc_clat_percentiles(ts->io_u_plat[ddir],
994 ts->clat_stat[ddir].samples,
995 ts->percentile_list, &ovals, &maxv,
1000 percentile_object = json_create_object();
1001 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1002 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1004 json_object_add_value_int(percentile_object, "0.00", 0);
1007 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1008 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1011 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1012 clat_bins_object = json_create_object();
1013 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1014 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1015 if (ts->io_u_plat[ddir][i]) {
1016 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1017 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1022 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1026 tmp_object = json_create_object();
1027 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1028 json_object_add_value_int(tmp_object, "min", min);
1029 json_object_add_value_int(tmp_object, "max", max);
1030 json_object_add_value_float(tmp_object, "mean", mean);
1031 json_object_add_value_float(tmp_object, "stddev", dev);
1035 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1036 if (rs->agg[ddir]) {
1037 p_of_agg = mean * 100 / (double) rs->agg[ddir];
1038 if (p_of_agg > 100.0)
1043 p_of_agg = mean = dev = 0.0;
1045 json_object_add_value_int(dir_object, "bw_min", min);
1046 json_object_add_value_int(dir_object, "bw_max", max);
1047 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1048 json_object_add_value_float(dir_object, "bw_mean", mean);
1049 json_object_add_value_float(dir_object, "bw_dev", dev);
1052 static void show_thread_status_terse_v2(struct thread_stat *ts,
1053 struct group_run_stats *rs,
1054 struct buf_output *out)
1056 double io_u_dist[FIO_IO_U_MAP_NR];
1057 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1058 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1059 double usr_cpu, sys_cpu;
1063 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1064 /* Log Read Status */
1065 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1066 /* Log Write Status */
1067 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1068 /* Log Trim Status */
1069 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1072 if (ts->total_run_time) {
1073 double runt = (double) ts->total_run_time;
1075 usr_cpu = (double) ts->usr_time * 100 / runt;
1076 sys_cpu = (double) ts->sys_time * 100 / runt;
1082 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1083 (unsigned long long) ts->ctx,
1084 (unsigned long long) ts->majf,
1085 (unsigned long long) ts->minf);
1087 /* Calc % distribution of IO depths, usecond, msecond latency */
1088 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1089 stat_calc_lat_nu(ts, io_u_lat_u);
1090 stat_calc_lat_m(ts, io_u_lat_m);
1092 /* Only show fixed 7 I/O depth levels*/
1093 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1094 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1095 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1097 /* Microsecond latency */
1098 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1099 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1100 /* Millisecond latency */
1101 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1102 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1103 /* Additional output if continue_on_error set - default off*/
1104 if (ts->continue_on_error)
1105 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1108 /* Additional output if description is set */
1109 if (strlen(ts->description))
1110 log_buf(out, ";%s", ts->description);
1115 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1116 struct group_run_stats *rs, int ver,
1117 struct buf_output *out)
1119 double io_u_dist[FIO_IO_U_MAP_NR];
1120 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1121 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1122 double usr_cpu, sys_cpu;
1126 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1127 ts->name, ts->groupid, ts->error);
1128 /* Log Read Status */
1129 show_ddir_status_terse(ts, rs, DDIR_READ, out);
1130 /* Log Write Status */
1131 show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1132 /* Log Trim Status */
1134 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1137 if (ts->total_run_time) {
1138 double runt = (double) ts->total_run_time;
1140 usr_cpu = (double) ts->usr_time * 100 / runt;
1141 sys_cpu = (double) ts->sys_time * 100 / runt;
1147 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1148 (unsigned long long) ts->ctx,
1149 (unsigned long long) ts->majf,
1150 (unsigned long long) ts->minf);
1152 /* Calc % distribution of IO depths, usecond, msecond latency */
1153 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1154 stat_calc_lat_nu(ts, io_u_lat_u);
1155 stat_calc_lat_m(ts, io_u_lat_m);
1157 /* Only show fixed 7 I/O depth levels*/
1158 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1159 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1160 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1162 /* Microsecond latency */
1163 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1164 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1165 /* Millisecond latency */
1166 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1167 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1169 /* disk util stats, if any */
1170 show_disk_util(1, NULL, out);
1172 /* Additional output if continue_on_error set - default off*/
1173 if (ts->continue_on_error)
1174 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1176 /* Additional output if description is set */
1177 if (strlen(ts->description))
1178 log_buf(out, ";%s", ts->description);
1183 static void json_add_job_opts(struct json_object *root, const char *name,
1184 struct flist_head *opt_list, bool num_jobs)
1186 struct json_object *dir_object;
1187 struct flist_head *entry;
1188 struct print_option *p;
1190 if (flist_empty(opt_list))
1193 dir_object = json_create_object();
1194 json_object_add_value_object(root, name, dir_object);
1196 flist_for_each(entry, opt_list) {
1197 const char *pos = "";
1199 p = flist_entry(entry, struct print_option, list);
1200 if (!num_jobs && !strcmp(p->name, "numjobs"))
1204 json_object_add_value_string(dir_object, p->name, pos);
1208 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1209 struct group_run_stats *rs,
1210 struct flist_head *opt_list)
1212 struct json_object *root, *tmp;
1213 struct jobs_eta *je;
1214 double io_u_dist[FIO_IO_U_MAP_NR];
1215 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1216 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1217 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1218 double usr_cpu, sys_cpu;
1222 root = json_create_object();
1223 json_object_add_value_string(root, "jobname", ts->name);
1224 json_object_add_value_int(root, "groupid", ts->groupid);
1225 json_object_add_value_int(root, "error", ts->error);
1228 je = get_jobs_eta(true, &size);
1230 json_object_add_value_int(root, "eta", je->eta_sec);
1231 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1235 json_add_job_opts(root, "job options", opt_list, true);
1237 add_ddir_status_json(ts, rs, DDIR_READ, root);
1238 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1239 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1242 if (ts->total_run_time) {
1243 double runt = (double) ts->total_run_time;
1245 usr_cpu = (double) ts->usr_time * 100 / runt;
1246 sys_cpu = (double) ts->sys_time * 100 / runt;
1251 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1252 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1253 json_object_add_value_int(root, "ctx", ts->ctx);
1254 json_object_add_value_int(root, "majf", ts->majf);
1255 json_object_add_value_int(root, "minf", ts->minf);
1258 /* Calc % distribution of IO depths, usecond, msecond latency */
1259 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1260 stat_calc_lat_n(ts, io_u_lat_n);
1261 stat_calc_lat_u(ts, io_u_lat_u);
1262 stat_calc_lat_m(ts, io_u_lat_m);
1264 tmp = json_create_object();
1265 json_object_add_value_object(root, "iodepth_level", tmp);
1266 /* Only show fixed 7 I/O depth levels*/
1267 for (i = 0; i < 7; i++) {
1270 snprintf(name, 20, "%d", 1 << i);
1272 snprintf(name, 20, ">=%d", 1 << i);
1273 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1276 /* Nanosecond latency */
1277 tmp = json_create_object();
1278 json_object_add_value_object(root, "latency_ns", tmp);
1279 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1280 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1281 "250", "500", "750", "1000", };
1282 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1284 /* Microsecond latency */
1285 tmp = json_create_object();
1286 json_object_add_value_object(root, "latency_us", tmp);
1287 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1288 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1289 "250", "500", "750", "1000", };
1290 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1292 /* Millisecond latency */
1293 tmp = json_create_object();
1294 json_object_add_value_object(root, "latency_ms", tmp);
1295 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1296 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1297 "250", "500", "750", "1000", "2000",
1299 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1302 /* Additional output if continue_on_error set - default off*/
1303 if (ts->continue_on_error) {
1304 json_object_add_value_int(root, "total_err", ts->total_err_count);
1305 json_object_add_value_int(root, "first_error", ts->first_error);
1308 if (ts->latency_depth) {
1309 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1310 json_object_add_value_int(root, "latency_target", ts->latency_target);
1311 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1312 json_object_add_value_int(root, "latency_window", ts->latency_window);
1315 /* Additional output if description is set */
1316 if (strlen(ts->description))
1317 json_object_add_value_string(root, "desc", ts->description);
1319 if (ts->nr_block_infos) {
1320 /* Block error histogram and types */
1322 unsigned int *percentiles = NULL;
1323 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1325 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1326 ts->percentile_list,
1327 &percentiles, block_state_counts);
1330 struct json_object *block, *percentile_object, *states;
1332 block = json_create_object();
1333 json_object_add_value_object(root, "block", block);
1335 percentile_object = json_create_object();
1336 json_object_add_value_object(block, "percentiles",
1338 for (i = 0; i < len; i++) {
1340 snprintf(buf, sizeof(buf), "%f",
1341 ts->percentile_list[i].u.f);
1342 json_object_add_value_int(percentile_object,
1347 states = json_create_object();
1348 json_object_add_value_object(block, "states", states);
1349 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1350 json_object_add_value_int(states,
1351 block_state_names[state],
1352 block_state_counts[state]);
1359 struct json_object *data;
1360 struct json_array *iops, *bw;
1364 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1365 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1366 ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1367 (float) ts->ss_limit.u.f,
1368 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1370 tmp = json_create_object();
1371 json_object_add_value_object(root, "steadystate", tmp);
1372 json_object_add_value_string(tmp, "ss", ss_buf);
1373 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1374 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1376 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1377 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1378 json_object_add_value_string(tmp, "criterion", ss_buf);
1379 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1380 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1382 data = json_create_object();
1383 json_object_add_value_object(tmp, "data", data);
1384 bw = json_create_array();
1385 iops = json_create_array();
1388 ** if ss was attained or the buffer is not full,
1389 ** ss->head points to the first element in the list.
1390 ** otherwise it actually points to the second element
1393 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1396 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1397 for (i = 0; i < ts->ss_dur; i++) {
1398 k = (j + i) % ts->ss_dur;
1399 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1400 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1402 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1403 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1404 json_object_add_value_array(data, "iops", iops);
1405 json_object_add_value_array(data, "bw", bw);
1411 static void show_thread_status_terse(struct thread_stat *ts,
1412 struct group_run_stats *rs,
1413 struct buf_output *out)
1415 if (terse_version == 2)
1416 show_thread_status_terse_v2(ts, rs, out);
1417 else if (terse_version == 3 || terse_version == 4)
1418 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1420 log_err("fio: bad terse version!? %d\n", terse_version);
1423 struct json_object *show_thread_status(struct thread_stat *ts,
1424 struct group_run_stats *rs,
1425 struct flist_head *opt_list,
1426 struct buf_output *out)
1428 struct json_object *ret = NULL;
1430 if (output_format & FIO_OUTPUT_TERSE)
1431 show_thread_status_terse(ts, rs, out);
1432 if (output_format & FIO_OUTPUT_JSON)
1433 ret = show_thread_status_json(ts, rs, opt_list);
1434 if (output_format & FIO_OUTPUT_NORMAL)
1435 show_thread_status_normal(ts, rs, out);
1440 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1444 if (src->samples == 0)
1447 dst->min_val = min(dst->min_val, src->min_val);
1448 dst->max_val = max(dst->max_val, src->max_val);
1451 * Compute new mean and S after the merge
1452 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1453 * #Parallel_algorithm>
1456 mean = src->mean.u.f;
1459 double delta = src->mean.u.f - dst->mean.u.f;
1461 mean = ((src->mean.u.f * src->samples) +
1462 (dst->mean.u.f * dst->samples)) /
1463 (dst->samples + src->samples);
1465 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1466 (dst->samples * src->samples) /
1467 (dst->samples + src->samples);
1470 dst->samples += src->samples;
1471 dst->mean.u.f = mean;
1475 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1479 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1480 if (dst->max_run[i] < src->max_run[i])
1481 dst->max_run[i] = src->max_run[i];
1482 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1483 dst->min_run[i] = src->min_run[i];
1484 if (dst->max_bw[i] < src->max_bw[i])
1485 dst->max_bw[i] = src->max_bw[i];
1486 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1487 dst->min_bw[i] = src->min_bw[i];
1489 dst->iobytes[i] += src->iobytes[i];
1490 dst->agg[i] += src->agg[i];
1494 dst->kb_base = src->kb_base;
1495 if (!dst->unit_base)
1496 dst->unit_base = src->unit_base;
1499 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1504 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1505 if (!dst->unified_rw_rep) {
1506 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1507 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1508 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1509 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1511 dst->io_bytes[l] += src->io_bytes[l];
1513 if (dst->runtime[l] < src->runtime[l])
1514 dst->runtime[l] = src->runtime[l];
1516 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1517 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1518 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1519 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1521 dst->io_bytes[0] += src->io_bytes[l];
1523 if (dst->runtime[0] < src->runtime[l])
1524 dst->runtime[0] = src->runtime[l];
1527 * We're summing to the same destination, so override
1528 * 'first' after the first iteration of the loop
1534 dst->usr_time += src->usr_time;
1535 dst->sys_time += src->sys_time;
1536 dst->ctx += src->ctx;
1537 dst->majf += src->majf;
1538 dst->minf += src->minf;
1540 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1541 dst->io_u_map[k] += src->io_u_map[k];
1542 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1543 dst->io_u_submit[k] += src->io_u_submit[k];
1544 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1545 dst->io_u_complete[k] += src->io_u_complete[k];
1546 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1547 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1548 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1549 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1550 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1551 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1553 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1554 if (!dst->unified_rw_rep) {
1555 dst->total_io_u[k] += src->total_io_u[k];
1556 dst->short_io_u[k] += src->short_io_u[k];
1557 dst->drop_io_u[k] += src->drop_io_u[k];
1559 dst->total_io_u[0] += src->total_io_u[k];
1560 dst->short_io_u[0] += src->short_io_u[k];
1561 dst->drop_io_u[0] += src->drop_io_u[k];
1565 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1568 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1569 if (!dst->unified_rw_rep)
1570 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1572 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1576 dst->total_run_time += src->total_run_time;
1577 dst->total_submit += src->total_submit;
1578 dst->total_complete += src->total_complete;
1581 void init_group_run_stat(struct group_run_stats *gs)
1584 memset(gs, 0, sizeof(*gs));
1586 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1587 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1590 void init_thread_stat(struct thread_stat *ts)
1594 memset(ts, 0, sizeof(*ts));
1596 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1597 ts->lat_stat[j].min_val = -1UL;
1598 ts->clat_stat[j].min_val = -1UL;
1599 ts->slat_stat[j].min_val = -1UL;
1600 ts->bw_stat[j].min_val = -1UL;
1605 void __show_run_stats(void)
1607 struct group_run_stats *runstats, *rs;
1608 struct thread_data *td;
1609 struct thread_stat *threadstats, *ts;
1610 int i, j, k, nr_ts, last_ts, idx;
1611 int kb_base_warned = 0;
1612 int unit_base_warned = 0;
1613 struct json_object *root = NULL;
1614 struct json_array *array = NULL;
1615 struct buf_output output[FIO_OUTPUT_NR];
1616 struct flist_head **opt_lists;
1618 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1620 for (i = 0; i < groupid + 1; i++)
1621 init_group_run_stat(&runstats[i]);
1624 * find out how many threads stats we need. if group reporting isn't
1625 * enabled, it's one-per-td.
1629 for_each_td(td, i) {
1630 if (!td->o.group_reporting) {
1634 if (last_ts == td->groupid)
1639 last_ts = td->groupid;
1643 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1644 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1646 for (i = 0; i < nr_ts; i++) {
1647 init_thread_stat(&threadstats[i]);
1648 opt_lists[i] = NULL;
1654 for_each_td(td, i) {
1657 if (idx && (!td->o.group_reporting ||
1658 (td->o.group_reporting && last_ts != td->groupid))) {
1663 last_ts = td->groupid;
1665 ts = &threadstats[j];
1667 ts->clat_percentiles = td->o.clat_percentiles;
1668 ts->percentile_precision = td->o.percentile_precision;
1669 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1670 opt_lists[j] = &td->opt_list;
1675 if (ts->groupid == -1) {
1677 * These are per-group shared already
1679 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1680 if (td->o.description)
1681 strncpy(ts->description, td->o.description,
1682 FIO_JOBDESC_SIZE - 1);
1684 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1687 * If multiple entries in this group, this is
1690 ts->thread_number = td->thread_number;
1691 ts->groupid = td->groupid;
1694 * first pid in group, not very useful...
1698 ts->kb_base = td->o.kb_base;
1699 ts->unit_base = td->o.unit_base;
1700 ts->unified_rw_rep = td->o.unified_rw_rep;
1701 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1702 log_info("fio: kb_base differs for jobs in group, using"
1703 " %u as the base\n", ts->kb_base);
1705 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1706 log_info("fio: unit_base differs for jobs in group, using"
1707 " %u as the base\n", ts->unit_base);
1708 unit_base_warned = 1;
1711 ts->continue_on_error = td->o.continue_on_error;
1712 ts->total_err_count += td->total_err_count;
1713 ts->first_error = td->first_error;
1715 if (!td->error && td->o.continue_on_error &&
1717 ts->error = td->first_error;
1718 ts->verror[sizeof(ts->verror) - 1] = '\0';
1719 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1720 } else if (td->error) {
1721 ts->error = td->error;
1722 ts->verror[sizeof(ts->verror) - 1] = '\0';
1723 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1727 ts->latency_depth = td->latency_qd;
1728 ts->latency_target = td->o.latency_target;
1729 ts->latency_percentile = td->o.latency_percentile;
1730 ts->latency_window = td->o.latency_window;
1732 ts->nr_block_infos = td->ts.nr_block_infos;
1733 for (k = 0; k < ts->nr_block_infos; k++)
1734 ts->block_infos[k] = td->ts.block_infos[k];
1736 sum_thread_stats(ts, &td->ts, idx == 1);
1739 ts->ss_state = td->ss.state;
1740 ts->ss_dur = td->ss.dur;
1741 ts->ss_head = td->ss.head;
1742 ts->ss_bw_data = td->ss.bw_data;
1743 ts->ss_iops_data = td->ss.iops_data;
1744 ts->ss_limit.u.f = td->ss.limit;
1745 ts->ss_slope.u.f = td->ss.slope;
1746 ts->ss_deviation.u.f = td->ss.deviation;
1747 ts->ss_criterion.u.f = td->ss.criterion;
1750 ts->ss_dur = ts->ss_state = 0;
1753 for (i = 0; i < nr_ts; i++) {
1754 unsigned long long bw;
1756 ts = &threadstats[i];
1757 if (ts->groupid == -1)
1759 rs = &runstats[ts->groupid];
1760 rs->kb_base = ts->kb_base;
1761 rs->unit_base = ts->unit_base;
1762 rs->unified_rw_rep += ts->unified_rw_rep;
1764 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1765 if (!ts->runtime[j])
1767 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1768 rs->min_run[j] = ts->runtime[j];
1769 if (ts->runtime[j] > rs->max_run[j])
1770 rs->max_run[j] = ts->runtime[j];
1774 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1775 if (bw < rs->min_bw[j])
1777 if (bw > rs->max_bw[j])
1780 rs->iobytes[j] += ts->io_bytes[j];
1784 for (i = 0; i < groupid + 1; i++) {
1789 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1790 if (rs->max_run[ddir])
1791 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1796 for (i = 0; i < FIO_OUTPUT_NR; i++)
1797 buf_output_init(&output[i]);
1800 * don't overwrite last signal output
1802 if (output_format & FIO_OUTPUT_NORMAL)
1803 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1804 if (output_format & FIO_OUTPUT_JSON) {
1805 struct thread_data *global;
1808 unsigned long long ms_since_epoch;
1810 gettimeofday(&now, NULL);
1811 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1812 (unsigned long long)(now.tv_usec) / 1000;
1814 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1816 if (time_buf[strlen(time_buf) - 1] == '\n')
1817 time_buf[strlen(time_buf) - 1] = '\0';
1819 root = json_create_object();
1820 json_object_add_value_string(root, "fio version", fio_version_string);
1821 json_object_add_value_int(root, "timestamp", now.tv_sec);
1822 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1823 json_object_add_value_string(root, "time", time_buf);
1824 global = get_global_options();
1825 json_add_job_opts(root, "global options", &global->opt_list, false);
1826 array = json_create_array();
1827 json_object_add_value_array(root, "jobs", array);
1831 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1833 for (i = 0; i < nr_ts; i++) {
1834 ts = &threadstats[i];
1835 rs = &runstats[ts->groupid];
1838 fio_server_send_job_options(opt_lists[i], i);
1839 fio_server_send_ts(ts, rs);
1841 if (output_format & FIO_OUTPUT_TERSE)
1842 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1843 if (output_format & FIO_OUTPUT_JSON) {
1844 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1845 json_array_add_value_object(array, tmp);
1847 if (output_format & FIO_OUTPUT_NORMAL)
1848 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1851 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1852 /* disk util stats, if any */
1853 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1855 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1857 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1858 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1859 json_free_object(root);
1862 for (i = 0; i < groupid + 1; i++) {
1867 fio_server_send_gs(rs);
1868 else if (output_format & FIO_OUTPUT_NORMAL)
1869 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1873 fio_server_send_du();
1874 else if (output_format & FIO_OUTPUT_NORMAL) {
1875 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1876 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1879 for (i = 0; i < FIO_OUTPUT_NR; i++) {
1880 struct buf_output *out = &output[i];
1882 log_info_buf(out->buf, out->buflen);
1883 buf_output_free(out);
1892 void show_run_stats(void)
1894 fio_mutex_down(stat_mutex);
1896 fio_mutex_up(stat_mutex);
1899 void __show_running_run_stats(void)
1901 struct thread_data *td;
1902 unsigned long long *rt;
1906 fio_mutex_down(stat_mutex);
1908 rt = malloc(thread_number * sizeof(unsigned long long));
1909 fio_gettime(&ts, NULL);
1911 for_each_td(td, i) {
1912 td->update_rusage = 1;
1913 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1914 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1915 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1916 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1918 rt[i] = mtime_since(&td->start, &ts);
1919 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1920 td->ts.runtime[DDIR_READ] += rt[i];
1921 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1922 td->ts.runtime[DDIR_WRITE] += rt[i];
1923 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1924 td->ts.runtime[DDIR_TRIM] += rt[i];
1927 for_each_td(td, i) {
1928 if (td->runstate >= TD_EXITED)
1930 if (td->rusage_sem) {
1931 td->update_rusage = 1;
1932 fio_mutex_down(td->rusage_sem);
1934 td->update_rusage = 0;
1939 for_each_td(td, i) {
1940 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1941 td->ts.runtime[DDIR_READ] -= rt[i];
1942 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1943 td->ts.runtime[DDIR_WRITE] -= rt[i];
1944 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1945 td->ts.runtime[DDIR_TRIM] -= rt[i];
1949 fio_mutex_up(stat_mutex);
1952 static int status_interval_init;
1953 static struct timespec status_time;
1954 static int status_file_disabled;
1956 #define FIO_STATUS_FILE "fio-dump-status"
1958 static int check_status_file(void)
1961 const char *temp_dir;
1962 char fio_status_file_path[PATH_MAX];
1964 if (status_file_disabled)
1967 temp_dir = getenv("TMPDIR");
1968 if (temp_dir == NULL) {
1969 temp_dir = getenv("TEMP");
1970 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1973 if (temp_dir == NULL)
1976 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1978 if (stat(fio_status_file_path, &sb))
1981 if (unlink(fio_status_file_path) < 0) {
1982 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1984 log_err("fio: disabling status file updates\n");
1985 status_file_disabled = 1;
1991 void check_for_running_stats(void)
1993 if (status_interval) {
1994 if (!status_interval_init) {
1995 fio_gettime(&status_time, NULL);
1996 status_interval_init = 1;
1997 } else if (mtime_since_now(&status_time) >= status_interval) {
1998 show_running_run_stats();
1999 fio_gettime(&status_time, NULL);
2003 if (check_status_file()) {
2004 show_running_run_stats();
2009 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2014 if (data > is->max_val)
2016 if (data < is->min_val)
2019 delta = val - is->mean.u.f;
2021 is->mean.u.f += delta / (is->samples + 1.0);
2022 is->S.u.f += delta * (val - is->mean.u.f);
2029 * Return a struct io_logs, which is added to the tail of the log
2032 static struct io_logs *get_new_log(struct io_log *iolog)
2034 size_t new_size, new_samples;
2035 struct io_logs *cur_log;
2038 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2041 if (!iolog->cur_log_max)
2042 new_samples = DEF_LOG_ENTRIES;
2044 new_samples = iolog->cur_log_max * 2;
2045 if (new_samples > MAX_LOG_ENTRIES)
2046 new_samples = MAX_LOG_ENTRIES;
2049 new_size = new_samples * log_entry_sz(iolog);
2051 cur_log = smalloc(sizeof(*cur_log));
2053 INIT_FLIST_HEAD(&cur_log->list);
2054 cur_log->log = malloc(new_size);
2056 cur_log->nr_samples = 0;
2057 cur_log->max_samples = new_samples;
2058 flist_add_tail(&cur_log->list, &iolog->io_logs);
2059 iolog->cur_log_max = new_samples;
2069 * Add and return a new log chunk, or return current log if big enough
2071 static struct io_logs *regrow_log(struct io_log *iolog)
2073 struct io_logs *cur_log;
2076 if (!iolog || iolog->disabled)
2079 cur_log = iolog_cur_log(iolog);
2081 cur_log = get_new_log(iolog);
2086 if (cur_log->nr_samples < cur_log->max_samples)
2090 * No room for a new sample. If we're compressing on the fly, flush
2091 * out the current chunk
2093 if (iolog->log_gz) {
2094 if (iolog_cur_flush(iolog, cur_log)) {
2095 log_err("fio: failed flushing iolog! Will stop logging.\n");
2101 * Get a new log array, and add to our list
2103 cur_log = get_new_log(iolog);
2105 log_err("fio: failed extending iolog! Will stop logging.\n");
2109 if (!iolog->pending || !iolog->pending->nr_samples)
2113 * Flush pending items to new log
2115 for (i = 0; i < iolog->pending->nr_samples; i++) {
2116 struct io_sample *src, *dst;
2118 src = get_sample(iolog, iolog->pending, i);
2119 dst = get_sample(iolog, cur_log, i);
2120 memcpy(dst, src, log_entry_sz(iolog));
2122 cur_log->nr_samples = iolog->pending->nr_samples;
2124 iolog->pending->nr_samples = 0;
2128 iolog->disabled = true;
2132 void regrow_logs(struct thread_data *td)
2134 regrow_log(td->slat_log);
2135 regrow_log(td->clat_log);
2136 regrow_log(td->clat_hist_log);
2137 regrow_log(td->lat_log);
2138 regrow_log(td->bw_log);
2139 regrow_log(td->iops_log);
2140 td->flags &= ~TD_F_REGROW_LOGS;
2143 static struct io_logs *get_cur_log(struct io_log *iolog)
2145 struct io_logs *cur_log;
2147 cur_log = iolog_cur_log(iolog);
2149 cur_log = get_new_log(iolog);
2154 if (cur_log->nr_samples < cur_log->max_samples)
2158 * Out of space. If we're in IO offload mode, or we're not doing
2159 * per unit logging (hence logging happens outside of the IO thread
2160 * as well), add a new log chunk inline. If we're doing inline
2161 * submissions, flag 'td' as needing a log regrow and we'll take
2162 * care of it on the submission side.
2164 if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2165 !per_unit_log(iolog))
2166 return regrow_log(iolog);
2168 iolog->td->flags |= TD_F_REGROW_LOGS;
2169 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2170 return iolog->pending;
2173 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2174 enum fio_ddir ddir, unsigned int bs,
2175 unsigned long t, uint64_t offset)
2177 struct io_logs *cur_log;
2179 if (iolog->disabled)
2181 if (flist_empty(&iolog->io_logs))
2182 iolog->avg_last = t;
2184 cur_log = get_cur_log(iolog);
2186 struct io_sample *s;
2188 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2191 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2192 io_sample_set_ddir(iolog, s, ddir);
2195 if (iolog->log_offset) {
2196 struct io_sample_offset *so = (void *) s;
2198 so->offset = offset;
2201 cur_log->nr_samples++;
2205 iolog->disabled = true;
2208 static inline void reset_io_stat(struct io_stat *ios)
2210 ios->max_val = ios->min_val = ios->samples = 0;
2211 ios->mean.u.f = ios->S.u.f = 0;
2214 void reset_io_stats(struct thread_data *td)
2216 struct thread_stat *ts = &td->ts;
2219 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2220 reset_io_stat(&ts->clat_stat[i]);
2221 reset_io_stat(&ts->slat_stat[i]);
2222 reset_io_stat(&ts->lat_stat[i]);
2223 reset_io_stat(&ts->bw_stat[i]);
2224 reset_io_stat(&ts->iops_stat[i]);
2226 ts->io_bytes[i] = 0;
2228 ts->total_io_u[i] = 0;
2229 ts->short_io_u[i] = 0;
2230 ts->drop_io_u[i] = 0;
2232 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2233 ts->io_u_plat[i][j] = 0;
2236 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2237 ts->io_u_map[i] = 0;
2238 ts->io_u_submit[i] = 0;
2239 ts->io_u_complete[i] = 0;
2242 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2243 ts->io_u_lat_n[i] = 0;
2244 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2245 ts->io_u_lat_u[i] = 0;
2246 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2247 ts->io_u_lat_m[i] = 0;
2249 ts->total_submit = 0;
2250 ts->total_complete = 0;
2253 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2254 unsigned long elapsed, bool log_max)
2257 * Note an entry in the log. Use the mean from the logged samples,
2258 * making sure to properly round up. Only write a log entry if we
2259 * had actual samples done.
2261 if (iolog->avg_window[ddir].samples) {
2262 union io_sample_data data;
2265 data.val = iolog->avg_window[ddir].max_val;
2267 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2269 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2272 reset_io_stat(&iolog->avg_window[ddir]);
2275 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2280 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2281 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2284 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2285 union io_sample_data data, enum fio_ddir ddir,
2286 unsigned int bs, uint64_t offset)
2288 unsigned long elapsed, this_window;
2293 elapsed = mtime_since_now(&td->epoch);
2296 * If no time averaging, just add the log sample.
2298 if (!iolog->avg_msec) {
2299 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2304 * Add the sample. If the time period has passed, then
2305 * add that entry to the log and clear.
2307 add_stat_sample(&iolog->avg_window[ddir], data.val);
2310 * If period hasn't passed, adding the above sample is all we
2313 this_window = elapsed - iolog->avg_last;
2314 if (elapsed < iolog->avg_last)
2315 return iolog->avg_last - elapsed;
2316 else if (this_window < iolog->avg_msec) {
2317 int diff = iolog->avg_msec - this_window;
2319 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2323 _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2325 iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2326 return iolog->avg_msec;
2329 void finalize_logs(struct thread_data *td, bool unit_logs)
2331 unsigned long elapsed;
2333 elapsed = mtime_since_now(&td->epoch);
2335 if (td->clat_log && unit_logs)
2336 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2337 if (td->slat_log && unit_logs)
2338 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2339 if (td->lat_log && unit_logs)
2340 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2341 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2342 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2343 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2344 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2347 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2349 struct io_log *iolog;
2354 iolog = agg_io_log[ddir];
2355 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2358 static void add_clat_percentile_sample(struct thread_stat *ts,
2359 unsigned long long nsec, enum fio_ddir ddir)
2361 unsigned int idx = plat_val_to_idx(nsec);
2362 assert(idx < FIO_IO_U_PLAT_NR);
2364 ts->io_u_plat[ddir][idx]++;
2367 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2368 unsigned long long nsec, unsigned int bs, uint64_t offset)
2370 unsigned long elapsed, this_window;
2371 struct thread_stat *ts = &td->ts;
2372 struct io_log *iolog = td->clat_hist_log;
2376 add_stat_sample(&ts->clat_stat[ddir], nsec);
2379 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2382 if (ts->clat_percentiles)
2383 add_clat_percentile_sample(ts, nsec, ddir);
2385 if (iolog && iolog->hist_msec) {
2386 struct io_hist *hw = &iolog->hist_window[ddir];
2389 elapsed = mtime_since_now(&td->epoch);
2391 hw->hist_last = elapsed;
2392 this_window = elapsed - hw->hist_last;
2394 if (this_window >= iolog->hist_msec) {
2395 unsigned int *io_u_plat;
2396 struct io_u_plat_entry *dst;
2399 * Make a byte-for-byte copy of the latency histogram
2400 * stored in td->ts.io_u_plat[ddir], recording it in a
2401 * log sample. Note that the matching call to free() is
2402 * located in iolog.c after printing this sample to the
2405 io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2406 dst = malloc(sizeof(struct io_u_plat_entry));
2407 memcpy(&(dst->io_u_plat), io_u_plat,
2408 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2409 flist_add(&dst->list, &hw->list);
2410 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2414 * Update the last time we recorded as being now, minus
2415 * any drift in time we encountered before actually
2416 * making the record.
2418 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2426 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2427 unsigned long usec, unsigned int bs, uint64_t offset)
2429 struct thread_stat *ts = &td->ts;
2436 add_stat_sample(&ts->slat_stat[ddir], usec);
2439 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2444 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2445 unsigned long long nsec, unsigned int bs, uint64_t offset)
2447 struct thread_stat *ts = &td->ts;
2454 add_stat_sample(&ts->lat_stat[ddir], nsec);
2457 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2463 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2464 unsigned int bytes, unsigned long long spent)
2466 struct thread_stat *ts = &td->ts;
2470 rate = (unsigned long) (bytes * 1000000ULL / spent);
2476 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2479 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2480 bytes, io_u->offset);
2482 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2486 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2487 struct timespec *t, unsigned int avg_time,
2488 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2489 struct io_stat *stat, struct io_log *log,
2492 unsigned long spent, rate;
2494 unsigned int next, next_log;
2496 next_log = avg_time;
2498 spent = mtime_since(parent_tv, t);
2499 if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2500 return avg_time - spent;
2505 * Compute both read and write rates for the interval.
2507 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2510 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2512 continue; /* No entries for interval */
2516 rate = delta * 1000 / spent / 1024; /* KiB/s */
2518 rate = (delta * 1000) / spent;
2522 add_stat_sample(&stat[ddir], rate);
2525 unsigned int bs = 0;
2527 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2528 bs = td->o.min_bs[ddir];
2530 next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2531 next_log = min(next_log, next);
2534 stat_io_bytes[ddir] = this_io_bytes[ddir];
2537 timespec_add_msec(parent_tv, avg_time);
2541 if (spent <= avg_time)
2544 next = avg_time - (1 + spent - avg_time);
2546 return min(next, next_log);
2549 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2551 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2552 td->this_io_bytes, td->stat_io_bytes,
2553 td->ts.bw_stat, td->bw_log, true);
2556 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2559 struct thread_stat *ts = &td->ts;
2563 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2566 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2567 bytes, io_u->offset);
2569 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2573 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2575 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2576 td->this_io_blocks, td->stat_io_blocks,
2577 td->ts.iops_stat, td->iops_log, false);
2581 * Returns msecs to next event
2583 int calc_log_samples(void)
2585 struct thread_data *td;
2586 unsigned int next = ~0U, tmp;
2587 struct timespec now;
2590 fio_gettime(&now, NULL);
2592 for_each_td(td, i) {
2595 if (in_ramp_time(td) ||
2596 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2597 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2601 (td->bw_log && !per_unit_log(td->bw_log))) {
2602 tmp = add_bw_samples(td, &now);
2606 if (!td->iops_log ||
2607 (td->iops_log && !per_unit_log(td->iops_log))) {
2608 tmp = add_iops_samples(td, &now);
2614 return next == ~0U ? 0 : next;
2617 void stat_init(void)
2619 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2622 void stat_exit(void)
2625 * When we have the mutex, we know out-of-band access to it
2628 fio_mutex_down(stat_mutex);
2629 fio_mutex_remove(stat_mutex);
2633 * Called from signal handler. Wake up status thread.
2635 void show_running_run_stats(void)
2640 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2642 /* Ignore io_u's which span multiple blocks--they will just get
2643 * inaccurate counts. */
2644 int idx = (io_u->offset - io_u->file->file_offset)
2645 / td->o.bs[DDIR_TRIM];
2646 uint32_t *info = &td->ts.block_infos[idx];
2647 assert(idx < td->ts.nr_block_infos);