10 #include "lib/ieee754.h"
12 #include "lib/getrusage.h"
15 #include "lib/output_buffer.h"
16 #include "helper_thread.h"
19 #include "oslib/asprintf.h"
22 #define LOG_MSEC_SLACK 2
24 #define LOG_MSEC_SLACK 1
27 struct fio_sem *stat_sem;
29 void clear_rusage_stat(struct thread_data *td)
31 struct thread_stat *ts = &td->ts;
33 fio_getrusage(&td->ru_start);
34 ts->usr_time = ts->sys_time = 0;
36 ts->minf = ts->majf = 0;
39 void update_rusage_stat(struct thread_data *td)
41 struct thread_stat *ts = &td->ts;
43 fio_getrusage(&td->ru_end);
44 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
45 &td->ru_end.ru_utime);
46 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
47 &td->ru_end.ru_stime);
48 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
49 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
50 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
51 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
53 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
57 * Given a latency, return the index of the corresponding bucket in
58 * the structure tracking percentiles.
60 * (1) find the group (and error bits) that the value (latency)
61 * belongs to by looking at its MSB. (2) find the bucket number in the
62 * group by looking at the index bits.
65 static unsigned int plat_val_to_idx(unsigned long long val)
67 unsigned int msb, error_bits, base, offset, idx;
69 /* Find MSB starting from bit 0 */
73 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
76 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
77 * all bits of the sample as index
79 if (msb <= FIO_IO_U_PLAT_BITS)
82 /* Compute the number of error bits to discard*/
83 error_bits = msb - FIO_IO_U_PLAT_BITS;
85 /* Compute the number of buckets before the group */
86 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
89 * Discard the error bits and apply the mask to find the
90 * index for the buckets in the group
92 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
94 /* Make sure the index does not exceed (array size - 1) */
95 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
96 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
102 * Convert the given index of the bucket array to the value
103 * represented by the bucket
105 static unsigned long long plat_idx_to_val(unsigned int idx)
107 unsigned int error_bits;
108 unsigned long long k, base;
110 assert(idx < FIO_IO_U_PLAT_NR);
112 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
113 * all bits of the sample as index */
114 if (idx < (FIO_IO_U_PLAT_VAL << 1))
117 /* Find the group and compute the minimum value of that group */
118 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
119 base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
121 /* Find its bucket number of the group */
122 k = idx % FIO_IO_U_PLAT_VAL;
124 /* Return the mean of the range of the bucket */
125 return base + ((k + 0.5) * (1 << error_bits));
128 static int double_cmp(const void *a, const void *b)
130 const fio_fp64_t fa = *(const fio_fp64_t *) a;
131 const fio_fp64_t fb = *(const fio_fp64_t *) b;
136 else if (fa.u.f < fb.u.f)
142 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
143 fio_fp64_t *plist, unsigned long long **output,
144 unsigned long long *maxv, unsigned long long *minv)
146 unsigned long long sum = 0;
147 unsigned int len, i, j = 0;
148 unsigned long long *ovals = NULL;
155 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
162 * Sort the percentile list. Note that it may already be sorted if
163 * we are using the default values, but since it's a short list this
164 * isn't a worry. Also note that this does not work for NaN values.
167 qsort(plist, len, sizeof(plist[0]), double_cmp);
169 ovals = malloc(len * sizeof(*ovals));
174 * Calculate bucket values, note down max and min values
177 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
179 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
180 assert(plist[j].u.f <= 100.0);
182 ovals[j] = plat_idx_to_val(i);
183 if (ovals[j] < *minv)
185 if (ovals[j] > *maxv)
188 is_last = (j == len - 1) != 0;
197 log_err("fio: error calculating latency percentiles\n");
204 * Find and display the p-th percentile of clat
206 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
207 fio_fp64_t *plist, unsigned int precision,
208 const char *pre, struct buf_output *out)
210 unsigned int divisor, len, i, j = 0;
211 unsigned long long minv, maxv;
212 unsigned long long *ovals;
213 int per_line, scale_down, time_width;
217 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
222 * We default to nsecs, but if the value range is such that we
223 * should scale down to usecs or msecs, do that.
225 if (minv > 2000000 && maxv > 99999999ULL) {
228 log_buf(out, " %s percentiles (msec):\n |", pre);
229 } else if (minv > 2000 && maxv > 99999) {
232 log_buf(out, " %s percentiles (usec):\n |", pre);
236 log_buf(out, " %s percentiles (nsec):\n |", pre);
240 time_width = max(5, (int) (log10(maxv / divisor) + 1));
241 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
242 precision, time_width);
243 /* fmt will be something like " %5.2fth=[%4llu]%c" */
244 per_line = (80 - 7) / (precision + 10 + time_width);
246 for (j = 0; j < len; j++) {
248 if (j != 0 && (j % per_line) == 0)
251 /* end of the list */
252 is_last = (j == len - 1) != 0;
254 for (i = 0; i < scale_down; i++)
255 ovals[j] = (ovals[j] + 999) / 1000;
257 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
262 if ((j % per_line) == per_line - 1) /* for formatting */
269 static int get_nr_prios_with_samples(struct thread_stat *ts, enum fio_ddir ddir)
271 int i, nr_prios_with_samples = 0;
273 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
274 if (ts->clat_prio[ddir][i].clat_stat.samples)
275 nr_prios_with_samples++;
278 return nr_prios_with_samples;
281 bool calc_lat(struct io_stat *is, unsigned long long *min,
282 unsigned long long *max, double *mean, double *dev)
284 double n = (double) is->samples;
291 *mean = is->mean.u.f;
294 *dev = sqrt(is->S.u.f / (n - 1.0));
301 void show_mixed_group_stats(struct group_run_stats *rs, struct buf_output *out)
303 char *io, *agg, *min, *max;
304 char *ioalt, *aggalt, *minalt, *maxalt;
305 uint64_t io_mix = 0, agg_mix = 0, min_mix = -1, max_mix = 0;
306 uint64_t min_run = -1, max_run = 0;
307 const int i2p = is_power_of_2(rs->kb_base);
310 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
313 io_mix += rs->iobytes[i];
314 agg_mix += rs->agg[i];
315 min_mix = min_mix < rs->min_bw[i] ? min_mix : rs->min_bw[i];
316 max_mix = max_mix > rs->max_bw[i] ? max_mix : rs->max_bw[i];
317 min_run = min_run < rs->min_run[i] ? min_run : rs->min_run[i];
318 max_run = max_run > rs->max_run[i] ? max_run : rs->max_run[i];
320 io = num2str(io_mix, rs->sig_figs, 1, i2p, N2S_BYTE);
321 ioalt = num2str(io_mix, rs->sig_figs, 1, !i2p, N2S_BYTE);
322 agg = num2str(agg_mix, rs->sig_figs, 1, i2p, rs->unit_base);
323 aggalt = num2str(agg_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
324 min = num2str(min_mix, rs->sig_figs, 1, i2p, rs->unit_base);
325 minalt = num2str(min_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
326 max = num2str(max_mix, rs->sig_figs, 1, i2p, rs->unit_base);
327 maxalt = num2str(max_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
328 log_buf(out, " MIXED: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
329 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
330 (unsigned long long) min_run,
331 (unsigned long long) max_run);
342 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
344 char *io, *agg, *min, *max;
345 char *ioalt, *aggalt, *minalt, *maxalt;
346 const char *str[] = { " READ", " WRITE" , " TRIM"};
349 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
351 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
352 const int i2p = is_power_of_2(rs->kb_base);
357 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
358 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
359 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
360 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
361 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
362 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
363 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
364 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
365 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
366 (rs->unified_rw_rep == UNIFIED_MIXED) ? " MIXED" : str[i],
367 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
368 (unsigned long long) rs->min_run[i],
369 (unsigned long long) rs->max_run[i]);
381 /* Need to aggregate statistics to show mixed values */
382 if (rs->unified_rw_rep == UNIFIED_BOTH)
383 show_mixed_group_stats(rs, out);
386 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
391 * Do depth distribution calculations
393 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
395 io_u_dist[i] = (double) map[i] / (double) total;
396 io_u_dist[i] *= 100.0;
397 if (io_u_dist[i] < 0.1 && map[i])
404 static void stat_calc_lat(struct thread_stat *ts, double *dst,
405 uint64_t *src, int nr)
407 unsigned long total = ddir_rw_sum(ts->total_io_u);
411 * Do latency distribution calculations
413 for (i = 0; i < nr; i++) {
415 dst[i] = (double) src[i] / (double) total;
417 if (dst[i] < 0.01 && src[i])
425 * To keep the terse format unaltered, add all of the ns latency
426 * buckets to the first us latency bucket
428 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
430 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
433 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
435 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
436 ntotal += ts->io_u_lat_n[i];
438 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
441 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
443 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
446 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
448 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
451 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
453 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
456 static void display_lat(const char *name, unsigned long long min,
457 unsigned long long max, double mean, double dev,
458 struct buf_output *out)
460 const char *base = "(nsec)";
463 if (nsec_to_msec(&min, &max, &mean, &dev))
465 else if (nsec_to_usec(&min, &max, &mean, &dev))
468 minp = num2str(min, 6, 1, 0, N2S_NONE);
469 maxp = num2str(max, 6, 1, 0, N2S_NONE);
471 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
472 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
478 static struct thread_stat *gen_mixed_ddir_stats_from_ts(struct thread_stat *ts)
480 struct thread_stat *ts_lcl;
483 * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
486 ts_lcl = malloc(sizeof(struct thread_stat));
488 log_err("fio: failed to allocate local thread stat\n");
492 init_thread_stat(ts_lcl);
494 /* calculate mixed stats */
495 ts_lcl->unified_rw_rep = UNIFIED_MIXED;
496 ts_lcl->lat_percentiles = ts->lat_percentiles;
497 ts_lcl->clat_percentiles = ts->clat_percentiles;
498 ts_lcl->slat_percentiles = ts->slat_percentiles;
499 ts_lcl->percentile_precision = ts->percentile_precision;
500 memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
502 sum_thread_stats(ts_lcl, ts);
507 static double convert_agg_kbytes_percent(struct group_run_stats *rs,
508 enum fio_ddir ddir, int mean)
510 double p_of_agg = 100.0;
511 if (rs && rs->agg[ddir] > 1024) {
512 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
514 if (p_of_agg > 100.0)
520 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
521 enum fio_ddir ddir, struct buf_output *out)
524 unsigned long long min, max, bw, iops;
526 char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
528 const char *clat_type = ts->lat_percentiles ? "lat" : "clat";
530 if (ddir_sync(ddir)) {
531 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
532 log_buf(out, " %s:\n", "fsync/fdatasync/sync_file_range");
533 display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
534 show_clat_percentiles(ts->io_u_sync_plat,
535 ts->sync_stat.samples,
537 ts->percentile_precision,
538 io_ddir_name(ddir), out);
543 assert(ddir_rw(ddir));
545 if (!ts->runtime[ddir])
548 i2p = is_power_of_2(rs->kb_base);
549 runt = ts->runtime[ddir];
551 bw = (1000 * ts->io_bytes[ddir]) / runt;
552 io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
553 bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
554 bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
556 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
557 iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
558 if (ddir == DDIR_WRITE || ddir == DDIR_TRIM)
559 post_st = zbd_write_status(ts);
560 else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
564 total = ts->cachehit + ts->cachemiss;
565 hit = (double) ts->cachehit / (double) total;
567 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
571 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
572 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
573 iops_p, bw_p, bw_p_alt, io_p,
574 (unsigned long long) ts->runtime[ddir],
583 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
584 display_lat("slat", min, max, mean, dev, out);
585 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
586 display_lat("clat", min, max, mean, dev, out);
587 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
588 display_lat(" lat", min, max, mean, dev, out);
590 /* Only print per prio stats if there are >= 2 prios with samples */
591 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
592 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
595 if (!calc_lat(&ts->clat_prio[ddir][i].clat_stat, &min,
599 snprintf(buf, sizeof(buf),
602 ioprio_class(ts->clat_prio[ddir][i].ioprio),
603 ioprio(ts->clat_prio[ddir][i].ioprio));
604 display_lat(buf, min, max, mean, dev, out);
608 if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
609 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
610 ts->slat_stat[ddir].samples,
612 ts->percentile_precision, "slat", out);
613 if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
614 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
615 ts->clat_stat[ddir].samples,
617 ts->percentile_precision, "clat", out);
618 if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
619 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
620 ts->lat_stat[ddir].samples,
622 ts->percentile_precision, "lat", out);
624 if (ts->clat_percentiles || ts->lat_percentiles) {
628 if (ts->lat_percentiles)
629 samples = ts->lat_stat[ddir].samples;
631 samples = ts->clat_stat[ddir].samples;
633 /* Only print per prio stats if there are >= 2 prios with samples */
634 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
635 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
636 uint64_t prio_samples =
637 ts->clat_prio[ddir][i].clat_stat.samples;
642 snprintf(prio_name, sizeof(prio_name),
643 "%s prio %u/%u (%.2f%% of IOs)",
645 ioprio_class(ts->clat_prio[ddir][i].ioprio),
646 ioprio(ts->clat_prio[ddir][i].ioprio),
647 100. * (double) prio_samples / (double) samples);
648 show_clat_percentiles(ts->clat_prio[ddir][i].io_u_plat,
649 prio_samples, ts->percentile_list,
650 ts->percentile_precision,
656 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
657 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
660 if ((rs->unit_base == 1) && i2p)
662 else if (rs->unit_base == 1)
669 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
671 if (rs->unit_base == 1) {
678 if (mean > fkb_base * fkb_base) {
683 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
686 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
687 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
688 bw_str, min, max, p_of_agg, mean, dev,
689 (&ts->bw_stat[ddir])->samples);
691 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
692 log_buf(out, " iops : min=%5llu, max=%5llu, "
693 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
694 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
698 static void show_mixed_ddir_status(struct group_run_stats *rs,
699 struct thread_stat *ts,
700 struct buf_output *out)
702 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
705 show_ddir_status(rs, ts_lcl, DDIR_READ, out);
707 free_clat_prio_stats(ts_lcl);
711 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
712 const char *msg, struct buf_output *out)
714 bool new_line = true, shown = false;
717 for (i = 0; i < nr; i++) {
718 if (io_u_lat[i] <= 0.0)
724 log_buf(out, " lat (%s) : ", msg);
730 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
742 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
744 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
745 "250=", "500=", "750=", "1000=", };
747 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
750 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
752 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
753 "250=", "500=", "750=", "1000=", };
755 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
758 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
760 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
761 "250=", "500=", "750=", "1000=", "2000=",
764 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
767 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
769 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
770 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
771 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
773 stat_calc_lat_n(ts, io_u_lat_n);
774 stat_calc_lat_u(ts, io_u_lat_u);
775 stat_calc_lat_m(ts, io_u_lat_m);
777 show_lat_n(io_u_lat_n, out);
778 show_lat_u(io_u_lat_u, out);
779 show_lat_m(io_u_lat_m, out);
782 static int block_state_category(int block_state)
784 switch (block_state) {
785 case BLOCK_STATE_UNINIT:
787 case BLOCK_STATE_TRIMMED:
788 case BLOCK_STATE_WRITTEN:
790 case BLOCK_STATE_WRITE_FAILURE:
791 case BLOCK_STATE_TRIM_FAILURE:
794 /* Silence compile warning on some BSDs and have a return */
800 static int compare_block_infos(const void *bs1, const void *bs2)
802 uint64_t block1 = *(uint64_t *)bs1;
803 uint64_t block2 = *(uint64_t *)bs2;
804 int state1 = BLOCK_INFO_STATE(block1);
805 int state2 = BLOCK_INFO_STATE(block2);
806 int bscat1 = block_state_category(state1);
807 int bscat2 = block_state_category(state2);
808 int cycles1 = BLOCK_INFO_TRIMS(block1);
809 int cycles2 = BLOCK_INFO_TRIMS(block2);
816 if (cycles1 < cycles2)
818 if (cycles1 > cycles2)
826 assert(block1 == block2);
830 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
831 fio_fp64_t *plist, unsigned int **percentiles,
837 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
839 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
846 * Sort the percentile list. Note that it may already be sorted if
847 * we are using the default values, but since it's a short list this
848 * isn't a worry. Also note that this does not work for NaN values.
851 qsort(plist, len, sizeof(plist[0]), double_cmp);
853 /* Start only after the uninit entries end */
855 nr_uninit < nr_block_infos
856 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
860 if (nr_uninit == nr_block_infos)
863 *percentiles = calloc(len, sizeof(**percentiles));
865 for (i = 0; i < len; i++) {
866 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
868 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
871 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
872 for (i = 0; i < nr_block_infos; i++)
873 types[BLOCK_INFO_STATE(block_infos[i])]++;
878 static const char *block_state_names[] = {
879 [BLOCK_STATE_UNINIT] = "unwritten",
880 [BLOCK_STATE_TRIMMED] = "trimmed",
881 [BLOCK_STATE_WRITTEN] = "written",
882 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
883 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
886 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
887 fio_fp64_t *plist, struct buf_output *out)
890 unsigned int *percentiles = NULL;
891 unsigned int block_state_counts[BLOCK_STATE_COUNT];
893 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
894 &percentiles, block_state_counts);
896 log_buf(out, " block lifetime percentiles :\n |");
898 for (i = 0; i < len; i++) {
899 uint32_t block_info = percentiles[i];
900 #define LINE_LENGTH 75
901 char str[LINE_LENGTH];
902 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
903 plist[i].u.f, block_info,
904 i == len - 1 ? '\n' : ',');
905 assert(strln < LINE_LENGTH);
906 if (pos + strln > LINE_LENGTH) {
908 log_buf(out, "\n |");
910 log_buf(out, "%s", str);
917 log_buf(out, " states :");
918 for (i = 0; i < BLOCK_STATE_COUNT; i++)
919 log_buf(out, " %s=%u%c",
920 block_state_names[i], block_state_counts[i],
921 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
924 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
926 char *p1, *p1alt, *p2;
927 unsigned long long bw_mean, iops_mean;
928 const int i2p = is_power_of_2(ts->kb_base);
933 bw_mean = steadystate_bw_mean(ts);
934 iops_mean = steadystate_iops_mean(ts);
936 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
937 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
938 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
940 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
941 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
943 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
944 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
945 ts->ss_criterion.u.f,
946 ts->ss_state & FIO_SS_PCT ? "%" : "");
953 static void show_agg_stats(struct disk_util_agg *agg, int terse,
954 struct buf_output *out)
956 if (!agg->slavecount)
960 log_buf(out, ", aggrios=%llu/%llu, aggsectors=%llu/%llu, "
961 "aggrmerge=%llu/%llu, aggrticks=%llu/%llu, "
962 "aggrin_queue=%llu, aggrutil=%3.2f%%",
963 (unsigned long long) agg->ios[0] / agg->slavecount,
964 (unsigned long long) agg->ios[1] / agg->slavecount,
965 (unsigned long long) agg->sectors[0] / agg->slavecount,
966 (unsigned long long) agg->sectors[1] / agg->slavecount,
967 (unsigned long long) agg->merges[0] / agg->slavecount,
968 (unsigned long long) agg->merges[1] / agg->slavecount,
969 (unsigned long long) agg->ticks[0] / agg->slavecount,
970 (unsigned long long) agg->ticks[1] / agg->slavecount,
971 (unsigned long long) agg->time_in_queue / agg->slavecount,
974 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
975 (unsigned long long) agg->ios[0] / agg->slavecount,
976 (unsigned long long) agg->ios[1] / agg->slavecount,
977 (unsigned long long) agg->merges[0] / agg->slavecount,
978 (unsigned long long) agg->merges[1] / agg->slavecount,
979 (unsigned long long) agg->ticks[0] / agg->slavecount,
980 (unsigned long long) agg->ticks[1] / agg->slavecount,
981 (unsigned long long) agg->time_in_queue / agg->slavecount,
986 static void aggregate_slaves_stats(struct disk_util *masterdu)
988 struct disk_util_agg *agg = &masterdu->agg;
989 struct disk_util_stat *dus;
990 struct flist_head *entry;
991 struct disk_util *slavedu;
994 flist_for_each(entry, &masterdu->slaves) {
995 slavedu = flist_entry(entry, struct disk_util, slavelist);
997 agg->ios[0] += dus->s.ios[0];
998 agg->ios[1] += dus->s.ios[1];
999 agg->merges[0] += dus->s.merges[0];
1000 agg->merges[1] += dus->s.merges[1];
1001 agg->sectors[0] += dus->s.sectors[0];
1002 agg->sectors[1] += dus->s.sectors[1];
1003 agg->ticks[0] += dus->s.ticks[0];
1004 agg->ticks[1] += dus->s.ticks[1];
1005 agg->time_in_queue += dus->s.time_in_queue;
1008 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
1009 /* System utilization is the utilization of the
1010 * component with the highest utilization.
1012 if (util > agg->max_util.u.f)
1013 agg->max_util.u.f = util;
1017 if (agg->max_util.u.f > 100.0)
1018 agg->max_util.u.f = 100.0;
1021 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
1022 int terse, struct buf_output *out)
1027 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1032 if (agg->slavecount)
1035 log_buf(out, " %s: ios=%llu/%llu, sectors=%llu/%llu, "
1036 "merge=%llu/%llu, ticks=%llu/%llu, in_queue=%llu, "
1039 (unsigned long long) dus->s.ios[0],
1040 (unsigned long long) dus->s.ios[1],
1041 (unsigned long long) dus->s.sectors[0],
1042 (unsigned long long) dus->s.sectors[1],
1043 (unsigned long long) dus->s.merges[0],
1044 (unsigned long long) dus->s.merges[1],
1045 (unsigned long long) dus->s.ticks[0],
1046 (unsigned long long) dus->s.ticks[1],
1047 (unsigned long long) dus->s.time_in_queue,
1050 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1052 (unsigned long long) dus->s.ios[0],
1053 (unsigned long long) dus->s.ios[1],
1054 (unsigned long long) dus->s.merges[0],
1055 (unsigned long long) dus->s.merges[1],
1056 (unsigned long long) dus->s.ticks[0],
1057 (unsigned long long) dus->s.ticks[1],
1058 (unsigned long long) dus->s.time_in_queue,
1063 * If the device has slaves, aggregate the stats for
1064 * those slave devices also.
1066 show_agg_stats(agg, terse, out);
1072 void json_array_add_disk_util(struct disk_util_stat *dus,
1073 struct disk_util_agg *agg, struct json_array *array)
1075 struct json_object *obj;
1079 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1083 obj = json_create_object();
1084 json_array_add_value_object(array, obj);
1086 json_object_add_value_string(obj, "name", (const char *)dus->name);
1087 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1088 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1089 json_object_add_value_int(obj, "read_sectors", dus->s.sectors[0]);
1090 json_object_add_value_int(obj, "write_sectors", dus->s.sectors[1]);
1091 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1092 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1093 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1094 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1095 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1096 json_object_add_value_float(obj, "util", util);
1099 * If the device has slaves, aggregate the stats for
1100 * those slave devices also.
1102 if (!agg->slavecount)
1104 json_object_add_value_int(obj, "aggr_read_ios",
1105 agg->ios[0] / agg->slavecount);
1106 json_object_add_value_int(obj, "aggr_write_ios",
1107 agg->ios[1] / agg->slavecount);
1108 json_object_add_value_int(obj, "aggr_read_sectors",
1109 agg->sectors[0] / agg->slavecount);
1110 json_object_add_value_int(obj, "aggr_write_sectors",
1111 agg->sectors[1] / agg->slavecount);
1112 json_object_add_value_int(obj, "aggr_read_merges",
1113 agg->merges[0] / agg->slavecount);
1114 json_object_add_value_int(obj, "aggr_write_merge",
1115 agg->merges[1] / agg->slavecount);
1116 json_object_add_value_int(obj, "aggr_read_ticks",
1117 agg->ticks[0] / agg->slavecount);
1118 json_object_add_value_int(obj, "aggr_write_ticks",
1119 agg->ticks[1] / agg->slavecount);
1120 json_object_add_value_int(obj, "aggr_in_queue",
1121 agg->time_in_queue / agg->slavecount);
1122 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1125 static void json_object_add_disk_utils(struct json_object *obj,
1126 struct flist_head *head)
1128 struct json_array *array = json_create_array();
1129 struct flist_head *entry;
1130 struct disk_util *du;
1132 json_object_add_value_array(obj, "disk_util", array);
1134 flist_for_each(entry, head) {
1135 du = flist_entry(entry, struct disk_util, list);
1137 aggregate_slaves_stats(du);
1138 json_array_add_disk_util(&du->dus, &du->agg, array);
1142 void show_disk_util(int terse, struct json_object *parent,
1143 struct buf_output *out)
1145 struct flist_head *entry;
1146 struct disk_util *du;
1149 if (!is_running_backend())
1152 if (flist_empty(&disk_list))
1155 if ((output_format & FIO_OUTPUT_JSON) && parent)
1160 if (!terse && !do_json)
1161 log_buf(out, "\nDisk stats (read/write):\n");
1164 json_object_add_disk_utils(parent, &disk_list);
1165 } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1166 flist_for_each(entry, &disk_list) {
1167 du = flist_entry(entry, struct disk_util, list);
1169 aggregate_slaves_stats(du);
1170 print_disk_util(&du->dus, &du->agg, terse, out);
1175 static void show_thread_status_normal(struct thread_stat *ts,
1176 struct group_run_stats *rs,
1177 struct buf_output *out)
1179 double usr_cpu, sys_cpu;
1180 unsigned long runtime;
1181 double io_u_dist[FIO_IO_U_MAP_NR];
1185 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1188 memset(time_buf, 0, sizeof(time_buf));
1191 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1194 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1195 ts->name, ts->groupid, ts->members,
1196 ts->error, (int) ts->pid, time_buf);
1198 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1199 ts->name, ts->groupid, ts->members,
1200 ts->error, ts->verror, (int) ts->pid,
1204 if (strlen(ts->description))
1205 log_buf(out, " Description : [%s]\n", ts->description);
1207 for_each_rw_ddir(ddir) {
1208 if (ts->io_bytes[ddir])
1209 show_ddir_status(rs, ts, ddir, out);
1212 if (ts->unified_rw_rep == UNIFIED_BOTH)
1213 show_mixed_ddir_status(rs, ts, out);
1215 show_latencies(ts, out);
1217 if (ts->sync_stat.samples)
1218 show_ddir_status(rs, ts, DDIR_SYNC, out);
1220 runtime = ts->total_run_time;
1222 double runt = (double) runtime;
1224 usr_cpu = (double) ts->usr_time * 100 / runt;
1225 sys_cpu = (double) ts->sys_time * 100 / runt;
1231 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1232 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1233 (unsigned long long) ts->ctx,
1234 (unsigned long long) ts->majf,
1235 (unsigned long long) ts->minf);
1237 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1238 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1239 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1240 io_u_dist[1], io_u_dist[2],
1241 io_u_dist[3], io_u_dist[4],
1242 io_u_dist[5], io_u_dist[6]);
1244 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1245 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1246 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1247 io_u_dist[1], io_u_dist[2],
1248 io_u_dist[3], io_u_dist[4],
1249 io_u_dist[5], io_u_dist[6]);
1250 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1251 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1252 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1253 io_u_dist[1], io_u_dist[2],
1254 io_u_dist[3], io_u_dist[4],
1255 io_u_dist[5], io_u_dist[6]);
1256 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
1257 " short=%llu,%llu,%llu,0"
1258 " dropped=%llu,%llu,%llu,0\n",
1259 (unsigned long long) ts->total_io_u[0],
1260 (unsigned long long) ts->total_io_u[1],
1261 (unsigned long long) ts->total_io_u[2],
1262 (unsigned long long) ts->total_io_u[3],
1263 (unsigned long long) ts->short_io_u[0],
1264 (unsigned long long) ts->short_io_u[1],
1265 (unsigned long long) ts->short_io_u[2],
1266 (unsigned long long) ts->drop_io_u[0],
1267 (unsigned long long) ts->drop_io_u[1],
1268 (unsigned long long) ts->drop_io_u[2]);
1269 if (ts->continue_on_error) {
1270 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
1271 (unsigned long long)ts->total_err_count,
1273 strerror(ts->first_error));
1275 if (ts->latency_depth) {
1276 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1277 (unsigned long long)ts->latency_target,
1278 (unsigned long long)ts->latency_window,
1279 ts->latency_percentile.u.f,
1283 if (ts->nr_block_infos)
1284 show_block_infos(ts->nr_block_infos, ts->block_infos,
1285 ts->percentile_list, out);
1288 show_ss_normal(ts, out);
1291 static void show_ddir_status_terse(struct thread_stat *ts,
1292 struct group_run_stats *rs,
1293 enum fio_ddir ddir, int ver,
1294 struct buf_output *out)
1296 unsigned long long min, max, minv, maxv, bw, iops;
1297 unsigned long long *ovals = NULL;
1302 assert(ddir_rw(ddir));
1305 if (ts->runtime[ddir]) {
1306 uint64_t runt = ts->runtime[ddir];
1308 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1309 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1312 log_buf(out, ";%llu;%llu;%llu;%llu",
1313 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1314 (unsigned long long) ts->runtime[ddir]);
1316 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1317 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1319 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1321 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1322 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1324 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1326 if (ts->lat_percentiles) {
1327 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1328 ts->lat_stat[ddir].samples,
1329 ts->percentile_list, &ovals, &maxv,
1331 } else if (ts->clat_percentiles) {
1332 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1333 ts->clat_stat[ddir].samples,
1334 ts->percentile_list, &ovals, &maxv,
1340 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1342 log_buf(out, ";0%%=0");
1345 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1348 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1349 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1351 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1355 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1357 double p_of_agg = 100.0;
1359 if (rs->agg[ddir]) {
1360 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1361 if (p_of_agg > 100.0)
1365 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1367 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1372 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1374 log_buf(out, ";%lu", 0UL);
1376 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1377 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1378 mean, dev, (&ts->iops_stat[ddir])->samples);
1380 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1384 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1385 struct group_run_stats *rs,
1386 int ver, struct buf_output *out)
1388 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1391 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1393 free_clat_prio_stats(ts_lcl);
1397 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1398 uint32_t percentiles,
1399 struct io_stat *lat_stat,
1400 uint64_t *io_u_plat)
1404 unsigned int i, len;
1405 struct json_object *lat_object, *percentile_object, *clat_bins_object;
1406 unsigned long long min, max, maxv, minv, *ovals = NULL;
1408 if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1412 lat_object = json_create_object();
1413 json_object_add_value_int(lat_object, "min", min);
1414 json_object_add_value_int(lat_object, "max", max);
1415 json_object_add_value_float(lat_object, "mean", mean);
1416 json_object_add_value_float(lat_object, "stddev", dev);
1417 json_object_add_value_int(lat_object, "N", lat_stat->samples);
1419 if (percentiles && lat_stat->samples) {
1420 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1421 ts->percentile_list, &ovals, &maxv, &minv);
1423 if (len > FIO_IO_U_LIST_MAX_LEN)
1424 len = FIO_IO_U_LIST_MAX_LEN;
1426 percentile_object = json_create_object();
1427 json_object_add_value_object(lat_object, "percentile", percentile_object);
1428 for (i = 0; i < len; i++) {
1429 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1430 json_object_add_value_int(percentile_object, buf, ovals[i]);
1434 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1435 clat_bins_object = json_create_object();
1436 json_object_add_value_object(lat_object, "bins", clat_bins_object);
1438 for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1440 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1441 json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1449 static void add_ddir_status_json(struct thread_stat *ts,
1450 struct group_run_stats *rs, enum fio_ddir ddir,
1451 struct json_object *parent)
1453 unsigned long long min, max;
1454 unsigned long long bw_bytes, bw;
1455 double mean, dev, iops;
1456 struct json_object *dir_object, *tmp_object;
1457 double p_of_agg = 100.0;
1459 assert(ddir_rw(ddir) || ddir_sync(ddir));
1461 if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1464 dir_object = json_create_object();
1465 json_object_add_value_object(parent,
1466 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1468 if (ddir_rw(ddir)) {
1472 if (ts->runtime[ddir]) {
1473 uint64_t runt = ts->runtime[ddir];
1475 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1476 bw = bw_bytes / 1024; /* KiB/s */
1477 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1480 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1481 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1482 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1483 json_object_add_value_int(dir_object, "bw", bw);
1484 json_object_add_value_float(dir_object, "iops", iops);
1485 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1486 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1487 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1488 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1490 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1491 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1492 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1494 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1495 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1496 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1498 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1499 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1500 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1502 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1503 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1504 &ts->sync_stat, ts->io_u_sync_plat);
1505 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1511 /* Only include per prio stats if there are >= 2 prios with samples */
1512 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
1513 struct json_array *array = json_create_array();
1514 const char *obj_name;
1517 if (ts->lat_percentiles)
1518 obj_name = "lat_ns";
1520 obj_name = "clat_ns";
1522 json_object_add_value_array(dir_object, "prios", array);
1524 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
1525 struct json_object *obj;
1527 if (!ts->clat_prio[ddir][i].clat_stat.samples)
1530 obj = json_create_object();
1532 json_object_add_value_int(obj, "prioclass",
1533 ioprio_class(ts->clat_prio[ddir][i].ioprio));
1534 json_object_add_value_int(obj, "prio",
1535 ioprio(ts->clat_prio[ddir][i].ioprio));
1537 tmp_object = add_ddir_lat_json(ts,
1538 ts->clat_percentiles | ts->lat_percentiles,
1539 &ts->clat_prio[ddir][i].clat_stat,
1540 ts->clat_prio[ddir][i].io_u_plat);
1541 json_object_add_value_object(obj, obj_name, tmp_object);
1542 json_array_add_value_object(array, obj);
1546 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1547 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1550 p_of_agg = mean = dev = 0.0;
1553 json_object_add_value_int(dir_object, "bw_min", min);
1554 json_object_add_value_int(dir_object, "bw_max", max);
1555 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1556 json_object_add_value_float(dir_object, "bw_mean", mean);
1557 json_object_add_value_float(dir_object, "bw_dev", dev);
1558 json_object_add_value_int(dir_object, "bw_samples",
1559 (&ts->bw_stat[ddir])->samples);
1561 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1565 json_object_add_value_int(dir_object, "iops_min", min);
1566 json_object_add_value_int(dir_object, "iops_max", max);
1567 json_object_add_value_float(dir_object, "iops_mean", mean);
1568 json_object_add_value_float(dir_object, "iops_stddev", dev);
1569 json_object_add_value_int(dir_object, "iops_samples",
1570 (&ts->iops_stat[ddir])->samples);
1572 if (ts->cachehit + ts->cachemiss) {
1576 total = ts->cachehit + ts->cachemiss;
1577 hit = (double) ts->cachehit / (double) total;
1579 json_object_add_value_float(dir_object, "cachehit", hit);
1583 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1584 struct group_run_stats *rs, struct json_object *parent)
1586 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1588 /* add the aggregated stats to json parent */
1590 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1592 free_clat_prio_stats(ts_lcl);
1596 static void show_thread_status_terse_all(struct thread_stat *ts,
1597 struct group_run_stats *rs, int ver,
1598 struct buf_output *out)
1600 double io_u_dist[FIO_IO_U_MAP_NR];
1601 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1602 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1603 double usr_cpu, sys_cpu;
1608 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1610 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1611 ts->name, ts->groupid, ts->error);
1613 /* Log Read Status, or mixed if unified_rw_rep = 1 */
1614 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1615 if (ts->unified_rw_rep != UNIFIED_MIXED) {
1616 /* Log Write Status */
1617 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1618 /* Log Trim Status */
1619 if (ver == 2 || ver == 4 || ver == 5)
1620 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1622 if (ts->unified_rw_rep == UNIFIED_BOTH)
1623 show_mixed_ddir_status_terse(ts, rs, ver, out);
1625 if (ts->total_run_time) {
1626 double runt = (double) ts->total_run_time;
1628 usr_cpu = (double) ts->usr_time * 100 / runt;
1629 sys_cpu = (double) ts->sys_time * 100 / runt;
1635 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1636 (unsigned long long) ts->ctx,
1637 (unsigned long long) ts->majf,
1638 (unsigned long long) ts->minf);
1640 /* Calc % distribution of IO depths, usecond, msecond latency */
1641 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1642 stat_calc_lat_nu(ts, io_u_lat_u);
1643 stat_calc_lat_m(ts, io_u_lat_m);
1645 /* Only show fixed 7 I/O depth levels*/
1646 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1647 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1648 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1650 /* Microsecond latency */
1651 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1652 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1653 /* Millisecond latency */
1654 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1655 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1657 /* disk util stats, if any */
1658 if (ver >= 3 && is_running_backend())
1659 show_disk_util(1, NULL, out);
1661 /* Additional output if continue_on_error set - default off*/
1662 if (ts->continue_on_error)
1663 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1665 /* Additional output if description is set */
1666 if (strlen(ts->description)) {
1669 log_buf(out, ";%s", ts->description);
1675 static void json_add_job_opts(struct json_object *root, const char *name,
1676 struct flist_head *opt_list)
1678 struct json_object *dir_object;
1679 struct flist_head *entry;
1680 struct print_option *p;
1682 if (flist_empty(opt_list))
1685 dir_object = json_create_object();
1686 json_object_add_value_object(root, name, dir_object);
1688 flist_for_each(entry, opt_list) {
1689 p = flist_entry(entry, struct print_option, list);
1690 json_object_add_value_string(dir_object, p->name, p->value);
1694 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1695 struct group_run_stats *rs,
1696 struct flist_head *opt_list)
1698 struct json_object *root, *tmp;
1699 struct jobs_eta *je;
1700 double io_u_dist[FIO_IO_U_MAP_NR];
1701 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1702 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1703 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1704 double usr_cpu, sys_cpu;
1708 root = json_create_object();
1709 json_object_add_value_string(root, "jobname", ts->name);
1710 json_object_add_value_int(root, "groupid", ts->groupid);
1711 json_object_add_value_int(root, "error", ts->error);
1714 je = get_jobs_eta(true, &size);
1716 json_object_add_value_int(root, "eta", je->eta_sec);
1717 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1722 json_add_job_opts(root, "job options", opt_list);
1724 add_ddir_status_json(ts, rs, DDIR_READ, root);
1725 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1726 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1727 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1729 if (ts->unified_rw_rep == UNIFIED_BOTH)
1730 add_mixed_ddir_status_json(ts, rs, root);
1733 if (ts->total_run_time) {
1734 double runt = (double) ts->total_run_time;
1736 usr_cpu = (double) ts->usr_time * 100 / runt;
1737 sys_cpu = (double) ts->sys_time * 100 / runt;
1742 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1743 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1744 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1745 json_object_add_value_int(root, "ctx", ts->ctx);
1746 json_object_add_value_int(root, "majf", ts->majf);
1747 json_object_add_value_int(root, "minf", ts->minf);
1749 /* Calc % distribution of IO depths */
1750 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1751 tmp = json_create_object();
1752 json_object_add_value_object(root, "iodepth_level", tmp);
1753 /* Only show fixed 7 I/O depth levels*/
1754 for (i = 0; i < 7; i++) {
1757 snprintf(name, 20, "%d", 1 << i);
1759 snprintf(name, 20, ">=%d", 1 << i);
1760 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1763 /* Calc % distribution of submit IO depths */
1764 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1765 tmp = json_create_object();
1766 json_object_add_value_object(root, "iodepth_submit", tmp);
1767 /* Only show fixed 7 I/O depth levels*/
1768 for (i = 0; i < 7; i++) {
1771 snprintf(name, 20, "0");
1773 snprintf(name, 20, "%d", 1 << (i+1));
1775 snprintf(name, 20, ">=%d", 1 << i);
1776 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1779 /* Calc % distribution of completion IO depths */
1780 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1781 tmp = json_create_object();
1782 json_object_add_value_object(root, "iodepth_complete", tmp);
1783 /* Only show fixed 7 I/O depth levels*/
1784 for (i = 0; i < 7; i++) {
1787 snprintf(name, 20, "0");
1789 snprintf(name, 20, "%d", 1 << (i+1));
1791 snprintf(name, 20, ">=%d", 1 << i);
1792 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1795 /* Calc % distribution of nsecond, usecond, msecond latency */
1796 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1797 stat_calc_lat_n(ts, io_u_lat_n);
1798 stat_calc_lat_u(ts, io_u_lat_u);
1799 stat_calc_lat_m(ts, io_u_lat_m);
1801 /* Nanosecond latency */
1802 tmp = json_create_object();
1803 json_object_add_value_object(root, "latency_ns", tmp);
1804 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1805 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1806 "250", "500", "750", "1000", };
1807 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1809 /* Microsecond latency */
1810 tmp = json_create_object();
1811 json_object_add_value_object(root, "latency_us", tmp);
1812 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1813 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1814 "250", "500", "750", "1000", };
1815 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1817 /* Millisecond latency */
1818 tmp = json_create_object();
1819 json_object_add_value_object(root, "latency_ms", tmp);
1820 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1821 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1822 "250", "500", "750", "1000", "2000",
1824 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1827 /* Additional output if continue_on_error set - default off*/
1828 if (ts->continue_on_error) {
1829 json_object_add_value_int(root, "total_err", ts->total_err_count);
1830 json_object_add_value_int(root, "first_error", ts->first_error);
1833 if (ts->latency_depth) {
1834 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1835 json_object_add_value_int(root, "latency_target", ts->latency_target);
1836 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1837 json_object_add_value_int(root, "latency_window", ts->latency_window);
1840 /* Additional output if description is set */
1841 if (strlen(ts->description))
1842 json_object_add_value_string(root, "desc", ts->description);
1844 if (ts->nr_block_infos) {
1845 /* Block error histogram and types */
1847 unsigned int *percentiles = NULL;
1848 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1850 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1851 ts->percentile_list,
1852 &percentiles, block_state_counts);
1855 struct json_object *block, *percentile_object, *states;
1857 block = json_create_object();
1858 json_object_add_value_object(root, "block", block);
1860 percentile_object = json_create_object();
1861 json_object_add_value_object(block, "percentiles",
1863 for (i = 0; i < len; i++) {
1865 snprintf(buf, sizeof(buf), "%f",
1866 ts->percentile_list[i].u.f);
1867 json_object_add_value_int(percentile_object,
1872 states = json_create_object();
1873 json_object_add_value_object(block, "states", states);
1874 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1875 json_object_add_value_int(states,
1876 block_state_names[state],
1877 block_state_counts[state]);
1884 struct json_object *data;
1885 struct json_array *iops, *bw;
1888 int intervals = ts->ss_dur / (ss_check_interval / 1000L);
1890 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1891 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1892 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1893 (float) ts->ss_limit.u.f,
1894 ts->ss_state & FIO_SS_PCT ? "%" : "");
1896 tmp = json_create_object();
1897 json_object_add_value_object(root, "steadystate", tmp);
1898 json_object_add_value_string(tmp, "ss", ss_buf);
1899 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1900 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1902 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1903 ts->ss_state & FIO_SS_PCT ? "%" : "");
1904 json_object_add_value_string(tmp, "criterion", ss_buf);
1905 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1906 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1908 data = json_create_object();
1909 json_object_add_value_object(tmp, "data", data);
1910 bw = json_create_array();
1911 iops = json_create_array();
1914 ** if ss was attained or the buffer is not full,
1915 ** ss->head points to the first element in the list.
1916 ** otherwise it actually points to the second element
1919 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1922 j = ts->ss_head == 0 ? intervals - 1 : ts->ss_head - 1;
1923 for (l = 0; l < intervals; l++) {
1924 k = (j + l) % intervals;
1925 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1926 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1928 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1929 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1930 json_object_add_value_array(data, "iops", iops);
1931 json_object_add_value_array(data, "bw", bw);
1937 static void show_thread_status_terse(struct thread_stat *ts,
1938 struct group_run_stats *rs,
1939 struct buf_output *out)
1941 if (terse_version >= 2 && terse_version <= 5)
1942 show_thread_status_terse_all(ts, rs, terse_version, out);
1944 log_err("fio: bad terse version!? %d\n", terse_version);
1947 struct json_object *show_thread_status(struct thread_stat *ts,
1948 struct group_run_stats *rs,
1949 struct flist_head *opt_list,
1950 struct buf_output *out)
1952 struct json_object *ret = NULL;
1954 if (output_format & FIO_OUTPUT_TERSE)
1955 show_thread_status_terse(ts, rs, out);
1956 if (output_format & FIO_OUTPUT_JSON)
1957 ret = show_thread_status_json(ts, rs, opt_list);
1958 if (output_format & FIO_OUTPUT_NORMAL)
1959 show_thread_status_normal(ts, rs, out);
1964 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1968 dst->min_val = min(dst->min_val, src->min_val);
1969 dst->max_val = max(dst->max_val, src->max_val);
1972 * Compute new mean and S after the merge
1973 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1974 * #Parallel_algorithm>
1977 mean = src->mean.u.f;
1980 double delta = src->mean.u.f - dst->mean.u.f;
1982 mean = ((src->mean.u.f * src->samples) +
1983 (dst->mean.u.f * dst->samples)) /
1984 (dst->samples + src->samples);
1986 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1987 (dst->samples * src->samples) /
1988 (dst->samples + src->samples);
1991 dst->samples += src->samples;
1992 dst->mean.u.f = mean;
1998 * We sum two kinds of stats - one that is time based, in which case we
1999 * apply the proper summing technique, and then one that is iops/bw
2000 * numbers. For group_reporting, we should just add those up, not make
2001 * them the mean of everything.
2003 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
2005 bool first = dst->samples == 0;
2007 if (src->samples == 0)
2011 __sum_stat(dst, src, first);
2016 dst->min_val = src->min_val;
2017 dst->max_val = src->max_val;
2018 dst->samples = src->samples;
2019 dst->mean.u.f = src->mean.u.f;
2020 dst->S.u.f = src->S.u.f;
2022 dst->min_val += src->min_val;
2023 dst->max_val += src->max_val;
2024 dst->samples += src->samples;
2025 dst->mean.u.f += src->mean.u.f;
2026 dst->S.u.f += src->S.u.f;
2030 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
2034 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2035 if (dst->max_run[i] < src->max_run[i])
2036 dst->max_run[i] = src->max_run[i];
2037 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
2038 dst->min_run[i] = src->min_run[i];
2039 if (dst->max_bw[i] < src->max_bw[i])
2040 dst->max_bw[i] = src->max_bw[i];
2041 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
2042 dst->min_bw[i] = src->min_bw[i];
2044 dst->iobytes[i] += src->iobytes[i];
2045 dst->agg[i] += src->agg[i];
2049 dst->kb_base = src->kb_base;
2050 if (!dst->unit_base)
2051 dst->unit_base = src->unit_base;
2053 dst->sig_figs = src->sig_figs;
2057 * Free the clat_prio_stat arrays allocated by alloc_clat_prio_stat_ddir().
2059 void free_clat_prio_stats(struct thread_stat *ts)
2066 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2067 sfree(ts->clat_prio[ddir]);
2068 ts->clat_prio[ddir] = NULL;
2069 ts->nr_clat_prio[ddir] = 0;
2074 * Allocate a clat_prio_stat array. The array has to be allocated/freed using
2075 * smalloc/sfree, so that it is accessible by the process/thread summing the
2078 int alloc_clat_prio_stat_ddir(struct thread_stat *ts, enum fio_ddir ddir,
2081 struct clat_prio_stat *clat_prio;
2084 clat_prio = scalloc(nr_prios, sizeof(*ts->clat_prio[ddir]));
2086 log_err("fio: failed to allocate ts clat data\n");
2090 for (i = 0; i < nr_prios; i++)
2091 clat_prio[i].clat_stat.min_val = ULONG_MAX;
2093 ts->clat_prio[ddir] = clat_prio;
2094 ts->nr_clat_prio[ddir] = nr_prios;
2099 static int grow_clat_prio_stat(struct thread_stat *dst, enum fio_ddir ddir)
2101 int curr_len = dst->nr_clat_prio[ddir];
2104 new_arr = scalloc(curr_len + 1, sizeof(*dst->clat_prio[ddir]));
2106 log_err("fio: failed to grow clat prio array\n");
2110 memcpy(new_arr, dst->clat_prio[ddir],
2111 curr_len * sizeof(*dst->clat_prio[ddir]));
2112 sfree(dst->clat_prio[ddir]);
2114 dst->clat_prio[ddir] = new_arr;
2115 dst->clat_prio[ddir][curr_len].clat_stat.min_val = ULONG_MAX;
2116 dst->nr_clat_prio[ddir]++;
2121 static int find_clat_prio_index(struct thread_stat *dst, enum fio_ddir ddir,
2124 int i, nr_prios = dst->nr_clat_prio[ddir];
2126 for (i = 0; i < nr_prios; i++) {
2127 if (dst->clat_prio[ddir][i].ioprio == ioprio)
2134 static int alloc_or_get_clat_prio_index(struct thread_stat *dst,
2135 enum fio_ddir ddir, uint32_t ioprio,
2138 int index = find_clat_prio_index(dst, ddir, ioprio);
2141 index = dst->nr_clat_prio[ddir];
2143 if (grow_clat_prio_stat(dst, ddir))
2146 dst->clat_prio[ddir][index].ioprio = ioprio;
2154 static int clat_prio_stats_copy(struct thread_stat *dst, struct thread_stat *src,
2155 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2157 size_t sz = sizeof(*src->clat_prio[src_ddir]) *
2158 src->nr_clat_prio[src_ddir];
2160 dst->clat_prio[dst_ddir] = smalloc(sz);
2161 if (!dst->clat_prio[dst_ddir]) {
2162 log_err("fio: failed to alloc clat prio array\n");
2166 memcpy(dst->clat_prio[dst_ddir], src->clat_prio[src_ddir], sz);
2167 dst->nr_clat_prio[dst_ddir] = src->nr_clat_prio[src_ddir];
2172 static int clat_prio_stat_add_samples(struct thread_stat *dst,
2173 enum fio_ddir dst_ddir, uint32_t ioprio,
2174 struct io_stat *io_stat,
2175 uint64_t *io_u_plat)
2179 if (!io_stat->samples)
2182 if (alloc_or_get_clat_prio_index(dst, dst_ddir, ioprio, &dst_index))
2185 sum_stat(&dst->clat_prio[dst_ddir][dst_index].clat_stat, io_stat,
2188 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
2189 dst->clat_prio[dst_ddir][dst_index].io_u_plat[i] += io_u_plat[i];
2194 static int sum_clat_prio_stats_src_single_prio(struct thread_stat *dst,
2195 struct thread_stat *src,
2196 enum fio_ddir dst_ddir,
2197 enum fio_ddir src_ddir)
2199 struct io_stat *io_stat;
2200 uint64_t *io_u_plat;
2203 * If src ts has no clat_prio_stat array, then all I/Os were submitted
2204 * using src->ioprio. Thus, the global samples in src->clat_stat (or
2205 * src->lat_stat) can be used as the 'per prio' samples for src->ioprio.
2207 assert(!src->clat_prio[src_ddir]);
2208 assert(src->nr_clat_prio[src_ddir] == 0);
2210 if (src->lat_percentiles) {
2211 io_u_plat = src->io_u_plat[FIO_LAT][src_ddir];
2212 io_stat = &src->lat_stat[src_ddir];
2214 io_u_plat = src->io_u_plat[FIO_CLAT][src_ddir];
2215 io_stat = &src->clat_stat[src_ddir];
2218 return clat_prio_stat_add_samples(dst, dst_ddir, src->ioprio, io_stat,
2222 static int sum_clat_prio_stats_src_multi_prio(struct thread_stat *dst,
2223 struct thread_stat *src,
2224 enum fio_ddir dst_ddir,
2225 enum fio_ddir src_ddir)
2230 * If src ts has a clat_prio_stat array, then there are multiple prios
2231 * in use (i.e. src ts had cmdprio_percentage or cmdprio_bssplit set).
2232 * The samples for the default prio will exist in the src->clat_prio
2233 * array, just like the samples for any other prio.
2235 assert(src->clat_prio[src_ddir]);
2236 assert(src->nr_clat_prio[src_ddir]);
2238 /* If the dst ts doesn't yet have a clat_prio array, simply memcpy. */
2239 if (!dst->clat_prio[dst_ddir])
2240 return clat_prio_stats_copy(dst, src, dst_ddir, src_ddir);
2242 /* The dst ts already has a clat_prio_array, add src stats into it. */
2243 for (i = 0; i < src->nr_clat_prio[src_ddir]; i++) {
2244 struct io_stat *io_stat = &src->clat_prio[src_ddir][i].clat_stat;
2245 uint64_t *io_u_plat = src->clat_prio[src_ddir][i].io_u_plat;
2246 uint32_t ioprio = src->clat_prio[src_ddir][i].ioprio;
2248 if (clat_prio_stat_add_samples(dst, dst_ddir, ioprio, io_stat, io_u_plat))
2255 static int sum_clat_prio_stats(struct thread_stat *dst, struct thread_stat *src,
2256 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2258 if (dst->disable_prio_stat)
2261 if (!src->clat_prio[src_ddir])
2262 return sum_clat_prio_stats_src_single_prio(dst, src, dst_ddir,
2265 return sum_clat_prio_stats_src_multi_prio(dst, src, dst_ddir, src_ddir);
2268 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2272 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2273 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2274 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2275 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2276 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2277 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2278 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2279 sum_clat_prio_stats(dst, src, l, l);
2281 dst->io_bytes[l] += src->io_bytes[l];
2283 if (dst->runtime[l] < src->runtime[l])
2284 dst->runtime[l] = src->runtime[l];
2286 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2287 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2288 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2289 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2290 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2291 sum_clat_prio_stats(dst, src, 0, l);
2293 dst->io_bytes[0] += src->io_bytes[l];
2295 if (dst->runtime[0] < src->runtime[l])
2296 dst->runtime[0] = src->runtime[l];
2300 sum_stat(&dst->sync_stat, &src->sync_stat, false);
2301 dst->usr_time += src->usr_time;
2302 dst->sys_time += src->sys_time;
2303 dst->ctx += src->ctx;
2304 dst->majf += src->majf;
2305 dst->minf += src->minf;
2307 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2308 dst->io_u_map[k] += src->io_u_map[k];
2309 dst->io_u_submit[k] += src->io_u_submit[k];
2310 dst->io_u_complete[k] += src->io_u_complete[k];
2313 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2314 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2315 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2316 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2317 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2318 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2320 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2321 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2322 dst->total_io_u[k] += src->total_io_u[k];
2323 dst->short_io_u[k] += src->short_io_u[k];
2324 dst->drop_io_u[k] += src->drop_io_u[k];
2326 dst->total_io_u[0] += src->total_io_u[k];
2327 dst->short_io_u[0] += src->short_io_u[k];
2328 dst->drop_io_u[0] += src->drop_io_u[k];
2332 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2334 for (k = 0; k < FIO_LAT_CNT; k++)
2335 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2336 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2337 if (dst->unified_rw_rep != UNIFIED_MIXED)
2338 dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2340 dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2342 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2343 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2345 dst->total_run_time += src->total_run_time;
2346 dst->total_submit += src->total_submit;
2347 dst->total_complete += src->total_complete;
2348 dst->nr_zone_resets += src->nr_zone_resets;
2349 dst->cachehit += src->cachehit;
2350 dst->cachemiss += src->cachemiss;
2353 void init_group_run_stat(struct group_run_stats *gs)
2356 memset(gs, 0, sizeof(*gs));
2358 for (i = 0; i < DDIR_RWDIR_CNT; i++)
2359 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2362 void init_thread_stat_min_vals(struct thread_stat *ts)
2366 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2367 ts->clat_stat[i].min_val = ULONG_MAX;
2368 ts->slat_stat[i].min_val = ULONG_MAX;
2369 ts->lat_stat[i].min_val = ULONG_MAX;
2370 ts->bw_stat[i].min_val = ULONG_MAX;
2371 ts->iops_stat[i].min_val = ULONG_MAX;
2373 ts->sync_stat.min_val = ULONG_MAX;
2376 void init_thread_stat(struct thread_stat *ts)
2378 memset(ts, 0, sizeof(*ts));
2380 init_thread_stat_min_vals(ts);
2384 static void init_per_prio_stats(struct thread_stat *threadstats, int nr_ts)
2386 struct thread_stat *ts;
2387 int i, j, last_ts, idx;
2395 * Loop through all tds, if a td requires per prio stats, temporarily
2396 * store a 1 in ts->disable_prio_stat, and then do an additional
2397 * loop at the end where we invert the ts->disable_prio_stat values.
2403 (!td->o.group_reporting ||
2404 (td->o.group_reporting && last_ts != td->groupid))) {
2409 last_ts = td->groupid;
2410 ts = &threadstats[j];
2412 /* idx == 0 means first td in group, or td is not in a group. */
2414 ts->ioprio = td->ioprio;
2415 else if (td->ioprio != ts->ioprio)
2416 ts->disable_prio_stat = 1;
2418 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2419 if (td->ts.clat_prio[ddir]) {
2420 ts->disable_prio_stat = 1;
2428 /* Loop through all dst threadstats and fixup the values. */
2429 for (i = 0; i < nr_ts; i++) {
2430 ts = &threadstats[i];
2431 ts->disable_prio_stat = !ts->disable_prio_stat;
2435 void __show_run_stats(void)
2437 struct group_run_stats *runstats, *rs;
2438 struct thread_stat *threadstats, *ts;
2439 int i, j, k, nr_ts, last_ts, idx;
2440 bool kb_base_warned = false;
2441 bool unit_base_warned = false;
2442 struct json_object *root = NULL;
2443 struct json_array *array = NULL;
2444 struct buf_output output[FIO_OUTPUT_NR];
2445 struct flist_head **opt_lists;
2447 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2449 for (i = 0; i < groupid + 1; i++)
2450 init_group_run_stat(&runstats[i]);
2453 * find out how many threads stats we need. if group reporting isn't
2454 * enabled, it's one-per-td.
2459 if (!td->o.group_reporting) {
2463 if (last_ts == td->groupid)
2468 last_ts = td->groupid;
2472 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2473 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2475 for (i = 0; i < nr_ts; i++) {
2476 init_thread_stat(&threadstats[i]);
2477 opt_lists[i] = NULL;
2480 init_per_prio_stats(threadstats, nr_ts);
2488 if (idx && (!td->o.group_reporting ||
2489 (td->o.group_reporting && last_ts != td->groupid))) {
2494 last_ts = td->groupid;
2496 ts = &threadstats[j];
2498 ts->clat_percentiles = td->o.clat_percentiles;
2499 ts->lat_percentiles = td->o.lat_percentiles;
2500 ts->slat_percentiles = td->o.slat_percentiles;
2501 ts->percentile_precision = td->o.percentile_precision;
2502 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2503 opt_lists[j] = &td->opt_list;
2507 if (ts->groupid == -1) {
2509 * These are per-group shared already
2511 snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2512 if (td->o.description)
2513 snprintf(ts->description,
2514 sizeof(ts->description), "%s",
2517 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2520 * If multiple entries in this group, this is
2523 ts->thread_number = td->thread_number;
2524 ts->groupid = td->groupid;
2527 * first pid in group, not very useful...
2531 ts->kb_base = td->o.kb_base;
2532 ts->unit_base = td->o.unit_base;
2533 ts->sig_figs = td->o.sig_figs;
2534 ts->unified_rw_rep = td->o.unified_rw_rep;
2535 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2536 log_info("fio: kb_base differs for jobs in group, using"
2537 " %u as the base\n", ts->kb_base);
2538 kb_base_warned = true;
2539 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2540 log_info("fio: unit_base differs for jobs in group, using"
2541 " %u as the base\n", ts->unit_base);
2542 unit_base_warned = true;
2545 ts->continue_on_error = td->o.continue_on_error;
2546 ts->total_err_count += td->total_err_count;
2547 ts->first_error = td->first_error;
2549 if (!td->error && td->o.continue_on_error &&
2551 ts->error = td->first_error;
2552 snprintf(ts->verror, sizeof(ts->verror), "%s",
2554 } else if (td->error) {
2555 ts->error = td->error;
2556 snprintf(ts->verror, sizeof(ts->verror), "%s",
2561 ts->latency_depth = td->latency_qd;
2562 ts->latency_target = td->o.latency_target;
2563 ts->latency_percentile = td->o.latency_percentile;
2564 ts->latency_window = td->o.latency_window;
2566 ts->nr_block_infos = td->ts.nr_block_infos;
2567 for (k = 0; k < ts->nr_block_infos; k++)
2568 ts->block_infos[k] = td->ts.block_infos[k];
2570 sum_thread_stats(ts, &td->ts);
2575 ts->ss_state = td->ss.state;
2576 ts->ss_dur = td->ss.dur;
2577 ts->ss_head = td->ss.head;
2578 ts->ss_bw_data = td->ss.bw_data;
2579 ts->ss_iops_data = td->ss.iops_data;
2580 ts->ss_limit.u.f = td->ss.limit;
2581 ts->ss_slope.u.f = td->ss.slope;
2582 ts->ss_deviation.u.f = td->ss.deviation;
2583 ts->ss_criterion.u.f = td->ss.criterion;
2586 ts->ss_dur = ts->ss_state = 0;
2589 for (i = 0; i < nr_ts; i++) {
2590 unsigned long long bw;
2592 ts = &threadstats[i];
2593 if (ts->groupid == -1)
2595 rs = &runstats[ts->groupid];
2596 rs->kb_base = ts->kb_base;
2597 rs->unit_base = ts->unit_base;
2598 rs->sig_figs = ts->sig_figs;
2599 rs->unified_rw_rep |= ts->unified_rw_rep;
2601 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2602 if (!ts->runtime[j])
2604 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2605 rs->min_run[j] = ts->runtime[j];
2606 if (ts->runtime[j] > rs->max_run[j])
2607 rs->max_run[j] = ts->runtime[j];
2611 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2612 if (bw < rs->min_bw[j])
2614 if (bw > rs->max_bw[j])
2617 rs->iobytes[j] += ts->io_bytes[j];
2621 for (i = 0; i < groupid + 1; i++) {
2626 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2627 if (rs->max_run[ddir])
2628 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2633 for (i = 0; i < FIO_OUTPUT_NR; i++)
2634 buf_output_init(&output[i]);
2637 * don't overwrite last signal output
2639 if (output_format & FIO_OUTPUT_NORMAL)
2640 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2641 if (output_format & FIO_OUTPUT_JSON) {
2642 struct thread_data *global;
2645 unsigned long long ms_since_epoch;
2648 gettimeofday(&now, NULL);
2649 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2650 (unsigned long long)(now.tv_usec) / 1000;
2652 tv_sec = now.tv_sec;
2653 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2654 if (time_buf[strlen(time_buf) - 1] == '\n')
2655 time_buf[strlen(time_buf) - 1] = '\0';
2657 root = json_create_object();
2658 json_object_add_value_string(root, "fio version", fio_version_string);
2659 json_object_add_value_int(root, "timestamp", now.tv_sec);
2660 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2661 json_object_add_value_string(root, "time", time_buf);
2662 global = get_global_options();
2663 json_add_job_opts(root, "global options", &global->opt_list);
2664 array = json_create_array();
2665 json_object_add_value_array(root, "jobs", array);
2669 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2671 for (i = 0; i < nr_ts; i++) {
2672 ts = &threadstats[i];
2673 rs = &runstats[ts->groupid];
2676 fio_server_send_job_options(opt_lists[i], i);
2677 fio_server_send_ts(ts, rs);
2679 if (output_format & FIO_OUTPUT_TERSE)
2680 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2681 if (output_format & FIO_OUTPUT_JSON) {
2682 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2683 json_array_add_value_object(array, tmp);
2685 if (output_format & FIO_OUTPUT_NORMAL)
2686 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2689 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2690 /* disk util stats, if any */
2691 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2693 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2695 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2696 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2697 json_free_object(root);
2700 for (i = 0; i < groupid + 1; i++) {
2705 fio_server_send_gs(rs);
2706 else if (output_format & FIO_OUTPUT_NORMAL)
2707 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2711 fio_server_send_du();
2712 else if (output_format & FIO_OUTPUT_NORMAL) {
2713 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2714 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2717 for (i = 0; i < FIO_OUTPUT_NR; i++) {
2718 struct buf_output *out = &output[i];
2720 log_info_buf(out->buf, out->buflen);
2721 buf_output_free(out);
2724 fio_idle_prof_cleanup();
2729 /* free arrays allocated by sum_thread_stats(), if any */
2730 for (i = 0; i < nr_ts; i++) {
2731 ts = &threadstats[i];
2732 free_clat_prio_stats(ts);
2738 int __show_running_run_stats(void)
2740 unsigned long long *rt;
2743 fio_sem_down(stat_sem);
2745 rt = malloc(thread_number * sizeof(unsigned long long));
2746 fio_gettime(&ts, NULL);
2749 if (td->runstate >= TD_EXITED)
2752 td->update_rusage = 1;
2753 for_each_rw_ddir(ddir) {
2754 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2756 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2758 rt[__td_index] = mtime_since(&td->start, &ts);
2759 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2760 td->ts.runtime[DDIR_READ] += rt[__td_index];
2761 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2762 td->ts.runtime[DDIR_WRITE] += rt[__td_index];
2763 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2764 td->ts.runtime[DDIR_TRIM] += rt[__td_index];
2768 if (td->runstate >= TD_EXITED)
2770 if (td->rusage_sem) {
2771 td->update_rusage = 1;
2772 fio_sem_down(td->rusage_sem);
2774 td->update_rusage = 0;
2780 if (td->runstate >= TD_EXITED)
2783 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2784 td->ts.runtime[DDIR_READ] -= rt[__td_index];
2785 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2786 td->ts.runtime[DDIR_WRITE] -= rt[__td_index];
2787 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2788 td->ts.runtime[DDIR_TRIM] -= rt[__td_index];
2792 fio_sem_up(stat_sem);
2797 static bool status_file_disabled;
2799 #define FIO_STATUS_FILE "fio-dump-status"
2801 static int check_status_file(void)
2804 const char *temp_dir;
2805 char fio_status_file_path[PATH_MAX];
2807 if (status_file_disabled)
2810 temp_dir = getenv("TMPDIR");
2811 if (temp_dir == NULL) {
2812 temp_dir = getenv("TEMP");
2813 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2816 if (temp_dir == NULL)
2819 __coverity_tainted_data_sanitize__(temp_dir);
2822 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2824 if (stat(fio_status_file_path, &sb))
2827 if (unlink(fio_status_file_path) < 0) {
2828 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2830 log_err("fio: disabling status file updates\n");
2831 status_file_disabled = true;
2837 void check_for_running_stats(void)
2839 if (check_status_file()) {
2840 show_running_run_stats();
2845 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2850 if (data > is->max_val)
2852 if (data < is->min_val)
2855 delta = val - is->mean.u.f;
2857 is->mean.u.f += delta / (is->samples + 1.0);
2858 is->S.u.f += delta * (val - is->mean.u.f);
2864 static inline void add_stat_prio_sample(struct clat_prio_stat *clat_prio,
2865 unsigned short clat_prio_index,
2866 unsigned long long nsec)
2869 add_stat_sample(&clat_prio[clat_prio_index].clat_stat, nsec);
2873 * Return a struct io_logs, which is added to the tail of the log
2876 static struct io_logs *get_new_log(struct io_log *iolog)
2879 struct io_logs *cur_log;
2882 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2885 if (!iolog->cur_log_max) {
2887 new_samples = iolog->td->o.log_entries;
2889 new_samples = DEF_LOG_ENTRIES;
2891 new_samples = iolog->cur_log_max * 2;
2892 if (new_samples > MAX_LOG_ENTRIES)
2893 new_samples = MAX_LOG_ENTRIES;
2896 cur_log = smalloc(sizeof(*cur_log));
2898 INIT_FLIST_HEAD(&cur_log->list);
2899 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2901 cur_log->nr_samples = 0;
2902 cur_log->max_samples = new_samples;
2903 flist_add_tail(&cur_log->list, &iolog->io_logs);
2904 iolog->cur_log_max = new_samples;
2914 * Add and return a new log chunk, or return current log if big enough
2916 static struct io_logs *regrow_log(struct io_log *iolog)
2918 struct io_logs *cur_log;
2921 if (!iolog || iolog->disabled)
2924 cur_log = iolog_cur_log(iolog);
2926 cur_log = get_new_log(iolog);
2931 if (cur_log->nr_samples < cur_log->max_samples)
2935 * No room for a new sample. If we're compressing on the fly, flush
2936 * out the current chunk
2938 if (iolog->log_gz) {
2939 if (iolog_cur_flush(iolog, cur_log)) {
2940 log_err("fio: failed flushing iolog! Will stop logging.\n");
2946 * Get a new log array, and add to our list
2948 cur_log = get_new_log(iolog);
2950 log_err("fio: failed extending iolog! Will stop logging.\n");
2954 if (!iolog->pending || !iolog->pending->nr_samples)
2958 * Flush pending items to new log
2960 for (i = 0; i < iolog->pending->nr_samples; i++) {
2961 struct io_sample *src, *dst;
2963 src = get_sample(iolog, iolog->pending, i);
2964 dst = get_sample(iolog, cur_log, i);
2965 memcpy(dst, src, log_entry_sz(iolog));
2967 cur_log->nr_samples = iolog->pending->nr_samples;
2969 iolog->pending->nr_samples = 0;
2973 iolog->disabled = true;
2977 void regrow_logs(struct thread_data *td)
2979 regrow_log(td->slat_log);
2980 regrow_log(td->clat_log);
2981 regrow_log(td->clat_hist_log);
2982 regrow_log(td->lat_log);
2983 regrow_log(td->bw_log);
2984 regrow_log(td->iops_log);
2985 td->flags &= ~TD_F_REGROW_LOGS;
2988 void regrow_agg_logs(void)
2992 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2993 regrow_log(agg_io_log[ddir]);
2996 static struct io_logs *get_cur_log(struct io_log *iolog)
2998 struct io_logs *cur_log;
3000 cur_log = iolog_cur_log(iolog);
3002 cur_log = get_new_log(iolog);
3007 if (cur_log->nr_samples < cur_log->max_samples)
3011 * Out of space. If we're in IO offload mode, or we're not doing
3012 * per unit logging (hence logging happens outside of the IO thread
3013 * as well), add a new log chunk inline. If we're doing inline
3014 * submissions, flag 'td' as needing a log regrow and we'll take
3015 * care of it on the submission side.
3017 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
3018 !per_unit_log(iolog))
3019 return regrow_log(iolog);
3022 iolog->td->flags |= TD_F_REGROW_LOGS;
3024 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
3025 return iolog->pending;
3028 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
3029 enum fio_ddir ddir, unsigned long long bs,
3030 unsigned long t, uint64_t offset,
3031 unsigned int priority)
3033 struct io_logs *cur_log;
3035 if (iolog->disabled)
3037 if (flist_empty(&iolog->io_logs))
3038 iolog->avg_last[ddir] = t;
3040 cur_log = get_cur_log(iolog);
3042 struct io_sample *s;
3044 s = get_sample(iolog, cur_log, cur_log->nr_samples);
3047 s->time = t + (iolog->td ? iolog->td->alternate_epoch : 0);
3048 io_sample_set_ddir(iolog, s, ddir);
3050 s->priority = priority;
3052 if (iolog->log_offset) {
3053 struct io_sample_offset *so = (void *) s;
3055 so->offset = offset;
3058 cur_log->nr_samples++;
3062 iolog->disabled = true;
3065 static inline void reset_io_stat(struct io_stat *ios)
3067 ios->min_val = -1ULL;
3068 ios->max_val = ios->samples = 0;
3069 ios->mean.u.f = ios->S.u.f = 0;
3072 static inline void reset_io_u_plat(uint64_t *io_u_plat)
3076 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
3080 static inline void reset_clat_prio_stats(struct thread_stat *ts)
3085 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3086 if (!ts->clat_prio[ddir])
3089 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
3090 reset_io_stat(&ts->clat_prio[ddir][i].clat_stat);
3091 reset_io_u_plat(ts->clat_prio[ddir][i].io_u_plat);
3096 void reset_io_stats(struct thread_data *td)
3098 struct thread_stat *ts = &td->ts;
3101 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
3102 reset_io_stat(&ts->clat_stat[i]);
3103 reset_io_stat(&ts->slat_stat[i]);
3104 reset_io_stat(&ts->lat_stat[i]);
3105 reset_io_stat(&ts->bw_stat[i]);
3106 reset_io_stat(&ts->iops_stat[i]);
3108 ts->io_bytes[i] = 0;
3110 ts->total_io_u[i] = 0;
3111 ts->short_io_u[i] = 0;
3112 ts->drop_io_u[i] = 0;
3115 for (i = 0; i < FIO_LAT_CNT; i++)
3116 for (j = 0; j < DDIR_RWDIR_CNT; j++)
3117 reset_io_u_plat(ts->io_u_plat[i][j]);
3119 reset_clat_prio_stats(ts);
3121 ts->total_io_u[DDIR_SYNC] = 0;
3122 reset_io_u_plat(ts->io_u_sync_plat);
3124 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
3125 ts->io_u_map[i] = 0;
3126 ts->io_u_submit[i] = 0;
3127 ts->io_u_complete[i] = 0;
3130 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
3131 ts->io_u_lat_n[i] = 0;
3132 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
3133 ts->io_u_lat_u[i] = 0;
3134 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
3135 ts->io_u_lat_m[i] = 0;
3137 ts->total_submit = 0;
3138 ts->total_complete = 0;
3139 ts->nr_zone_resets = 0;
3140 ts->cachehit = ts->cachemiss = 0;
3143 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
3144 unsigned long elapsed, bool log_max)
3147 * Note an entry in the log. Use the mean from the logged samples,
3148 * making sure to properly round up. Only write a log entry if we
3149 * had actual samples done.
3151 if (iolog->avg_window[ddir].samples) {
3152 union io_sample_data data;
3155 data.val = iolog->avg_window[ddir].max_val;
3157 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
3159 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
3162 reset_io_stat(&iolog->avg_window[ddir]);
3165 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
3170 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
3171 __add_stat_to_log(iolog, ddir, elapsed, log_max);
3174 static unsigned long add_log_sample(struct thread_data *td,
3175 struct io_log *iolog,
3176 union io_sample_data data,
3177 enum fio_ddir ddir, unsigned long long bs,
3178 uint64_t offset, unsigned int ioprio)
3180 unsigned long elapsed, this_window;
3185 elapsed = mtime_since_now(&td->epoch);
3188 * If no time averaging, just add the log sample.
3190 if (!iolog->avg_msec) {
3191 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
3197 * Add the sample. If the time period has passed, then
3198 * add that entry to the log and clear.
3200 add_stat_sample(&iolog->avg_window[ddir], data.val);
3203 * If period hasn't passed, adding the above sample is all we
3206 this_window = elapsed - iolog->avg_last[ddir];
3207 if (elapsed < iolog->avg_last[ddir])
3208 return iolog->avg_last[ddir] - elapsed;
3209 else if (this_window < iolog->avg_msec) {
3210 unsigned long diff = iolog->avg_msec - this_window;
3212 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
3216 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
3218 iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
3220 return iolog->avg_msec;
3223 void finalize_logs(struct thread_data *td, bool unit_logs)
3225 unsigned long elapsed;
3227 elapsed = mtime_since_now(&td->epoch);
3229 if (td->clat_log && unit_logs)
3230 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
3231 if (td->slat_log && unit_logs)
3232 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
3233 if (td->lat_log && unit_logs)
3234 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
3235 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
3236 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
3237 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
3238 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
3241 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
3242 unsigned long long bs)
3244 struct io_log *iolog;
3249 iolog = agg_io_log[ddir];
3250 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
3253 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
3255 unsigned int idx = plat_val_to_idx(nsec);
3256 assert(idx < FIO_IO_U_PLAT_NR);
3258 ts->io_u_sync_plat[idx]++;
3259 add_stat_sample(&ts->sync_stat, nsec);
3262 static inline void add_lat_percentile_sample(struct thread_stat *ts,
3263 unsigned long long nsec,
3267 unsigned int idx = plat_val_to_idx(nsec);
3268 assert(idx < FIO_IO_U_PLAT_NR);
3270 ts->io_u_plat[lat][ddir][idx]++;
3274 add_lat_percentile_prio_sample(struct thread_stat *ts, unsigned long long nsec,
3276 unsigned short clat_prio_index)
3278 unsigned int idx = plat_val_to_idx(nsec);
3280 if (ts->clat_prio[ddir])
3281 ts->clat_prio[ddir][clat_prio_index].io_u_plat[idx]++;
3284 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
3285 unsigned long long nsec, unsigned long long bs,
3286 uint64_t offset, unsigned int ioprio,
3287 unsigned short clat_prio_index)
3289 const bool needs_lock = td_async_processing(td);
3290 unsigned long elapsed, this_window;
3291 struct thread_stat *ts = &td->ts;
3292 struct io_log *iolog = td->clat_hist_log;
3297 add_stat_sample(&ts->clat_stat[ddir], nsec);
3300 * When lat_percentiles=1 (default 0), the reported per priority
3301 * percentiles and stats are used for describing total latency values,
3302 * even though the variable names themselves start with clat_.
3304 * Because of the above definition, add a prio stat sample only when
3305 * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3306 * when lat_percentiles=1.
3308 if (!ts->lat_percentiles)
3309 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3313 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3316 if (ts->clat_percentiles) {
3318 * Because of the above definition, add a prio lat percentile
3319 * sample only when lat_percentiles=0. add_lat_sample() will add
3320 * the prio lat percentile sample when lat_percentiles=1.
3322 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3323 if (!ts->lat_percentiles)
3324 add_lat_percentile_prio_sample(ts, nsec, ddir,
3328 if (iolog && iolog->hist_msec) {
3329 struct io_hist *hw = &iolog->hist_window[ddir];
3332 elapsed = mtime_since_now(&td->epoch);
3334 hw->hist_last = elapsed;
3335 this_window = elapsed - hw->hist_last;
3337 if (this_window >= iolog->hist_msec) {
3338 uint64_t *io_u_plat;
3339 struct io_u_plat_entry *dst;
3342 * Make a byte-for-byte copy of the latency histogram
3343 * stored in td->ts.io_u_plat[ddir], recording it in a
3344 * log sample. Note that the matching call to free() is
3345 * located in iolog.c after printing this sample to the
3348 io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3349 dst = malloc(sizeof(struct io_u_plat_entry));
3350 memcpy(&(dst->io_u_plat), io_u_plat,
3351 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3352 flist_add(&dst->list, &hw->list);
3353 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3354 elapsed, offset, ioprio);
3357 * Update the last time we recorded as being now, minus
3358 * any drift in time we encountered before actually
3359 * making the record.
3361 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3367 __td_io_u_unlock(td);
3370 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3371 unsigned long long nsec, unsigned long long bs,
3372 uint64_t offset, unsigned int ioprio)
3374 const bool needs_lock = td_async_processing(td);
3375 struct thread_stat *ts = &td->ts;
3383 add_stat_sample(&ts->slat_stat[ddir], nsec);
3386 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3389 if (ts->slat_percentiles)
3390 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3393 __td_io_u_unlock(td);
3396 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3397 unsigned long long nsec, unsigned long long bs,
3398 uint64_t offset, unsigned int ioprio,
3399 unsigned short clat_prio_index)
3401 const bool needs_lock = td_async_processing(td);
3402 struct thread_stat *ts = &td->ts;
3410 add_stat_sample(&ts->lat_stat[ddir], nsec);
3413 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3417 * When lat_percentiles=1 (default 0), the reported per priority
3418 * percentiles and stats are used for describing total latency values,
3419 * even though the variable names themselves start with clat_.
3421 * Because of the above definition, add a prio stat and prio lat
3422 * percentile sample only when lat_percentiles=1. add_clat_sample() will
3423 * add the prio stat and prio lat percentile sample when
3424 * lat_percentiles=0.
3426 if (ts->lat_percentiles) {
3427 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3428 add_lat_percentile_prio_sample(ts, nsec, ddir, clat_prio_index);
3429 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3433 __td_io_u_unlock(td);
3436 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3437 unsigned int bytes, unsigned long long spent)
3439 const bool needs_lock = td_async_processing(td);
3440 struct thread_stat *ts = &td->ts;
3444 rate = (unsigned long) (bytes * 1000000ULL / spent);
3451 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3454 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3455 bytes, io_u->offset, io_u->ioprio);
3457 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3460 __td_io_u_unlock(td);
3463 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3464 struct timespec *t, unsigned int avg_time,
3465 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3466 struct io_stat *stat, struct io_log *log,
3469 const bool needs_lock = td_async_processing(td);
3470 unsigned long spent, rate;
3472 unsigned long next, next_log;
3474 next_log = avg_time;
3476 spent = mtime_since(parent_tv, t);
3477 if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3478 return avg_time - spent;
3484 * Compute both read and write rates for the interval.
3486 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3489 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3491 continue; /* No entries for interval */
3495 rate = delta * 1000 / spent / 1024; /* KiB/s */
3497 rate = (delta * 1000) / spent;
3501 add_stat_sample(&stat[ddir], rate);
3504 unsigned long long bs = 0;
3506 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3507 bs = td->o.min_bs[ddir];
3509 next = add_log_sample(td, log, sample_val(rate), ddir,
3511 next_log = min(next_log, next);
3514 stat_io_bytes[ddir] = this_io_bytes[ddir];
3520 __td_io_u_unlock(td);
3522 if (spent <= avg_time)
3525 next = avg_time - (1 + spent - avg_time);
3527 return min(next, next_log);
3530 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3532 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3533 td->this_io_bytes, td->stat_io_bytes,
3534 td->ts.bw_stat, td->bw_log, true);
3537 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3540 const bool needs_lock = td_async_processing(td);
3541 struct thread_stat *ts = &td->ts;
3546 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3549 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3550 bytes, io_u->offset, io_u->ioprio);
3552 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3555 __td_io_u_unlock(td);
3558 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3560 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3561 td->this_io_blocks, td->stat_io_blocks,
3562 td->ts.iops_stat, td->iops_log, false);
3566 * Returns msecs to next event
3568 int calc_log_samples(void)
3570 unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3571 struct timespec now;
3572 long elapsed_time = 0;
3574 fio_gettime(&now, NULL);
3577 elapsed_time = mtime_since_now(&td->epoch);
3581 if (in_ramp_time(td) ||
3582 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3583 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3587 (td->bw_log && !per_unit_log(td->bw_log))) {
3588 tmp = add_bw_samples(td, &now);
3591 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3593 if (!td->iops_log ||
3594 (td->iops_log && !per_unit_log(td->iops_log))) {
3595 tmp = add_iops_samples(td, &now);
3598 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3605 /* if log_avg_msec_min has not been changed, set it to 0 */
3606 if (log_avg_msec_min == -1U)
3607 log_avg_msec_min = 0;
3609 if (log_avg_msec_min == 0)
3610 next_mod = elapsed_time;
3612 next_mod = elapsed_time % log_avg_msec_min;
3614 /* correction to keep the time on the log avg msec boundary */
3615 next = min(next, (log_avg_msec_min - next_mod));
3617 return next == ~0U ? 0 : next;
3620 void stat_init(void)
3622 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3625 void stat_exit(void)
3628 * When we have the mutex, we know out-of-band access to it
3631 fio_sem_down(stat_sem);
3632 fio_sem_remove(stat_sem);
3636 * Called from signal handler. Wake up status thread.
3638 void show_running_run_stats(void)
3643 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3645 /* Ignore io_u's which span multiple blocks--they will just get
3646 * inaccurate counts. */
3647 int idx = (io_u->offset - io_u->file->file_offset)
3648 / td->o.bs[DDIR_TRIM];
3649 uint32_t *info = &td->ts.block_infos[idx];
3650 assert(idx < td->ts.nr_block_infos);