9 #include "lib/ieee754.h"
11 #include "lib/getrusage.h"
14 #include "lib/output_buffer.h"
15 #include "helper_thread.h"
18 #include "oslib/asprintf.h"
21 #define LOG_MSEC_SLACK 2
23 #define LOG_MSEC_SLACK 1
26 struct fio_sem *stat_sem;
28 void clear_rusage_stat(struct thread_data *td)
30 struct thread_stat *ts = &td->ts;
32 fio_getrusage(&td->ru_start);
33 ts->usr_time = ts->sys_time = 0;
35 ts->minf = ts->majf = 0;
38 void update_rusage_stat(struct thread_data *td)
40 struct thread_stat *ts = &td->ts;
42 fio_getrusage(&td->ru_end);
43 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
44 &td->ru_end.ru_utime);
45 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
46 &td->ru_end.ru_stime);
47 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
48 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
49 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
50 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
52 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
56 * Given a latency, return the index of the corresponding bucket in
57 * the structure tracking percentiles.
59 * (1) find the group (and error bits) that the value (latency)
60 * belongs to by looking at its MSB. (2) find the bucket number in the
61 * group by looking at the index bits.
64 static unsigned int plat_val_to_idx(unsigned long long val)
66 unsigned int msb, error_bits, base, offset, idx;
68 /* Find MSB starting from bit 0 */
72 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
75 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
76 * all bits of the sample as index
78 if (msb <= FIO_IO_U_PLAT_BITS)
81 /* Compute the number of error bits to discard*/
82 error_bits = msb - FIO_IO_U_PLAT_BITS;
84 /* Compute the number of buckets before the group */
85 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
88 * Discard the error bits and apply the mask to find the
89 * index for the buckets in the group
91 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
93 /* Make sure the index does not exceed (array size - 1) */
94 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
95 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
101 * Convert the given index of the bucket array to the value
102 * represented by the bucket
104 static unsigned long long plat_idx_to_val(unsigned int idx)
106 unsigned int error_bits;
107 unsigned long long k, base;
109 assert(idx < FIO_IO_U_PLAT_NR);
111 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
112 * all bits of the sample as index */
113 if (idx < (FIO_IO_U_PLAT_VAL << 1))
116 /* Find the group and compute the minimum value of that group */
117 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
118 base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
120 /* Find its bucket number of the group */
121 k = idx % FIO_IO_U_PLAT_VAL;
123 /* Return the mean of the range of the bucket */
124 return base + ((k + 0.5) * (1 << error_bits));
127 static int double_cmp(const void *a, const void *b)
129 const fio_fp64_t fa = *(const fio_fp64_t *) a;
130 const fio_fp64_t fb = *(const fio_fp64_t *) b;
135 else if (fa.u.f < fb.u.f)
141 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
142 fio_fp64_t *plist, unsigned long long **output,
143 unsigned long long *maxv, unsigned long long *minv)
145 unsigned long long sum = 0;
146 unsigned int len, i, j = 0;
147 unsigned long long *ovals = NULL;
154 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
161 * Sort the percentile list. Note that it may already be sorted if
162 * we are using the default values, but since it's a short list this
163 * isn't a worry. Also note that this does not work for NaN values.
166 qsort(plist, len, sizeof(plist[0]), double_cmp);
168 ovals = malloc(len * sizeof(*ovals));
173 * Calculate bucket values, note down max and min values
176 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
178 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
179 assert(plist[j].u.f <= 100.0);
181 ovals[j] = plat_idx_to_val(i);
182 if (ovals[j] < *minv)
184 if (ovals[j] > *maxv)
187 is_last = (j == len - 1) != 0;
196 log_err("fio: error calculating latency percentiles\n");
203 * Find and display the p-th percentile of clat
205 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
206 fio_fp64_t *plist, unsigned int precision,
207 const char *pre, struct buf_output *out)
209 unsigned int divisor, len, i, j = 0;
210 unsigned long long minv, maxv;
211 unsigned long long *ovals;
212 int per_line, scale_down, time_width;
216 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
221 * We default to nsecs, but if the value range is such that we
222 * should scale down to usecs or msecs, do that.
224 if (minv > 2000000 && maxv > 99999999ULL) {
227 log_buf(out, " %s percentiles (msec):\n |", pre);
228 } else if (minv > 2000 && maxv > 99999) {
231 log_buf(out, " %s percentiles (usec):\n |", pre);
235 log_buf(out, " %s percentiles (nsec):\n |", pre);
239 time_width = max(5, (int) (log10(maxv / divisor) + 1));
240 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
241 precision, time_width);
242 /* fmt will be something like " %5.2fth=[%4llu]%c" */
243 per_line = (80 - 7) / (precision + 10 + time_width);
245 for (j = 0; j < len; j++) {
247 if (j != 0 && (j % per_line) == 0)
250 /* end of the list */
251 is_last = (j == len - 1) != 0;
253 for (i = 0; i < scale_down; i++)
254 ovals[j] = (ovals[j] + 999) / 1000;
256 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
261 if ((j % per_line) == per_line - 1) /* for formatting */
268 bool calc_lat(struct io_stat *is, unsigned long long *min,
269 unsigned long long *max, double *mean, double *dev)
271 double n = (double) is->samples;
278 *mean = is->mean.u.f;
281 *dev = sqrt(is->S.u.f / (n - 1.0));
288 void show_mixed_group_stats(struct group_run_stats *rs, struct buf_output *out)
290 char *io, *agg, *min, *max;
291 char *ioalt, *aggalt, *minalt, *maxalt;
292 uint64_t io_mix = 0, agg_mix = 0, min_mix = -1, max_mix = 0;
293 uint64_t min_run = -1, max_run = 0;
294 const int i2p = is_power_of_2(rs->kb_base);
297 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
300 io_mix += rs->iobytes[i];
301 agg_mix += rs->agg[i];
302 min_mix = min_mix < rs->min_bw[i] ? min_mix : rs->min_bw[i];
303 max_mix = max_mix > rs->max_bw[i] ? max_mix : rs->max_bw[i];
304 min_run = min_run < rs->min_run[i] ? min_run : rs->min_run[i];
305 max_run = max_run > rs->max_run[i] ? max_run : rs->max_run[i];
307 io = num2str(io_mix, rs->sig_figs, 1, i2p, N2S_BYTE);
308 ioalt = num2str(io_mix, rs->sig_figs, 1, !i2p, N2S_BYTE);
309 agg = num2str(agg_mix, rs->sig_figs, 1, i2p, rs->unit_base);
310 aggalt = num2str(agg_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
311 min = num2str(min_mix, rs->sig_figs, 1, i2p, rs->unit_base);
312 minalt = num2str(min_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
313 max = num2str(max_mix, rs->sig_figs, 1, i2p, rs->unit_base);
314 maxalt = num2str(max_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
315 log_buf(out, " MIXED: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
316 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
317 (unsigned long long) min_run,
318 (unsigned long long) max_run);
329 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
331 char *io, *agg, *min, *max;
332 char *ioalt, *aggalt, *minalt, *maxalt;
333 const char *str[] = { " READ", " WRITE" , " TRIM"};
336 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
338 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
339 const int i2p = is_power_of_2(rs->kb_base);
344 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
345 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
346 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
347 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
348 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
349 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
350 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
351 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
352 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
353 (rs->unified_rw_rep == UNIFIED_MIXED) ? " MIXED" : str[i],
354 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
355 (unsigned long long) rs->min_run[i],
356 (unsigned long long) rs->max_run[i]);
368 /* Need to aggregate statisitics to show mixed values */
369 if (rs->unified_rw_rep == UNIFIED_BOTH)
370 show_mixed_group_stats(rs, out);
373 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
378 * Do depth distribution calculations
380 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
382 io_u_dist[i] = (double) map[i] / (double) total;
383 io_u_dist[i] *= 100.0;
384 if (io_u_dist[i] < 0.1 && map[i])
391 static void stat_calc_lat(struct thread_stat *ts, double *dst,
392 uint64_t *src, int nr)
394 unsigned long total = ddir_rw_sum(ts->total_io_u);
398 * Do latency distribution calculations
400 for (i = 0; i < nr; i++) {
402 dst[i] = (double) src[i] / (double) total;
404 if (dst[i] < 0.01 && src[i])
412 * To keep the terse format unaltered, add all of the ns latency
413 * buckets to the first us latency bucket
415 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
417 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
420 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
422 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
423 ntotal += ts->io_u_lat_n[i];
425 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
428 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
430 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
433 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
435 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
438 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
440 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
443 static void display_lat(const char *name, unsigned long long min,
444 unsigned long long max, double mean, double dev,
445 struct buf_output *out)
447 const char *base = "(nsec)";
450 if (nsec_to_msec(&min, &max, &mean, &dev))
452 else if (nsec_to_usec(&min, &max, &mean, &dev))
455 minp = num2str(min, 6, 1, 0, N2S_NONE);
456 maxp = num2str(max, 6, 1, 0, N2S_NONE);
458 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
459 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
465 static struct thread_stat *gen_mixed_ddir_stats_from_ts(struct thread_stat *ts)
467 struct thread_stat *ts_lcl;
470 * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
473 ts_lcl = malloc(sizeof(struct thread_stat));
475 log_err("fio: failed to allocate local thread stat\n");
479 init_thread_stat(ts_lcl);
481 /* calculate mixed stats */
482 ts_lcl->unified_rw_rep = UNIFIED_MIXED;
483 ts_lcl->lat_percentiles = ts->lat_percentiles;
484 ts_lcl->clat_percentiles = ts->clat_percentiles;
485 ts_lcl->slat_percentiles = ts->slat_percentiles;
486 ts_lcl->percentile_precision = ts->percentile_precision;
487 memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
489 sum_thread_stats(ts_lcl, ts);
494 static double convert_agg_kbytes_percent(struct group_run_stats *rs,
495 enum fio_ddir ddir, int mean)
497 double p_of_agg = 100.0;
498 if (rs && rs->agg[ddir] > 1024) {
499 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
501 if (p_of_agg > 100.0)
507 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
508 enum fio_ddir ddir, struct buf_output *out)
511 unsigned long long min, max, bw, iops;
513 char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
516 if (ddir_sync(ddir)) {
517 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
518 log_buf(out, " %s:\n", "fsync/fdatasync/sync_file_range");
519 display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
520 show_clat_percentiles(ts->io_u_sync_plat,
521 ts->sync_stat.samples,
523 ts->percentile_precision,
524 io_ddir_name(ddir), out);
529 assert(ddir_rw(ddir));
531 if (!ts->runtime[ddir])
534 i2p = is_power_of_2(rs->kb_base);
535 runt = ts->runtime[ddir];
537 bw = (1000 * ts->io_bytes[ddir]) / runt;
538 io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
539 bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
540 bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
542 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
543 iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
544 if (ddir == DDIR_WRITE)
545 post_st = zbd_write_status(ts);
546 else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
550 total = ts->cachehit + ts->cachemiss;
551 hit = (double) ts->cachehit / (double) total;
553 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
557 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
558 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
559 iops_p, bw_p, bw_p_alt, io_p,
560 (unsigned long long) ts->runtime[ddir],
569 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
570 display_lat("slat", min, max, mean, dev, out);
571 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
572 display_lat("clat", min, max, mean, dev, out);
573 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
574 display_lat(" lat", min, max, mean, dev, out);
575 if (calc_lat(&ts->clat_high_prio_stat[ddir], &min, &max, &mean, &dev)) {
576 display_lat(ts->lat_percentiles ? "high prio_lat" : "high prio_clat",
577 min, max, mean, dev, out);
578 if (calc_lat(&ts->clat_low_prio_stat[ddir], &min, &max, &mean, &dev))
579 display_lat(ts->lat_percentiles ? "low prio_lat" : "low prio_clat",
580 min, max, mean, dev, out);
583 if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
584 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
585 ts->slat_stat[ddir].samples,
587 ts->percentile_precision, "slat", out);
588 if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
589 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
590 ts->clat_stat[ddir].samples,
592 ts->percentile_precision, "clat", out);
593 if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
594 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
595 ts->lat_stat[ddir].samples,
597 ts->percentile_precision, "lat", out);
599 if (ts->clat_percentiles || ts->lat_percentiles) {
600 const char *name = ts->lat_percentiles ? "lat" : "clat";
604 if (ts->lat_percentiles)
605 samples = ts->lat_stat[ddir].samples;
607 samples = ts->clat_stat[ddir].samples;
609 /* Only print this if some high and low priority stats were collected */
610 if (ts->clat_high_prio_stat[ddir].samples > 0 &&
611 ts->clat_low_prio_stat[ddir].samples > 0)
613 sprintf(prio_name, "high prio (%.2f%%) %s",
614 100. * (double) ts->clat_high_prio_stat[ddir].samples / (double) samples,
616 show_clat_percentiles(ts->io_u_plat_high_prio[ddir],
617 ts->clat_high_prio_stat[ddir].samples,
619 ts->percentile_precision, prio_name, out);
621 sprintf(prio_name, "low prio (%.2f%%) %s",
622 100. * (double) ts->clat_low_prio_stat[ddir].samples / (double) samples,
624 show_clat_percentiles(ts->io_u_plat_low_prio[ddir],
625 ts->clat_low_prio_stat[ddir].samples,
627 ts->percentile_precision, prio_name, out);
631 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
632 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
635 if ((rs->unit_base == 1) && i2p)
637 else if (rs->unit_base == 1)
644 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
646 if (rs->unit_base == 1) {
653 if (mean > fkb_base * fkb_base) {
658 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
661 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
662 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
663 bw_str, min, max, p_of_agg, mean, dev,
664 (&ts->bw_stat[ddir])->samples);
666 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
667 log_buf(out, " iops : min=%5llu, max=%5llu, "
668 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
669 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
673 static void show_mixed_ddir_status(struct group_run_stats *rs,
674 struct thread_stat *ts,
675 struct buf_output *out)
677 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
680 show_ddir_status(rs, ts_lcl, DDIR_READ, out);
685 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
686 const char *msg, struct buf_output *out)
688 bool new_line = true, shown = false;
691 for (i = 0; i < nr; i++) {
692 if (io_u_lat[i] <= 0.0)
698 log_buf(out, " lat (%s) : ", msg);
704 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
716 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
718 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
719 "250=", "500=", "750=", "1000=", };
721 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
724 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
726 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
727 "250=", "500=", "750=", "1000=", };
729 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
732 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
734 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
735 "250=", "500=", "750=", "1000=", "2000=",
738 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
741 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
743 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
744 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
745 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
747 stat_calc_lat_n(ts, io_u_lat_n);
748 stat_calc_lat_u(ts, io_u_lat_u);
749 stat_calc_lat_m(ts, io_u_lat_m);
751 show_lat_n(io_u_lat_n, out);
752 show_lat_u(io_u_lat_u, out);
753 show_lat_m(io_u_lat_m, out);
756 static int block_state_category(int block_state)
758 switch (block_state) {
759 case BLOCK_STATE_UNINIT:
761 case BLOCK_STATE_TRIMMED:
762 case BLOCK_STATE_WRITTEN:
764 case BLOCK_STATE_WRITE_FAILURE:
765 case BLOCK_STATE_TRIM_FAILURE:
768 /* Silence compile warning on some BSDs and have a return */
774 static int compare_block_infos(const void *bs1, const void *bs2)
776 uint64_t block1 = *(uint64_t *)bs1;
777 uint64_t block2 = *(uint64_t *)bs2;
778 int state1 = BLOCK_INFO_STATE(block1);
779 int state2 = BLOCK_INFO_STATE(block2);
780 int bscat1 = block_state_category(state1);
781 int bscat2 = block_state_category(state2);
782 int cycles1 = BLOCK_INFO_TRIMS(block1);
783 int cycles2 = BLOCK_INFO_TRIMS(block2);
790 if (cycles1 < cycles2)
792 if (cycles1 > cycles2)
800 assert(block1 == block2);
804 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
805 fio_fp64_t *plist, unsigned int **percentiles,
811 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
813 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
820 * Sort the percentile list. Note that it may already be sorted if
821 * we are using the default values, but since it's a short list this
822 * isn't a worry. Also note that this does not work for NaN values.
825 qsort(plist, len, sizeof(plist[0]), double_cmp);
827 /* Start only after the uninit entries end */
829 nr_uninit < nr_block_infos
830 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
834 if (nr_uninit == nr_block_infos)
837 *percentiles = calloc(len, sizeof(**percentiles));
839 for (i = 0; i < len; i++) {
840 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
842 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
845 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
846 for (i = 0; i < nr_block_infos; i++)
847 types[BLOCK_INFO_STATE(block_infos[i])]++;
852 static const char *block_state_names[] = {
853 [BLOCK_STATE_UNINIT] = "unwritten",
854 [BLOCK_STATE_TRIMMED] = "trimmed",
855 [BLOCK_STATE_WRITTEN] = "written",
856 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
857 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
860 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
861 fio_fp64_t *plist, struct buf_output *out)
864 unsigned int *percentiles = NULL;
865 unsigned int block_state_counts[BLOCK_STATE_COUNT];
867 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
868 &percentiles, block_state_counts);
870 log_buf(out, " block lifetime percentiles :\n |");
872 for (i = 0; i < len; i++) {
873 uint32_t block_info = percentiles[i];
874 #define LINE_LENGTH 75
875 char str[LINE_LENGTH];
876 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
877 plist[i].u.f, block_info,
878 i == len - 1 ? '\n' : ',');
879 assert(strln < LINE_LENGTH);
880 if (pos + strln > LINE_LENGTH) {
882 log_buf(out, "\n |");
884 log_buf(out, "%s", str);
891 log_buf(out, " states :");
892 for (i = 0; i < BLOCK_STATE_COUNT; i++)
893 log_buf(out, " %s=%u%c",
894 block_state_names[i], block_state_counts[i],
895 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
898 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
900 char *p1, *p1alt, *p2;
901 unsigned long long bw_mean, iops_mean;
902 const int i2p = is_power_of_2(ts->kb_base);
907 bw_mean = steadystate_bw_mean(ts);
908 iops_mean = steadystate_iops_mean(ts);
910 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
911 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
912 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
914 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
915 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
917 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
918 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
919 ts->ss_criterion.u.f,
920 ts->ss_state & FIO_SS_PCT ? "%" : "");
927 static void show_agg_stats(struct disk_util_agg *agg, int terse,
928 struct buf_output *out)
930 if (!agg->slavecount)
934 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
935 "aggrticks=%llu/%llu, aggrin_queue=%llu, "
937 (unsigned long long) agg->ios[0] / agg->slavecount,
938 (unsigned long long) agg->ios[1] / agg->slavecount,
939 (unsigned long long) agg->merges[0] / agg->slavecount,
940 (unsigned long long) agg->merges[1] / agg->slavecount,
941 (unsigned long long) agg->ticks[0] / agg->slavecount,
942 (unsigned long long) agg->ticks[1] / agg->slavecount,
943 (unsigned long long) agg->time_in_queue / agg->slavecount,
946 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
947 (unsigned long long) agg->ios[0] / agg->slavecount,
948 (unsigned long long) agg->ios[1] / agg->slavecount,
949 (unsigned long long) agg->merges[0] / agg->slavecount,
950 (unsigned long long) agg->merges[1] / agg->slavecount,
951 (unsigned long long) agg->ticks[0] / agg->slavecount,
952 (unsigned long long) agg->ticks[1] / agg->slavecount,
953 (unsigned long long) agg->time_in_queue / agg->slavecount,
958 static void aggregate_slaves_stats(struct disk_util *masterdu)
960 struct disk_util_agg *agg = &masterdu->agg;
961 struct disk_util_stat *dus;
962 struct flist_head *entry;
963 struct disk_util *slavedu;
966 flist_for_each(entry, &masterdu->slaves) {
967 slavedu = flist_entry(entry, struct disk_util, slavelist);
969 agg->ios[0] += dus->s.ios[0];
970 agg->ios[1] += dus->s.ios[1];
971 agg->merges[0] += dus->s.merges[0];
972 agg->merges[1] += dus->s.merges[1];
973 agg->sectors[0] += dus->s.sectors[0];
974 agg->sectors[1] += dus->s.sectors[1];
975 agg->ticks[0] += dus->s.ticks[0];
976 agg->ticks[1] += dus->s.ticks[1];
977 agg->time_in_queue += dus->s.time_in_queue;
980 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
981 /* System utilization is the utilization of the
982 * component with the highest utilization.
984 if (util > agg->max_util.u.f)
985 agg->max_util.u.f = util;
989 if (agg->max_util.u.f > 100.0)
990 agg->max_util.u.f = 100.0;
993 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
994 int terse, struct buf_output *out)
999 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1004 if (agg->slavecount)
1007 log_buf(out, " %s: ios=%llu/%llu, merge=%llu/%llu, "
1008 "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
1010 (unsigned long long) dus->s.ios[0],
1011 (unsigned long long) dus->s.ios[1],
1012 (unsigned long long) dus->s.merges[0],
1013 (unsigned long long) dus->s.merges[1],
1014 (unsigned long long) dus->s.ticks[0],
1015 (unsigned long long) dus->s.ticks[1],
1016 (unsigned long long) dus->s.time_in_queue,
1019 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1021 (unsigned long long) dus->s.ios[0],
1022 (unsigned long long) dus->s.ios[1],
1023 (unsigned long long) dus->s.merges[0],
1024 (unsigned long long) dus->s.merges[1],
1025 (unsigned long long) dus->s.ticks[0],
1026 (unsigned long long) dus->s.ticks[1],
1027 (unsigned long long) dus->s.time_in_queue,
1032 * If the device has slaves, aggregate the stats for
1033 * those slave devices also.
1035 show_agg_stats(agg, terse, out);
1041 void json_array_add_disk_util(struct disk_util_stat *dus,
1042 struct disk_util_agg *agg, struct json_array *array)
1044 struct json_object *obj;
1048 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1052 obj = json_create_object();
1053 json_array_add_value_object(array, obj);
1055 json_object_add_value_string(obj, "name", (const char *)dus->name);
1056 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1057 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1058 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1059 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1060 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1061 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1062 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1063 json_object_add_value_float(obj, "util", util);
1066 * If the device has slaves, aggregate the stats for
1067 * those slave devices also.
1069 if (!agg->slavecount)
1071 json_object_add_value_int(obj, "aggr_read_ios",
1072 agg->ios[0] / agg->slavecount);
1073 json_object_add_value_int(obj, "aggr_write_ios",
1074 agg->ios[1] / agg->slavecount);
1075 json_object_add_value_int(obj, "aggr_read_merges",
1076 agg->merges[0] / agg->slavecount);
1077 json_object_add_value_int(obj, "aggr_write_merge",
1078 agg->merges[1] / agg->slavecount);
1079 json_object_add_value_int(obj, "aggr_read_ticks",
1080 agg->ticks[0] / agg->slavecount);
1081 json_object_add_value_int(obj, "aggr_write_ticks",
1082 agg->ticks[1] / agg->slavecount);
1083 json_object_add_value_int(obj, "aggr_in_queue",
1084 agg->time_in_queue / agg->slavecount);
1085 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1088 static void json_object_add_disk_utils(struct json_object *obj,
1089 struct flist_head *head)
1091 struct json_array *array = json_create_array();
1092 struct flist_head *entry;
1093 struct disk_util *du;
1095 json_object_add_value_array(obj, "disk_util", array);
1097 flist_for_each(entry, head) {
1098 du = flist_entry(entry, struct disk_util, list);
1100 aggregate_slaves_stats(du);
1101 json_array_add_disk_util(&du->dus, &du->agg, array);
1105 void show_disk_util(int terse, struct json_object *parent,
1106 struct buf_output *out)
1108 struct flist_head *entry;
1109 struct disk_util *du;
1112 if (!is_running_backend())
1115 if (flist_empty(&disk_list))
1118 if ((output_format & FIO_OUTPUT_JSON) && parent)
1123 if (!terse && !do_json)
1124 log_buf(out, "\nDisk stats (read/write):\n");
1127 json_object_add_disk_utils(parent, &disk_list);
1128 } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1129 flist_for_each(entry, &disk_list) {
1130 du = flist_entry(entry, struct disk_util, list);
1132 aggregate_slaves_stats(du);
1133 print_disk_util(&du->dus, &du->agg, terse, out);
1138 static void show_thread_status_normal(struct thread_stat *ts,
1139 struct group_run_stats *rs,
1140 struct buf_output *out)
1142 double usr_cpu, sys_cpu;
1143 unsigned long runtime;
1144 double io_u_dist[FIO_IO_U_MAP_NR];
1148 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1151 memset(time_buf, 0, sizeof(time_buf));
1154 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1157 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1158 ts->name, ts->groupid, ts->members,
1159 ts->error, (int) ts->pid, time_buf);
1161 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1162 ts->name, ts->groupid, ts->members,
1163 ts->error, ts->verror, (int) ts->pid,
1167 if (strlen(ts->description))
1168 log_buf(out, " Description : [%s]\n", ts->description);
1170 for_each_rw_ddir(ddir) {
1171 if (ts->io_bytes[ddir])
1172 show_ddir_status(rs, ts, ddir, out);
1175 if (ts->unified_rw_rep == UNIFIED_BOTH)
1176 show_mixed_ddir_status(rs, ts, out);
1178 show_latencies(ts, out);
1180 if (ts->sync_stat.samples)
1181 show_ddir_status(rs, ts, DDIR_SYNC, out);
1183 runtime = ts->total_run_time;
1185 double runt = (double) runtime;
1187 usr_cpu = (double) ts->usr_time * 100 / runt;
1188 sys_cpu = (double) ts->sys_time * 100 / runt;
1194 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1195 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1196 (unsigned long long) ts->ctx,
1197 (unsigned long long) ts->majf,
1198 (unsigned long long) ts->minf);
1200 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1201 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1202 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1203 io_u_dist[1], io_u_dist[2],
1204 io_u_dist[3], io_u_dist[4],
1205 io_u_dist[5], io_u_dist[6]);
1207 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1208 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1209 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1210 io_u_dist[1], io_u_dist[2],
1211 io_u_dist[3], io_u_dist[4],
1212 io_u_dist[5], io_u_dist[6]);
1213 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1214 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1215 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1216 io_u_dist[1], io_u_dist[2],
1217 io_u_dist[3], io_u_dist[4],
1218 io_u_dist[5], io_u_dist[6]);
1219 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
1220 " short=%llu,%llu,%llu,0"
1221 " dropped=%llu,%llu,%llu,0\n",
1222 (unsigned long long) ts->total_io_u[0],
1223 (unsigned long long) ts->total_io_u[1],
1224 (unsigned long long) ts->total_io_u[2],
1225 (unsigned long long) ts->total_io_u[3],
1226 (unsigned long long) ts->short_io_u[0],
1227 (unsigned long long) ts->short_io_u[1],
1228 (unsigned long long) ts->short_io_u[2],
1229 (unsigned long long) ts->drop_io_u[0],
1230 (unsigned long long) ts->drop_io_u[1],
1231 (unsigned long long) ts->drop_io_u[2]);
1232 if (ts->continue_on_error) {
1233 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
1234 (unsigned long long)ts->total_err_count,
1236 strerror(ts->first_error));
1238 if (ts->latency_depth) {
1239 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1240 (unsigned long long)ts->latency_target,
1241 (unsigned long long)ts->latency_window,
1242 ts->latency_percentile.u.f,
1246 if (ts->nr_block_infos)
1247 show_block_infos(ts->nr_block_infos, ts->block_infos,
1248 ts->percentile_list, out);
1251 show_ss_normal(ts, out);
1254 static void show_ddir_status_terse(struct thread_stat *ts,
1255 struct group_run_stats *rs,
1256 enum fio_ddir ddir, int ver,
1257 struct buf_output *out)
1259 unsigned long long min, max, minv, maxv, bw, iops;
1260 unsigned long long *ovals = NULL;
1265 assert(ddir_rw(ddir));
1268 if (ts->runtime[ddir]) {
1269 uint64_t runt = ts->runtime[ddir];
1271 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1272 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1275 log_buf(out, ";%llu;%llu;%llu;%llu",
1276 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1277 (unsigned long long) ts->runtime[ddir]);
1279 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1280 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1282 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1284 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1285 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1287 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1289 if (ts->lat_percentiles) {
1290 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1291 ts->lat_stat[ddir].samples,
1292 ts->percentile_list, &ovals, &maxv,
1294 } else if (ts->clat_percentiles) {
1295 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1296 ts->clat_stat[ddir].samples,
1297 ts->percentile_list, &ovals, &maxv,
1303 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1305 log_buf(out, ";0%%=0");
1308 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1311 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1312 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1314 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1318 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1320 double p_of_agg = 100.0;
1322 if (rs->agg[ddir]) {
1323 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1324 if (p_of_agg > 100.0)
1328 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1330 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1335 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1337 log_buf(out, ";%lu", 0UL);
1339 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1340 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1341 mean, dev, (&ts->iops_stat[ddir])->samples);
1343 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1347 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1348 struct group_run_stats *rs,
1349 int ver, struct buf_output *out)
1351 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1354 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1359 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1360 uint32_t percentiles,
1361 struct io_stat *lat_stat,
1362 uint64_t *io_u_plat)
1366 unsigned int i, len;
1367 struct json_object *lat_object, *percentile_object, *clat_bins_object;
1368 unsigned long long min, max, maxv, minv, *ovals = NULL;
1370 if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1374 lat_object = json_create_object();
1375 json_object_add_value_int(lat_object, "min", min);
1376 json_object_add_value_int(lat_object, "max", max);
1377 json_object_add_value_float(lat_object, "mean", mean);
1378 json_object_add_value_float(lat_object, "stddev", dev);
1379 json_object_add_value_int(lat_object, "N", lat_stat->samples);
1381 if (percentiles && lat_stat->samples) {
1382 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1383 ts->percentile_list, &ovals, &maxv, &minv);
1385 if (len > FIO_IO_U_LIST_MAX_LEN)
1386 len = FIO_IO_U_LIST_MAX_LEN;
1388 percentile_object = json_create_object();
1389 json_object_add_value_object(lat_object, "percentile", percentile_object);
1390 for (i = 0; i < len; i++) {
1391 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1392 json_object_add_value_int(percentile_object, buf, ovals[i]);
1396 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1397 clat_bins_object = json_create_object();
1398 json_object_add_value_object(lat_object, "bins", clat_bins_object);
1400 for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1402 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1403 json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1411 static void add_ddir_status_json(struct thread_stat *ts,
1412 struct group_run_stats *rs, enum fio_ddir ddir,
1413 struct json_object *parent)
1415 unsigned long long min, max;
1416 unsigned long long bw_bytes, bw;
1417 double mean, dev, iops;
1418 struct json_object *dir_object, *tmp_object;
1419 double p_of_agg = 100.0;
1421 assert(ddir_rw(ddir) || ddir_sync(ddir));
1423 if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1426 dir_object = json_create_object();
1427 json_object_add_value_object(parent,
1428 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1430 if (ddir_rw(ddir)) {
1434 if (ts->runtime[ddir]) {
1435 uint64_t runt = ts->runtime[ddir];
1437 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1438 bw = bw_bytes / 1024; /* KiB/s */
1439 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1442 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1443 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1444 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1445 json_object_add_value_int(dir_object, "bw", bw);
1446 json_object_add_value_float(dir_object, "iops", iops);
1447 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1448 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1449 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1450 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1452 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1453 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1454 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1456 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1457 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1458 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1460 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1461 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1462 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1464 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1465 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1466 &ts->sync_stat, ts->io_u_sync_plat);
1467 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1473 /* Only print PRIO latencies if some high priority samples were gathered */
1474 if (ts->clat_high_prio_stat[ddir].samples > 0) {
1475 const char *high, *low;
1477 if (ts->lat_percentiles) {
1478 high = "lat_high_prio";
1479 low = "lat_low_prio";
1481 high = "clat_high_prio";
1482 low = "clat_low_prio";
1485 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1486 &ts->clat_high_prio_stat[ddir], ts->io_u_plat_high_prio[ddir]);
1487 json_object_add_value_object(dir_object, high, tmp_object);
1489 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1490 &ts->clat_low_prio_stat[ddir], ts->io_u_plat_low_prio[ddir]);
1491 json_object_add_value_object(dir_object, low, tmp_object);
1494 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1495 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1498 p_of_agg = mean = dev = 0.0;
1501 json_object_add_value_int(dir_object, "bw_min", min);
1502 json_object_add_value_int(dir_object, "bw_max", max);
1503 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1504 json_object_add_value_float(dir_object, "bw_mean", mean);
1505 json_object_add_value_float(dir_object, "bw_dev", dev);
1506 json_object_add_value_int(dir_object, "bw_samples",
1507 (&ts->bw_stat[ddir])->samples);
1509 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1513 json_object_add_value_int(dir_object, "iops_min", min);
1514 json_object_add_value_int(dir_object, "iops_max", max);
1515 json_object_add_value_float(dir_object, "iops_mean", mean);
1516 json_object_add_value_float(dir_object, "iops_stddev", dev);
1517 json_object_add_value_int(dir_object, "iops_samples",
1518 (&ts->iops_stat[ddir])->samples);
1520 if (ts->cachehit + ts->cachemiss) {
1524 total = ts->cachehit + ts->cachemiss;
1525 hit = (double) ts->cachehit / (double) total;
1527 json_object_add_value_float(dir_object, "cachehit", hit);
1531 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1532 struct group_run_stats *rs, struct json_object *parent)
1534 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1536 /* add the aggregated stats to json parent */
1538 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1543 static void show_thread_status_terse_all(struct thread_stat *ts,
1544 struct group_run_stats *rs, int ver,
1545 struct buf_output *out)
1547 double io_u_dist[FIO_IO_U_MAP_NR];
1548 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1549 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1550 double usr_cpu, sys_cpu;
1555 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1557 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1558 ts->name, ts->groupid, ts->error);
1560 /* Log Read Status, or mixed if unified_rw_rep = 1 */
1561 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1562 if (ts->unified_rw_rep != UNIFIED_MIXED) {
1563 /* Log Write Status */
1564 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1565 /* Log Trim Status */
1566 if (ver == 2 || ver == 4 || ver == 5)
1567 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1569 if (ts->unified_rw_rep == UNIFIED_BOTH)
1570 show_mixed_ddir_status_terse(ts, rs, ver, out);
1572 if (ts->total_run_time) {
1573 double runt = (double) ts->total_run_time;
1575 usr_cpu = (double) ts->usr_time * 100 / runt;
1576 sys_cpu = (double) ts->sys_time * 100 / runt;
1582 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1583 (unsigned long long) ts->ctx,
1584 (unsigned long long) ts->majf,
1585 (unsigned long long) ts->minf);
1587 /* Calc % distribution of IO depths, usecond, msecond latency */
1588 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1589 stat_calc_lat_nu(ts, io_u_lat_u);
1590 stat_calc_lat_m(ts, io_u_lat_m);
1592 /* Only show fixed 7 I/O depth levels*/
1593 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1594 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1595 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1597 /* Microsecond latency */
1598 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1599 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1600 /* Millisecond latency */
1601 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1602 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1604 /* disk util stats, if any */
1605 if (ver >= 3 && is_running_backend())
1606 show_disk_util(1, NULL, out);
1608 /* Additional output if continue_on_error set - default off*/
1609 if (ts->continue_on_error)
1610 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1612 /* Additional output if description is set */
1613 if (strlen(ts->description)) {
1616 log_buf(out, ";%s", ts->description);
1622 static void json_add_job_opts(struct json_object *root, const char *name,
1623 struct flist_head *opt_list)
1625 struct json_object *dir_object;
1626 struct flist_head *entry;
1627 struct print_option *p;
1629 if (flist_empty(opt_list))
1632 dir_object = json_create_object();
1633 json_object_add_value_object(root, name, dir_object);
1635 flist_for_each(entry, opt_list) {
1636 p = flist_entry(entry, struct print_option, list);
1637 json_object_add_value_string(dir_object, p->name, p->value);
1641 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1642 struct group_run_stats *rs,
1643 struct flist_head *opt_list)
1645 struct json_object *root, *tmp;
1646 struct jobs_eta *je;
1647 double io_u_dist[FIO_IO_U_MAP_NR];
1648 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1649 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1650 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1651 double usr_cpu, sys_cpu;
1655 root = json_create_object();
1656 json_object_add_value_string(root, "jobname", ts->name);
1657 json_object_add_value_int(root, "groupid", ts->groupid);
1658 json_object_add_value_int(root, "error", ts->error);
1661 je = get_jobs_eta(true, &size);
1663 json_object_add_value_int(root, "eta", je->eta_sec);
1664 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1668 json_add_job_opts(root, "job options", opt_list);
1670 add_ddir_status_json(ts, rs, DDIR_READ, root);
1671 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1672 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1673 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1675 if (ts->unified_rw_rep == UNIFIED_BOTH)
1676 add_mixed_ddir_status_json(ts, rs, root);
1679 if (ts->total_run_time) {
1680 double runt = (double) ts->total_run_time;
1682 usr_cpu = (double) ts->usr_time * 100 / runt;
1683 sys_cpu = (double) ts->sys_time * 100 / runt;
1688 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1689 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1690 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1691 json_object_add_value_int(root, "ctx", ts->ctx);
1692 json_object_add_value_int(root, "majf", ts->majf);
1693 json_object_add_value_int(root, "minf", ts->minf);
1695 /* Calc % distribution of IO depths */
1696 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1697 tmp = json_create_object();
1698 json_object_add_value_object(root, "iodepth_level", tmp);
1699 /* Only show fixed 7 I/O depth levels*/
1700 for (i = 0; i < 7; i++) {
1703 snprintf(name, 20, "%d", 1 << i);
1705 snprintf(name, 20, ">=%d", 1 << i);
1706 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1709 /* Calc % distribution of submit IO depths */
1710 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1711 tmp = json_create_object();
1712 json_object_add_value_object(root, "iodepth_submit", tmp);
1713 /* Only show fixed 7 I/O depth levels*/
1714 for (i = 0; i < 7; i++) {
1717 snprintf(name, 20, "0");
1719 snprintf(name, 20, "%d", 1 << (i+1));
1721 snprintf(name, 20, ">=%d", 1 << i);
1722 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1725 /* Calc % distribution of completion IO depths */
1726 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1727 tmp = json_create_object();
1728 json_object_add_value_object(root, "iodepth_complete", tmp);
1729 /* Only show fixed 7 I/O depth levels*/
1730 for (i = 0; i < 7; i++) {
1733 snprintf(name, 20, "0");
1735 snprintf(name, 20, "%d", 1 << (i+1));
1737 snprintf(name, 20, ">=%d", 1 << i);
1738 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1741 /* Calc % distribution of nsecond, usecond, msecond latency */
1742 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1743 stat_calc_lat_n(ts, io_u_lat_n);
1744 stat_calc_lat_u(ts, io_u_lat_u);
1745 stat_calc_lat_m(ts, io_u_lat_m);
1747 /* Nanosecond latency */
1748 tmp = json_create_object();
1749 json_object_add_value_object(root, "latency_ns", tmp);
1750 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1751 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1752 "250", "500", "750", "1000", };
1753 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1755 /* Microsecond latency */
1756 tmp = json_create_object();
1757 json_object_add_value_object(root, "latency_us", tmp);
1758 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1759 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1760 "250", "500", "750", "1000", };
1761 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1763 /* Millisecond latency */
1764 tmp = json_create_object();
1765 json_object_add_value_object(root, "latency_ms", tmp);
1766 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1767 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1768 "250", "500", "750", "1000", "2000",
1770 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1773 /* Additional output if continue_on_error set - default off*/
1774 if (ts->continue_on_error) {
1775 json_object_add_value_int(root, "total_err", ts->total_err_count);
1776 json_object_add_value_int(root, "first_error", ts->first_error);
1779 if (ts->latency_depth) {
1780 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1781 json_object_add_value_int(root, "latency_target", ts->latency_target);
1782 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1783 json_object_add_value_int(root, "latency_window", ts->latency_window);
1786 /* Additional output if description is set */
1787 if (strlen(ts->description))
1788 json_object_add_value_string(root, "desc", ts->description);
1790 if (ts->nr_block_infos) {
1791 /* Block error histogram and types */
1793 unsigned int *percentiles = NULL;
1794 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1796 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1797 ts->percentile_list,
1798 &percentiles, block_state_counts);
1801 struct json_object *block, *percentile_object, *states;
1803 block = json_create_object();
1804 json_object_add_value_object(root, "block", block);
1806 percentile_object = json_create_object();
1807 json_object_add_value_object(block, "percentiles",
1809 for (i = 0; i < len; i++) {
1811 snprintf(buf, sizeof(buf), "%f",
1812 ts->percentile_list[i].u.f);
1813 json_object_add_value_int(percentile_object,
1818 states = json_create_object();
1819 json_object_add_value_object(block, "states", states);
1820 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1821 json_object_add_value_int(states,
1822 block_state_names[state],
1823 block_state_counts[state]);
1830 struct json_object *data;
1831 struct json_array *iops, *bw;
1835 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1836 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1837 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1838 (float) ts->ss_limit.u.f,
1839 ts->ss_state & FIO_SS_PCT ? "%" : "");
1841 tmp = json_create_object();
1842 json_object_add_value_object(root, "steadystate", tmp);
1843 json_object_add_value_string(tmp, "ss", ss_buf);
1844 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1845 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1847 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1848 ts->ss_state & FIO_SS_PCT ? "%" : "");
1849 json_object_add_value_string(tmp, "criterion", ss_buf);
1850 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1851 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1853 data = json_create_object();
1854 json_object_add_value_object(tmp, "data", data);
1855 bw = json_create_array();
1856 iops = json_create_array();
1859 ** if ss was attained or the buffer is not full,
1860 ** ss->head points to the first element in the list.
1861 ** otherwise it actually points to the second element
1864 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1867 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1868 for (l = 0; l < ts->ss_dur; l++) {
1869 k = (j + l) % ts->ss_dur;
1870 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1871 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1873 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1874 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1875 json_object_add_value_array(data, "iops", iops);
1876 json_object_add_value_array(data, "bw", bw);
1882 static void show_thread_status_terse(struct thread_stat *ts,
1883 struct group_run_stats *rs,
1884 struct buf_output *out)
1886 if (terse_version >= 2 && terse_version <= 5)
1887 show_thread_status_terse_all(ts, rs, terse_version, out);
1889 log_err("fio: bad terse version!? %d\n", terse_version);
1892 struct json_object *show_thread_status(struct thread_stat *ts,
1893 struct group_run_stats *rs,
1894 struct flist_head *opt_list,
1895 struct buf_output *out)
1897 struct json_object *ret = NULL;
1899 if (output_format & FIO_OUTPUT_TERSE)
1900 show_thread_status_terse(ts, rs, out);
1901 if (output_format & FIO_OUTPUT_JSON)
1902 ret = show_thread_status_json(ts, rs, opt_list);
1903 if (output_format & FIO_OUTPUT_NORMAL)
1904 show_thread_status_normal(ts, rs, out);
1909 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1913 dst->min_val = min(dst->min_val, src->min_val);
1914 dst->max_val = max(dst->max_val, src->max_val);
1917 * Compute new mean and S after the merge
1918 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1919 * #Parallel_algorithm>
1922 mean = src->mean.u.f;
1925 double delta = src->mean.u.f - dst->mean.u.f;
1927 mean = ((src->mean.u.f * src->samples) +
1928 (dst->mean.u.f * dst->samples)) /
1929 (dst->samples + src->samples);
1931 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1932 (dst->samples * src->samples) /
1933 (dst->samples + src->samples);
1936 dst->samples += src->samples;
1937 dst->mean.u.f = mean;
1943 * We sum two kinds of stats - one that is time based, in which case we
1944 * apply the proper summing technique, and then one that is iops/bw
1945 * numbers. For group_reporting, we should just add those up, not make
1946 * them the mean of everything.
1948 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
1950 bool first = dst->samples == 0;
1952 if (src->samples == 0)
1956 __sum_stat(dst, src, first);
1961 dst->min_val = src->min_val;
1962 dst->max_val = src->max_val;
1963 dst->samples = src->samples;
1964 dst->mean.u.f = src->mean.u.f;
1965 dst->S.u.f = src->S.u.f;
1967 dst->min_val += src->min_val;
1968 dst->max_val += src->max_val;
1969 dst->samples += src->samples;
1970 dst->mean.u.f += src->mean.u.f;
1971 dst->S.u.f += src->S.u.f;
1975 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1979 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1980 if (dst->max_run[i] < src->max_run[i])
1981 dst->max_run[i] = src->max_run[i];
1982 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1983 dst->min_run[i] = src->min_run[i];
1984 if (dst->max_bw[i] < src->max_bw[i])
1985 dst->max_bw[i] = src->max_bw[i];
1986 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1987 dst->min_bw[i] = src->min_bw[i];
1989 dst->iobytes[i] += src->iobytes[i];
1990 dst->agg[i] += src->agg[i];
1994 dst->kb_base = src->kb_base;
1995 if (!dst->unit_base)
1996 dst->unit_base = src->unit_base;
1998 dst->sig_figs = src->sig_figs;
2002 * Free the clat_prio_stat arrays allocated by alloc_clat_prio_stat_ddir().
2004 void free_clat_prio_stats(struct thread_stat *ts)
2008 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2009 sfree(ts->clat_prio[ddir]);
2010 ts->clat_prio[ddir] = NULL;
2011 ts->nr_clat_prio[ddir] = 0;
2016 * Allocate a clat_prio_stat array. The array has to be allocated/freed using
2017 * smalloc/sfree, so that it is accessible by the process/thread summing the
2020 int alloc_clat_prio_stat_ddir(struct thread_stat *ts, enum fio_ddir ddir,
2023 struct clat_prio_stat *clat_prio;
2026 clat_prio = scalloc(nr_prios, sizeof(*ts->clat_prio[ddir]));
2028 log_err("fio: failed to allocate ts clat data\n");
2032 for (i = 0; i < nr_prios; i++)
2033 clat_prio[i].clat_stat.min_val = ULONG_MAX;
2035 ts->clat_prio[ddir] = clat_prio;
2036 ts->nr_clat_prio[ddir] = nr_prios;
2041 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2045 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2046 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2047 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2048 sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], false);
2049 sum_stat(&dst->clat_low_prio_stat[l], &src->clat_low_prio_stat[l], false);
2050 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2051 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2052 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2053 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2055 dst->io_bytes[l] += src->io_bytes[l];
2057 if (dst->runtime[l] < src->runtime[l])
2058 dst->runtime[l] = src->runtime[l];
2060 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2061 sum_stat(&dst->clat_high_prio_stat[0], &src->clat_high_prio_stat[l], false);
2062 sum_stat(&dst->clat_low_prio_stat[0], &src->clat_low_prio_stat[l], false);
2063 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2064 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2065 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2066 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2068 dst->io_bytes[0] += src->io_bytes[l];
2070 if (dst->runtime[0] < src->runtime[l])
2071 dst->runtime[0] = src->runtime[l];
2075 sum_stat(&dst->sync_stat, &src->sync_stat, false);
2076 dst->usr_time += src->usr_time;
2077 dst->sys_time += src->sys_time;
2078 dst->ctx += src->ctx;
2079 dst->majf += src->majf;
2080 dst->minf += src->minf;
2082 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2083 dst->io_u_map[k] += src->io_u_map[k];
2084 dst->io_u_submit[k] += src->io_u_submit[k];
2085 dst->io_u_complete[k] += src->io_u_complete[k];
2088 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2089 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2090 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2091 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2092 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2093 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2095 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2096 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2097 dst->total_io_u[k] += src->total_io_u[k];
2098 dst->short_io_u[k] += src->short_io_u[k];
2099 dst->drop_io_u[k] += src->drop_io_u[k];
2101 dst->total_io_u[0] += src->total_io_u[k];
2102 dst->short_io_u[0] += src->short_io_u[k];
2103 dst->drop_io_u[0] += src->drop_io_u[k];
2107 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2109 for (k = 0; k < FIO_LAT_CNT; k++)
2110 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2111 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2112 if (dst->unified_rw_rep != UNIFIED_MIXED)
2113 dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2115 dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2117 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2118 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2120 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2121 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
2122 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2123 dst->io_u_plat_high_prio[k][m] += src->io_u_plat_high_prio[k][m];
2124 dst->io_u_plat_low_prio[k][m] += src->io_u_plat_low_prio[k][m];
2126 dst->io_u_plat_high_prio[0][m] += src->io_u_plat_high_prio[k][m];
2127 dst->io_u_plat_low_prio[0][m] += src->io_u_plat_low_prio[k][m];
2133 dst->total_run_time += src->total_run_time;
2134 dst->total_submit += src->total_submit;
2135 dst->total_complete += src->total_complete;
2136 dst->nr_zone_resets += src->nr_zone_resets;
2137 dst->cachehit += src->cachehit;
2138 dst->cachemiss += src->cachemiss;
2141 void init_group_run_stat(struct group_run_stats *gs)
2144 memset(gs, 0, sizeof(*gs));
2146 for (i = 0; i < DDIR_RWDIR_CNT; i++)
2147 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2150 void init_thread_stat_min_vals(struct thread_stat *ts)
2154 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2155 ts->clat_stat[i].min_val = ULONG_MAX;
2156 ts->slat_stat[i].min_val = ULONG_MAX;
2157 ts->lat_stat[i].min_val = ULONG_MAX;
2158 ts->bw_stat[i].min_val = ULONG_MAX;
2159 ts->iops_stat[i].min_val = ULONG_MAX;
2160 ts->clat_high_prio_stat[i].min_val = ULONG_MAX;
2161 ts->clat_low_prio_stat[i].min_val = ULONG_MAX;
2163 ts->sync_stat.min_val = ULONG_MAX;
2166 void init_thread_stat(struct thread_stat *ts)
2168 memset(ts, 0, sizeof(*ts));
2170 init_thread_stat_min_vals(ts);
2174 void __show_run_stats(void)
2176 struct group_run_stats *runstats, *rs;
2177 struct thread_data *td;
2178 struct thread_stat *threadstats, *ts;
2179 int i, j, k, nr_ts, last_ts, idx;
2180 bool kb_base_warned = false;
2181 bool unit_base_warned = false;
2182 struct json_object *root = NULL;
2183 struct json_array *array = NULL;
2184 struct buf_output output[FIO_OUTPUT_NR];
2185 struct flist_head **opt_lists;
2187 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2189 for (i = 0; i < groupid + 1; i++)
2190 init_group_run_stat(&runstats[i]);
2193 * find out how many threads stats we need. if group reporting isn't
2194 * enabled, it's one-per-td.
2198 for_each_td(td, i) {
2199 if (!td->o.group_reporting) {
2203 if (last_ts == td->groupid)
2208 last_ts = td->groupid;
2212 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2213 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2215 for (i = 0; i < nr_ts; i++) {
2216 init_thread_stat(&threadstats[i]);
2217 opt_lists[i] = NULL;
2223 for_each_td(td, i) {
2226 if (idx && (!td->o.group_reporting ||
2227 (td->o.group_reporting && last_ts != td->groupid))) {
2232 last_ts = td->groupid;
2234 ts = &threadstats[j];
2236 ts->clat_percentiles = td->o.clat_percentiles;
2237 ts->lat_percentiles = td->o.lat_percentiles;
2238 ts->slat_percentiles = td->o.slat_percentiles;
2239 ts->percentile_precision = td->o.percentile_precision;
2240 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2241 opt_lists[j] = &td->opt_list;
2245 if (ts->groupid == -1) {
2247 * These are per-group shared already
2249 snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2250 if (td->o.description)
2251 snprintf(ts->description,
2252 sizeof(ts->description), "%s",
2255 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2258 * If multiple entries in this group, this is
2261 ts->thread_number = td->thread_number;
2262 ts->groupid = td->groupid;
2265 * first pid in group, not very useful...
2269 ts->kb_base = td->o.kb_base;
2270 ts->unit_base = td->o.unit_base;
2271 ts->sig_figs = td->o.sig_figs;
2272 ts->unified_rw_rep = td->o.unified_rw_rep;
2273 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2274 log_info("fio: kb_base differs for jobs in group, using"
2275 " %u as the base\n", ts->kb_base);
2276 kb_base_warned = true;
2277 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2278 log_info("fio: unit_base differs for jobs in group, using"
2279 " %u as the base\n", ts->unit_base);
2280 unit_base_warned = true;
2283 ts->continue_on_error = td->o.continue_on_error;
2284 ts->total_err_count += td->total_err_count;
2285 ts->first_error = td->first_error;
2287 if (!td->error && td->o.continue_on_error &&
2289 ts->error = td->first_error;
2290 snprintf(ts->verror, sizeof(ts->verror), "%s",
2292 } else if (td->error) {
2293 ts->error = td->error;
2294 snprintf(ts->verror, sizeof(ts->verror), "%s",
2299 ts->latency_depth = td->latency_qd;
2300 ts->latency_target = td->o.latency_target;
2301 ts->latency_percentile = td->o.latency_percentile;
2302 ts->latency_window = td->o.latency_window;
2304 ts->nr_block_infos = td->ts.nr_block_infos;
2305 for (k = 0; k < ts->nr_block_infos; k++)
2306 ts->block_infos[k] = td->ts.block_infos[k];
2308 sum_thread_stats(ts, &td->ts);
2313 ts->ss_state = td->ss.state;
2314 ts->ss_dur = td->ss.dur;
2315 ts->ss_head = td->ss.head;
2316 ts->ss_bw_data = td->ss.bw_data;
2317 ts->ss_iops_data = td->ss.iops_data;
2318 ts->ss_limit.u.f = td->ss.limit;
2319 ts->ss_slope.u.f = td->ss.slope;
2320 ts->ss_deviation.u.f = td->ss.deviation;
2321 ts->ss_criterion.u.f = td->ss.criterion;
2324 ts->ss_dur = ts->ss_state = 0;
2327 for (i = 0; i < nr_ts; i++) {
2328 unsigned long long bw;
2330 ts = &threadstats[i];
2331 if (ts->groupid == -1)
2333 rs = &runstats[ts->groupid];
2334 rs->kb_base = ts->kb_base;
2335 rs->unit_base = ts->unit_base;
2336 rs->sig_figs = ts->sig_figs;
2337 rs->unified_rw_rep |= ts->unified_rw_rep;
2339 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2340 if (!ts->runtime[j])
2342 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2343 rs->min_run[j] = ts->runtime[j];
2344 if (ts->runtime[j] > rs->max_run[j])
2345 rs->max_run[j] = ts->runtime[j];
2349 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2350 if (bw < rs->min_bw[j])
2352 if (bw > rs->max_bw[j])
2355 rs->iobytes[j] += ts->io_bytes[j];
2359 for (i = 0; i < groupid + 1; i++) {
2364 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2365 if (rs->max_run[ddir])
2366 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2371 for (i = 0; i < FIO_OUTPUT_NR; i++)
2372 buf_output_init(&output[i]);
2375 * don't overwrite last signal output
2377 if (output_format & FIO_OUTPUT_NORMAL)
2378 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2379 if (output_format & FIO_OUTPUT_JSON) {
2380 struct thread_data *global;
2383 unsigned long long ms_since_epoch;
2386 gettimeofday(&now, NULL);
2387 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2388 (unsigned long long)(now.tv_usec) / 1000;
2390 tv_sec = now.tv_sec;
2391 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2392 if (time_buf[strlen(time_buf) - 1] == '\n')
2393 time_buf[strlen(time_buf) - 1] = '\0';
2395 root = json_create_object();
2396 json_object_add_value_string(root, "fio version", fio_version_string);
2397 json_object_add_value_int(root, "timestamp", now.tv_sec);
2398 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2399 json_object_add_value_string(root, "time", time_buf);
2400 global = get_global_options();
2401 json_add_job_opts(root, "global options", &global->opt_list);
2402 array = json_create_array();
2403 json_object_add_value_array(root, "jobs", array);
2407 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2409 for (i = 0; i < nr_ts; i++) {
2410 ts = &threadstats[i];
2411 rs = &runstats[ts->groupid];
2414 fio_server_send_job_options(opt_lists[i], i);
2415 fio_server_send_ts(ts, rs);
2417 if (output_format & FIO_OUTPUT_TERSE)
2418 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2419 if (output_format & FIO_OUTPUT_JSON) {
2420 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2421 json_array_add_value_object(array, tmp);
2423 if (output_format & FIO_OUTPUT_NORMAL)
2424 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2427 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2428 /* disk util stats, if any */
2429 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2431 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2433 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2434 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2435 json_free_object(root);
2438 for (i = 0; i < groupid + 1; i++) {
2443 fio_server_send_gs(rs);
2444 else if (output_format & FIO_OUTPUT_NORMAL)
2445 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2449 fio_server_send_du();
2450 else if (output_format & FIO_OUTPUT_NORMAL) {
2451 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2452 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2455 for (i = 0; i < FIO_OUTPUT_NR; i++) {
2456 struct buf_output *out = &output[i];
2458 log_info_buf(out->buf, out->buflen);
2459 buf_output_free(out);
2462 fio_idle_prof_cleanup();
2470 int __show_running_run_stats(void)
2472 struct thread_data *td;
2473 unsigned long long *rt;
2477 fio_sem_down(stat_sem);
2479 rt = malloc(thread_number * sizeof(unsigned long long));
2480 fio_gettime(&ts, NULL);
2482 for_each_td(td, i) {
2483 td->update_rusage = 1;
2484 for_each_rw_ddir(ddir) {
2485 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2487 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2489 rt[i] = mtime_since(&td->start, &ts);
2490 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2491 td->ts.runtime[DDIR_READ] += rt[i];
2492 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2493 td->ts.runtime[DDIR_WRITE] += rt[i];
2494 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2495 td->ts.runtime[DDIR_TRIM] += rt[i];
2498 for_each_td(td, i) {
2499 if (td->runstate >= TD_EXITED)
2501 if (td->rusage_sem) {
2502 td->update_rusage = 1;
2503 fio_sem_down(td->rusage_sem);
2505 td->update_rusage = 0;
2510 for_each_td(td, i) {
2511 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2512 td->ts.runtime[DDIR_READ] -= rt[i];
2513 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2514 td->ts.runtime[DDIR_WRITE] -= rt[i];
2515 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2516 td->ts.runtime[DDIR_TRIM] -= rt[i];
2520 fio_sem_up(stat_sem);
2525 static bool status_file_disabled;
2527 #define FIO_STATUS_FILE "fio-dump-status"
2529 static int check_status_file(void)
2532 const char *temp_dir;
2533 char fio_status_file_path[PATH_MAX];
2535 if (status_file_disabled)
2538 temp_dir = getenv("TMPDIR");
2539 if (temp_dir == NULL) {
2540 temp_dir = getenv("TEMP");
2541 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2544 if (temp_dir == NULL)
2547 __coverity_tainted_data_sanitize__(temp_dir);
2550 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2552 if (stat(fio_status_file_path, &sb))
2555 if (unlink(fio_status_file_path) < 0) {
2556 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2558 log_err("fio: disabling status file updates\n");
2559 status_file_disabled = true;
2565 void check_for_running_stats(void)
2567 if (check_status_file()) {
2568 show_running_run_stats();
2573 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2578 if (data > is->max_val)
2580 if (data < is->min_val)
2583 delta = val - is->mean.u.f;
2585 is->mean.u.f += delta / (is->samples + 1.0);
2586 is->S.u.f += delta * (val - is->mean.u.f);
2593 * Return a struct io_logs, which is added to the tail of the log
2596 static struct io_logs *get_new_log(struct io_log *iolog)
2599 struct io_logs *cur_log;
2602 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2605 if (!iolog->cur_log_max) {
2606 new_samples = iolog->td->o.log_entries;
2608 new_samples = iolog->cur_log_max * 2;
2609 if (new_samples > MAX_LOG_ENTRIES)
2610 new_samples = MAX_LOG_ENTRIES;
2613 cur_log = smalloc(sizeof(*cur_log));
2615 INIT_FLIST_HEAD(&cur_log->list);
2616 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2618 cur_log->nr_samples = 0;
2619 cur_log->max_samples = new_samples;
2620 flist_add_tail(&cur_log->list, &iolog->io_logs);
2621 iolog->cur_log_max = new_samples;
2631 * Add and return a new log chunk, or return current log if big enough
2633 static struct io_logs *regrow_log(struct io_log *iolog)
2635 struct io_logs *cur_log;
2638 if (!iolog || iolog->disabled)
2641 cur_log = iolog_cur_log(iolog);
2643 cur_log = get_new_log(iolog);
2648 if (cur_log->nr_samples < cur_log->max_samples)
2652 * No room for a new sample. If we're compressing on the fly, flush
2653 * out the current chunk
2655 if (iolog->log_gz) {
2656 if (iolog_cur_flush(iolog, cur_log)) {
2657 log_err("fio: failed flushing iolog! Will stop logging.\n");
2663 * Get a new log array, and add to our list
2665 cur_log = get_new_log(iolog);
2667 log_err("fio: failed extending iolog! Will stop logging.\n");
2671 if (!iolog->pending || !iolog->pending->nr_samples)
2675 * Flush pending items to new log
2677 for (i = 0; i < iolog->pending->nr_samples; i++) {
2678 struct io_sample *src, *dst;
2680 src = get_sample(iolog, iolog->pending, i);
2681 dst = get_sample(iolog, cur_log, i);
2682 memcpy(dst, src, log_entry_sz(iolog));
2684 cur_log->nr_samples = iolog->pending->nr_samples;
2686 iolog->pending->nr_samples = 0;
2690 iolog->disabled = true;
2694 void regrow_logs(struct thread_data *td)
2696 regrow_log(td->slat_log);
2697 regrow_log(td->clat_log);
2698 regrow_log(td->clat_hist_log);
2699 regrow_log(td->lat_log);
2700 regrow_log(td->bw_log);
2701 regrow_log(td->iops_log);
2702 td->flags &= ~TD_F_REGROW_LOGS;
2705 void regrow_agg_logs(void)
2709 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2710 regrow_log(agg_io_log[ddir]);
2713 static struct io_logs *get_cur_log(struct io_log *iolog)
2715 struct io_logs *cur_log;
2717 cur_log = iolog_cur_log(iolog);
2719 cur_log = get_new_log(iolog);
2724 if (cur_log->nr_samples < cur_log->max_samples)
2728 * Out of space. If we're in IO offload mode, or we're not doing
2729 * per unit logging (hence logging happens outside of the IO thread
2730 * as well), add a new log chunk inline. If we're doing inline
2731 * submissions, flag 'td' as needing a log regrow and we'll take
2732 * care of it on the submission side.
2734 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2735 !per_unit_log(iolog))
2736 return regrow_log(iolog);
2739 iolog->td->flags |= TD_F_REGROW_LOGS;
2741 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2742 return iolog->pending;
2745 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2746 enum fio_ddir ddir, unsigned long long bs,
2747 unsigned long t, uint64_t offset,
2748 unsigned int priority)
2750 struct io_logs *cur_log;
2752 if (iolog->disabled)
2754 if (flist_empty(&iolog->io_logs))
2755 iolog->avg_last[ddir] = t;
2757 cur_log = get_cur_log(iolog);
2759 struct io_sample *s;
2761 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2764 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2765 io_sample_set_ddir(iolog, s, ddir);
2767 s->priority = priority;
2769 if (iolog->log_offset) {
2770 struct io_sample_offset *so = (void *) s;
2772 so->offset = offset;
2775 cur_log->nr_samples++;
2779 iolog->disabled = true;
2782 static inline void reset_io_stat(struct io_stat *ios)
2784 ios->min_val = -1ULL;
2785 ios->max_val = ios->samples = 0;
2786 ios->mean.u.f = ios->S.u.f = 0;
2789 static inline void reset_io_u_plat(uint64_t *io_u_plat)
2793 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
2797 void reset_io_stats(struct thread_data *td)
2799 struct thread_stat *ts = &td->ts;
2802 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2803 reset_io_stat(&ts->clat_high_prio_stat[i]);
2804 reset_io_stat(&ts->clat_low_prio_stat[i]);
2805 reset_io_stat(&ts->clat_stat[i]);
2806 reset_io_stat(&ts->slat_stat[i]);
2807 reset_io_stat(&ts->lat_stat[i]);
2808 reset_io_stat(&ts->bw_stat[i]);
2809 reset_io_stat(&ts->iops_stat[i]);
2811 ts->io_bytes[i] = 0;
2813 ts->total_io_u[i] = 0;
2814 ts->short_io_u[i] = 0;
2815 ts->drop_io_u[i] = 0;
2817 reset_io_u_plat(ts->io_u_plat_high_prio[i]);
2818 reset_io_u_plat(ts->io_u_plat_low_prio[i]);
2821 for (i = 0; i < FIO_LAT_CNT; i++)
2822 for (j = 0; j < DDIR_RWDIR_CNT; j++)
2823 reset_io_u_plat(ts->io_u_plat[i][j]);
2825 ts->total_io_u[DDIR_SYNC] = 0;
2826 reset_io_u_plat(ts->io_u_sync_plat);
2828 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2829 ts->io_u_map[i] = 0;
2830 ts->io_u_submit[i] = 0;
2831 ts->io_u_complete[i] = 0;
2834 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2835 ts->io_u_lat_n[i] = 0;
2836 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2837 ts->io_u_lat_u[i] = 0;
2838 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2839 ts->io_u_lat_m[i] = 0;
2841 ts->total_submit = 0;
2842 ts->total_complete = 0;
2843 ts->nr_zone_resets = 0;
2844 ts->cachehit = ts->cachemiss = 0;
2847 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2848 unsigned long elapsed, bool log_max)
2851 * Note an entry in the log. Use the mean from the logged samples,
2852 * making sure to properly round up. Only write a log entry if we
2853 * had actual samples done.
2855 if (iolog->avg_window[ddir].samples) {
2856 union io_sample_data data;
2859 data.val = iolog->avg_window[ddir].max_val;
2861 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2863 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
2866 reset_io_stat(&iolog->avg_window[ddir]);
2869 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2874 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2875 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2878 static unsigned long add_log_sample(struct thread_data *td,
2879 struct io_log *iolog,
2880 union io_sample_data data,
2881 enum fio_ddir ddir, unsigned long long bs,
2882 uint64_t offset, unsigned int ioprio)
2884 unsigned long elapsed, this_window;
2889 elapsed = mtime_since_now(&td->epoch);
2892 * If no time averaging, just add the log sample.
2894 if (!iolog->avg_msec) {
2895 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
2901 * Add the sample. If the time period has passed, then
2902 * add that entry to the log and clear.
2904 add_stat_sample(&iolog->avg_window[ddir], data.val);
2907 * If period hasn't passed, adding the above sample is all we
2910 this_window = elapsed - iolog->avg_last[ddir];
2911 if (elapsed < iolog->avg_last[ddir])
2912 return iolog->avg_last[ddir] - elapsed;
2913 else if (this_window < iolog->avg_msec) {
2914 unsigned long diff = iolog->avg_msec - this_window;
2916 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2920 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2922 iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
2924 return iolog->avg_msec;
2927 void finalize_logs(struct thread_data *td, bool unit_logs)
2929 unsigned long elapsed;
2931 elapsed = mtime_since_now(&td->epoch);
2933 if (td->clat_log && unit_logs)
2934 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2935 if (td->slat_log && unit_logs)
2936 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2937 if (td->lat_log && unit_logs)
2938 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2939 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2940 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2941 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2942 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2945 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
2946 unsigned long long bs)
2948 struct io_log *iolog;
2953 iolog = agg_io_log[ddir];
2954 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
2957 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2959 unsigned int idx = plat_val_to_idx(nsec);
2960 assert(idx < FIO_IO_U_PLAT_NR);
2962 ts->io_u_sync_plat[idx]++;
2963 add_stat_sample(&ts->sync_stat, nsec);
2966 static inline void add_lat_percentile_sample(struct thread_stat *ts,
2967 unsigned long long nsec,
2971 unsigned int idx = plat_val_to_idx(nsec);
2972 assert(idx < FIO_IO_U_PLAT_NR);
2974 ts->io_u_plat[lat][ddir][idx]++;
2977 static inline void add_lat_percentile_prio_sample(struct thread_stat *ts,
2978 unsigned long long nsec,
2982 unsigned int idx = plat_val_to_idx(nsec);
2985 ts->io_u_plat_low_prio[ddir][idx]++;
2987 ts->io_u_plat_high_prio[ddir][idx]++;
2990 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2991 unsigned long long nsec, unsigned long long bs,
2992 uint64_t offset, unsigned int ioprio, bool high_prio)
2994 const bool needs_lock = td_async_processing(td);
2995 unsigned long elapsed, this_window;
2996 struct thread_stat *ts = &td->ts;
2997 struct io_log *iolog = td->clat_hist_log;
3002 add_stat_sample(&ts->clat_stat[ddir], nsec);
3005 * When lat_percentiles=1 (default 0), the reported high/low priority
3006 * percentiles and stats are used for describing total latency values,
3007 * even though the variable names themselves start with clat_.
3009 * Because of the above definition, add a prio stat sample only when
3010 * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3011 * when lat_percentiles=1.
3013 if (!ts->lat_percentiles) {
3015 add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
3017 add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
3021 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3024 if (ts->clat_percentiles) {
3026 * Because of the above definition, add a prio lat percentile
3027 * sample only when lat_percentiles=0. add_lat_sample() will add
3028 * the prio lat percentile sample when lat_percentiles=1.
3030 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3031 if (!ts->lat_percentiles)
3032 add_lat_percentile_prio_sample(ts, nsec, ddir,
3036 if (iolog && iolog->hist_msec) {
3037 struct io_hist *hw = &iolog->hist_window[ddir];
3040 elapsed = mtime_since_now(&td->epoch);
3042 hw->hist_last = elapsed;
3043 this_window = elapsed - hw->hist_last;
3045 if (this_window >= iolog->hist_msec) {
3046 uint64_t *io_u_plat;
3047 struct io_u_plat_entry *dst;
3050 * Make a byte-for-byte copy of the latency histogram
3051 * stored in td->ts.io_u_plat[ddir], recording it in a
3052 * log sample. Note that the matching call to free() is
3053 * located in iolog.c after printing this sample to the
3056 io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3057 dst = malloc(sizeof(struct io_u_plat_entry));
3058 memcpy(&(dst->io_u_plat), io_u_plat,
3059 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3060 flist_add(&dst->list, &hw->list);
3061 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3062 elapsed, offset, ioprio);
3065 * Update the last time we recorded as being now, minus
3066 * any drift in time we encountered before actually
3067 * making the record.
3069 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3075 __td_io_u_unlock(td);
3078 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3079 unsigned long long nsec, unsigned long long bs,
3080 uint64_t offset, unsigned int ioprio)
3082 const bool needs_lock = td_async_processing(td);
3083 struct thread_stat *ts = &td->ts;
3091 add_stat_sample(&ts->slat_stat[ddir], nsec);
3094 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3097 if (ts->slat_percentiles)
3098 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3101 __td_io_u_unlock(td);
3104 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3105 unsigned long long nsec, unsigned long long bs,
3106 uint64_t offset, unsigned int ioprio, bool high_prio)
3108 const bool needs_lock = td_async_processing(td);
3109 struct thread_stat *ts = &td->ts;
3117 add_stat_sample(&ts->lat_stat[ddir], nsec);
3120 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3124 * When lat_percentiles=1 (default 0), the reported high/low priority
3125 * percentiles and stats are used for describing total latency values,
3126 * even though the variable names themselves start with clat_.
3128 * Because of the above definition, add a prio stat and prio lat
3129 * percentile sample only when lat_percentiles=1. add_clat_sample() will
3130 * add the prio stat and prio lat percentile sample when
3131 * lat_percentiles=0.
3133 if (ts->lat_percentiles) {
3134 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3135 add_lat_percentile_prio_sample(ts, nsec, ddir, high_prio);
3137 add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
3139 add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
3143 __td_io_u_unlock(td);
3146 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3147 unsigned int bytes, unsigned long long spent)
3149 const bool needs_lock = td_async_processing(td);
3150 struct thread_stat *ts = &td->ts;
3154 rate = (unsigned long) (bytes * 1000000ULL / spent);
3161 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3164 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3165 bytes, io_u->offset, io_u->ioprio);
3167 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3170 __td_io_u_unlock(td);
3173 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3174 struct timespec *t, unsigned int avg_time,
3175 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3176 struct io_stat *stat, struct io_log *log,
3179 const bool needs_lock = td_async_processing(td);
3180 unsigned long spent, rate;
3182 unsigned long next, next_log;
3184 next_log = avg_time;
3186 spent = mtime_since(parent_tv, t);
3187 if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3188 return avg_time - spent;
3194 * Compute both read and write rates for the interval.
3196 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3199 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3201 continue; /* No entries for interval */
3205 rate = delta * 1000 / spent / 1024; /* KiB/s */
3207 rate = (delta * 1000) / spent;
3211 add_stat_sample(&stat[ddir], rate);
3214 unsigned long long bs = 0;
3216 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3217 bs = td->o.min_bs[ddir];
3219 next = add_log_sample(td, log, sample_val(rate), ddir,
3221 next_log = min(next_log, next);
3224 stat_io_bytes[ddir] = this_io_bytes[ddir];
3230 __td_io_u_unlock(td);
3232 if (spent <= avg_time)
3235 next = avg_time - (1 + spent - avg_time);
3237 return min(next, next_log);
3240 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3242 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3243 td->this_io_bytes, td->stat_io_bytes,
3244 td->ts.bw_stat, td->bw_log, true);
3247 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3250 const bool needs_lock = td_async_processing(td);
3251 struct thread_stat *ts = &td->ts;
3256 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3259 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3260 bytes, io_u->offset, io_u->ioprio);
3262 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3265 __td_io_u_unlock(td);
3268 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3270 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3271 td->this_io_blocks, td->stat_io_blocks,
3272 td->ts.iops_stat, td->iops_log, false);
3276 * Returns msecs to next event
3278 int calc_log_samples(void)
3280 struct thread_data *td;
3281 unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3282 struct timespec now;
3284 long elapsed_time = 0;
3286 fio_gettime(&now, NULL);
3288 for_each_td(td, i) {
3289 elapsed_time = mtime_since_now(&td->epoch);
3293 if (in_ramp_time(td) ||
3294 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3295 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3299 (td->bw_log && !per_unit_log(td->bw_log))) {
3300 tmp = add_bw_samples(td, &now);
3303 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3305 if (!td->iops_log ||
3306 (td->iops_log && !per_unit_log(td->iops_log))) {
3307 tmp = add_iops_samples(td, &now);
3310 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3317 /* if log_avg_msec_min has not been changed, set it to 0 */
3318 if (log_avg_msec_min == -1U)
3319 log_avg_msec_min = 0;
3321 if (log_avg_msec_min == 0)
3322 next_mod = elapsed_time;
3324 next_mod = elapsed_time % log_avg_msec_min;
3326 /* correction to keep the time on the log avg msec boundary */
3327 next = min(next, (log_avg_msec_min - next_mod));
3329 return next == ~0U ? 0 : next;
3332 void stat_init(void)
3334 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3337 void stat_exit(void)
3340 * When we have the mutex, we know out-of-band access to it
3343 fio_sem_down(stat_sem);
3344 fio_sem_remove(stat_sem);
3348 * Called from signal handler. Wake up status thread.
3350 void show_running_run_stats(void)
3355 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3357 /* Ignore io_u's which span multiple blocks--they will just get
3358 * inaccurate counts. */
3359 int idx = (io_u->offset - io_u->file->file_offset)
3360 / td->o.bs[DDIR_TRIM];
3361 uint32_t *info = &td->ts.block_infos[idx];
3362 assert(idx < td->ts.nr_block_infos);