9 #include "lib/ieee754.h"
11 #include "lib/getrusage.h"
14 #include "lib/output_buffer.h"
15 #include "helper_thread.h"
18 #define LOG_MSEC_SLACK 1
20 struct fio_sem *stat_sem;
22 void clear_rusage_stat(struct thread_data *td)
24 struct thread_stat *ts = &td->ts;
26 fio_getrusage(&td->ru_start);
27 ts->usr_time = ts->sys_time = 0;
29 ts->minf = ts->majf = 0;
32 void update_rusage_stat(struct thread_data *td)
34 struct thread_stat *ts = &td->ts;
36 fio_getrusage(&td->ru_end);
37 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
38 &td->ru_end.ru_utime);
39 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
40 &td->ru_end.ru_stime);
41 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
42 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
43 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
44 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
46 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
50 * Given a latency, return the index of the corresponding bucket in
51 * the structure tracking percentiles.
53 * (1) find the group (and error bits) that the value (latency)
54 * belongs to by looking at its MSB. (2) find the bucket number in the
55 * group by looking at the index bits.
58 static unsigned int plat_val_to_idx(unsigned long long val)
60 unsigned int msb, error_bits, base, offset, idx;
62 /* Find MSB starting from bit 0 */
66 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
69 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
70 * all bits of the sample as index
72 if (msb <= FIO_IO_U_PLAT_BITS)
75 /* Compute the number of error bits to discard*/
76 error_bits = msb - FIO_IO_U_PLAT_BITS;
78 /* Compute the number of buckets before the group */
79 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
82 * Discard the error bits and apply the mask to find the
83 * index for the buckets in the group
85 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
87 /* Make sure the index does not exceed (array size - 1) */
88 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
89 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
95 * Convert the given index of the bucket array to the value
96 * represented by the bucket
98 static unsigned long long plat_idx_to_val(unsigned int idx)
100 unsigned int error_bits;
101 unsigned long long k, base;
103 assert(idx < FIO_IO_U_PLAT_NR);
105 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
106 * all bits of the sample as index */
107 if (idx < (FIO_IO_U_PLAT_VAL << 1))
110 /* Find the group and compute the minimum value of that group */
111 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
112 base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
114 /* Find its bucket number of the group */
115 k = idx % FIO_IO_U_PLAT_VAL;
117 /* Return the mean of the range of the bucket */
118 return base + ((k + 0.5) * (1 << error_bits));
121 static int double_cmp(const void *a, const void *b)
123 const fio_fp64_t fa = *(const fio_fp64_t *) a;
124 const fio_fp64_t fb = *(const fio_fp64_t *) b;
129 else if (fa.u.f < fb.u.f)
135 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
136 fio_fp64_t *plist, unsigned long long **output,
137 unsigned long long *maxv, unsigned long long *minv)
139 unsigned long long sum = 0;
140 unsigned int len, i, j = 0;
141 unsigned int oval_len = 0;
142 unsigned long long *ovals = NULL;
149 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
156 * Sort the percentile list. Note that it may already be sorted if
157 * we are using the default values, but since it's a short list this
158 * isn't a worry. Also note that this does not work for NaN values.
161 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
164 * Calculate bucket values, note down max and min values
167 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
169 while (sum >= (plist[j].u.f / 100.0 * nr)) {
170 assert(plist[j].u.f <= 100.0);
174 ovals = realloc(ovals, oval_len * sizeof(*ovals));
177 ovals[j] = plat_idx_to_val(i);
178 if (ovals[j] < *minv)
180 if (ovals[j] > *maxv)
183 is_last = (j == len - 1) != 0;
196 * Find and display the p-th percentile of clat
198 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
199 fio_fp64_t *plist, unsigned int precision,
200 const char *pre, struct buf_output *out)
202 unsigned int divisor, len, i, j = 0;
203 unsigned long long minv, maxv;
204 unsigned long long *ovals;
205 int per_line, scale_down, time_width;
209 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
214 * We default to nsecs, but if the value range is such that we
215 * should scale down to usecs or msecs, do that.
217 if (minv > 2000000 && maxv > 99999999ULL) {
220 log_buf(out, " %s percentiles (msec):\n |", pre);
221 } else if (minv > 2000 && maxv > 99999) {
224 log_buf(out, " %s percentiles (usec):\n |", pre);
228 log_buf(out, " %s percentiles (nsec):\n |", pre);
232 time_width = max(5, (int) (log10(maxv / divisor) + 1));
233 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
234 precision, time_width);
235 /* fmt will be something like " %5.2fth=[%4llu]%c" */
236 per_line = (80 - 7) / (precision + 10 + time_width);
238 for (j = 0; j < len; j++) {
240 if (j != 0 && (j % per_line) == 0)
243 /* end of the list */
244 is_last = (j == len - 1) != 0;
246 for (i = 0; i < scale_down; i++)
247 ovals[j] = (ovals[j] + 999) / 1000;
249 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
254 if ((j % per_line) == per_line - 1) /* for formatting */
263 bool calc_lat(struct io_stat *is, unsigned long long *min,
264 unsigned long long *max, double *mean, double *dev)
266 double n = (double) is->samples;
273 *mean = is->mean.u.f;
276 *dev = sqrt(is->S.u.f / (n - 1.0));
283 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
285 char *io, *agg, *min, *max;
286 char *ioalt, *aggalt, *minalt, *maxalt;
287 const char *str[] = { " READ", " WRITE" , " TRIM"};
290 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
292 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
293 const int i2p = is_power_of_2(rs->kb_base);
298 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
299 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
300 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
301 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
302 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
303 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
304 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
305 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
306 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
307 rs->unified_rw_rep ? " MIXED" : str[i],
308 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
309 (unsigned long long) rs->min_run[i],
310 (unsigned long long) rs->max_run[i]);
323 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
328 * Do depth distribution calculations
330 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
332 io_u_dist[i] = (double) map[i] / (double) total;
333 io_u_dist[i] *= 100.0;
334 if (io_u_dist[i] < 0.1 && map[i])
341 static void stat_calc_lat(struct thread_stat *ts, double *dst,
342 uint64_t *src, int nr)
344 unsigned long total = ddir_rw_sum(ts->total_io_u);
348 * Do latency distribution calculations
350 for (i = 0; i < nr; i++) {
352 dst[i] = (double) src[i] / (double) total;
354 if (dst[i] < 0.01 && src[i])
362 * To keep the terse format unaltered, add all of the ns latency
363 * buckets to the first us latency bucket
365 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
367 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
370 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
372 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
373 ntotal += ts->io_u_lat_n[i];
375 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
378 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
380 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
383 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
385 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
388 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
390 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
393 static void display_lat(const char *name, unsigned long long min,
394 unsigned long long max, double mean, double dev,
395 struct buf_output *out)
397 const char *base = "(nsec)";
400 if (nsec_to_msec(&min, &max, &mean, &dev))
402 else if (nsec_to_usec(&min, &max, &mean, &dev))
405 minp = num2str(min, 6, 1, 0, N2S_NONE);
406 maxp = num2str(max, 6, 1, 0, N2S_NONE);
408 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
409 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
415 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
416 int ddir, struct buf_output *out)
418 const char *str[] = { " read", "write", " trim", "sync" };
420 unsigned long long min, max, bw, iops;
422 char *io_p, *bw_p, *bw_p_alt, *iops_p;
425 if (ddir_sync(ddir)) {
426 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
427 log_buf(out, " %s:\n", "fsync/fdatasync/sync_file_range");
428 display_lat(str[ddir], min, max, mean, dev, out);
429 show_clat_percentiles(ts->io_u_sync_plat,
430 ts->sync_stat.samples,
432 ts->percentile_precision,
438 assert(ddir_rw(ddir));
440 if (!ts->runtime[ddir])
443 i2p = is_power_of_2(rs->kb_base);
444 runt = ts->runtime[ddir];
446 bw = (1000 * ts->io_bytes[ddir]) / runt;
447 io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
448 bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
449 bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
451 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
452 iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
454 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
455 rs->unified_rw_rep ? "mixed" : str[ddir],
456 iops_p, bw_p, bw_p_alt, io_p,
457 (unsigned long long) ts->runtime[ddir]);
464 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
465 display_lat("slat", min, max, mean, dev, out);
466 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
467 display_lat("clat", min, max, mean, dev, out);
468 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
469 display_lat(" lat", min, max, mean, dev, out);
471 if (ts->clat_percentiles || ts->lat_percentiles) {
472 const char *name = ts->clat_percentiles ? "clat" : " lat";
475 if (ts->clat_percentiles)
476 samples = ts->clat_stat[ddir].samples;
478 samples = ts->lat_stat[ddir].samples;
480 show_clat_percentiles(ts->io_u_plat[ddir],
483 ts->percentile_precision, name, out);
485 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
486 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
489 if ((rs->unit_base == 1) && i2p)
491 else if (rs->unit_base == 1)
499 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
500 if (p_of_agg > 100.0)
504 if (rs->unit_base == 1) {
511 if (mean > fkb_base * fkb_base) {
516 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
519 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
520 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
521 bw_str, min, max, p_of_agg, mean, dev,
522 (&ts->bw_stat[ddir])->samples);
524 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
525 log_buf(out, " iops : min=%5llu, max=%5llu, "
526 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
527 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
531 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
532 const char *msg, struct buf_output *out)
534 bool new_line = true, shown = false;
537 for (i = 0; i < nr; i++) {
538 if (io_u_lat[i] <= 0.0)
544 log_buf(out, " lat (%s) : ", msg);
550 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
562 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
564 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
565 "250=", "500=", "750=", "1000=", };
567 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
570 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
572 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
573 "250=", "500=", "750=", "1000=", };
575 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
578 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
580 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
581 "250=", "500=", "750=", "1000=", "2000=",
584 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
587 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
589 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
590 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
591 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
593 stat_calc_lat_n(ts, io_u_lat_n);
594 stat_calc_lat_u(ts, io_u_lat_u);
595 stat_calc_lat_m(ts, io_u_lat_m);
597 show_lat_n(io_u_lat_n, out);
598 show_lat_u(io_u_lat_u, out);
599 show_lat_m(io_u_lat_m, out);
602 static int block_state_category(int block_state)
604 switch (block_state) {
605 case BLOCK_STATE_UNINIT:
607 case BLOCK_STATE_TRIMMED:
608 case BLOCK_STATE_WRITTEN:
610 case BLOCK_STATE_WRITE_FAILURE:
611 case BLOCK_STATE_TRIM_FAILURE:
614 /* Silence compile warning on some BSDs and have a return */
620 static int compare_block_infos(const void *bs1, const void *bs2)
622 uint32_t block1 = *(uint32_t *)bs1;
623 uint32_t block2 = *(uint32_t *)bs2;
624 int state1 = BLOCK_INFO_STATE(block1);
625 int state2 = BLOCK_INFO_STATE(block2);
626 int bscat1 = block_state_category(state1);
627 int bscat2 = block_state_category(state2);
628 int cycles1 = BLOCK_INFO_TRIMS(block1);
629 int cycles2 = BLOCK_INFO_TRIMS(block2);
636 if (cycles1 < cycles2)
638 if (cycles1 > cycles2)
646 assert(block1 == block2);
650 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
651 fio_fp64_t *plist, unsigned int **percentiles,
657 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
659 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
666 * Sort the percentile list. Note that it may already be sorted if
667 * we are using the default values, but since it's a short list this
668 * isn't a worry. Also note that this does not work for NaN values.
671 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
673 /* Start only after the uninit entries end */
675 nr_uninit < nr_block_infos
676 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
680 if (nr_uninit == nr_block_infos)
683 *percentiles = calloc(len, sizeof(**percentiles));
685 for (i = 0; i < len; i++) {
686 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
688 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
691 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
692 for (i = 0; i < nr_block_infos; i++)
693 types[BLOCK_INFO_STATE(block_infos[i])]++;
698 static const char *block_state_names[] = {
699 [BLOCK_STATE_UNINIT] = "unwritten",
700 [BLOCK_STATE_TRIMMED] = "trimmed",
701 [BLOCK_STATE_WRITTEN] = "written",
702 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
703 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
706 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
707 fio_fp64_t *plist, struct buf_output *out)
710 unsigned int *percentiles = NULL;
711 unsigned int block_state_counts[BLOCK_STATE_COUNT];
713 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
714 &percentiles, block_state_counts);
716 log_buf(out, " block lifetime percentiles :\n |");
718 for (i = 0; i < len; i++) {
719 uint32_t block_info = percentiles[i];
720 #define LINE_LENGTH 75
721 char str[LINE_LENGTH];
722 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
723 plist[i].u.f, block_info,
724 i == len - 1 ? '\n' : ',');
725 assert(strln < LINE_LENGTH);
726 if (pos + strln > LINE_LENGTH) {
728 log_buf(out, "\n |");
730 log_buf(out, "%s", str);
737 log_buf(out, " states :");
738 for (i = 0; i < BLOCK_STATE_COUNT; i++)
739 log_buf(out, " %s=%u%c",
740 block_state_names[i], block_state_counts[i],
741 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
744 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
746 char *p1, *p1alt, *p2;
747 unsigned long long bw_mean, iops_mean;
748 const int i2p = is_power_of_2(ts->kb_base);
753 bw_mean = steadystate_bw_mean(ts);
754 iops_mean = steadystate_iops_mean(ts);
756 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
757 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
758 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
760 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
761 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
763 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
764 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
765 ts->ss_criterion.u.f,
766 ts->ss_state & FIO_SS_PCT ? "%" : "");
773 static void show_thread_status_normal(struct thread_stat *ts,
774 struct group_run_stats *rs,
775 struct buf_output *out)
777 double usr_cpu, sys_cpu;
778 unsigned long runtime;
779 double io_u_dist[FIO_IO_U_MAP_NR];
783 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
786 memset(time_buf, 0, sizeof(time_buf));
789 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
792 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
793 ts->name, ts->groupid, ts->members,
794 ts->error, (int) ts->pid, time_buf);
796 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
797 ts->name, ts->groupid, ts->members,
798 ts->error, ts->verror, (int) ts->pid,
802 if (strlen(ts->description))
803 log_buf(out, " Description : [%s]\n", ts->description);
805 if (ts->io_bytes[DDIR_READ])
806 show_ddir_status(rs, ts, DDIR_READ, out);
807 if (ts->io_bytes[DDIR_WRITE])
808 show_ddir_status(rs, ts, DDIR_WRITE, out);
809 if (ts->io_bytes[DDIR_TRIM])
810 show_ddir_status(rs, ts, DDIR_TRIM, out);
812 show_latencies(ts, out);
814 if (ts->sync_stat.samples)
815 show_ddir_status(rs, ts, DDIR_SYNC, out);
817 runtime = ts->total_run_time;
819 double runt = (double) runtime;
821 usr_cpu = (double) ts->usr_time * 100 / runt;
822 sys_cpu = (double) ts->sys_time * 100 / runt;
828 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
829 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
830 (unsigned long long) ts->ctx,
831 (unsigned long long) ts->majf,
832 (unsigned long long) ts->minf);
834 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
835 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
836 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
837 io_u_dist[1], io_u_dist[2],
838 io_u_dist[3], io_u_dist[4],
839 io_u_dist[5], io_u_dist[6]);
841 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
842 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
843 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
844 io_u_dist[1], io_u_dist[2],
845 io_u_dist[3], io_u_dist[4],
846 io_u_dist[5], io_u_dist[6]);
847 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
848 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
849 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
850 io_u_dist[1], io_u_dist[2],
851 io_u_dist[3], io_u_dist[4],
852 io_u_dist[5], io_u_dist[6]);
853 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
854 " short=%llu,%llu,%llu,0"
855 " dropped=%llu,%llu,%llu,0\n",
856 (unsigned long long) ts->total_io_u[0],
857 (unsigned long long) ts->total_io_u[1],
858 (unsigned long long) ts->total_io_u[2],
859 (unsigned long long) ts->total_io_u[3],
860 (unsigned long long) ts->short_io_u[0],
861 (unsigned long long) ts->short_io_u[1],
862 (unsigned long long) ts->short_io_u[2],
863 (unsigned long long) ts->drop_io_u[0],
864 (unsigned long long) ts->drop_io_u[1],
865 (unsigned long long) ts->drop_io_u[2]);
866 if (ts->continue_on_error) {
867 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
868 (unsigned long long)ts->total_err_count,
870 strerror(ts->first_error));
872 if (ts->latency_depth) {
873 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
874 (unsigned long long)ts->latency_target,
875 (unsigned long long)ts->latency_window,
876 ts->latency_percentile.u.f,
880 if (ts->nr_block_infos)
881 show_block_infos(ts->nr_block_infos, ts->block_infos,
882 ts->percentile_list, out);
885 show_ss_normal(ts, out);
888 static void show_ddir_status_terse(struct thread_stat *ts,
889 struct group_run_stats *rs, int ddir,
890 int ver, struct buf_output *out)
892 unsigned long long min, max, minv, maxv, bw, iops;
893 unsigned long long *ovals = NULL;
898 assert(ddir_rw(ddir));
901 if (ts->runtime[ddir]) {
902 uint64_t runt = ts->runtime[ddir];
904 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
905 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
908 log_buf(out, ";%llu;%llu;%llu;%llu",
909 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
910 (unsigned long long) ts->runtime[ddir]);
912 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
913 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
915 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
917 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
918 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
920 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
922 if (ts->clat_percentiles || ts->lat_percentiles) {
923 len = calc_clat_percentiles(ts->io_u_plat[ddir],
924 ts->clat_stat[ddir].samples,
925 ts->percentile_list, &ovals, &maxv,
930 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
932 log_buf(out, ";0%%=0");
935 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
938 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
939 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
941 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
946 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
948 double p_of_agg = 100.0;
951 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
952 if (p_of_agg > 100.0)
956 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
958 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
962 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
964 log_buf(out, ";%lu", 0UL);
966 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
967 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
968 mean, dev, (&ts->iops_stat[ddir])->samples);
970 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
974 static void add_ddir_status_json(struct thread_stat *ts,
975 struct group_run_stats *rs, int ddir, struct json_object *parent)
977 unsigned long long min, max, minv, maxv;
978 unsigned long long bw_bytes, bw;
979 unsigned long long *ovals = NULL;
980 double mean, dev, iops;
983 const char *ddirname[] = { "read", "write", "trim", "sync" };
984 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object = NULL;
986 double p_of_agg = 100.0;
988 assert(ddir_rw(ddir) || ddir_sync(ddir));
990 if (ts->unified_rw_rep && ddir != DDIR_READ)
993 dir_object = json_create_object();
994 json_object_add_value_object(parent,
995 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
1001 if (ts->runtime[ddir]) {
1002 uint64_t runt = ts->runtime[ddir];
1004 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1005 bw = bw_bytes / 1024; /* KiB/s */
1006 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1009 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1010 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1011 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1012 json_object_add_value_int(dir_object, "bw", bw);
1013 json_object_add_value_float(dir_object, "iops", iops);
1014 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1015 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1016 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1017 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1019 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
1023 tmp_object = json_create_object();
1024 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1025 json_object_add_value_int(tmp_object, "min", min);
1026 json_object_add_value_int(tmp_object, "max", max);
1027 json_object_add_value_float(tmp_object, "mean", mean);
1028 json_object_add_value_float(tmp_object, "stddev", dev);
1030 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1034 tmp_object = json_create_object();
1035 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1036 json_object_add_value_int(tmp_object, "min", min);
1037 json_object_add_value_int(tmp_object, "max", max);
1038 json_object_add_value_float(tmp_object, "mean", mean);
1039 json_object_add_value_float(tmp_object, "stddev", dev);
1041 if (!calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
1046 tmp_object = json_create_object();
1047 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1048 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1049 json_object_add_value_int(tmp_object, "min", min);
1050 json_object_add_value_int(tmp_object, "max", max);
1051 json_object_add_value_float(tmp_object, "mean", mean);
1052 json_object_add_value_float(tmp_object, "stddev", dev);
1055 if (ts->clat_percentiles || ts->lat_percentiles) {
1056 if (ddir_rw(ddir)) {
1057 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1058 ts->clat_stat[ddir].samples,
1059 ts->percentile_list, &ovals, &maxv,
1062 len = calc_clat_percentiles(ts->io_u_sync_plat,
1063 ts->sync_stat.samples,
1064 ts->percentile_list, &ovals, &maxv,
1068 if (len > FIO_IO_U_LIST_MAX_LEN)
1069 len = FIO_IO_U_LIST_MAX_LEN;
1073 percentile_object = json_create_object();
1074 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1075 for (i = 0; i < len; i++) {
1076 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1077 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1080 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1081 clat_bins_object = json_create_object();
1082 if (ts->clat_percentiles)
1083 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1085 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1086 if (ddir_rw(ddir)) {
1087 if (ts->io_u_plat[ddir][i]) {
1088 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1089 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1092 if (ts->io_u_sync_plat[i]) {
1093 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1094 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_sync_plat[i]);
1103 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1107 tmp_object = json_create_object();
1108 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1109 json_object_add_value_int(tmp_object, "min", min);
1110 json_object_add_value_int(tmp_object, "max", max);
1111 json_object_add_value_float(tmp_object, "mean", mean);
1112 json_object_add_value_float(tmp_object, "stddev", dev);
1113 if (output_format & FIO_OUTPUT_JSON_PLUS && ts->lat_percentiles)
1114 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1119 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1120 if (rs->agg[ddir]) {
1121 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1122 if (p_of_agg > 100.0)
1127 p_of_agg = mean = dev = 0.0;
1129 json_object_add_value_int(dir_object, "bw_min", min);
1130 json_object_add_value_int(dir_object, "bw_max", max);
1131 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1132 json_object_add_value_float(dir_object, "bw_mean", mean);
1133 json_object_add_value_float(dir_object, "bw_dev", dev);
1134 json_object_add_value_int(dir_object, "bw_samples",
1135 (&ts->bw_stat[ddir])->samples);
1137 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1141 json_object_add_value_int(dir_object, "iops_min", min);
1142 json_object_add_value_int(dir_object, "iops_max", max);
1143 json_object_add_value_float(dir_object, "iops_mean", mean);
1144 json_object_add_value_float(dir_object, "iops_stddev", dev);
1145 json_object_add_value_int(dir_object, "iops_samples",
1146 (&ts->iops_stat[ddir])->samples);
1149 static void show_thread_status_terse_all(struct thread_stat *ts,
1150 struct group_run_stats *rs, int ver,
1151 struct buf_output *out)
1153 double io_u_dist[FIO_IO_U_MAP_NR];
1154 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1155 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1156 double usr_cpu, sys_cpu;
1161 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1163 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1164 ts->name, ts->groupid, ts->error);
1166 /* Log Read Status */
1167 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1168 /* Log Write Status */
1169 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1170 /* Log Trim Status */
1171 if (ver == 2 || ver == 4 || ver == 5)
1172 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1175 if (ts->total_run_time) {
1176 double runt = (double) ts->total_run_time;
1178 usr_cpu = (double) ts->usr_time * 100 / runt;
1179 sys_cpu = (double) ts->sys_time * 100 / runt;
1185 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1186 (unsigned long long) ts->ctx,
1187 (unsigned long long) ts->majf,
1188 (unsigned long long) ts->minf);
1190 /* Calc % distribution of IO depths, usecond, msecond latency */
1191 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1192 stat_calc_lat_nu(ts, io_u_lat_u);
1193 stat_calc_lat_m(ts, io_u_lat_m);
1195 /* Only show fixed 7 I/O depth levels*/
1196 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1197 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1198 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1200 /* Microsecond latency */
1201 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1202 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1203 /* Millisecond latency */
1204 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1205 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1207 /* disk util stats, if any */
1209 show_disk_util(1, NULL, out);
1211 /* Additional output if continue_on_error set - default off*/
1212 if (ts->continue_on_error)
1213 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1217 /* Additional output if description is set */
1218 if (strlen(ts->description))
1219 log_buf(out, ";%s", ts->description);
1224 static void json_add_job_opts(struct json_object *root, const char *name,
1225 struct flist_head *opt_list)
1227 struct json_object *dir_object;
1228 struct flist_head *entry;
1229 struct print_option *p;
1231 if (flist_empty(opt_list))
1234 dir_object = json_create_object();
1235 json_object_add_value_object(root, name, dir_object);
1237 flist_for_each(entry, opt_list) {
1238 const char *pos = "";
1240 p = flist_entry(entry, struct print_option, list);
1243 json_object_add_value_string(dir_object, p->name, pos);
1247 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1248 struct group_run_stats *rs,
1249 struct flist_head *opt_list)
1251 struct json_object *root, *tmp;
1252 struct jobs_eta *je;
1253 double io_u_dist[FIO_IO_U_MAP_NR];
1254 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1255 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1256 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1257 double usr_cpu, sys_cpu;
1261 root = json_create_object();
1262 json_object_add_value_string(root, "jobname", ts->name);
1263 json_object_add_value_int(root, "groupid", ts->groupid);
1264 json_object_add_value_int(root, "error", ts->error);
1267 je = get_jobs_eta(true, &size);
1269 json_object_add_value_int(root, "eta", je->eta_sec);
1270 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1274 json_add_job_opts(root, "job options", opt_list);
1276 add_ddir_status_json(ts, rs, DDIR_READ, root);
1277 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1278 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1279 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1282 if (ts->total_run_time) {
1283 double runt = (double) ts->total_run_time;
1285 usr_cpu = (double) ts->usr_time * 100 / runt;
1286 sys_cpu = (double) ts->sys_time * 100 / runt;
1291 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1292 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1293 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1294 json_object_add_value_int(root, "ctx", ts->ctx);
1295 json_object_add_value_int(root, "majf", ts->majf);
1296 json_object_add_value_int(root, "minf", ts->minf);
1299 /* Calc % distribution of IO depths, usecond, msecond latency */
1300 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1301 stat_calc_lat_n(ts, io_u_lat_n);
1302 stat_calc_lat_u(ts, io_u_lat_u);
1303 stat_calc_lat_m(ts, io_u_lat_m);
1305 tmp = json_create_object();
1306 json_object_add_value_object(root, "iodepth_level", tmp);
1307 /* Only show fixed 7 I/O depth levels*/
1308 for (i = 0; i < 7; i++) {
1311 snprintf(name, 20, "%d", 1 << i);
1313 snprintf(name, 20, ">=%d", 1 << i);
1314 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1317 /* Nanosecond latency */
1318 tmp = json_create_object();
1319 json_object_add_value_object(root, "latency_ns", tmp);
1320 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1321 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1322 "250", "500", "750", "1000", };
1323 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1325 /* Microsecond latency */
1326 tmp = json_create_object();
1327 json_object_add_value_object(root, "latency_us", tmp);
1328 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1329 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1330 "250", "500", "750", "1000", };
1331 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1333 /* Millisecond latency */
1334 tmp = json_create_object();
1335 json_object_add_value_object(root, "latency_ms", tmp);
1336 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1337 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1338 "250", "500", "750", "1000", "2000",
1340 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1343 /* Additional output if continue_on_error set - default off*/
1344 if (ts->continue_on_error) {
1345 json_object_add_value_int(root, "total_err", ts->total_err_count);
1346 json_object_add_value_int(root, "first_error", ts->first_error);
1349 if (ts->latency_depth) {
1350 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1351 json_object_add_value_int(root, "latency_target", ts->latency_target);
1352 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1353 json_object_add_value_int(root, "latency_window", ts->latency_window);
1356 /* Additional output if description is set */
1357 if (strlen(ts->description))
1358 json_object_add_value_string(root, "desc", ts->description);
1360 if (ts->nr_block_infos) {
1361 /* Block error histogram and types */
1363 unsigned int *percentiles = NULL;
1364 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1366 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1367 ts->percentile_list,
1368 &percentiles, block_state_counts);
1371 struct json_object *block, *percentile_object, *states;
1373 block = json_create_object();
1374 json_object_add_value_object(root, "block", block);
1376 percentile_object = json_create_object();
1377 json_object_add_value_object(block, "percentiles",
1379 for (i = 0; i < len; i++) {
1381 snprintf(buf, sizeof(buf), "%f",
1382 ts->percentile_list[i].u.f);
1383 json_object_add_value_int(percentile_object,
1388 states = json_create_object();
1389 json_object_add_value_object(block, "states", states);
1390 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1391 json_object_add_value_int(states,
1392 block_state_names[state],
1393 block_state_counts[state]);
1400 struct json_object *data;
1401 struct json_array *iops, *bw;
1405 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1406 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1407 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1408 (float) ts->ss_limit.u.f,
1409 ts->ss_state & FIO_SS_PCT ? "%" : "");
1411 tmp = json_create_object();
1412 json_object_add_value_object(root, "steadystate", tmp);
1413 json_object_add_value_string(tmp, "ss", ss_buf);
1414 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1415 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1417 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1418 ts->ss_state & FIO_SS_PCT ? "%" : "");
1419 json_object_add_value_string(tmp, "criterion", ss_buf);
1420 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1421 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1423 data = json_create_object();
1424 json_object_add_value_object(tmp, "data", data);
1425 bw = json_create_array();
1426 iops = json_create_array();
1429 ** if ss was attained or the buffer is not full,
1430 ** ss->head points to the first element in the list.
1431 ** otherwise it actually points to the second element
1434 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1437 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1438 for (l = 0; l < ts->ss_dur; l++) {
1439 k = (j + l) % ts->ss_dur;
1440 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1441 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1443 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1444 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1445 json_object_add_value_array(data, "iops", iops);
1446 json_object_add_value_array(data, "bw", bw);
1452 static void show_thread_status_terse(struct thread_stat *ts,
1453 struct group_run_stats *rs,
1454 struct buf_output *out)
1456 if (terse_version >= 2 && terse_version <= 5)
1457 show_thread_status_terse_all(ts, rs, terse_version, out);
1459 log_err("fio: bad terse version!? %d\n", terse_version);
1462 struct json_object *show_thread_status(struct thread_stat *ts,
1463 struct group_run_stats *rs,
1464 struct flist_head *opt_list,
1465 struct buf_output *out)
1467 struct json_object *ret = NULL;
1469 if (output_format & FIO_OUTPUT_TERSE)
1470 show_thread_status_terse(ts, rs, out);
1471 if (output_format & FIO_OUTPUT_JSON)
1472 ret = show_thread_status_json(ts, rs, opt_list);
1473 if (output_format & FIO_OUTPUT_NORMAL)
1474 show_thread_status_normal(ts, rs, out);
1479 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1483 if (src->samples == 0)
1486 dst->min_val = min(dst->min_val, src->min_val);
1487 dst->max_val = max(dst->max_val, src->max_val);
1490 * Compute new mean and S after the merge
1491 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1492 * #Parallel_algorithm>
1495 mean = src->mean.u.f;
1498 double delta = src->mean.u.f - dst->mean.u.f;
1500 mean = ((src->mean.u.f * src->samples) +
1501 (dst->mean.u.f * dst->samples)) /
1502 (dst->samples + src->samples);
1504 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1505 (dst->samples * src->samples) /
1506 (dst->samples + src->samples);
1509 dst->samples += src->samples;
1510 dst->mean.u.f = mean;
1514 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1518 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1519 if (dst->max_run[i] < src->max_run[i])
1520 dst->max_run[i] = src->max_run[i];
1521 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1522 dst->min_run[i] = src->min_run[i];
1523 if (dst->max_bw[i] < src->max_bw[i])
1524 dst->max_bw[i] = src->max_bw[i];
1525 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1526 dst->min_bw[i] = src->min_bw[i];
1528 dst->iobytes[i] += src->iobytes[i];
1529 dst->agg[i] += src->agg[i];
1533 dst->kb_base = src->kb_base;
1534 if (!dst->unit_base)
1535 dst->unit_base = src->unit_base;
1537 dst->sig_figs = src->sig_figs;
1540 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1545 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1546 if (!dst->unified_rw_rep) {
1547 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1548 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1549 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1550 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1551 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1553 dst->io_bytes[l] += src->io_bytes[l];
1555 if (dst->runtime[l] < src->runtime[l])
1556 dst->runtime[l] = src->runtime[l];
1558 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1559 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1560 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1561 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1562 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1564 dst->io_bytes[0] += src->io_bytes[l];
1566 if (dst->runtime[0] < src->runtime[l])
1567 dst->runtime[0] = src->runtime[l];
1570 * We're summing to the same destination, so override
1571 * 'first' after the first iteration of the loop
1577 sum_stat(&dst->sync_stat, &src->sync_stat, first);
1578 dst->usr_time += src->usr_time;
1579 dst->sys_time += src->sys_time;
1580 dst->ctx += src->ctx;
1581 dst->majf += src->majf;
1582 dst->minf += src->minf;
1584 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
1585 dst->io_u_map[k] += src->io_u_map[k];
1586 dst->io_u_submit[k] += src->io_u_submit[k];
1587 dst->io_u_complete[k] += src->io_u_complete[k];
1589 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++) {
1590 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1591 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1592 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1594 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1595 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
1597 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1598 if (!dst->unified_rw_rep) {
1599 dst->total_io_u[k] += src->total_io_u[k];
1600 dst->short_io_u[k] += src->short_io_u[k];
1601 dst->drop_io_u[k] += src->drop_io_u[k];
1603 dst->total_io_u[0] += src->total_io_u[k];
1604 dst->short_io_u[0] += src->short_io_u[k];
1605 dst->drop_io_u[0] += src->drop_io_u[k];
1609 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
1611 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1614 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1615 if (!dst->unified_rw_rep)
1616 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1618 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1622 dst->total_run_time += src->total_run_time;
1623 dst->total_submit += src->total_submit;
1624 dst->total_complete += src->total_complete;
1627 void init_group_run_stat(struct group_run_stats *gs)
1630 memset(gs, 0, sizeof(*gs));
1632 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1633 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1636 void init_thread_stat(struct thread_stat *ts)
1640 memset(ts, 0, sizeof(*ts));
1642 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1643 ts->lat_stat[j].min_val = -1UL;
1644 ts->clat_stat[j].min_val = -1UL;
1645 ts->slat_stat[j].min_val = -1UL;
1646 ts->bw_stat[j].min_val = -1UL;
1647 ts->iops_stat[j].min_val = -1UL;
1649 ts->sync_stat.min_val = -1UL;
1653 void __show_run_stats(void)
1655 struct group_run_stats *runstats, *rs;
1656 struct thread_data *td;
1657 struct thread_stat *threadstats, *ts;
1658 int i, j, k, nr_ts, last_ts, idx;
1659 bool kb_base_warned = false;
1660 bool unit_base_warned = false;
1661 struct json_object *root = NULL;
1662 struct json_array *array = NULL;
1663 struct buf_output output[FIO_OUTPUT_NR];
1664 struct flist_head **opt_lists;
1666 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1668 for (i = 0; i < groupid + 1; i++)
1669 init_group_run_stat(&runstats[i]);
1672 * find out how many threads stats we need. if group reporting isn't
1673 * enabled, it's one-per-td.
1677 for_each_td(td, i) {
1678 if (!td->o.group_reporting) {
1682 if (last_ts == td->groupid)
1687 last_ts = td->groupid;
1691 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1692 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1694 for (i = 0; i < nr_ts; i++) {
1695 init_thread_stat(&threadstats[i]);
1696 opt_lists[i] = NULL;
1702 for_each_td(td, i) {
1705 if (idx && (!td->o.group_reporting ||
1706 (td->o.group_reporting && last_ts != td->groupid))) {
1711 last_ts = td->groupid;
1713 ts = &threadstats[j];
1715 ts->clat_percentiles = td->o.clat_percentiles;
1716 ts->lat_percentiles = td->o.lat_percentiles;
1717 ts->percentile_precision = td->o.percentile_precision;
1718 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1719 opt_lists[j] = &td->opt_list;
1724 if (ts->groupid == -1) {
1726 * These are per-group shared already
1728 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1729 if (td->o.description)
1730 strncpy(ts->description, td->o.description,
1731 FIO_JOBDESC_SIZE - 1);
1733 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1736 * If multiple entries in this group, this is
1739 ts->thread_number = td->thread_number;
1740 ts->groupid = td->groupid;
1743 * first pid in group, not very useful...
1747 ts->kb_base = td->o.kb_base;
1748 ts->unit_base = td->o.unit_base;
1749 ts->sig_figs = td->o.sig_figs;
1750 ts->unified_rw_rep = td->o.unified_rw_rep;
1751 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1752 log_info("fio: kb_base differs for jobs in group, using"
1753 " %u as the base\n", ts->kb_base);
1754 kb_base_warned = true;
1755 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1756 log_info("fio: unit_base differs for jobs in group, using"
1757 " %u as the base\n", ts->unit_base);
1758 unit_base_warned = true;
1761 ts->continue_on_error = td->o.continue_on_error;
1762 ts->total_err_count += td->total_err_count;
1763 ts->first_error = td->first_error;
1765 if (!td->error && td->o.continue_on_error &&
1767 ts->error = td->first_error;
1768 ts->verror[sizeof(ts->verror) - 1] = '\0';
1769 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1770 } else if (td->error) {
1771 ts->error = td->error;
1772 ts->verror[sizeof(ts->verror) - 1] = '\0';
1773 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1777 ts->latency_depth = td->latency_qd;
1778 ts->latency_target = td->o.latency_target;
1779 ts->latency_percentile = td->o.latency_percentile;
1780 ts->latency_window = td->o.latency_window;
1782 ts->nr_block_infos = td->ts.nr_block_infos;
1783 for (k = 0; k < ts->nr_block_infos; k++)
1784 ts->block_infos[k] = td->ts.block_infos[k];
1786 sum_thread_stats(ts, &td->ts, idx == 1);
1789 ts->ss_state = td->ss.state;
1790 ts->ss_dur = td->ss.dur;
1791 ts->ss_head = td->ss.head;
1792 ts->ss_bw_data = td->ss.bw_data;
1793 ts->ss_iops_data = td->ss.iops_data;
1794 ts->ss_limit.u.f = td->ss.limit;
1795 ts->ss_slope.u.f = td->ss.slope;
1796 ts->ss_deviation.u.f = td->ss.deviation;
1797 ts->ss_criterion.u.f = td->ss.criterion;
1800 ts->ss_dur = ts->ss_state = 0;
1803 for (i = 0; i < nr_ts; i++) {
1804 unsigned long long bw;
1806 ts = &threadstats[i];
1807 if (ts->groupid == -1)
1809 rs = &runstats[ts->groupid];
1810 rs->kb_base = ts->kb_base;
1811 rs->unit_base = ts->unit_base;
1812 rs->sig_figs = ts->sig_figs;
1813 rs->unified_rw_rep += ts->unified_rw_rep;
1815 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1816 if (!ts->runtime[j])
1818 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1819 rs->min_run[j] = ts->runtime[j];
1820 if (ts->runtime[j] > rs->max_run[j])
1821 rs->max_run[j] = ts->runtime[j];
1825 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1826 if (bw < rs->min_bw[j])
1828 if (bw > rs->max_bw[j])
1831 rs->iobytes[j] += ts->io_bytes[j];
1835 for (i = 0; i < groupid + 1; i++) {
1840 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1841 if (rs->max_run[ddir])
1842 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1847 for (i = 0; i < FIO_OUTPUT_NR; i++)
1848 buf_output_init(&output[i]);
1851 * don't overwrite last signal output
1853 if (output_format & FIO_OUTPUT_NORMAL)
1854 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1855 if (output_format & FIO_OUTPUT_JSON) {
1856 struct thread_data *global;
1859 unsigned long long ms_since_epoch;
1862 gettimeofday(&now, NULL);
1863 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1864 (unsigned long long)(now.tv_usec) / 1000;
1866 tv_sec = now.tv_sec;
1867 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
1868 if (time_buf[strlen(time_buf) - 1] == '\n')
1869 time_buf[strlen(time_buf) - 1] = '\0';
1871 root = json_create_object();
1872 json_object_add_value_string(root, "fio version", fio_version_string);
1873 json_object_add_value_int(root, "timestamp", now.tv_sec);
1874 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1875 json_object_add_value_string(root, "time", time_buf);
1876 global = get_global_options();
1877 json_add_job_opts(root, "global options", &global->opt_list);
1878 array = json_create_array();
1879 json_object_add_value_array(root, "jobs", array);
1883 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1885 for (i = 0; i < nr_ts; i++) {
1886 ts = &threadstats[i];
1887 rs = &runstats[ts->groupid];
1890 fio_server_send_job_options(opt_lists[i], i);
1891 fio_server_send_ts(ts, rs);
1893 if (output_format & FIO_OUTPUT_TERSE)
1894 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1895 if (output_format & FIO_OUTPUT_JSON) {
1896 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1897 json_array_add_value_object(array, tmp);
1899 if (output_format & FIO_OUTPUT_NORMAL)
1900 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1903 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1904 /* disk util stats, if any */
1905 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1907 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1909 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1910 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1911 json_free_object(root);
1914 for (i = 0; i < groupid + 1; i++) {
1919 fio_server_send_gs(rs);
1920 else if (output_format & FIO_OUTPUT_NORMAL)
1921 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1925 fio_server_send_du();
1926 else if (output_format & FIO_OUTPUT_NORMAL) {
1927 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1928 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1931 for (i = 0; i < FIO_OUTPUT_NR; i++) {
1932 struct buf_output *out = &output[i];
1934 log_info_buf(out->buf, out->buflen);
1935 buf_output_free(out);
1938 fio_idle_prof_cleanup();
1946 void __show_running_run_stats(void)
1948 struct thread_data *td;
1949 unsigned long long *rt;
1953 fio_sem_down(stat_sem);
1955 rt = malloc(thread_number * sizeof(unsigned long long));
1956 fio_gettime(&ts, NULL);
1958 for_each_td(td, i) {
1959 td->update_rusage = 1;
1960 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1961 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1962 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1963 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1965 rt[i] = mtime_since(&td->start, &ts);
1966 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1967 td->ts.runtime[DDIR_READ] += rt[i];
1968 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1969 td->ts.runtime[DDIR_WRITE] += rt[i];
1970 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1971 td->ts.runtime[DDIR_TRIM] += rt[i];
1974 for_each_td(td, i) {
1975 if (td->runstate >= TD_EXITED)
1977 if (td->rusage_sem) {
1978 td->update_rusage = 1;
1979 fio_sem_down(td->rusage_sem);
1981 td->update_rusage = 0;
1986 for_each_td(td, i) {
1987 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1988 td->ts.runtime[DDIR_READ] -= rt[i];
1989 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1990 td->ts.runtime[DDIR_WRITE] -= rt[i];
1991 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1992 td->ts.runtime[DDIR_TRIM] -= rt[i];
1996 fio_sem_up(stat_sem);
1999 static bool status_interval_init;
2000 static struct timespec status_time;
2001 static bool status_file_disabled;
2003 #define FIO_STATUS_FILE "fio-dump-status"
2005 static int check_status_file(void)
2008 const char *temp_dir;
2009 char fio_status_file_path[PATH_MAX];
2011 if (status_file_disabled)
2014 temp_dir = getenv("TMPDIR");
2015 if (temp_dir == NULL) {
2016 temp_dir = getenv("TEMP");
2017 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2020 if (temp_dir == NULL)
2023 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2025 if (stat(fio_status_file_path, &sb))
2028 if (unlink(fio_status_file_path) < 0) {
2029 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2031 log_err("fio: disabling status file updates\n");
2032 status_file_disabled = true;
2038 void check_for_running_stats(void)
2040 if (status_interval) {
2041 if (!status_interval_init) {
2042 fio_gettime(&status_time, NULL);
2043 status_interval_init = true;
2044 } else if (mtime_since_now(&status_time) >= status_interval) {
2045 show_running_run_stats();
2046 fio_gettime(&status_time, NULL);
2050 if (check_status_file()) {
2051 show_running_run_stats();
2056 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2061 if (data > is->max_val)
2063 if (data < is->min_val)
2066 delta = val - is->mean.u.f;
2068 is->mean.u.f += delta / (is->samples + 1.0);
2069 is->S.u.f += delta * (val - is->mean.u.f);
2076 * Return a struct io_logs, which is added to the tail of the log
2079 static struct io_logs *get_new_log(struct io_log *iolog)
2081 size_t new_size, new_samples;
2082 struct io_logs *cur_log;
2085 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2088 if (!iolog->cur_log_max)
2089 new_samples = DEF_LOG_ENTRIES;
2091 new_samples = iolog->cur_log_max * 2;
2092 if (new_samples > MAX_LOG_ENTRIES)
2093 new_samples = MAX_LOG_ENTRIES;
2096 new_size = new_samples * log_entry_sz(iolog);
2098 cur_log = smalloc(sizeof(*cur_log));
2100 INIT_FLIST_HEAD(&cur_log->list);
2101 cur_log->log = malloc(new_size);
2103 cur_log->nr_samples = 0;
2104 cur_log->max_samples = new_samples;
2105 flist_add_tail(&cur_log->list, &iolog->io_logs);
2106 iolog->cur_log_max = new_samples;
2116 * Add and return a new log chunk, or return current log if big enough
2118 static struct io_logs *regrow_log(struct io_log *iolog)
2120 struct io_logs *cur_log;
2123 if (!iolog || iolog->disabled)
2126 cur_log = iolog_cur_log(iolog);
2128 cur_log = get_new_log(iolog);
2133 if (cur_log->nr_samples < cur_log->max_samples)
2137 * No room for a new sample. If we're compressing on the fly, flush
2138 * out the current chunk
2140 if (iolog->log_gz) {
2141 if (iolog_cur_flush(iolog, cur_log)) {
2142 log_err("fio: failed flushing iolog! Will stop logging.\n");
2148 * Get a new log array, and add to our list
2150 cur_log = get_new_log(iolog);
2152 log_err("fio: failed extending iolog! Will stop logging.\n");
2156 if (!iolog->pending || !iolog->pending->nr_samples)
2160 * Flush pending items to new log
2162 for (i = 0; i < iolog->pending->nr_samples; i++) {
2163 struct io_sample *src, *dst;
2165 src = get_sample(iolog, iolog->pending, i);
2166 dst = get_sample(iolog, cur_log, i);
2167 memcpy(dst, src, log_entry_sz(iolog));
2169 cur_log->nr_samples = iolog->pending->nr_samples;
2171 iolog->pending->nr_samples = 0;
2175 iolog->disabled = true;
2179 void regrow_logs(struct thread_data *td)
2181 regrow_log(td->slat_log);
2182 regrow_log(td->clat_log);
2183 regrow_log(td->clat_hist_log);
2184 regrow_log(td->lat_log);
2185 regrow_log(td->bw_log);
2186 regrow_log(td->iops_log);
2187 td->flags &= ~TD_F_REGROW_LOGS;
2190 static struct io_logs *get_cur_log(struct io_log *iolog)
2192 struct io_logs *cur_log;
2194 cur_log = iolog_cur_log(iolog);
2196 cur_log = get_new_log(iolog);
2201 if (cur_log->nr_samples < cur_log->max_samples)
2205 * Out of space. If we're in IO offload mode, or we're not doing
2206 * per unit logging (hence logging happens outside of the IO thread
2207 * as well), add a new log chunk inline. If we're doing inline
2208 * submissions, flag 'td' as needing a log regrow and we'll take
2209 * care of it on the submission side.
2211 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2212 !per_unit_log(iolog))
2213 return regrow_log(iolog);
2216 iolog->td->flags |= TD_F_REGROW_LOGS;
2218 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2219 return iolog->pending;
2222 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2223 enum fio_ddir ddir, unsigned int bs,
2224 unsigned long t, uint64_t offset)
2226 struct io_logs *cur_log;
2228 if (iolog->disabled)
2230 if (flist_empty(&iolog->io_logs))
2231 iolog->avg_last[ddir] = t;
2233 cur_log = get_cur_log(iolog);
2235 struct io_sample *s;
2237 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2240 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2241 io_sample_set_ddir(iolog, s, ddir);
2244 if (iolog->log_offset) {
2245 struct io_sample_offset *so = (void *) s;
2247 so->offset = offset;
2250 cur_log->nr_samples++;
2254 iolog->disabled = true;
2257 static inline void reset_io_stat(struct io_stat *ios)
2259 ios->max_val = ios->min_val = ios->samples = 0;
2260 ios->mean.u.f = ios->S.u.f = 0;
2263 void reset_io_stats(struct thread_data *td)
2265 struct thread_stat *ts = &td->ts;
2268 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2269 reset_io_stat(&ts->clat_stat[i]);
2270 reset_io_stat(&ts->slat_stat[i]);
2271 reset_io_stat(&ts->lat_stat[i]);
2272 reset_io_stat(&ts->bw_stat[i]);
2273 reset_io_stat(&ts->iops_stat[i]);
2275 ts->io_bytes[i] = 0;
2277 ts->total_io_u[i] = 0;
2278 ts->short_io_u[i] = 0;
2279 ts->drop_io_u[i] = 0;
2281 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2282 ts->io_u_plat[i][j] = 0;
2284 ts->io_u_sync_plat[j] = 0;
2288 ts->total_io_u[DDIR_SYNC] = 0;
2290 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2291 ts->io_u_map[i] = 0;
2292 ts->io_u_submit[i] = 0;
2293 ts->io_u_complete[i] = 0;
2296 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2297 ts->io_u_lat_n[i] = 0;
2298 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2299 ts->io_u_lat_u[i] = 0;
2300 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2301 ts->io_u_lat_m[i] = 0;
2303 ts->total_submit = 0;
2304 ts->total_complete = 0;
2307 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2308 unsigned long elapsed, bool log_max)
2311 * Note an entry in the log. Use the mean from the logged samples,
2312 * making sure to properly round up. Only write a log entry if we
2313 * had actual samples done.
2315 if (iolog->avg_window[ddir].samples) {
2316 union io_sample_data data;
2319 data.val = iolog->avg_window[ddir].max_val;
2321 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2323 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2326 reset_io_stat(&iolog->avg_window[ddir]);
2329 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2334 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2335 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2338 static unsigned long add_log_sample(struct thread_data *td,
2339 struct io_log *iolog,
2340 union io_sample_data data,
2341 enum fio_ddir ddir, unsigned int bs,
2344 unsigned long elapsed, this_window;
2349 elapsed = mtime_since_now(&td->epoch);
2352 * If no time averaging, just add the log sample.
2354 if (!iolog->avg_msec) {
2355 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2360 * Add the sample. If the time period has passed, then
2361 * add that entry to the log and clear.
2363 add_stat_sample(&iolog->avg_window[ddir], data.val);
2366 * If period hasn't passed, adding the above sample is all we
2369 this_window = elapsed - iolog->avg_last[ddir];
2370 if (elapsed < iolog->avg_last[ddir])
2371 return iolog->avg_last[ddir] - elapsed;
2372 else if (this_window < iolog->avg_msec) {
2373 unsigned long diff = iolog->avg_msec - this_window;
2375 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2379 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2381 iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2382 return iolog->avg_msec;
2385 void finalize_logs(struct thread_data *td, bool unit_logs)
2387 unsigned long elapsed;
2389 elapsed = mtime_since_now(&td->epoch);
2391 if (td->clat_log && unit_logs)
2392 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2393 if (td->slat_log && unit_logs)
2394 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2395 if (td->lat_log && unit_logs)
2396 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2397 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2398 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2399 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2400 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2403 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2405 struct io_log *iolog;
2410 iolog = agg_io_log[ddir];
2411 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2414 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2416 unsigned int idx = plat_val_to_idx(nsec);
2417 assert(idx < FIO_IO_U_PLAT_NR);
2419 ts->io_u_sync_plat[idx]++;
2420 add_stat_sample(&ts->sync_stat, nsec);
2423 static void add_clat_percentile_sample(struct thread_stat *ts,
2424 unsigned long long nsec, enum fio_ddir ddir)
2426 unsigned int idx = plat_val_to_idx(nsec);
2427 assert(idx < FIO_IO_U_PLAT_NR);
2429 ts->io_u_plat[ddir][idx]++;
2432 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2433 unsigned long long nsec, unsigned int bs, uint64_t offset)
2435 unsigned long elapsed, this_window;
2436 struct thread_stat *ts = &td->ts;
2437 struct io_log *iolog = td->clat_hist_log;
2441 add_stat_sample(&ts->clat_stat[ddir], nsec);
2444 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2447 if (ts->clat_percentiles)
2448 add_clat_percentile_sample(ts, nsec, ddir);
2450 if (iolog && iolog->hist_msec) {
2451 struct io_hist *hw = &iolog->hist_window[ddir];
2454 elapsed = mtime_since_now(&td->epoch);
2456 hw->hist_last = elapsed;
2457 this_window = elapsed - hw->hist_last;
2459 if (this_window >= iolog->hist_msec) {
2460 uint64_t *io_u_plat;
2461 struct io_u_plat_entry *dst;
2464 * Make a byte-for-byte copy of the latency histogram
2465 * stored in td->ts.io_u_plat[ddir], recording it in a
2466 * log sample. Note that the matching call to free() is
2467 * located in iolog.c after printing this sample to the
2470 io_u_plat = (uint64_t *) td->ts.io_u_plat[ddir];
2471 dst = malloc(sizeof(struct io_u_plat_entry));
2472 memcpy(&(dst->io_u_plat), io_u_plat,
2473 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2474 flist_add(&dst->list, &hw->list);
2475 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2479 * Update the last time we recorded as being now, minus
2480 * any drift in time we encountered before actually
2481 * making the record.
2483 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2491 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2492 unsigned long usec, unsigned int bs, uint64_t offset)
2494 struct thread_stat *ts = &td->ts;
2501 add_stat_sample(&ts->slat_stat[ddir], usec);
2504 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2509 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2510 unsigned long long nsec, unsigned int bs, uint64_t offset)
2512 struct thread_stat *ts = &td->ts;
2519 add_stat_sample(&ts->lat_stat[ddir], nsec);
2522 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2525 if (ts->lat_percentiles)
2526 add_clat_percentile_sample(ts, nsec, ddir);
2531 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2532 unsigned int bytes, unsigned long long spent)
2534 struct thread_stat *ts = &td->ts;
2538 rate = (unsigned long) (bytes * 1000000ULL / spent);
2544 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2547 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2548 bytes, io_u->offset);
2550 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2554 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2555 struct timespec *t, unsigned int avg_time,
2556 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2557 struct io_stat *stat, struct io_log *log,
2560 unsigned long spent, rate;
2562 unsigned long next, next_log;
2564 next_log = avg_time;
2566 spent = mtime_since(parent_tv, t);
2567 if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2568 return avg_time - spent;
2573 * Compute both read and write rates for the interval.
2575 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2578 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2580 continue; /* No entries for interval */
2584 rate = delta * 1000 / spent / 1024; /* KiB/s */
2586 rate = (delta * 1000) / spent;
2590 add_stat_sample(&stat[ddir], rate);
2593 unsigned int bs = 0;
2595 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2596 bs = td->o.min_bs[ddir];
2598 next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2599 next_log = min(next_log, next);
2602 stat_io_bytes[ddir] = this_io_bytes[ddir];
2605 timespec_add_msec(parent_tv, avg_time);
2609 if (spent <= avg_time)
2612 next = avg_time - (1 + spent - avg_time);
2614 return min(next, next_log);
2617 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2619 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2620 td->this_io_bytes, td->stat_io_bytes,
2621 td->ts.bw_stat, td->bw_log, true);
2624 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2627 struct thread_stat *ts = &td->ts;
2631 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2634 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2635 bytes, io_u->offset);
2637 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2641 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2643 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2644 td->this_io_blocks, td->stat_io_blocks,
2645 td->ts.iops_stat, td->iops_log, false);
2649 * Returns msecs to next event
2651 int calc_log_samples(void)
2653 struct thread_data *td;
2654 unsigned int next = ~0U, tmp;
2655 struct timespec now;
2658 fio_gettime(&now, NULL);
2660 for_each_td(td, i) {
2663 if (in_ramp_time(td) ||
2664 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2665 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2669 (td->bw_log && !per_unit_log(td->bw_log))) {
2670 tmp = add_bw_samples(td, &now);
2674 if (!td->iops_log ||
2675 (td->iops_log && !per_unit_log(td->iops_log))) {
2676 tmp = add_iops_samples(td, &now);
2682 return next == ~0U ? 0 : next;
2685 void stat_init(void)
2687 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
2690 void stat_exit(void)
2693 * When we have the mutex, we know out-of-band access to it
2696 fio_sem_down(stat_sem);
2697 fio_sem_remove(stat_sem);
2701 * Called from signal handler. Wake up status thread.
2703 void show_running_run_stats(void)
2708 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2710 /* Ignore io_u's which span multiple blocks--they will just get
2711 * inaccurate counts. */
2712 int idx = (io_u->offset - io_u->file->file_offset)
2713 / td->o.bs[DDIR_TRIM];
2714 uint32_t *info = &td->ts.block_infos[idx];
2715 assert(idx < td->ts.nr_block_infos);