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, aggrmerge=%llu/%llu, "
961 "aggrticks=%llu/%llu, aggrin_queue=%llu, "
963 (unsigned long long) agg->ios[0] / agg->slavecount,
964 (unsigned long long) agg->ios[1] / agg->slavecount,
965 (unsigned long long) agg->merges[0] / agg->slavecount,
966 (unsigned long long) agg->merges[1] / agg->slavecount,
967 (unsigned long long) agg->ticks[0] / agg->slavecount,
968 (unsigned long long) agg->ticks[1] / agg->slavecount,
969 (unsigned long long) agg->time_in_queue / agg->slavecount,
972 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
973 (unsigned long long) agg->ios[0] / agg->slavecount,
974 (unsigned long long) agg->ios[1] / agg->slavecount,
975 (unsigned long long) agg->merges[0] / agg->slavecount,
976 (unsigned long long) agg->merges[1] / agg->slavecount,
977 (unsigned long long) agg->ticks[0] / agg->slavecount,
978 (unsigned long long) agg->ticks[1] / agg->slavecount,
979 (unsigned long long) agg->time_in_queue / agg->slavecount,
984 static void aggregate_slaves_stats(struct disk_util *masterdu)
986 struct disk_util_agg *agg = &masterdu->agg;
987 struct disk_util_stat *dus;
988 struct flist_head *entry;
989 struct disk_util *slavedu;
992 flist_for_each(entry, &masterdu->slaves) {
993 slavedu = flist_entry(entry, struct disk_util, slavelist);
995 agg->ios[0] += dus->s.ios[0];
996 agg->ios[1] += dus->s.ios[1];
997 agg->merges[0] += dus->s.merges[0];
998 agg->merges[1] += dus->s.merges[1];
999 agg->sectors[0] += dus->s.sectors[0];
1000 agg->sectors[1] += dus->s.sectors[1];
1001 agg->ticks[0] += dus->s.ticks[0];
1002 agg->ticks[1] += dus->s.ticks[1];
1003 agg->time_in_queue += dus->s.time_in_queue;
1006 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
1007 /* System utilization is the utilization of the
1008 * component with the highest utilization.
1010 if (util > agg->max_util.u.f)
1011 agg->max_util.u.f = util;
1015 if (agg->max_util.u.f > 100.0)
1016 agg->max_util.u.f = 100.0;
1019 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
1020 int terse, struct buf_output *out)
1025 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1030 if (agg->slavecount)
1033 log_buf(out, " %s: ios=%llu/%llu, sectors=%llu/%llu, "
1034 "merge=%llu/%llu, ticks=%llu/%llu, in_queue=%llu, "
1037 (unsigned long long) dus->s.ios[0],
1038 (unsigned long long) dus->s.ios[1],
1039 (unsigned long long) dus->s.sectors[0],
1040 (unsigned long long) dus->s.sectors[1],
1041 (unsigned long long) dus->s.merges[0],
1042 (unsigned long long) dus->s.merges[1],
1043 (unsigned long long) dus->s.ticks[0],
1044 (unsigned long long) dus->s.ticks[1],
1045 (unsigned long long) dus->s.time_in_queue,
1048 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1050 (unsigned long long) dus->s.ios[0],
1051 (unsigned long long) dus->s.ios[1],
1052 (unsigned long long) dus->s.merges[0],
1053 (unsigned long long) dus->s.merges[1],
1054 (unsigned long long) dus->s.ticks[0],
1055 (unsigned long long) dus->s.ticks[1],
1056 (unsigned long long) dus->s.time_in_queue,
1061 * If the device has slaves, aggregate the stats for
1062 * those slave devices also.
1064 show_agg_stats(agg, terse, out);
1070 void json_array_add_disk_util(struct disk_util_stat *dus,
1071 struct disk_util_agg *agg, struct json_array *array)
1073 struct json_object *obj;
1077 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1081 obj = json_create_object();
1082 json_array_add_value_object(array, obj);
1084 json_object_add_value_string(obj, "name", (const char *)dus->name);
1085 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1086 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1087 json_object_add_value_int(obj, "read_sectors", dus->s.sectors[0]);
1088 json_object_add_value_int(obj, "write_sectors", dus->s.sectors[1]);
1089 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1090 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1091 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1092 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1093 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1094 json_object_add_value_float(obj, "util", util);
1097 * If the device has slaves, aggregate the stats for
1098 * those slave devices also.
1100 if (!agg->slavecount)
1102 json_object_add_value_int(obj, "aggr_read_ios",
1103 agg->ios[0] / agg->slavecount);
1104 json_object_add_value_int(obj, "aggr_write_ios",
1105 agg->ios[1] / agg->slavecount);
1106 json_object_add_value_int(obj, "aggr_read_sectors",
1107 agg->sectors[0] / agg->slavecount);
1108 json_object_add_value_int(obj, "aggr_write_sectors",
1109 agg->sectors[1] / agg->slavecount);
1110 json_object_add_value_int(obj, "aggr_read_merges",
1111 agg->merges[0] / agg->slavecount);
1112 json_object_add_value_int(obj, "aggr_write_merge",
1113 agg->merges[1] / agg->slavecount);
1114 json_object_add_value_int(obj, "aggr_read_ticks",
1115 agg->ticks[0] / agg->slavecount);
1116 json_object_add_value_int(obj, "aggr_write_ticks",
1117 agg->ticks[1] / agg->slavecount);
1118 json_object_add_value_int(obj, "aggr_in_queue",
1119 agg->time_in_queue / agg->slavecount);
1120 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1123 static void json_object_add_disk_utils(struct json_object *obj,
1124 struct flist_head *head)
1126 struct json_array *array = json_create_array();
1127 struct flist_head *entry;
1128 struct disk_util *du;
1130 json_object_add_value_array(obj, "disk_util", array);
1132 flist_for_each(entry, head) {
1133 du = flist_entry(entry, struct disk_util, list);
1135 aggregate_slaves_stats(du);
1136 json_array_add_disk_util(&du->dus, &du->agg, array);
1140 void show_disk_util(int terse, struct json_object *parent,
1141 struct buf_output *out)
1143 struct flist_head *entry;
1144 struct disk_util *du;
1147 if (!is_running_backend())
1150 if (flist_empty(&disk_list))
1153 if ((output_format & FIO_OUTPUT_JSON) && parent)
1158 if (!terse && !do_json)
1159 log_buf(out, "\nDisk stats (read/write):\n");
1162 json_object_add_disk_utils(parent, &disk_list);
1163 } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1164 flist_for_each(entry, &disk_list) {
1165 du = flist_entry(entry, struct disk_util, list);
1167 aggregate_slaves_stats(du);
1168 print_disk_util(&du->dus, &du->agg, terse, out);
1173 static void show_thread_status_normal(struct thread_stat *ts,
1174 struct group_run_stats *rs,
1175 struct buf_output *out)
1177 double usr_cpu, sys_cpu;
1178 unsigned long runtime;
1179 double io_u_dist[FIO_IO_U_MAP_NR];
1183 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1186 memset(time_buf, 0, sizeof(time_buf));
1189 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1192 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1193 ts->name, ts->groupid, ts->members,
1194 ts->error, (int) ts->pid, time_buf);
1196 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1197 ts->name, ts->groupid, ts->members,
1198 ts->error, ts->verror, (int) ts->pid,
1202 if (strlen(ts->description))
1203 log_buf(out, " Description : [%s]\n", ts->description);
1205 for_each_rw_ddir(ddir) {
1206 if (ts->io_bytes[ddir])
1207 show_ddir_status(rs, ts, ddir, out);
1210 if (ts->unified_rw_rep == UNIFIED_BOTH)
1211 show_mixed_ddir_status(rs, ts, out);
1213 show_latencies(ts, out);
1215 if (ts->sync_stat.samples)
1216 show_ddir_status(rs, ts, DDIR_SYNC, out);
1218 runtime = ts->total_run_time;
1220 double runt = (double) runtime;
1222 usr_cpu = (double) ts->usr_time * 100 / runt;
1223 sys_cpu = (double) ts->sys_time * 100 / runt;
1229 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1230 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1231 (unsigned long long) ts->ctx,
1232 (unsigned long long) ts->majf,
1233 (unsigned long long) ts->minf);
1235 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1236 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1237 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1238 io_u_dist[1], io_u_dist[2],
1239 io_u_dist[3], io_u_dist[4],
1240 io_u_dist[5], io_u_dist[6]);
1242 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1243 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1244 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1245 io_u_dist[1], io_u_dist[2],
1246 io_u_dist[3], io_u_dist[4],
1247 io_u_dist[5], io_u_dist[6]);
1248 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1249 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1250 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1251 io_u_dist[1], io_u_dist[2],
1252 io_u_dist[3], io_u_dist[4],
1253 io_u_dist[5], io_u_dist[6]);
1254 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
1255 " short=%llu,%llu,%llu,0"
1256 " dropped=%llu,%llu,%llu,0\n",
1257 (unsigned long long) ts->total_io_u[0],
1258 (unsigned long long) ts->total_io_u[1],
1259 (unsigned long long) ts->total_io_u[2],
1260 (unsigned long long) ts->total_io_u[3],
1261 (unsigned long long) ts->short_io_u[0],
1262 (unsigned long long) ts->short_io_u[1],
1263 (unsigned long long) ts->short_io_u[2],
1264 (unsigned long long) ts->drop_io_u[0],
1265 (unsigned long long) ts->drop_io_u[1],
1266 (unsigned long long) ts->drop_io_u[2]);
1267 if (ts->continue_on_error) {
1268 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
1269 (unsigned long long)ts->total_err_count,
1271 strerror(ts->first_error));
1273 if (ts->latency_depth) {
1274 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1275 (unsigned long long)ts->latency_target,
1276 (unsigned long long)ts->latency_window,
1277 ts->latency_percentile.u.f,
1281 if (ts->nr_block_infos)
1282 show_block_infos(ts->nr_block_infos, ts->block_infos,
1283 ts->percentile_list, out);
1286 show_ss_normal(ts, out);
1289 static void show_ddir_status_terse(struct thread_stat *ts,
1290 struct group_run_stats *rs,
1291 enum fio_ddir ddir, int ver,
1292 struct buf_output *out)
1294 unsigned long long min, max, minv, maxv, bw, iops;
1295 unsigned long long *ovals = NULL;
1300 assert(ddir_rw(ddir));
1303 if (ts->runtime[ddir]) {
1304 uint64_t runt = ts->runtime[ddir];
1306 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1307 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1310 log_buf(out, ";%llu;%llu;%llu;%llu",
1311 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1312 (unsigned long long) ts->runtime[ddir]);
1314 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1315 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1317 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1319 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1320 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1322 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1324 if (ts->lat_percentiles) {
1325 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1326 ts->lat_stat[ddir].samples,
1327 ts->percentile_list, &ovals, &maxv,
1329 } else if (ts->clat_percentiles) {
1330 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1331 ts->clat_stat[ddir].samples,
1332 ts->percentile_list, &ovals, &maxv,
1338 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1340 log_buf(out, ";0%%=0");
1343 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1346 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1347 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1349 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1353 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1355 double p_of_agg = 100.0;
1357 if (rs->agg[ddir]) {
1358 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1359 if (p_of_agg > 100.0)
1363 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1365 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1370 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1372 log_buf(out, ";%lu", 0UL);
1374 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1375 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1376 mean, dev, (&ts->iops_stat[ddir])->samples);
1378 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1382 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1383 struct group_run_stats *rs,
1384 int ver, struct buf_output *out)
1386 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1389 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1391 free_clat_prio_stats(ts_lcl);
1395 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1396 uint32_t percentiles,
1397 struct io_stat *lat_stat,
1398 uint64_t *io_u_plat)
1402 unsigned int i, len;
1403 struct json_object *lat_object, *percentile_object, *clat_bins_object;
1404 unsigned long long min, max, maxv, minv, *ovals = NULL;
1406 if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1410 lat_object = json_create_object();
1411 json_object_add_value_int(lat_object, "min", min);
1412 json_object_add_value_int(lat_object, "max", max);
1413 json_object_add_value_float(lat_object, "mean", mean);
1414 json_object_add_value_float(lat_object, "stddev", dev);
1415 json_object_add_value_int(lat_object, "N", lat_stat->samples);
1417 if (percentiles && lat_stat->samples) {
1418 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1419 ts->percentile_list, &ovals, &maxv, &minv);
1421 if (len > FIO_IO_U_LIST_MAX_LEN)
1422 len = FIO_IO_U_LIST_MAX_LEN;
1424 percentile_object = json_create_object();
1425 json_object_add_value_object(lat_object, "percentile", percentile_object);
1426 for (i = 0; i < len; i++) {
1427 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1428 json_object_add_value_int(percentile_object, buf, ovals[i]);
1432 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1433 clat_bins_object = json_create_object();
1434 json_object_add_value_object(lat_object, "bins", clat_bins_object);
1436 for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1438 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1439 json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1447 static void add_ddir_status_json(struct thread_stat *ts,
1448 struct group_run_stats *rs, enum fio_ddir ddir,
1449 struct json_object *parent)
1451 unsigned long long min, max;
1452 unsigned long long bw_bytes, bw;
1453 double mean, dev, iops;
1454 struct json_object *dir_object, *tmp_object;
1455 double p_of_agg = 100.0;
1457 assert(ddir_rw(ddir) || ddir_sync(ddir));
1459 if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1462 dir_object = json_create_object();
1463 json_object_add_value_object(parent,
1464 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1466 if (ddir_rw(ddir)) {
1470 if (ts->runtime[ddir]) {
1471 uint64_t runt = ts->runtime[ddir];
1473 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1474 bw = bw_bytes / 1024; /* KiB/s */
1475 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1478 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1479 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1480 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1481 json_object_add_value_int(dir_object, "bw", bw);
1482 json_object_add_value_float(dir_object, "iops", iops);
1483 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1484 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1485 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1486 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1488 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1489 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1490 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1492 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1493 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1494 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1496 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1497 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1498 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1500 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1501 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1502 &ts->sync_stat, ts->io_u_sync_plat);
1503 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1509 /* Only include per prio stats if there are >= 2 prios with samples */
1510 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
1511 struct json_array *array = json_create_array();
1512 const char *obj_name;
1515 if (ts->lat_percentiles)
1516 obj_name = "lat_ns";
1518 obj_name = "clat_ns";
1520 json_object_add_value_array(dir_object, "prios", array);
1522 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
1523 struct json_object *obj;
1525 if (!ts->clat_prio[ddir][i].clat_stat.samples)
1528 obj = json_create_object();
1530 json_object_add_value_int(obj, "prioclass",
1531 ioprio_class(ts->clat_prio[ddir][i].ioprio));
1532 json_object_add_value_int(obj, "prio",
1533 ioprio(ts->clat_prio[ddir][i].ioprio));
1535 tmp_object = add_ddir_lat_json(ts,
1536 ts->clat_percentiles | ts->lat_percentiles,
1537 &ts->clat_prio[ddir][i].clat_stat,
1538 ts->clat_prio[ddir][i].io_u_plat);
1539 json_object_add_value_object(obj, obj_name, tmp_object);
1540 json_array_add_value_object(array, obj);
1544 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1545 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1548 p_of_agg = mean = dev = 0.0;
1551 json_object_add_value_int(dir_object, "bw_min", min);
1552 json_object_add_value_int(dir_object, "bw_max", max);
1553 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1554 json_object_add_value_float(dir_object, "bw_mean", mean);
1555 json_object_add_value_float(dir_object, "bw_dev", dev);
1556 json_object_add_value_int(dir_object, "bw_samples",
1557 (&ts->bw_stat[ddir])->samples);
1559 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1563 json_object_add_value_int(dir_object, "iops_min", min);
1564 json_object_add_value_int(dir_object, "iops_max", max);
1565 json_object_add_value_float(dir_object, "iops_mean", mean);
1566 json_object_add_value_float(dir_object, "iops_stddev", dev);
1567 json_object_add_value_int(dir_object, "iops_samples",
1568 (&ts->iops_stat[ddir])->samples);
1570 if (ts->cachehit + ts->cachemiss) {
1574 total = ts->cachehit + ts->cachemiss;
1575 hit = (double) ts->cachehit / (double) total;
1577 json_object_add_value_float(dir_object, "cachehit", hit);
1581 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1582 struct group_run_stats *rs, struct json_object *parent)
1584 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1586 /* add the aggregated stats to json parent */
1588 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1590 free_clat_prio_stats(ts_lcl);
1594 static void show_thread_status_terse_all(struct thread_stat *ts,
1595 struct group_run_stats *rs, int ver,
1596 struct buf_output *out)
1598 double io_u_dist[FIO_IO_U_MAP_NR];
1599 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1600 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1601 double usr_cpu, sys_cpu;
1606 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1608 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1609 ts->name, ts->groupid, ts->error);
1611 /* Log Read Status, or mixed if unified_rw_rep = 1 */
1612 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1613 if (ts->unified_rw_rep != UNIFIED_MIXED) {
1614 /* Log Write Status */
1615 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1616 /* Log Trim Status */
1617 if (ver == 2 || ver == 4 || ver == 5)
1618 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1620 if (ts->unified_rw_rep == UNIFIED_BOTH)
1621 show_mixed_ddir_status_terse(ts, rs, ver, out);
1623 if (ts->total_run_time) {
1624 double runt = (double) ts->total_run_time;
1626 usr_cpu = (double) ts->usr_time * 100 / runt;
1627 sys_cpu = (double) ts->sys_time * 100 / runt;
1633 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1634 (unsigned long long) ts->ctx,
1635 (unsigned long long) ts->majf,
1636 (unsigned long long) ts->minf);
1638 /* Calc % distribution of IO depths, usecond, msecond latency */
1639 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1640 stat_calc_lat_nu(ts, io_u_lat_u);
1641 stat_calc_lat_m(ts, io_u_lat_m);
1643 /* Only show fixed 7 I/O depth levels*/
1644 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1645 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1646 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1648 /* Microsecond latency */
1649 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1650 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1651 /* Millisecond latency */
1652 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1653 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1655 /* disk util stats, if any */
1656 if (ver >= 3 && is_running_backend())
1657 show_disk_util(1, NULL, out);
1659 /* Additional output if continue_on_error set - default off*/
1660 if (ts->continue_on_error)
1661 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1663 /* Additional output if description is set */
1664 if (strlen(ts->description)) {
1667 log_buf(out, ";%s", ts->description);
1673 static void json_add_job_opts(struct json_object *root, const char *name,
1674 struct flist_head *opt_list)
1676 struct json_object *dir_object;
1677 struct flist_head *entry;
1678 struct print_option *p;
1680 if (flist_empty(opt_list))
1683 dir_object = json_create_object();
1684 json_object_add_value_object(root, name, dir_object);
1686 flist_for_each(entry, opt_list) {
1687 p = flist_entry(entry, struct print_option, list);
1688 json_object_add_value_string(dir_object, p->name, p->value);
1692 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1693 struct group_run_stats *rs,
1694 struct flist_head *opt_list)
1696 struct json_object *root, *tmp;
1697 struct jobs_eta *je;
1698 double io_u_dist[FIO_IO_U_MAP_NR];
1699 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1700 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1701 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1702 double usr_cpu, sys_cpu;
1706 root = json_create_object();
1707 json_object_add_value_string(root, "jobname", ts->name);
1708 json_object_add_value_int(root, "groupid", ts->groupid);
1709 json_object_add_value_int(root, "error", ts->error);
1712 je = get_jobs_eta(true, &size);
1714 json_object_add_value_int(root, "eta", je->eta_sec);
1715 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1720 json_add_job_opts(root, "job options", opt_list);
1722 add_ddir_status_json(ts, rs, DDIR_READ, root);
1723 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1724 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1725 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1727 if (ts->unified_rw_rep == UNIFIED_BOTH)
1728 add_mixed_ddir_status_json(ts, rs, root);
1731 if (ts->total_run_time) {
1732 double runt = (double) ts->total_run_time;
1734 usr_cpu = (double) ts->usr_time * 100 / runt;
1735 sys_cpu = (double) ts->sys_time * 100 / runt;
1740 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1741 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1742 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1743 json_object_add_value_int(root, "ctx", ts->ctx);
1744 json_object_add_value_int(root, "majf", ts->majf);
1745 json_object_add_value_int(root, "minf", ts->minf);
1747 /* Calc % distribution of IO depths */
1748 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1749 tmp = json_create_object();
1750 json_object_add_value_object(root, "iodepth_level", tmp);
1751 /* Only show fixed 7 I/O depth levels*/
1752 for (i = 0; i < 7; i++) {
1755 snprintf(name, 20, "%d", 1 << i);
1757 snprintf(name, 20, ">=%d", 1 << i);
1758 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1761 /* Calc % distribution of submit IO depths */
1762 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1763 tmp = json_create_object();
1764 json_object_add_value_object(root, "iodepth_submit", tmp);
1765 /* Only show fixed 7 I/O depth levels*/
1766 for (i = 0; i < 7; i++) {
1769 snprintf(name, 20, "0");
1771 snprintf(name, 20, "%d", 1 << (i+1));
1773 snprintf(name, 20, ">=%d", 1 << i);
1774 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1777 /* Calc % distribution of completion IO depths */
1778 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1779 tmp = json_create_object();
1780 json_object_add_value_object(root, "iodepth_complete", tmp);
1781 /* Only show fixed 7 I/O depth levels*/
1782 for (i = 0; i < 7; i++) {
1785 snprintf(name, 20, "0");
1787 snprintf(name, 20, "%d", 1 << (i+1));
1789 snprintf(name, 20, ">=%d", 1 << i);
1790 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1793 /* Calc % distribution of nsecond, usecond, msecond latency */
1794 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1795 stat_calc_lat_n(ts, io_u_lat_n);
1796 stat_calc_lat_u(ts, io_u_lat_u);
1797 stat_calc_lat_m(ts, io_u_lat_m);
1799 /* Nanosecond latency */
1800 tmp = json_create_object();
1801 json_object_add_value_object(root, "latency_ns", tmp);
1802 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1803 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1804 "250", "500", "750", "1000", };
1805 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1807 /* Microsecond latency */
1808 tmp = json_create_object();
1809 json_object_add_value_object(root, "latency_us", tmp);
1810 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1811 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1812 "250", "500", "750", "1000", };
1813 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1815 /* Millisecond latency */
1816 tmp = json_create_object();
1817 json_object_add_value_object(root, "latency_ms", tmp);
1818 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1819 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1820 "250", "500", "750", "1000", "2000",
1822 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1825 /* Additional output if continue_on_error set - default off*/
1826 if (ts->continue_on_error) {
1827 json_object_add_value_int(root, "total_err", ts->total_err_count);
1828 json_object_add_value_int(root, "first_error", ts->first_error);
1831 if (ts->latency_depth) {
1832 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1833 json_object_add_value_int(root, "latency_target", ts->latency_target);
1834 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1835 json_object_add_value_int(root, "latency_window", ts->latency_window);
1838 /* Additional output if description is set */
1839 if (strlen(ts->description))
1840 json_object_add_value_string(root, "desc", ts->description);
1842 if (ts->nr_block_infos) {
1843 /* Block error histogram and types */
1845 unsigned int *percentiles = NULL;
1846 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1848 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1849 ts->percentile_list,
1850 &percentiles, block_state_counts);
1853 struct json_object *block, *percentile_object, *states;
1855 block = json_create_object();
1856 json_object_add_value_object(root, "block", block);
1858 percentile_object = json_create_object();
1859 json_object_add_value_object(block, "percentiles",
1861 for (i = 0; i < len; i++) {
1863 snprintf(buf, sizeof(buf), "%f",
1864 ts->percentile_list[i].u.f);
1865 json_object_add_value_int(percentile_object,
1870 states = json_create_object();
1871 json_object_add_value_object(block, "states", states);
1872 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1873 json_object_add_value_int(states,
1874 block_state_names[state],
1875 block_state_counts[state]);
1882 struct json_object *data;
1883 struct json_array *iops, *bw;
1886 int intervals = ts->ss_dur / (ss_check_interval / 1000L);
1888 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1889 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1890 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1891 (float) ts->ss_limit.u.f,
1892 ts->ss_state & FIO_SS_PCT ? "%" : "");
1894 tmp = json_create_object();
1895 json_object_add_value_object(root, "steadystate", tmp);
1896 json_object_add_value_string(tmp, "ss", ss_buf);
1897 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1898 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1900 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1901 ts->ss_state & FIO_SS_PCT ? "%" : "");
1902 json_object_add_value_string(tmp, "criterion", ss_buf);
1903 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1904 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1906 data = json_create_object();
1907 json_object_add_value_object(tmp, "data", data);
1908 bw = json_create_array();
1909 iops = json_create_array();
1912 ** if ss was attained or the buffer is not full,
1913 ** ss->head points to the first element in the list.
1914 ** otherwise it actually points to the second element
1917 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1920 j = ts->ss_head == 0 ? intervals - 1 : ts->ss_head - 1;
1921 for (l = 0; l < intervals; l++) {
1922 k = (j + l) % intervals;
1923 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1924 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1926 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1927 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1928 json_object_add_value_array(data, "iops", iops);
1929 json_object_add_value_array(data, "bw", bw);
1935 static void show_thread_status_terse(struct thread_stat *ts,
1936 struct group_run_stats *rs,
1937 struct buf_output *out)
1939 if (terse_version >= 2 && terse_version <= 5)
1940 show_thread_status_terse_all(ts, rs, terse_version, out);
1942 log_err("fio: bad terse version!? %d\n", terse_version);
1945 struct json_object *show_thread_status(struct thread_stat *ts,
1946 struct group_run_stats *rs,
1947 struct flist_head *opt_list,
1948 struct buf_output *out)
1950 struct json_object *ret = NULL;
1952 if (output_format & FIO_OUTPUT_TERSE)
1953 show_thread_status_terse(ts, rs, out);
1954 if (output_format & FIO_OUTPUT_JSON)
1955 ret = show_thread_status_json(ts, rs, opt_list);
1956 if (output_format & FIO_OUTPUT_NORMAL)
1957 show_thread_status_normal(ts, rs, out);
1962 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1966 dst->min_val = min(dst->min_val, src->min_val);
1967 dst->max_val = max(dst->max_val, src->max_val);
1970 * Compute new mean and S after the merge
1971 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1972 * #Parallel_algorithm>
1975 mean = src->mean.u.f;
1978 double delta = src->mean.u.f - dst->mean.u.f;
1980 mean = ((src->mean.u.f * src->samples) +
1981 (dst->mean.u.f * dst->samples)) /
1982 (dst->samples + src->samples);
1984 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1985 (dst->samples * src->samples) /
1986 (dst->samples + src->samples);
1989 dst->samples += src->samples;
1990 dst->mean.u.f = mean;
1996 * We sum two kinds of stats - one that is time based, in which case we
1997 * apply the proper summing technique, and then one that is iops/bw
1998 * numbers. For group_reporting, we should just add those up, not make
1999 * them the mean of everything.
2001 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
2003 bool first = dst->samples == 0;
2005 if (src->samples == 0)
2009 __sum_stat(dst, src, first);
2014 dst->min_val = src->min_val;
2015 dst->max_val = src->max_val;
2016 dst->samples = src->samples;
2017 dst->mean.u.f = src->mean.u.f;
2018 dst->S.u.f = src->S.u.f;
2020 dst->min_val += src->min_val;
2021 dst->max_val += src->max_val;
2022 dst->samples += src->samples;
2023 dst->mean.u.f += src->mean.u.f;
2024 dst->S.u.f += src->S.u.f;
2028 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
2032 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2033 if (dst->max_run[i] < src->max_run[i])
2034 dst->max_run[i] = src->max_run[i];
2035 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
2036 dst->min_run[i] = src->min_run[i];
2037 if (dst->max_bw[i] < src->max_bw[i])
2038 dst->max_bw[i] = src->max_bw[i];
2039 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
2040 dst->min_bw[i] = src->min_bw[i];
2042 dst->iobytes[i] += src->iobytes[i];
2043 dst->agg[i] += src->agg[i];
2047 dst->kb_base = src->kb_base;
2048 if (!dst->unit_base)
2049 dst->unit_base = src->unit_base;
2051 dst->sig_figs = src->sig_figs;
2055 * Free the clat_prio_stat arrays allocated by alloc_clat_prio_stat_ddir().
2057 void free_clat_prio_stats(struct thread_stat *ts)
2064 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2065 sfree(ts->clat_prio[ddir]);
2066 ts->clat_prio[ddir] = NULL;
2067 ts->nr_clat_prio[ddir] = 0;
2072 * Allocate a clat_prio_stat array. The array has to be allocated/freed using
2073 * smalloc/sfree, so that it is accessible by the process/thread summing the
2076 int alloc_clat_prio_stat_ddir(struct thread_stat *ts, enum fio_ddir ddir,
2079 struct clat_prio_stat *clat_prio;
2082 clat_prio = scalloc(nr_prios, sizeof(*ts->clat_prio[ddir]));
2084 log_err("fio: failed to allocate ts clat data\n");
2088 for (i = 0; i < nr_prios; i++)
2089 clat_prio[i].clat_stat.min_val = ULONG_MAX;
2091 ts->clat_prio[ddir] = clat_prio;
2092 ts->nr_clat_prio[ddir] = nr_prios;
2097 static int grow_clat_prio_stat(struct thread_stat *dst, enum fio_ddir ddir)
2099 int curr_len = dst->nr_clat_prio[ddir];
2102 new_arr = scalloc(curr_len + 1, sizeof(*dst->clat_prio[ddir]));
2104 log_err("fio: failed to grow clat prio array\n");
2108 memcpy(new_arr, dst->clat_prio[ddir],
2109 curr_len * sizeof(*dst->clat_prio[ddir]));
2110 sfree(dst->clat_prio[ddir]);
2112 dst->clat_prio[ddir] = new_arr;
2113 dst->clat_prio[ddir][curr_len].clat_stat.min_val = ULONG_MAX;
2114 dst->nr_clat_prio[ddir]++;
2119 static int find_clat_prio_index(struct thread_stat *dst, enum fio_ddir ddir,
2122 int i, nr_prios = dst->nr_clat_prio[ddir];
2124 for (i = 0; i < nr_prios; i++) {
2125 if (dst->clat_prio[ddir][i].ioprio == ioprio)
2132 static int alloc_or_get_clat_prio_index(struct thread_stat *dst,
2133 enum fio_ddir ddir, uint32_t ioprio,
2136 int index = find_clat_prio_index(dst, ddir, ioprio);
2139 index = dst->nr_clat_prio[ddir];
2141 if (grow_clat_prio_stat(dst, ddir))
2144 dst->clat_prio[ddir][index].ioprio = ioprio;
2152 static int clat_prio_stats_copy(struct thread_stat *dst, struct thread_stat *src,
2153 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2155 size_t sz = sizeof(*src->clat_prio[src_ddir]) *
2156 src->nr_clat_prio[src_ddir];
2158 dst->clat_prio[dst_ddir] = smalloc(sz);
2159 if (!dst->clat_prio[dst_ddir]) {
2160 log_err("fio: failed to alloc clat prio array\n");
2164 memcpy(dst->clat_prio[dst_ddir], src->clat_prio[src_ddir], sz);
2165 dst->nr_clat_prio[dst_ddir] = src->nr_clat_prio[src_ddir];
2170 static int clat_prio_stat_add_samples(struct thread_stat *dst,
2171 enum fio_ddir dst_ddir, uint32_t ioprio,
2172 struct io_stat *io_stat,
2173 uint64_t *io_u_plat)
2177 if (!io_stat->samples)
2180 if (alloc_or_get_clat_prio_index(dst, dst_ddir, ioprio, &dst_index))
2183 sum_stat(&dst->clat_prio[dst_ddir][dst_index].clat_stat, io_stat,
2186 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
2187 dst->clat_prio[dst_ddir][dst_index].io_u_plat[i] += io_u_plat[i];
2192 static int sum_clat_prio_stats_src_single_prio(struct thread_stat *dst,
2193 struct thread_stat *src,
2194 enum fio_ddir dst_ddir,
2195 enum fio_ddir src_ddir)
2197 struct io_stat *io_stat;
2198 uint64_t *io_u_plat;
2201 * If src ts has no clat_prio_stat array, then all I/Os were submitted
2202 * using src->ioprio. Thus, the global samples in src->clat_stat (or
2203 * src->lat_stat) can be used as the 'per prio' samples for src->ioprio.
2205 assert(!src->clat_prio[src_ddir]);
2206 assert(src->nr_clat_prio[src_ddir] == 0);
2208 if (src->lat_percentiles) {
2209 io_u_plat = src->io_u_plat[FIO_LAT][src_ddir];
2210 io_stat = &src->lat_stat[src_ddir];
2212 io_u_plat = src->io_u_plat[FIO_CLAT][src_ddir];
2213 io_stat = &src->clat_stat[src_ddir];
2216 return clat_prio_stat_add_samples(dst, dst_ddir, src->ioprio, io_stat,
2220 static int sum_clat_prio_stats_src_multi_prio(struct thread_stat *dst,
2221 struct thread_stat *src,
2222 enum fio_ddir dst_ddir,
2223 enum fio_ddir src_ddir)
2228 * If src ts has a clat_prio_stat array, then there are multiple prios
2229 * in use (i.e. src ts had cmdprio_percentage or cmdprio_bssplit set).
2230 * The samples for the default prio will exist in the src->clat_prio
2231 * array, just like the samples for any other prio.
2233 assert(src->clat_prio[src_ddir]);
2234 assert(src->nr_clat_prio[src_ddir]);
2236 /* If the dst ts doesn't yet have a clat_prio array, simply memcpy. */
2237 if (!dst->clat_prio[dst_ddir])
2238 return clat_prio_stats_copy(dst, src, dst_ddir, src_ddir);
2240 /* The dst ts already has a clat_prio_array, add src stats into it. */
2241 for (i = 0; i < src->nr_clat_prio[src_ddir]; i++) {
2242 struct io_stat *io_stat = &src->clat_prio[src_ddir][i].clat_stat;
2243 uint64_t *io_u_plat = src->clat_prio[src_ddir][i].io_u_plat;
2244 uint32_t ioprio = src->clat_prio[src_ddir][i].ioprio;
2246 if (clat_prio_stat_add_samples(dst, dst_ddir, ioprio, io_stat, io_u_plat))
2253 static int sum_clat_prio_stats(struct thread_stat *dst, struct thread_stat *src,
2254 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2256 if (dst->disable_prio_stat)
2259 if (!src->clat_prio[src_ddir])
2260 return sum_clat_prio_stats_src_single_prio(dst, src, dst_ddir,
2263 return sum_clat_prio_stats_src_multi_prio(dst, src, dst_ddir, src_ddir);
2266 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2270 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2271 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2272 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2273 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2274 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2275 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2276 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2277 sum_clat_prio_stats(dst, src, l, l);
2279 dst->io_bytes[l] += src->io_bytes[l];
2281 if (dst->runtime[l] < src->runtime[l])
2282 dst->runtime[l] = src->runtime[l];
2284 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2285 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2286 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2287 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2288 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2289 sum_clat_prio_stats(dst, src, 0, l);
2291 dst->io_bytes[0] += src->io_bytes[l];
2293 if (dst->runtime[0] < src->runtime[l])
2294 dst->runtime[0] = src->runtime[l];
2298 sum_stat(&dst->sync_stat, &src->sync_stat, false);
2299 dst->usr_time += src->usr_time;
2300 dst->sys_time += src->sys_time;
2301 dst->ctx += src->ctx;
2302 dst->majf += src->majf;
2303 dst->minf += src->minf;
2305 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2306 dst->io_u_map[k] += src->io_u_map[k];
2307 dst->io_u_submit[k] += src->io_u_submit[k];
2308 dst->io_u_complete[k] += src->io_u_complete[k];
2311 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2312 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2313 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2314 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2315 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2316 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2318 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2319 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2320 dst->total_io_u[k] += src->total_io_u[k];
2321 dst->short_io_u[k] += src->short_io_u[k];
2322 dst->drop_io_u[k] += src->drop_io_u[k];
2324 dst->total_io_u[0] += src->total_io_u[k];
2325 dst->short_io_u[0] += src->short_io_u[k];
2326 dst->drop_io_u[0] += src->drop_io_u[k];
2330 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2332 for (k = 0; k < FIO_LAT_CNT; k++)
2333 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2334 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2335 if (dst->unified_rw_rep != UNIFIED_MIXED)
2336 dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2338 dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2340 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2341 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2343 dst->total_run_time += src->total_run_time;
2344 dst->total_submit += src->total_submit;
2345 dst->total_complete += src->total_complete;
2346 dst->nr_zone_resets += src->nr_zone_resets;
2347 dst->cachehit += src->cachehit;
2348 dst->cachemiss += src->cachemiss;
2351 void init_group_run_stat(struct group_run_stats *gs)
2354 memset(gs, 0, sizeof(*gs));
2356 for (i = 0; i < DDIR_RWDIR_CNT; i++)
2357 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2360 void init_thread_stat_min_vals(struct thread_stat *ts)
2364 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2365 ts->clat_stat[i].min_val = ULONG_MAX;
2366 ts->slat_stat[i].min_val = ULONG_MAX;
2367 ts->lat_stat[i].min_val = ULONG_MAX;
2368 ts->bw_stat[i].min_val = ULONG_MAX;
2369 ts->iops_stat[i].min_val = ULONG_MAX;
2371 ts->sync_stat.min_val = ULONG_MAX;
2374 void init_thread_stat(struct thread_stat *ts)
2376 memset(ts, 0, sizeof(*ts));
2378 init_thread_stat_min_vals(ts);
2382 static void init_per_prio_stats(struct thread_stat *threadstats, int nr_ts)
2384 struct thread_stat *ts;
2385 int i, j, last_ts, idx;
2393 * Loop through all tds, if a td requires per prio stats, temporarily
2394 * store a 1 in ts->disable_prio_stat, and then do an additional
2395 * loop at the end where we invert the ts->disable_prio_stat values.
2401 (!td->o.group_reporting ||
2402 (td->o.group_reporting && last_ts != td->groupid))) {
2407 last_ts = td->groupid;
2408 ts = &threadstats[j];
2410 /* idx == 0 means first td in group, or td is not in a group. */
2412 ts->ioprio = td->ioprio;
2413 else if (td->ioprio != ts->ioprio)
2414 ts->disable_prio_stat = 1;
2416 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2417 if (td->ts.clat_prio[ddir]) {
2418 ts->disable_prio_stat = 1;
2426 /* Loop through all dst threadstats and fixup the values. */
2427 for (i = 0; i < nr_ts; i++) {
2428 ts = &threadstats[i];
2429 ts->disable_prio_stat = !ts->disable_prio_stat;
2433 void __show_run_stats(void)
2435 struct group_run_stats *runstats, *rs;
2436 struct thread_stat *threadstats, *ts;
2437 int i, j, k, nr_ts, last_ts, idx;
2438 bool kb_base_warned = false;
2439 bool unit_base_warned = false;
2440 struct json_object *root = NULL;
2441 struct json_array *array = NULL;
2442 struct buf_output output[FIO_OUTPUT_NR];
2443 struct flist_head **opt_lists;
2445 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2447 for (i = 0; i < groupid + 1; i++)
2448 init_group_run_stat(&runstats[i]);
2451 * find out how many threads stats we need. if group reporting isn't
2452 * enabled, it's one-per-td.
2457 if (!td->o.group_reporting) {
2461 if (last_ts == td->groupid)
2466 last_ts = td->groupid;
2470 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2471 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2473 for (i = 0; i < nr_ts; i++) {
2474 init_thread_stat(&threadstats[i]);
2475 opt_lists[i] = NULL;
2478 init_per_prio_stats(threadstats, nr_ts);
2486 if (idx && (!td->o.group_reporting ||
2487 (td->o.group_reporting && last_ts != td->groupid))) {
2492 last_ts = td->groupid;
2494 ts = &threadstats[j];
2496 ts->clat_percentiles = td->o.clat_percentiles;
2497 ts->lat_percentiles = td->o.lat_percentiles;
2498 ts->slat_percentiles = td->o.slat_percentiles;
2499 ts->percentile_precision = td->o.percentile_precision;
2500 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2501 opt_lists[j] = &td->opt_list;
2505 if (ts->groupid == -1) {
2507 * These are per-group shared already
2509 snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2510 if (td->o.description)
2511 snprintf(ts->description,
2512 sizeof(ts->description), "%s",
2515 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2518 * If multiple entries in this group, this is
2521 ts->thread_number = td->thread_number;
2522 ts->groupid = td->groupid;
2525 * first pid in group, not very useful...
2529 ts->kb_base = td->o.kb_base;
2530 ts->unit_base = td->o.unit_base;
2531 ts->sig_figs = td->o.sig_figs;
2532 ts->unified_rw_rep = td->o.unified_rw_rep;
2533 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2534 log_info("fio: kb_base differs for jobs in group, using"
2535 " %u as the base\n", ts->kb_base);
2536 kb_base_warned = true;
2537 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2538 log_info("fio: unit_base differs for jobs in group, using"
2539 " %u as the base\n", ts->unit_base);
2540 unit_base_warned = true;
2543 ts->continue_on_error = td->o.continue_on_error;
2544 ts->total_err_count += td->total_err_count;
2545 ts->first_error = td->first_error;
2547 if (!td->error && td->o.continue_on_error &&
2549 ts->error = td->first_error;
2550 snprintf(ts->verror, sizeof(ts->verror), "%s",
2552 } else if (td->error) {
2553 ts->error = td->error;
2554 snprintf(ts->verror, sizeof(ts->verror), "%s",
2559 ts->latency_depth = td->latency_qd;
2560 ts->latency_target = td->o.latency_target;
2561 ts->latency_percentile = td->o.latency_percentile;
2562 ts->latency_window = td->o.latency_window;
2564 ts->nr_block_infos = td->ts.nr_block_infos;
2565 for (k = 0; k < ts->nr_block_infos; k++)
2566 ts->block_infos[k] = td->ts.block_infos[k];
2568 sum_thread_stats(ts, &td->ts);
2573 ts->ss_state = td->ss.state;
2574 ts->ss_dur = td->ss.dur;
2575 ts->ss_head = td->ss.head;
2576 ts->ss_bw_data = td->ss.bw_data;
2577 ts->ss_iops_data = td->ss.iops_data;
2578 ts->ss_limit.u.f = td->ss.limit;
2579 ts->ss_slope.u.f = td->ss.slope;
2580 ts->ss_deviation.u.f = td->ss.deviation;
2581 ts->ss_criterion.u.f = td->ss.criterion;
2584 ts->ss_dur = ts->ss_state = 0;
2587 for (i = 0; i < nr_ts; i++) {
2588 unsigned long long bw;
2590 ts = &threadstats[i];
2591 if (ts->groupid == -1)
2593 rs = &runstats[ts->groupid];
2594 rs->kb_base = ts->kb_base;
2595 rs->unit_base = ts->unit_base;
2596 rs->sig_figs = ts->sig_figs;
2597 rs->unified_rw_rep |= ts->unified_rw_rep;
2599 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2600 if (!ts->runtime[j])
2602 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2603 rs->min_run[j] = ts->runtime[j];
2604 if (ts->runtime[j] > rs->max_run[j])
2605 rs->max_run[j] = ts->runtime[j];
2609 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2610 if (bw < rs->min_bw[j])
2612 if (bw > rs->max_bw[j])
2615 rs->iobytes[j] += ts->io_bytes[j];
2619 for (i = 0; i < groupid + 1; i++) {
2624 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2625 if (rs->max_run[ddir])
2626 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2631 for (i = 0; i < FIO_OUTPUT_NR; i++)
2632 buf_output_init(&output[i]);
2635 * don't overwrite last signal output
2637 if (output_format & FIO_OUTPUT_NORMAL)
2638 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2639 if (output_format & FIO_OUTPUT_JSON) {
2640 struct thread_data *global;
2643 unsigned long long ms_since_epoch;
2646 gettimeofday(&now, NULL);
2647 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2648 (unsigned long long)(now.tv_usec) / 1000;
2650 tv_sec = now.tv_sec;
2651 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2652 if (time_buf[strlen(time_buf) - 1] == '\n')
2653 time_buf[strlen(time_buf) - 1] = '\0';
2655 root = json_create_object();
2656 json_object_add_value_string(root, "fio version", fio_version_string);
2657 json_object_add_value_int(root, "timestamp", now.tv_sec);
2658 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2659 json_object_add_value_string(root, "time", time_buf);
2660 global = get_global_options();
2661 json_add_job_opts(root, "global options", &global->opt_list);
2662 array = json_create_array();
2663 json_object_add_value_array(root, "jobs", array);
2667 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2669 for (i = 0; i < nr_ts; i++) {
2670 ts = &threadstats[i];
2671 rs = &runstats[ts->groupid];
2674 fio_server_send_job_options(opt_lists[i], i);
2675 fio_server_send_ts(ts, rs);
2677 if (output_format & FIO_OUTPUT_TERSE)
2678 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2679 if (output_format & FIO_OUTPUT_JSON) {
2680 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2681 json_array_add_value_object(array, tmp);
2683 if (output_format & FIO_OUTPUT_NORMAL)
2684 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2687 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2688 /* disk util stats, if any */
2689 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2691 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2693 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2694 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2695 json_free_object(root);
2698 for (i = 0; i < groupid + 1; i++) {
2703 fio_server_send_gs(rs);
2704 else if (output_format & FIO_OUTPUT_NORMAL)
2705 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2709 fio_server_send_du();
2710 else if (output_format & FIO_OUTPUT_NORMAL) {
2711 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2712 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2715 for (i = 0; i < FIO_OUTPUT_NR; i++) {
2716 struct buf_output *out = &output[i];
2718 log_info_buf(out->buf, out->buflen);
2719 buf_output_free(out);
2722 fio_idle_prof_cleanup();
2727 /* free arrays allocated by sum_thread_stats(), if any */
2728 for (i = 0; i < nr_ts; i++) {
2729 ts = &threadstats[i];
2730 free_clat_prio_stats(ts);
2736 int __show_running_run_stats(void)
2738 unsigned long long *rt;
2741 fio_sem_down(stat_sem);
2743 rt = malloc(thread_number * sizeof(unsigned long long));
2744 fio_gettime(&ts, NULL);
2747 if (td->runstate >= TD_EXITED)
2750 td->update_rusage = 1;
2751 for_each_rw_ddir(ddir) {
2752 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2754 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2756 rt[__td_index] = mtime_since(&td->start, &ts);
2757 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2758 td->ts.runtime[DDIR_READ] += rt[__td_index];
2759 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2760 td->ts.runtime[DDIR_WRITE] += rt[__td_index];
2761 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2762 td->ts.runtime[DDIR_TRIM] += rt[__td_index];
2766 if (td->runstate >= TD_EXITED)
2768 if (td->rusage_sem) {
2769 td->update_rusage = 1;
2770 fio_sem_down(td->rusage_sem);
2772 td->update_rusage = 0;
2778 if (td->runstate >= TD_EXITED)
2781 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2782 td->ts.runtime[DDIR_READ] -= rt[__td_index];
2783 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2784 td->ts.runtime[DDIR_WRITE] -= rt[__td_index];
2785 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2786 td->ts.runtime[DDIR_TRIM] -= rt[__td_index];
2790 fio_sem_up(stat_sem);
2795 static bool status_file_disabled;
2797 #define FIO_STATUS_FILE "fio-dump-status"
2799 static int check_status_file(void)
2802 const char *temp_dir;
2803 char fio_status_file_path[PATH_MAX];
2805 if (status_file_disabled)
2808 temp_dir = getenv("TMPDIR");
2809 if (temp_dir == NULL) {
2810 temp_dir = getenv("TEMP");
2811 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2814 if (temp_dir == NULL)
2817 __coverity_tainted_data_sanitize__(temp_dir);
2820 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2822 if (stat(fio_status_file_path, &sb))
2825 if (unlink(fio_status_file_path) < 0) {
2826 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2828 log_err("fio: disabling status file updates\n");
2829 status_file_disabled = true;
2835 void check_for_running_stats(void)
2837 if (check_status_file()) {
2838 show_running_run_stats();
2843 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2848 if (data > is->max_val)
2850 if (data < is->min_val)
2853 delta = val - is->mean.u.f;
2855 is->mean.u.f += delta / (is->samples + 1.0);
2856 is->S.u.f += delta * (val - is->mean.u.f);
2862 static inline void add_stat_prio_sample(struct clat_prio_stat *clat_prio,
2863 unsigned short clat_prio_index,
2864 unsigned long long nsec)
2867 add_stat_sample(&clat_prio[clat_prio_index].clat_stat, nsec);
2871 * Return a struct io_logs, which is added to the tail of the log
2874 static struct io_logs *get_new_log(struct io_log *iolog)
2877 struct io_logs *cur_log;
2880 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2883 if (!iolog->cur_log_max) {
2885 new_samples = iolog->td->o.log_entries;
2887 new_samples = DEF_LOG_ENTRIES;
2889 new_samples = iolog->cur_log_max * 2;
2890 if (new_samples > MAX_LOG_ENTRIES)
2891 new_samples = MAX_LOG_ENTRIES;
2894 cur_log = smalloc(sizeof(*cur_log));
2896 INIT_FLIST_HEAD(&cur_log->list);
2897 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2899 cur_log->nr_samples = 0;
2900 cur_log->max_samples = new_samples;
2901 flist_add_tail(&cur_log->list, &iolog->io_logs);
2902 iolog->cur_log_max = new_samples;
2912 * Add and return a new log chunk, or return current log if big enough
2914 static struct io_logs *regrow_log(struct io_log *iolog)
2916 struct io_logs *cur_log;
2919 if (!iolog || iolog->disabled)
2922 cur_log = iolog_cur_log(iolog);
2924 cur_log = get_new_log(iolog);
2929 if (cur_log->nr_samples < cur_log->max_samples)
2933 * No room for a new sample. If we're compressing on the fly, flush
2934 * out the current chunk
2936 if (iolog->log_gz) {
2937 if (iolog_cur_flush(iolog, cur_log)) {
2938 log_err("fio: failed flushing iolog! Will stop logging.\n");
2944 * Get a new log array, and add to our list
2946 cur_log = get_new_log(iolog);
2948 log_err("fio: failed extending iolog! Will stop logging.\n");
2952 if (!iolog->pending || !iolog->pending->nr_samples)
2956 * Flush pending items to new log
2958 for (i = 0; i < iolog->pending->nr_samples; i++) {
2959 struct io_sample *src, *dst;
2961 src = get_sample(iolog, iolog->pending, i);
2962 dst = get_sample(iolog, cur_log, i);
2963 memcpy(dst, src, log_entry_sz(iolog));
2965 cur_log->nr_samples = iolog->pending->nr_samples;
2967 iolog->pending->nr_samples = 0;
2971 iolog->disabled = true;
2975 void regrow_logs(struct thread_data *td)
2977 regrow_log(td->slat_log);
2978 regrow_log(td->clat_log);
2979 regrow_log(td->clat_hist_log);
2980 regrow_log(td->lat_log);
2981 regrow_log(td->bw_log);
2982 regrow_log(td->iops_log);
2983 td->flags &= ~TD_F_REGROW_LOGS;
2986 void regrow_agg_logs(void)
2990 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2991 regrow_log(agg_io_log[ddir]);
2994 static struct io_logs *get_cur_log(struct io_log *iolog)
2996 struct io_logs *cur_log;
2998 cur_log = iolog_cur_log(iolog);
3000 cur_log = get_new_log(iolog);
3005 if (cur_log->nr_samples < cur_log->max_samples)
3009 * Out of space. If we're in IO offload mode, or we're not doing
3010 * per unit logging (hence logging happens outside of the IO thread
3011 * as well), add a new log chunk inline. If we're doing inline
3012 * submissions, flag 'td' as needing a log regrow and we'll take
3013 * care of it on the submission side.
3015 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
3016 !per_unit_log(iolog))
3017 return regrow_log(iolog);
3020 iolog->td->flags |= TD_F_REGROW_LOGS;
3022 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
3023 return iolog->pending;
3026 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
3027 enum fio_ddir ddir, unsigned long long bs,
3028 unsigned long t, uint64_t offset,
3029 unsigned int priority)
3031 struct io_logs *cur_log;
3033 if (iolog->disabled)
3035 if (flist_empty(&iolog->io_logs))
3036 iolog->avg_last[ddir] = t;
3038 cur_log = get_cur_log(iolog);
3040 struct io_sample *s;
3042 s = get_sample(iolog, cur_log, cur_log->nr_samples);
3045 s->time = t + (iolog->td ? iolog->td->alternate_epoch : 0);
3046 io_sample_set_ddir(iolog, s, ddir);
3048 s->priority = priority;
3050 if (iolog->log_offset) {
3051 struct io_sample_offset *so = (void *) s;
3053 so->offset = offset;
3056 cur_log->nr_samples++;
3060 iolog->disabled = true;
3063 static inline void reset_io_stat(struct io_stat *ios)
3065 ios->min_val = -1ULL;
3066 ios->max_val = ios->samples = 0;
3067 ios->mean.u.f = ios->S.u.f = 0;
3070 static inline void reset_io_u_plat(uint64_t *io_u_plat)
3074 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
3078 static inline void reset_clat_prio_stats(struct thread_stat *ts)
3083 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3084 if (!ts->clat_prio[ddir])
3087 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
3088 reset_io_stat(&ts->clat_prio[ddir][i].clat_stat);
3089 reset_io_u_plat(ts->clat_prio[ddir][i].io_u_plat);
3094 void reset_io_stats(struct thread_data *td)
3096 struct thread_stat *ts = &td->ts;
3099 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
3100 reset_io_stat(&ts->clat_stat[i]);
3101 reset_io_stat(&ts->slat_stat[i]);
3102 reset_io_stat(&ts->lat_stat[i]);
3103 reset_io_stat(&ts->bw_stat[i]);
3104 reset_io_stat(&ts->iops_stat[i]);
3106 ts->io_bytes[i] = 0;
3108 ts->total_io_u[i] = 0;
3109 ts->short_io_u[i] = 0;
3110 ts->drop_io_u[i] = 0;
3113 for (i = 0; i < FIO_LAT_CNT; i++)
3114 for (j = 0; j < DDIR_RWDIR_CNT; j++)
3115 reset_io_u_plat(ts->io_u_plat[i][j]);
3117 reset_clat_prio_stats(ts);
3119 ts->total_io_u[DDIR_SYNC] = 0;
3120 reset_io_u_plat(ts->io_u_sync_plat);
3122 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
3123 ts->io_u_map[i] = 0;
3124 ts->io_u_submit[i] = 0;
3125 ts->io_u_complete[i] = 0;
3128 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
3129 ts->io_u_lat_n[i] = 0;
3130 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
3131 ts->io_u_lat_u[i] = 0;
3132 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
3133 ts->io_u_lat_m[i] = 0;
3135 ts->total_submit = 0;
3136 ts->total_complete = 0;
3137 ts->nr_zone_resets = 0;
3138 ts->cachehit = ts->cachemiss = 0;
3141 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
3142 unsigned long elapsed, bool log_max)
3145 * Note an entry in the log. Use the mean from the logged samples,
3146 * making sure to properly round up. Only write a log entry if we
3147 * had actual samples done.
3149 if (iolog->avg_window[ddir].samples) {
3150 union io_sample_data data;
3153 data.val = iolog->avg_window[ddir].max_val;
3155 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
3157 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
3160 reset_io_stat(&iolog->avg_window[ddir]);
3163 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
3168 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
3169 __add_stat_to_log(iolog, ddir, elapsed, log_max);
3172 static unsigned long add_log_sample(struct thread_data *td,
3173 struct io_log *iolog,
3174 union io_sample_data data,
3175 enum fio_ddir ddir, unsigned long long bs,
3176 uint64_t offset, unsigned int ioprio)
3178 unsigned long elapsed, this_window;
3183 elapsed = mtime_since_now(&td->epoch);
3186 * If no time averaging, just add the log sample.
3188 if (!iolog->avg_msec) {
3189 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
3195 * Add the sample. If the time period has passed, then
3196 * add that entry to the log and clear.
3198 add_stat_sample(&iolog->avg_window[ddir], data.val);
3201 * If period hasn't passed, adding the above sample is all we
3204 this_window = elapsed - iolog->avg_last[ddir];
3205 if (elapsed < iolog->avg_last[ddir])
3206 return iolog->avg_last[ddir] - elapsed;
3207 else if (this_window < iolog->avg_msec) {
3208 unsigned long diff = iolog->avg_msec - this_window;
3210 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
3214 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
3216 iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
3218 return iolog->avg_msec;
3221 void finalize_logs(struct thread_data *td, bool unit_logs)
3223 unsigned long elapsed;
3225 elapsed = mtime_since_now(&td->epoch);
3227 if (td->clat_log && unit_logs)
3228 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
3229 if (td->slat_log && unit_logs)
3230 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
3231 if (td->lat_log && unit_logs)
3232 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
3233 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
3234 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
3235 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
3236 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
3239 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
3240 unsigned long long bs)
3242 struct io_log *iolog;
3247 iolog = agg_io_log[ddir];
3248 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
3251 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
3253 unsigned int idx = plat_val_to_idx(nsec);
3254 assert(idx < FIO_IO_U_PLAT_NR);
3256 ts->io_u_sync_plat[idx]++;
3257 add_stat_sample(&ts->sync_stat, nsec);
3260 static inline void add_lat_percentile_sample(struct thread_stat *ts,
3261 unsigned long long nsec,
3265 unsigned int idx = plat_val_to_idx(nsec);
3266 assert(idx < FIO_IO_U_PLAT_NR);
3268 ts->io_u_plat[lat][ddir][idx]++;
3272 add_lat_percentile_prio_sample(struct thread_stat *ts, unsigned long long nsec,
3274 unsigned short clat_prio_index)
3276 unsigned int idx = plat_val_to_idx(nsec);
3278 if (ts->clat_prio[ddir])
3279 ts->clat_prio[ddir][clat_prio_index].io_u_plat[idx]++;
3282 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
3283 unsigned long long nsec, unsigned long long bs,
3284 uint64_t offset, unsigned int ioprio,
3285 unsigned short clat_prio_index)
3287 const bool needs_lock = td_async_processing(td);
3288 unsigned long elapsed, this_window;
3289 struct thread_stat *ts = &td->ts;
3290 struct io_log *iolog = td->clat_hist_log;
3295 add_stat_sample(&ts->clat_stat[ddir], nsec);
3298 * When lat_percentiles=1 (default 0), the reported per priority
3299 * percentiles and stats are used for describing total latency values,
3300 * even though the variable names themselves start with clat_.
3302 * Because of the above definition, add a prio stat sample only when
3303 * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3304 * when lat_percentiles=1.
3306 if (!ts->lat_percentiles)
3307 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3311 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3314 if (ts->clat_percentiles) {
3316 * Because of the above definition, add a prio lat percentile
3317 * sample only when lat_percentiles=0. add_lat_sample() will add
3318 * the prio lat percentile sample when lat_percentiles=1.
3320 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3321 if (!ts->lat_percentiles)
3322 add_lat_percentile_prio_sample(ts, nsec, ddir,
3326 if (iolog && iolog->hist_msec) {
3327 struct io_hist *hw = &iolog->hist_window[ddir];
3330 elapsed = mtime_since_now(&td->epoch);
3332 hw->hist_last = elapsed;
3333 this_window = elapsed - hw->hist_last;
3335 if (this_window >= iolog->hist_msec) {
3336 uint64_t *io_u_plat;
3337 struct io_u_plat_entry *dst;
3340 * Make a byte-for-byte copy of the latency histogram
3341 * stored in td->ts.io_u_plat[ddir], recording it in a
3342 * log sample. Note that the matching call to free() is
3343 * located in iolog.c after printing this sample to the
3346 io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3347 dst = malloc(sizeof(struct io_u_plat_entry));
3348 memcpy(&(dst->io_u_plat), io_u_plat,
3349 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3350 flist_add(&dst->list, &hw->list);
3351 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3352 elapsed, offset, ioprio);
3355 * Update the last time we recorded as being now, minus
3356 * any drift in time we encountered before actually
3357 * making the record.
3359 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3365 __td_io_u_unlock(td);
3368 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3369 unsigned long long nsec, unsigned long long bs,
3370 uint64_t offset, unsigned int ioprio)
3372 const bool needs_lock = td_async_processing(td);
3373 struct thread_stat *ts = &td->ts;
3381 add_stat_sample(&ts->slat_stat[ddir], nsec);
3384 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3387 if (ts->slat_percentiles)
3388 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3391 __td_io_u_unlock(td);
3394 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3395 unsigned long long nsec, unsigned long long bs,
3396 uint64_t offset, unsigned int ioprio,
3397 unsigned short clat_prio_index)
3399 const bool needs_lock = td_async_processing(td);
3400 struct thread_stat *ts = &td->ts;
3408 add_stat_sample(&ts->lat_stat[ddir], nsec);
3411 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3415 * When lat_percentiles=1 (default 0), the reported per priority
3416 * percentiles and stats are used for describing total latency values,
3417 * even though the variable names themselves start with clat_.
3419 * Because of the above definition, add a prio stat and prio lat
3420 * percentile sample only when lat_percentiles=1. add_clat_sample() will
3421 * add the prio stat and prio lat percentile sample when
3422 * lat_percentiles=0.
3424 if (ts->lat_percentiles) {
3425 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3426 add_lat_percentile_prio_sample(ts, nsec, ddir, clat_prio_index);
3427 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3431 __td_io_u_unlock(td);
3434 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3435 unsigned int bytes, unsigned long long spent)
3437 const bool needs_lock = td_async_processing(td);
3438 struct thread_stat *ts = &td->ts;
3442 rate = (unsigned long) (bytes * 1000000ULL / spent);
3449 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3452 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3453 bytes, io_u->offset, io_u->ioprio);
3455 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3458 __td_io_u_unlock(td);
3461 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3462 struct timespec *t, unsigned int avg_time,
3463 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3464 struct io_stat *stat, struct io_log *log,
3467 const bool needs_lock = td_async_processing(td);
3468 unsigned long spent, rate;
3470 unsigned long next, next_log;
3472 next_log = avg_time;
3474 spent = mtime_since(parent_tv, t);
3475 if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3476 return avg_time - spent;
3482 * Compute both read and write rates for the interval.
3484 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3487 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3489 continue; /* No entries for interval */
3493 rate = delta * 1000 / spent / 1024; /* KiB/s */
3495 rate = (delta * 1000) / spent;
3499 add_stat_sample(&stat[ddir], rate);
3502 unsigned long long bs = 0;
3504 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3505 bs = td->o.min_bs[ddir];
3507 next = add_log_sample(td, log, sample_val(rate), ddir,
3509 next_log = min(next_log, next);
3512 stat_io_bytes[ddir] = this_io_bytes[ddir];
3518 __td_io_u_unlock(td);
3520 if (spent <= avg_time)
3523 next = avg_time - (1 + spent - avg_time);
3525 return min(next, next_log);
3528 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3530 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3531 td->this_io_bytes, td->stat_io_bytes,
3532 td->ts.bw_stat, td->bw_log, true);
3535 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3538 const bool needs_lock = td_async_processing(td);
3539 struct thread_stat *ts = &td->ts;
3544 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3547 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3548 bytes, io_u->offset, io_u->ioprio);
3550 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3553 __td_io_u_unlock(td);
3556 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3558 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3559 td->this_io_blocks, td->stat_io_blocks,
3560 td->ts.iops_stat, td->iops_log, false);
3564 * Returns msecs to next event
3566 int calc_log_samples(void)
3568 unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3569 struct timespec now;
3570 long elapsed_time = 0;
3572 fio_gettime(&now, NULL);
3575 elapsed_time = mtime_since_now(&td->epoch);
3579 if (in_ramp_time(td) ||
3580 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3581 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3585 (td->bw_log && !per_unit_log(td->bw_log))) {
3586 tmp = add_bw_samples(td, &now);
3589 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3591 if (!td->iops_log ||
3592 (td->iops_log && !per_unit_log(td->iops_log))) {
3593 tmp = add_iops_samples(td, &now);
3596 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3603 /* if log_avg_msec_min has not been changed, set it to 0 */
3604 if (log_avg_msec_min == -1U)
3605 log_avg_msec_min = 0;
3607 if (log_avg_msec_min == 0)
3608 next_mod = elapsed_time;
3610 next_mod = elapsed_time % log_avg_msec_min;
3612 /* correction to keep the time on the log avg msec boundary */
3613 next = min(next, (log_avg_msec_min - next_mod));
3615 return next == ~0U ? 0 : next;
3618 void stat_init(void)
3620 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3623 void stat_exit(void)
3626 * When we have the mutex, we know out-of-band access to it
3629 fio_sem_down(stat_sem);
3630 fio_sem_remove(stat_sem);
3634 * Called from signal handler. Wake up status thread.
3636 void show_running_run_stats(void)
3641 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3643 /* Ignore io_u's which span multiple blocks--they will just get
3644 * inaccurate counts. */
3645 int idx = (io_u->offset - io_u->file->file_offset)
3646 / td->o.bs[DDIR_TRIM];
3647 uint32_t *info = &td->ts.block_infos[idx];
3648 assert(idx < td->ts.nr_block_infos);