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_merges", dus->s.merges[0]);
1088 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1089 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1090 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1091 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1092 json_object_add_value_float(obj, "util", util);
1095 * If the device has slaves, aggregate the stats for
1096 * those slave devices also.
1098 if (!agg->slavecount)
1100 json_object_add_value_int(obj, "aggr_read_ios",
1101 agg->ios[0] / agg->slavecount);
1102 json_object_add_value_int(obj, "aggr_write_ios",
1103 agg->ios[1] / agg->slavecount);
1104 json_object_add_value_int(obj, "aggr_read_merges",
1105 agg->merges[0] / agg->slavecount);
1106 json_object_add_value_int(obj, "aggr_write_merge",
1107 agg->merges[1] / agg->slavecount);
1108 json_object_add_value_int(obj, "aggr_read_ticks",
1109 agg->ticks[0] / agg->slavecount);
1110 json_object_add_value_int(obj, "aggr_write_ticks",
1111 agg->ticks[1] / agg->slavecount);
1112 json_object_add_value_int(obj, "aggr_in_queue",
1113 agg->time_in_queue / agg->slavecount);
1114 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1117 static void json_object_add_disk_utils(struct json_object *obj,
1118 struct flist_head *head)
1120 struct json_array *array = json_create_array();
1121 struct flist_head *entry;
1122 struct disk_util *du;
1124 json_object_add_value_array(obj, "disk_util", array);
1126 flist_for_each(entry, head) {
1127 du = flist_entry(entry, struct disk_util, list);
1129 aggregate_slaves_stats(du);
1130 json_array_add_disk_util(&du->dus, &du->agg, array);
1134 void show_disk_util(int terse, struct json_object *parent,
1135 struct buf_output *out)
1137 struct flist_head *entry;
1138 struct disk_util *du;
1141 if (!is_running_backend())
1144 if (flist_empty(&disk_list))
1147 if ((output_format & FIO_OUTPUT_JSON) && parent)
1152 if (!terse && !do_json)
1153 log_buf(out, "\nDisk stats (read/write):\n");
1156 json_object_add_disk_utils(parent, &disk_list);
1157 } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1158 flist_for_each(entry, &disk_list) {
1159 du = flist_entry(entry, struct disk_util, list);
1161 aggregate_slaves_stats(du);
1162 print_disk_util(&du->dus, &du->agg, terse, out);
1167 static void show_thread_status_normal(struct thread_stat *ts,
1168 struct group_run_stats *rs,
1169 struct buf_output *out)
1171 double usr_cpu, sys_cpu;
1172 unsigned long runtime;
1173 double io_u_dist[FIO_IO_U_MAP_NR];
1177 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1180 memset(time_buf, 0, sizeof(time_buf));
1183 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1186 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1187 ts->name, ts->groupid, ts->members,
1188 ts->error, (int) ts->pid, time_buf);
1190 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1191 ts->name, ts->groupid, ts->members,
1192 ts->error, ts->verror, (int) ts->pid,
1196 if (strlen(ts->description))
1197 log_buf(out, " Description : [%s]\n", ts->description);
1199 for_each_rw_ddir(ddir) {
1200 if (ts->io_bytes[ddir])
1201 show_ddir_status(rs, ts, ddir, out);
1204 if (ts->unified_rw_rep == UNIFIED_BOTH)
1205 show_mixed_ddir_status(rs, ts, out);
1207 show_latencies(ts, out);
1209 if (ts->sync_stat.samples)
1210 show_ddir_status(rs, ts, DDIR_SYNC, out);
1212 runtime = ts->total_run_time;
1214 double runt = (double) runtime;
1216 usr_cpu = (double) ts->usr_time * 100 / runt;
1217 sys_cpu = (double) ts->sys_time * 100 / runt;
1223 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1224 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1225 (unsigned long long) ts->ctx,
1226 (unsigned long long) ts->majf,
1227 (unsigned long long) ts->minf);
1229 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1230 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1231 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1232 io_u_dist[1], io_u_dist[2],
1233 io_u_dist[3], io_u_dist[4],
1234 io_u_dist[5], io_u_dist[6]);
1236 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1237 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1238 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1239 io_u_dist[1], io_u_dist[2],
1240 io_u_dist[3], io_u_dist[4],
1241 io_u_dist[5], io_u_dist[6]);
1242 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1243 log_buf(out, " complete : 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 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
1249 " short=%llu,%llu,%llu,0"
1250 " dropped=%llu,%llu,%llu,0\n",
1251 (unsigned long long) ts->total_io_u[0],
1252 (unsigned long long) ts->total_io_u[1],
1253 (unsigned long long) ts->total_io_u[2],
1254 (unsigned long long) ts->total_io_u[3],
1255 (unsigned long long) ts->short_io_u[0],
1256 (unsigned long long) ts->short_io_u[1],
1257 (unsigned long long) ts->short_io_u[2],
1258 (unsigned long long) ts->drop_io_u[0],
1259 (unsigned long long) ts->drop_io_u[1],
1260 (unsigned long long) ts->drop_io_u[2]);
1261 if (ts->continue_on_error) {
1262 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
1263 (unsigned long long)ts->total_err_count,
1265 strerror(ts->first_error));
1267 if (ts->latency_depth) {
1268 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1269 (unsigned long long)ts->latency_target,
1270 (unsigned long long)ts->latency_window,
1271 ts->latency_percentile.u.f,
1275 if (ts->nr_block_infos)
1276 show_block_infos(ts->nr_block_infos, ts->block_infos,
1277 ts->percentile_list, out);
1280 show_ss_normal(ts, out);
1283 static void show_ddir_status_terse(struct thread_stat *ts,
1284 struct group_run_stats *rs,
1285 enum fio_ddir ddir, int ver,
1286 struct buf_output *out)
1288 unsigned long long min, max, minv, maxv, bw, iops;
1289 unsigned long long *ovals = NULL;
1294 assert(ddir_rw(ddir));
1297 if (ts->runtime[ddir]) {
1298 uint64_t runt = ts->runtime[ddir];
1300 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1301 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1304 log_buf(out, ";%llu;%llu;%llu;%llu",
1305 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1306 (unsigned long long) ts->runtime[ddir]);
1308 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1309 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1311 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1313 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1314 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1316 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1318 if (ts->lat_percentiles) {
1319 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1320 ts->lat_stat[ddir].samples,
1321 ts->percentile_list, &ovals, &maxv,
1323 } else if (ts->clat_percentiles) {
1324 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1325 ts->clat_stat[ddir].samples,
1326 ts->percentile_list, &ovals, &maxv,
1332 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1334 log_buf(out, ";0%%=0");
1337 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1340 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1341 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1343 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1347 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1349 double p_of_agg = 100.0;
1351 if (rs->agg[ddir]) {
1352 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1353 if (p_of_agg > 100.0)
1357 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1359 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1364 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1366 log_buf(out, ";%lu", 0UL);
1368 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1369 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1370 mean, dev, (&ts->iops_stat[ddir])->samples);
1372 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1376 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1377 struct group_run_stats *rs,
1378 int ver, struct buf_output *out)
1380 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1383 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1385 free_clat_prio_stats(ts_lcl);
1389 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1390 uint32_t percentiles,
1391 struct io_stat *lat_stat,
1392 uint64_t *io_u_plat)
1396 unsigned int i, len;
1397 struct json_object *lat_object, *percentile_object, *clat_bins_object;
1398 unsigned long long min, max, maxv, minv, *ovals = NULL;
1400 if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1404 lat_object = json_create_object();
1405 json_object_add_value_int(lat_object, "min", min);
1406 json_object_add_value_int(lat_object, "max", max);
1407 json_object_add_value_float(lat_object, "mean", mean);
1408 json_object_add_value_float(lat_object, "stddev", dev);
1409 json_object_add_value_int(lat_object, "N", lat_stat->samples);
1411 if (percentiles && lat_stat->samples) {
1412 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1413 ts->percentile_list, &ovals, &maxv, &minv);
1415 if (len > FIO_IO_U_LIST_MAX_LEN)
1416 len = FIO_IO_U_LIST_MAX_LEN;
1418 percentile_object = json_create_object();
1419 json_object_add_value_object(lat_object, "percentile", percentile_object);
1420 for (i = 0; i < len; i++) {
1421 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1422 json_object_add_value_int(percentile_object, buf, ovals[i]);
1426 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1427 clat_bins_object = json_create_object();
1428 json_object_add_value_object(lat_object, "bins", clat_bins_object);
1430 for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1432 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1433 json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1441 static void add_ddir_status_json(struct thread_stat *ts,
1442 struct group_run_stats *rs, enum fio_ddir ddir,
1443 struct json_object *parent)
1445 unsigned long long min, max;
1446 unsigned long long bw_bytes, bw;
1447 double mean, dev, iops;
1448 struct json_object *dir_object, *tmp_object;
1449 double p_of_agg = 100.0;
1451 assert(ddir_rw(ddir) || ddir_sync(ddir));
1453 if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1456 dir_object = json_create_object();
1457 json_object_add_value_object(parent,
1458 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1460 if (ddir_rw(ddir)) {
1464 if (ts->runtime[ddir]) {
1465 uint64_t runt = ts->runtime[ddir];
1467 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1468 bw = bw_bytes / 1024; /* KiB/s */
1469 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1472 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1473 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1474 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1475 json_object_add_value_int(dir_object, "bw", bw);
1476 json_object_add_value_float(dir_object, "iops", iops);
1477 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1478 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1479 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1480 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1482 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1483 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1484 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1486 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1487 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1488 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1490 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1491 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1492 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1494 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1495 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1496 &ts->sync_stat, ts->io_u_sync_plat);
1497 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1503 /* Only include per prio stats if there are >= 2 prios with samples */
1504 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
1505 struct json_array *array = json_create_array();
1506 const char *obj_name;
1509 if (ts->lat_percentiles)
1510 obj_name = "lat_ns";
1512 obj_name = "clat_ns";
1514 json_object_add_value_array(dir_object, "prios", array);
1516 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
1517 struct json_object *obj;
1519 if (!ts->clat_prio[ddir][i].clat_stat.samples)
1522 obj = json_create_object();
1524 json_object_add_value_int(obj, "prioclass",
1525 ioprio_class(ts->clat_prio[ddir][i].ioprio));
1526 json_object_add_value_int(obj, "prio",
1527 ioprio(ts->clat_prio[ddir][i].ioprio));
1529 tmp_object = add_ddir_lat_json(ts,
1530 ts->clat_percentiles | ts->lat_percentiles,
1531 &ts->clat_prio[ddir][i].clat_stat,
1532 ts->clat_prio[ddir][i].io_u_plat);
1533 json_object_add_value_object(obj, obj_name, tmp_object);
1534 json_array_add_value_object(array, obj);
1538 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1539 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1542 p_of_agg = mean = dev = 0.0;
1545 json_object_add_value_int(dir_object, "bw_min", min);
1546 json_object_add_value_int(dir_object, "bw_max", max);
1547 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1548 json_object_add_value_float(dir_object, "bw_mean", mean);
1549 json_object_add_value_float(dir_object, "bw_dev", dev);
1550 json_object_add_value_int(dir_object, "bw_samples",
1551 (&ts->bw_stat[ddir])->samples);
1553 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1557 json_object_add_value_int(dir_object, "iops_min", min);
1558 json_object_add_value_int(dir_object, "iops_max", max);
1559 json_object_add_value_float(dir_object, "iops_mean", mean);
1560 json_object_add_value_float(dir_object, "iops_stddev", dev);
1561 json_object_add_value_int(dir_object, "iops_samples",
1562 (&ts->iops_stat[ddir])->samples);
1564 if (ts->cachehit + ts->cachemiss) {
1568 total = ts->cachehit + ts->cachemiss;
1569 hit = (double) ts->cachehit / (double) total;
1571 json_object_add_value_float(dir_object, "cachehit", hit);
1575 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1576 struct group_run_stats *rs, struct json_object *parent)
1578 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1580 /* add the aggregated stats to json parent */
1582 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1584 free_clat_prio_stats(ts_lcl);
1588 static void show_thread_status_terse_all(struct thread_stat *ts,
1589 struct group_run_stats *rs, int ver,
1590 struct buf_output *out)
1592 double io_u_dist[FIO_IO_U_MAP_NR];
1593 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1594 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1595 double usr_cpu, sys_cpu;
1600 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1602 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1603 ts->name, ts->groupid, ts->error);
1605 /* Log Read Status, or mixed if unified_rw_rep = 1 */
1606 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1607 if (ts->unified_rw_rep != UNIFIED_MIXED) {
1608 /* Log Write Status */
1609 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1610 /* Log Trim Status */
1611 if (ver == 2 || ver == 4 || ver == 5)
1612 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1614 if (ts->unified_rw_rep == UNIFIED_BOTH)
1615 show_mixed_ddir_status_terse(ts, rs, ver, out);
1617 if (ts->total_run_time) {
1618 double runt = (double) ts->total_run_time;
1620 usr_cpu = (double) ts->usr_time * 100 / runt;
1621 sys_cpu = (double) ts->sys_time * 100 / runt;
1627 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1628 (unsigned long long) ts->ctx,
1629 (unsigned long long) ts->majf,
1630 (unsigned long long) ts->minf);
1632 /* Calc % distribution of IO depths, usecond, msecond latency */
1633 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1634 stat_calc_lat_nu(ts, io_u_lat_u);
1635 stat_calc_lat_m(ts, io_u_lat_m);
1637 /* Only show fixed 7 I/O depth levels*/
1638 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1639 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1640 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1642 /* Microsecond latency */
1643 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1644 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1645 /* Millisecond latency */
1646 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1647 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1649 /* disk util stats, if any */
1650 if (ver >= 3 && is_running_backend())
1651 show_disk_util(1, NULL, out);
1653 /* Additional output if continue_on_error set - default off*/
1654 if (ts->continue_on_error)
1655 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1657 /* Additional output if description is set */
1658 if (strlen(ts->description)) {
1661 log_buf(out, ";%s", ts->description);
1667 static void json_add_job_opts(struct json_object *root, const char *name,
1668 struct flist_head *opt_list)
1670 struct json_object *dir_object;
1671 struct flist_head *entry;
1672 struct print_option *p;
1674 if (flist_empty(opt_list))
1677 dir_object = json_create_object();
1678 json_object_add_value_object(root, name, dir_object);
1680 flist_for_each(entry, opt_list) {
1681 p = flist_entry(entry, struct print_option, list);
1682 json_object_add_value_string(dir_object, p->name, p->value);
1686 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1687 struct group_run_stats *rs,
1688 struct flist_head *opt_list)
1690 struct json_object *root, *tmp;
1691 struct jobs_eta *je;
1692 double io_u_dist[FIO_IO_U_MAP_NR];
1693 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1694 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1695 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1696 double usr_cpu, sys_cpu;
1700 root = json_create_object();
1701 json_object_add_value_string(root, "jobname", ts->name);
1702 json_object_add_value_int(root, "groupid", ts->groupid);
1703 json_object_add_value_int(root, "error", ts->error);
1706 je = get_jobs_eta(true, &size);
1708 json_object_add_value_int(root, "eta", je->eta_sec);
1709 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1714 json_add_job_opts(root, "job options", opt_list);
1716 add_ddir_status_json(ts, rs, DDIR_READ, root);
1717 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1718 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1719 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1721 if (ts->unified_rw_rep == UNIFIED_BOTH)
1722 add_mixed_ddir_status_json(ts, rs, root);
1725 if (ts->total_run_time) {
1726 double runt = (double) ts->total_run_time;
1728 usr_cpu = (double) ts->usr_time * 100 / runt;
1729 sys_cpu = (double) ts->sys_time * 100 / runt;
1734 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1735 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1736 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1737 json_object_add_value_int(root, "ctx", ts->ctx);
1738 json_object_add_value_int(root, "majf", ts->majf);
1739 json_object_add_value_int(root, "minf", ts->minf);
1741 /* Calc % distribution of IO depths */
1742 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1743 tmp = json_create_object();
1744 json_object_add_value_object(root, "iodepth_level", tmp);
1745 /* Only show fixed 7 I/O depth levels*/
1746 for (i = 0; i < 7; i++) {
1749 snprintf(name, 20, "%d", 1 << i);
1751 snprintf(name, 20, ">=%d", 1 << i);
1752 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1755 /* Calc % distribution of submit IO depths */
1756 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1757 tmp = json_create_object();
1758 json_object_add_value_object(root, "iodepth_submit", tmp);
1759 /* Only show fixed 7 I/O depth levels*/
1760 for (i = 0; i < 7; i++) {
1763 snprintf(name, 20, "0");
1765 snprintf(name, 20, "%d", 1 << (i+1));
1767 snprintf(name, 20, ">=%d", 1 << i);
1768 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1771 /* Calc % distribution of completion IO depths */
1772 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1773 tmp = json_create_object();
1774 json_object_add_value_object(root, "iodepth_complete", tmp);
1775 /* Only show fixed 7 I/O depth levels*/
1776 for (i = 0; i < 7; i++) {
1779 snprintf(name, 20, "0");
1781 snprintf(name, 20, "%d", 1 << (i+1));
1783 snprintf(name, 20, ">=%d", 1 << i);
1784 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1787 /* Calc % distribution of nsecond, usecond, msecond latency */
1788 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1789 stat_calc_lat_n(ts, io_u_lat_n);
1790 stat_calc_lat_u(ts, io_u_lat_u);
1791 stat_calc_lat_m(ts, io_u_lat_m);
1793 /* Nanosecond latency */
1794 tmp = json_create_object();
1795 json_object_add_value_object(root, "latency_ns", tmp);
1796 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1797 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1798 "250", "500", "750", "1000", };
1799 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1801 /* Microsecond latency */
1802 tmp = json_create_object();
1803 json_object_add_value_object(root, "latency_us", tmp);
1804 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1805 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1806 "250", "500", "750", "1000", };
1807 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1809 /* Millisecond latency */
1810 tmp = json_create_object();
1811 json_object_add_value_object(root, "latency_ms", tmp);
1812 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1813 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1814 "250", "500", "750", "1000", "2000",
1816 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1819 /* Additional output if continue_on_error set - default off*/
1820 if (ts->continue_on_error) {
1821 json_object_add_value_int(root, "total_err", ts->total_err_count);
1822 json_object_add_value_int(root, "first_error", ts->first_error);
1825 if (ts->latency_depth) {
1826 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1827 json_object_add_value_int(root, "latency_target", ts->latency_target);
1828 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1829 json_object_add_value_int(root, "latency_window", ts->latency_window);
1832 /* Additional output if description is set */
1833 if (strlen(ts->description))
1834 json_object_add_value_string(root, "desc", ts->description);
1836 if (ts->nr_block_infos) {
1837 /* Block error histogram and types */
1839 unsigned int *percentiles = NULL;
1840 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1842 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1843 ts->percentile_list,
1844 &percentiles, block_state_counts);
1847 struct json_object *block, *percentile_object, *states;
1849 block = json_create_object();
1850 json_object_add_value_object(root, "block", block);
1852 percentile_object = json_create_object();
1853 json_object_add_value_object(block, "percentiles",
1855 for (i = 0; i < len; i++) {
1857 snprintf(buf, sizeof(buf), "%f",
1858 ts->percentile_list[i].u.f);
1859 json_object_add_value_int(percentile_object,
1864 states = json_create_object();
1865 json_object_add_value_object(block, "states", states);
1866 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1867 json_object_add_value_int(states,
1868 block_state_names[state],
1869 block_state_counts[state]);
1876 struct json_object *data;
1877 struct json_array *iops, *bw;
1880 int intervals = ts->ss_dur / (ss_check_interval / 1000L);
1882 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1883 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1884 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1885 (float) ts->ss_limit.u.f,
1886 ts->ss_state & FIO_SS_PCT ? "%" : "");
1888 tmp = json_create_object();
1889 json_object_add_value_object(root, "steadystate", tmp);
1890 json_object_add_value_string(tmp, "ss", ss_buf);
1891 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1892 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1894 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1895 ts->ss_state & FIO_SS_PCT ? "%" : "");
1896 json_object_add_value_string(tmp, "criterion", ss_buf);
1897 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1898 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1900 data = json_create_object();
1901 json_object_add_value_object(tmp, "data", data);
1902 bw = json_create_array();
1903 iops = json_create_array();
1906 ** if ss was attained or the buffer is not full,
1907 ** ss->head points to the first element in the list.
1908 ** otherwise it actually points to the second element
1911 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1914 j = ts->ss_head == 0 ? intervals - 1 : ts->ss_head - 1;
1915 for (l = 0; l < intervals; l++) {
1916 k = (j + l) % intervals;
1917 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1918 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1920 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1921 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1922 json_object_add_value_array(data, "iops", iops);
1923 json_object_add_value_array(data, "bw", bw);
1929 static void show_thread_status_terse(struct thread_stat *ts,
1930 struct group_run_stats *rs,
1931 struct buf_output *out)
1933 if (terse_version >= 2 && terse_version <= 5)
1934 show_thread_status_terse_all(ts, rs, terse_version, out);
1936 log_err("fio: bad terse version!? %d\n", terse_version);
1939 struct json_object *show_thread_status(struct thread_stat *ts,
1940 struct group_run_stats *rs,
1941 struct flist_head *opt_list,
1942 struct buf_output *out)
1944 struct json_object *ret = NULL;
1946 if (output_format & FIO_OUTPUT_TERSE)
1947 show_thread_status_terse(ts, rs, out);
1948 if (output_format & FIO_OUTPUT_JSON)
1949 ret = show_thread_status_json(ts, rs, opt_list);
1950 if (output_format & FIO_OUTPUT_NORMAL)
1951 show_thread_status_normal(ts, rs, out);
1956 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1960 dst->min_val = min(dst->min_val, src->min_val);
1961 dst->max_val = max(dst->max_val, src->max_val);
1964 * Compute new mean and S after the merge
1965 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1966 * #Parallel_algorithm>
1969 mean = src->mean.u.f;
1972 double delta = src->mean.u.f - dst->mean.u.f;
1974 mean = ((src->mean.u.f * src->samples) +
1975 (dst->mean.u.f * dst->samples)) /
1976 (dst->samples + src->samples);
1978 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1979 (dst->samples * src->samples) /
1980 (dst->samples + src->samples);
1983 dst->samples += src->samples;
1984 dst->mean.u.f = mean;
1990 * We sum two kinds of stats - one that is time based, in which case we
1991 * apply the proper summing technique, and then one that is iops/bw
1992 * numbers. For group_reporting, we should just add those up, not make
1993 * them the mean of everything.
1995 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
1997 bool first = dst->samples == 0;
1999 if (src->samples == 0)
2003 __sum_stat(dst, src, first);
2008 dst->min_val = src->min_val;
2009 dst->max_val = src->max_val;
2010 dst->samples = src->samples;
2011 dst->mean.u.f = src->mean.u.f;
2012 dst->S.u.f = src->S.u.f;
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;
2022 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
2026 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2027 if (dst->max_run[i] < src->max_run[i])
2028 dst->max_run[i] = src->max_run[i];
2029 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
2030 dst->min_run[i] = src->min_run[i];
2031 if (dst->max_bw[i] < src->max_bw[i])
2032 dst->max_bw[i] = src->max_bw[i];
2033 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
2034 dst->min_bw[i] = src->min_bw[i];
2036 dst->iobytes[i] += src->iobytes[i];
2037 dst->agg[i] += src->agg[i];
2041 dst->kb_base = src->kb_base;
2042 if (!dst->unit_base)
2043 dst->unit_base = src->unit_base;
2045 dst->sig_figs = src->sig_figs;
2049 * Free the clat_prio_stat arrays allocated by alloc_clat_prio_stat_ddir().
2051 void free_clat_prio_stats(struct thread_stat *ts)
2058 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2059 sfree(ts->clat_prio[ddir]);
2060 ts->clat_prio[ddir] = NULL;
2061 ts->nr_clat_prio[ddir] = 0;
2066 * Allocate a clat_prio_stat array. The array has to be allocated/freed using
2067 * smalloc/sfree, so that it is accessible by the process/thread summing the
2070 int alloc_clat_prio_stat_ddir(struct thread_stat *ts, enum fio_ddir ddir,
2073 struct clat_prio_stat *clat_prio;
2076 clat_prio = scalloc(nr_prios, sizeof(*ts->clat_prio[ddir]));
2078 log_err("fio: failed to allocate ts clat data\n");
2082 for (i = 0; i < nr_prios; i++)
2083 clat_prio[i].clat_stat.min_val = ULONG_MAX;
2085 ts->clat_prio[ddir] = clat_prio;
2086 ts->nr_clat_prio[ddir] = nr_prios;
2091 static int grow_clat_prio_stat(struct thread_stat *dst, enum fio_ddir ddir)
2093 int curr_len = dst->nr_clat_prio[ddir];
2096 new_arr = scalloc(curr_len + 1, sizeof(*dst->clat_prio[ddir]));
2098 log_err("fio: failed to grow clat prio array\n");
2102 memcpy(new_arr, dst->clat_prio[ddir],
2103 curr_len * sizeof(*dst->clat_prio[ddir]));
2104 sfree(dst->clat_prio[ddir]);
2106 dst->clat_prio[ddir] = new_arr;
2107 dst->clat_prio[ddir][curr_len].clat_stat.min_val = ULONG_MAX;
2108 dst->nr_clat_prio[ddir]++;
2113 static int find_clat_prio_index(struct thread_stat *dst, enum fio_ddir ddir,
2116 int i, nr_prios = dst->nr_clat_prio[ddir];
2118 for (i = 0; i < nr_prios; i++) {
2119 if (dst->clat_prio[ddir][i].ioprio == ioprio)
2126 static int alloc_or_get_clat_prio_index(struct thread_stat *dst,
2127 enum fio_ddir ddir, uint32_t ioprio,
2130 int index = find_clat_prio_index(dst, ddir, ioprio);
2133 index = dst->nr_clat_prio[ddir];
2135 if (grow_clat_prio_stat(dst, ddir))
2138 dst->clat_prio[ddir][index].ioprio = ioprio;
2146 static int clat_prio_stats_copy(struct thread_stat *dst, struct thread_stat *src,
2147 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2149 size_t sz = sizeof(*src->clat_prio[src_ddir]) *
2150 src->nr_clat_prio[src_ddir];
2152 dst->clat_prio[dst_ddir] = smalloc(sz);
2153 if (!dst->clat_prio[dst_ddir]) {
2154 log_err("fio: failed to alloc clat prio array\n");
2158 memcpy(dst->clat_prio[dst_ddir], src->clat_prio[src_ddir], sz);
2159 dst->nr_clat_prio[dst_ddir] = src->nr_clat_prio[src_ddir];
2164 static int clat_prio_stat_add_samples(struct thread_stat *dst,
2165 enum fio_ddir dst_ddir, uint32_t ioprio,
2166 struct io_stat *io_stat,
2167 uint64_t *io_u_plat)
2171 if (!io_stat->samples)
2174 if (alloc_or_get_clat_prio_index(dst, dst_ddir, ioprio, &dst_index))
2177 sum_stat(&dst->clat_prio[dst_ddir][dst_index].clat_stat, io_stat,
2180 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
2181 dst->clat_prio[dst_ddir][dst_index].io_u_plat[i] += io_u_plat[i];
2186 static int sum_clat_prio_stats_src_single_prio(struct thread_stat *dst,
2187 struct thread_stat *src,
2188 enum fio_ddir dst_ddir,
2189 enum fio_ddir src_ddir)
2191 struct io_stat *io_stat;
2192 uint64_t *io_u_plat;
2195 * If src ts has no clat_prio_stat array, then all I/Os were submitted
2196 * using src->ioprio. Thus, the global samples in src->clat_stat (or
2197 * src->lat_stat) can be used as the 'per prio' samples for src->ioprio.
2199 assert(!src->clat_prio[src_ddir]);
2200 assert(src->nr_clat_prio[src_ddir] == 0);
2202 if (src->lat_percentiles) {
2203 io_u_plat = src->io_u_plat[FIO_LAT][src_ddir];
2204 io_stat = &src->lat_stat[src_ddir];
2206 io_u_plat = src->io_u_plat[FIO_CLAT][src_ddir];
2207 io_stat = &src->clat_stat[src_ddir];
2210 return clat_prio_stat_add_samples(dst, dst_ddir, src->ioprio, io_stat,
2214 static int sum_clat_prio_stats_src_multi_prio(struct thread_stat *dst,
2215 struct thread_stat *src,
2216 enum fio_ddir dst_ddir,
2217 enum fio_ddir src_ddir)
2222 * If src ts has a clat_prio_stat array, then there are multiple prios
2223 * in use (i.e. src ts had cmdprio_percentage or cmdprio_bssplit set).
2224 * The samples for the default prio will exist in the src->clat_prio
2225 * array, just like the samples for any other prio.
2227 assert(src->clat_prio[src_ddir]);
2228 assert(src->nr_clat_prio[src_ddir]);
2230 /* If the dst ts doesn't yet have a clat_prio array, simply memcpy. */
2231 if (!dst->clat_prio[dst_ddir])
2232 return clat_prio_stats_copy(dst, src, dst_ddir, src_ddir);
2234 /* The dst ts already has a clat_prio_array, add src stats into it. */
2235 for (i = 0; i < src->nr_clat_prio[src_ddir]; i++) {
2236 struct io_stat *io_stat = &src->clat_prio[src_ddir][i].clat_stat;
2237 uint64_t *io_u_plat = src->clat_prio[src_ddir][i].io_u_plat;
2238 uint32_t ioprio = src->clat_prio[src_ddir][i].ioprio;
2240 if (clat_prio_stat_add_samples(dst, dst_ddir, ioprio, io_stat, io_u_plat))
2247 static int sum_clat_prio_stats(struct thread_stat *dst, struct thread_stat *src,
2248 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2250 if (dst->disable_prio_stat)
2253 if (!src->clat_prio[src_ddir])
2254 return sum_clat_prio_stats_src_single_prio(dst, src, dst_ddir,
2257 return sum_clat_prio_stats_src_multi_prio(dst, src, dst_ddir, src_ddir);
2260 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2264 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2265 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2266 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2267 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2268 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2269 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2270 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2271 sum_clat_prio_stats(dst, src, l, l);
2273 dst->io_bytes[l] += src->io_bytes[l];
2275 if (dst->runtime[l] < src->runtime[l])
2276 dst->runtime[l] = src->runtime[l];
2278 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2279 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2280 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2281 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2282 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2283 sum_clat_prio_stats(dst, src, 0, l);
2285 dst->io_bytes[0] += src->io_bytes[l];
2287 if (dst->runtime[0] < src->runtime[l])
2288 dst->runtime[0] = src->runtime[l];
2292 sum_stat(&dst->sync_stat, &src->sync_stat, false);
2293 dst->usr_time += src->usr_time;
2294 dst->sys_time += src->sys_time;
2295 dst->ctx += src->ctx;
2296 dst->majf += src->majf;
2297 dst->minf += src->minf;
2299 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2300 dst->io_u_map[k] += src->io_u_map[k];
2301 dst->io_u_submit[k] += src->io_u_submit[k];
2302 dst->io_u_complete[k] += src->io_u_complete[k];
2305 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2306 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2307 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2308 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2309 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2310 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2312 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2313 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2314 dst->total_io_u[k] += src->total_io_u[k];
2315 dst->short_io_u[k] += src->short_io_u[k];
2316 dst->drop_io_u[k] += src->drop_io_u[k];
2318 dst->total_io_u[0] += src->total_io_u[k];
2319 dst->short_io_u[0] += src->short_io_u[k];
2320 dst->drop_io_u[0] += src->drop_io_u[k];
2324 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2326 for (k = 0; k < FIO_LAT_CNT; k++)
2327 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2328 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2329 if (dst->unified_rw_rep != UNIFIED_MIXED)
2330 dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2332 dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2334 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2335 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2337 dst->total_run_time += src->total_run_time;
2338 dst->total_submit += src->total_submit;
2339 dst->total_complete += src->total_complete;
2340 dst->nr_zone_resets += src->nr_zone_resets;
2341 dst->cachehit += src->cachehit;
2342 dst->cachemiss += src->cachemiss;
2345 void init_group_run_stat(struct group_run_stats *gs)
2348 memset(gs, 0, sizeof(*gs));
2350 for (i = 0; i < DDIR_RWDIR_CNT; i++)
2351 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2354 void init_thread_stat_min_vals(struct thread_stat *ts)
2358 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2359 ts->clat_stat[i].min_val = ULONG_MAX;
2360 ts->slat_stat[i].min_val = ULONG_MAX;
2361 ts->lat_stat[i].min_val = ULONG_MAX;
2362 ts->bw_stat[i].min_val = ULONG_MAX;
2363 ts->iops_stat[i].min_val = ULONG_MAX;
2365 ts->sync_stat.min_val = ULONG_MAX;
2368 void init_thread_stat(struct thread_stat *ts)
2370 memset(ts, 0, sizeof(*ts));
2372 init_thread_stat_min_vals(ts);
2376 static void init_per_prio_stats(struct thread_stat *threadstats, int nr_ts)
2378 struct thread_stat *ts;
2379 int i, j, last_ts, idx;
2387 * Loop through all tds, if a td requires per prio stats, temporarily
2388 * store a 1 in ts->disable_prio_stat, and then do an additional
2389 * loop at the end where we invert the ts->disable_prio_stat values.
2395 (!td->o.group_reporting ||
2396 (td->o.group_reporting && last_ts != td->groupid))) {
2401 last_ts = td->groupid;
2402 ts = &threadstats[j];
2404 /* idx == 0 means first td in group, or td is not in a group. */
2406 ts->ioprio = td->ioprio;
2407 else if (td->ioprio != ts->ioprio)
2408 ts->disable_prio_stat = 1;
2410 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2411 if (td->ts.clat_prio[ddir]) {
2412 ts->disable_prio_stat = 1;
2420 /* Loop through all dst threadstats and fixup the values. */
2421 for (i = 0; i < nr_ts; i++) {
2422 ts = &threadstats[i];
2423 ts->disable_prio_stat = !ts->disable_prio_stat;
2427 void __show_run_stats(void)
2429 struct group_run_stats *runstats, *rs;
2430 struct thread_stat *threadstats, *ts;
2431 int i, j, k, nr_ts, last_ts, idx;
2432 bool kb_base_warned = false;
2433 bool unit_base_warned = false;
2434 struct json_object *root = NULL;
2435 struct json_array *array = NULL;
2436 struct buf_output output[FIO_OUTPUT_NR];
2437 struct flist_head **opt_lists;
2439 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2441 for (i = 0; i < groupid + 1; i++)
2442 init_group_run_stat(&runstats[i]);
2445 * find out how many threads stats we need. if group reporting isn't
2446 * enabled, it's one-per-td.
2451 if (!td->o.group_reporting) {
2455 if (last_ts == td->groupid)
2460 last_ts = td->groupid;
2464 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2465 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2467 for (i = 0; i < nr_ts; i++) {
2468 init_thread_stat(&threadstats[i]);
2469 opt_lists[i] = NULL;
2472 init_per_prio_stats(threadstats, nr_ts);
2480 if (idx && (!td->o.group_reporting ||
2481 (td->o.group_reporting && last_ts != td->groupid))) {
2486 last_ts = td->groupid;
2488 ts = &threadstats[j];
2490 ts->clat_percentiles = td->o.clat_percentiles;
2491 ts->lat_percentiles = td->o.lat_percentiles;
2492 ts->slat_percentiles = td->o.slat_percentiles;
2493 ts->percentile_precision = td->o.percentile_precision;
2494 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2495 opt_lists[j] = &td->opt_list;
2499 if (ts->groupid == -1) {
2501 * These are per-group shared already
2503 snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2504 if (td->o.description)
2505 snprintf(ts->description,
2506 sizeof(ts->description), "%s",
2509 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2512 * If multiple entries in this group, this is
2515 ts->thread_number = td->thread_number;
2516 ts->groupid = td->groupid;
2519 * first pid in group, not very useful...
2523 ts->kb_base = td->o.kb_base;
2524 ts->unit_base = td->o.unit_base;
2525 ts->sig_figs = td->o.sig_figs;
2526 ts->unified_rw_rep = td->o.unified_rw_rep;
2527 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2528 log_info("fio: kb_base differs for jobs in group, using"
2529 " %u as the base\n", ts->kb_base);
2530 kb_base_warned = true;
2531 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2532 log_info("fio: unit_base differs for jobs in group, using"
2533 " %u as the base\n", ts->unit_base);
2534 unit_base_warned = true;
2537 ts->continue_on_error = td->o.continue_on_error;
2538 ts->total_err_count += td->total_err_count;
2539 ts->first_error = td->first_error;
2541 if (!td->error && td->o.continue_on_error &&
2543 ts->error = td->first_error;
2544 snprintf(ts->verror, sizeof(ts->verror), "%s",
2546 } else if (td->error) {
2547 ts->error = td->error;
2548 snprintf(ts->verror, sizeof(ts->verror), "%s",
2553 ts->latency_depth = td->latency_qd;
2554 ts->latency_target = td->o.latency_target;
2555 ts->latency_percentile = td->o.latency_percentile;
2556 ts->latency_window = td->o.latency_window;
2558 ts->nr_block_infos = td->ts.nr_block_infos;
2559 for (k = 0; k < ts->nr_block_infos; k++)
2560 ts->block_infos[k] = td->ts.block_infos[k];
2562 sum_thread_stats(ts, &td->ts);
2567 ts->ss_state = td->ss.state;
2568 ts->ss_dur = td->ss.dur;
2569 ts->ss_head = td->ss.head;
2570 ts->ss_bw_data = td->ss.bw_data;
2571 ts->ss_iops_data = td->ss.iops_data;
2572 ts->ss_limit.u.f = td->ss.limit;
2573 ts->ss_slope.u.f = td->ss.slope;
2574 ts->ss_deviation.u.f = td->ss.deviation;
2575 ts->ss_criterion.u.f = td->ss.criterion;
2578 ts->ss_dur = ts->ss_state = 0;
2581 for (i = 0; i < nr_ts; i++) {
2582 unsigned long long bw;
2584 ts = &threadstats[i];
2585 if (ts->groupid == -1)
2587 rs = &runstats[ts->groupid];
2588 rs->kb_base = ts->kb_base;
2589 rs->unit_base = ts->unit_base;
2590 rs->sig_figs = ts->sig_figs;
2591 rs->unified_rw_rep |= ts->unified_rw_rep;
2593 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2594 if (!ts->runtime[j])
2596 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2597 rs->min_run[j] = ts->runtime[j];
2598 if (ts->runtime[j] > rs->max_run[j])
2599 rs->max_run[j] = ts->runtime[j];
2603 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2604 if (bw < rs->min_bw[j])
2606 if (bw > rs->max_bw[j])
2609 rs->iobytes[j] += ts->io_bytes[j];
2613 for (i = 0; i < groupid + 1; i++) {
2618 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2619 if (rs->max_run[ddir])
2620 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2625 for (i = 0; i < FIO_OUTPUT_NR; i++)
2626 buf_output_init(&output[i]);
2629 * don't overwrite last signal output
2631 if (output_format & FIO_OUTPUT_NORMAL)
2632 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2633 if (output_format & FIO_OUTPUT_JSON) {
2634 struct thread_data *global;
2637 unsigned long long ms_since_epoch;
2640 gettimeofday(&now, NULL);
2641 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2642 (unsigned long long)(now.tv_usec) / 1000;
2644 tv_sec = now.tv_sec;
2645 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2646 if (time_buf[strlen(time_buf) - 1] == '\n')
2647 time_buf[strlen(time_buf) - 1] = '\0';
2649 root = json_create_object();
2650 json_object_add_value_string(root, "fio version", fio_version_string);
2651 json_object_add_value_int(root, "timestamp", now.tv_sec);
2652 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2653 json_object_add_value_string(root, "time", time_buf);
2654 global = get_global_options();
2655 json_add_job_opts(root, "global options", &global->opt_list);
2656 array = json_create_array();
2657 json_object_add_value_array(root, "jobs", array);
2661 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2663 for (i = 0; i < nr_ts; i++) {
2664 ts = &threadstats[i];
2665 rs = &runstats[ts->groupid];
2668 fio_server_send_job_options(opt_lists[i], i);
2669 fio_server_send_ts(ts, rs);
2671 if (output_format & FIO_OUTPUT_TERSE)
2672 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2673 if (output_format & FIO_OUTPUT_JSON) {
2674 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2675 json_array_add_value_object(array, tmp);
2677 if (output_format & FIO_OUTPUT_NORMAL)
2678 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2681 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2682 /* disk util stats, if any */
2683 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2685 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2687 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2688 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2689 json_free_object(root);
2692 for (i = 0; i < groupid + 1; i++) {
2697 fio_server_send_gs(rs);
2698 else if (output_format & FIO_OUTPUT_NORMAL)
2699 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2703 fio_server_send_du();
2704 else if (output_format & FIO_OUTPUT_NORMAL) {
2705 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2706 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2709 for (i = 0; i < FIO_OUTPUT_NR; i++) {
2710 struct buf_output *out = &output[i];
2712 log_info_buf(out->buf, out->buflen);
2713 buf_output_free(out);
2716 fio_idle_prof_cleanup();
2721 /* free arrays allocated by sum_thread_stats(), if any */
2722 for (i = 0; i < nr_ts; i++) {
2723 ts = &threadstats[i];
2724 free_clat_prio_stats(ts);
2730 int __show_running_run_stats(void)
2732 unsigned long long *rt;
2735 fio_sem_down(stat_sem);
2737 rt = malloc(thread_number * sizeof(unsigned long long));
2738 fio_gettime(&ts, NULL);
2741 if (td->runstate >= TD_EXITED)
2744 td->update_rusage = 1;
2745 for_each_rw_ddir(ddir) {
2746 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2748 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2750 rt[__td_index] = mtime_since(&td->start, &ts);
2751 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2752 td->ts.runtime[DDIR_READ] += rt[__td_index];
2753 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2754 td->ts.runtime[DDIR_WRITE] += rt[__td_index];
2755 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2756 td->ts.runtime[DDIR_TRIM] += rt[__td_index];
2760 if (td->runstate >= TD_EXITED)
2762 if (td->rusage_sem) {
2763 td->update_rusage = 1;
2764 fio_sem_down(td->rusage_sem);
2766 td->update_rusage = 0;
2772 if (td->runstate >= TD_EXITED)
2775 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2776 td->ts.runtime[DDIR_READ] -= rt[__td_index];
2777 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2778 td->ts.runtime[DDIR_WRITE] -= rt[__td_index];
2779 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2780 td->ts.runtime[DDIR_TRIM] -= rt[__td_index];
2784 fio_sem_up(stat_sem);
2789 static bool status_file_disabled;
2791 #define FIO_STATUS_FILE "fio-dump-status"
2793 static int check_status_file(void)
2796 const char *temp_dir;
2797 char fio_status_file_path[PATH_MAX];
2799 if (status_file_disabled)
2802 temp_dir = getenv("TMPDIR");
2803 if (temp_dir == NULL) {
2804 temp_dir = getenv("TEMP");
2805 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2808 if (temp_dir == NULL)
2811 __coverity_tainted_data_sanitize__(temp_dir);
2814 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2816 if (stat(fio_status_file_path, &sb))
2819 if (unlink(fio_status_file_path) < 0) {
2820 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2822 log_err("fio: disabling status file updates\n");
2823 status_file_disabled = true;
2829 void check_for_running_stats(void)
2831 if (check_status_file()) {
2832 show_running_run_stats();
2837 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2842 if (data > is->max_val)
2844 if (data < is->min_val)
2847 delta = val - is->mean.u.f;
2849 is->mean.u.f += delta / (is->samples + 1.0);
2850 is->S.u.f += delta * (val - is->mean.u.f);
2856 static inline void add_stat_prio_sample(struct clat_prio_stat *clat_prio,
2857 unsigned short clat_prio_index,
2858 unsigned long long nsec)
2861 add_stat_sample(&clat_prio[clat_prio_index].clat_stat, nsec);
2865 * Return a struct io_logs, which is added to the tail of the log
2868 static struct io_logs *get_new_log(struct io_log *iolog)
2871 struct io_logs *cur_log;
2874 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2877 if (!iolog->cur_log_max) {
2879 new_samples = iolog->td->o.log_entries;
2881 new_samples = DEF_LOG_ENTRIES;
2883 new_samples = iolog->cur_log_max * 2;
2884 if (new_samples > MAX_LOG_ENTRIES)
2885 new_samples = MAX_LOG_ENTRIES;
2888 cur_log = smalloc(sizeof(*cur_log));
2890 INIT_FLIST_HEAD(&cur_log->list);
2891 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2893 cur_log->nr_samples = 0;
2894 cur_log->max_samples = new_samples;
2895 flist_add_tail(&cur_log->list, &iolog->io_logs);
2896 iolog->cur_log_max = new_samples;
2906 * Add and return a new log chunk, or return current log if big enough
2908 static struct io_logs *regrow_log(struct io_log *iolog)
2910 struct io_logs *cur_log;
2913 if (!iolog || iolog->disabled)
2916 cur_log = iolog_cur_log(iolog);
2918 cur_log = get_new_log(iolog);
2923 if (cur_log->nr_samples < cur_log->max_samples)
2927 * No room for a new sample. If we're compressing on the fly, flush
2928 * out the current chunk
2930 if (iolog->log_gz) {
2931 if (iolog_cur_flush(iolog, cur_log)) {
2932 log_err("fio: failed flushing iolog! Will stop logging.\n");
2938 * Get a new log array, and add to our list
2940 cur_log = get_new_log(iolog);
2942 log_err("fio: failed extending iolog! Will stop logging.\n");
2946 if (!iolog->pending || !iolog->pending->nr_samples)
2950 * Flush pending items to new log
2952 for (i = 0; i < iolog->pending->nr_samples; i++) {
2953 struct io_sample *src, *dst;
2955 src = get_sample(iolog, iolog->pending, i);
2956 dst = get_sample(iolog, cur_log, i);
2957 memcpy(dst, src, log_entry_sz(iolog));
2959 cur_log->nr_samples = iolog->pending->nr_samples;
2961 iolog->pending->nr_samples = 0;
2965 iolog->disabled = true;
2969 void regrow_logs(struct thread_data *td)
2971 regrow_log(td->slat_log);
2972 regrow_log(td->clat_log);
2973 regrow_log(td->clat_hist_log);
2974 regrow_log(td->lat_log);
2975 regrow_log(td->bw_log);
2976 regrow_log(td->iops_log);
2977 td->flags &= ~TD_F_REGROW_LOGS;
2980 void regrow_agg_logs(void)
2984 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2985 regrow_log(agg_io_log[ddir]);
2988 static struct io_logs *get_cur_log(struct io_log *iolog)
2990 struct io_logs *cur_log;
2992 cur_log = iolog_cur_log(iolog);
2994 cur_log = get_new_log(iolog);
2999 if (cur_log->nr_samples < cur_log->max_samples)
3003 * Out of space. If we're in IO offload mode, or we're not doing
3004 * per unit logging (hence logging happens outside of the IO thread
3005 * as well), add a new log chunk inline. If we're doing inline
3006 * submissions, flag 'td' as needing a log regrow and we'll take
3007 * care of it on the submission side.
3009 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
3010 !per_unit_log(iolog))
3011 return regrow_log(iolog);
3014 iolog->td->flags |= TD_F_REGROW_LOGS;
3016 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
3017 return iolog->pending;
3020 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
3021 enum fio_ddir ddir, unsigned long long bs,
3022 unsigned long t, uint64_t offset,
3023 unsigned int priority)
3025 struct io_logs *cur_log;
3027 if (iolog->disabled)
3029 if (flist_empty(&iolog->io_logs))
3030 iolog->avg_last[ddir] = t;
3032 cur_log = get_cur_log(iolog);
3034 struct io_sample *s;
3036 s = get_sample(iolog, cur_log, cur_log->nr_samples);
3039 s->time = t + (iolog->td ? iolog->td->alternate_epoch : 0);
3040 io_sample_set_ddir(iolog, s, ddir);
3042 s->priority = priority;
3044 if (iolog->log_offset) {
3045 struct io_sample_offset *so = (void *) s;
3047 so->offset = offset;
3050 cur_log->nr_samples++;
3054 iolog->disabled = true;
3057 static inline void reset_io_stat(struct io_stat *ios)
3059 ios->min_val = -1ULL;
3060 ios->max_val = ios->samples = 0;
3061 ios->mean.u.f = ios->S.u.f = 0;
3064 static inline void reset_io_u_plat(uint64_t *io_u_plat)
3068 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
3072 static inline void reset_clat_prio_stats(struct thread_stat *ts)
3077 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3078 if (!ts->clat_prio[ddir])
3081 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
3082 reset_io_stat(&ts->clat_prio[ddir][i].clat_stat);
3083 reset_io_u_plat(ts->clat_prio[ddir][i].io_u_plat);
3088 void reset_io_stats(struct thread_data *td)
3090 struct thread_stat *ts = &td->ts;
3093 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
3094 reset_io_stat(&ts->clat_stat[i]);
3095 reset_io_stat(&ts->slat_stat[i]);
3096 reset_io_stat(&ts->lat_stat[i]);
3097 reset_io_stat(&ts->bw_stat[i]);
3098 reset_io_stat(&ts->iops_stat[i]);
3100 ts->io_bytes[i] = 0;
3102 ts->total_io_u[i] = 0;
3103 ts->short_io_u[i] = 0;
3104 ts->drop_io_u[i] = 0;
3107 for (i = 0; i < FIO_LAT_CNT; i++)
3108 for (j = 0; j < DDIR_RWDIR_CNT; j++)
3109 reset_io_u_plat(ts->io_u_plat[i][j]);
3111 reset_clat_prio_stats(ts);
3113 ts->total_io_u[DDIR_SYNC] = 0;
3114 reset_io_u_plat(ts->io_u_sync_plat);
3116 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
3117 ts->io_u_map[i] = 0;
3118 ts->io_u_submit[i] = 0;
3119 ts->io_u_complete[i] = 0;
3122 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
3123 ts->io_u_lat_n[i] = 0;
3124 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
3125 ts->io_u_lat_u[i] = 0;
3126 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
3127 ts->io_u_lat_m[i] = 0;
3129 ts->total_submit = 0;
3130 ts->total_complete = 0;
3131 ts->nr_zone_resets = 0;
3132 ts->cachehit = ts->cachemiss = 0;
3135 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
3136 unsigned long elapsed, bool log_max)
3139 * Note an entry in the log. Use the mean from the logged samples,
3140 * making sure to properly round up. Only write a log entry if we
3141 * had actual samples done.
3143 if (iolog->avg_window[ddir].samples) {
3144 union io_sample_data data;
3147 data.val = iolog->avg_window[ddir].max_val;
3149 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
3151 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
3154 reset_io_stat(&iolog->avg_window[ddir]);
3157 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
3162 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
3163 __add_stat_to_log(iolog, ddir, elapsed, log_max);
3166 static unsigned long add_log_sample(struct thread_data *td,
3167 struct io_log *iolog,
3168 union io_sample_data data,
3169 enum fio_ddir ddir, unsigned long long bs,
3170 uint64_t offset, unsigned int ioprio)
3172 unsigned long elapsed, this_window;
3177 elapsed = mtime_since_now(&td->epoch);
3180 * If no time averaging, just add the log sample.
3182 if (!iolog->avg_msec) {
3183 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
3189 * Add the sample. If the time period has passed, then
3190 * add that entry to the log and clear.
3192 add_stat_sample(&iolog->avg_window[ddir], data.val);
3195 * If period hasn't passed, adding the above sample is all we
3198 this_window = elapsed - iolog->avg_last[ddir];
3199 if (elapsed < iolog->avg_last[ddir])
3200 return iolog->avg_last[ddir] - elapsed;
3201 else if (this_window < iolog->avg_msec) {
3202 unsigned long diff = iolog->avg_msec - this_window;
3204 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
3208 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
3210 iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
3212 return iolog->avg_msec;
3215 void finalize_logs(struct thread_data *td, bool unit_logs)
3217 unsigned long elapsed;
3219 elapsed = mtime_since_now(&td->epoch);
3221 if (td->clat_log && unit_logs)
3222 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
3223 if (td->slat_log && unit_logs)
3224 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
3225 if (td->lat_log && unit_logs)
3226 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
3227 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
3228 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
3229 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
3230 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
3233 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
3234 unsigned long long bs)
3236 struct io_log *iolog;
3241 iolog = agg_io_log[ddir];
3242 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
3245 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
3247 unsigned int idx = plat_val_to_idx(nsec);
3248 assert(idx < FIO_IO_U_PLAT_NR);
3250 ts->io_u_sync_plat[idx]++;
3251 add_stat_sample(&ts->sync_stat, nsec);
3254 static inline void add_lat_percentile_sample(struct thread_stat *ts,
3255 unsigned long long nsec,
3259 unsigned int idx = plat_val_to_idx(nsec);
3260 assert(idx < FIO_IO_U_PLAT_NR);
3262 ts->io_u_plat[lat][ddir][idx]++;
3266 add_lat_percentile_prio_sample(struct thread_stat *ts, unsigned long long nsec,
3268 unsigned short clat_prio_index)
3270 unsigned int idx = plat_val_to_idx(nsec);
3272 if (ts->clat_prio[ddir])
3273 ts->clat_prio[ddir][clat_prio_index].io_u_plat[idx]++;
3276 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
3277 unsigned long long nsec, unsigned long long bs,
3278 uint64_t offset, unsigned int ioprio,
3279 unsigned short clat_prio_index)
3281 const bool needs_lock = td_async_processing(td);
3282 unsigned long elapsed, this_window;
3283 struct thread_stat *ts = &td->ts;
3284 struct io_log *iolog = td->clat_hist_log;
3289 add_stat_sample(&ts->clat_stat[ddir], nsec);
3292 * When lat_percentiles=1 (default 0), the reported per priority
3293 * percentiles and stats are used for describing total latency values,
3294 * even though the variable names themselves start with clat_.
3296 * Because of the above definition, add a prio stat sample only when
3297 * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3298 * when lat_percentiles=1.
3300 if (!ts->lat_percentiles)
3301 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3305 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3308 if (ts->clat_percentiles) {
3310 * Because of the above definition, add a prio lat percentile
3311 * sample only when lat_percentiles=0. add_lat_sample() will add
3312 * the prio lat percentile sample when lat_percentiles=1.
3314 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3315 if (!ts->lat_percentiles)
3316 add_lat_percentile_prio_sample(ts, nsec, ddir,
3320 if (iolog && iolog->hist_msec) {
3321 struct io_hist *hw = &iolog->hist_window[ddir];
3324 elapsed = mtime_since_now(&td->epoch);
3326 hw->hist_last = elapsed;
3327 this_window = elapsed - hw->hist_last;
3329 if (this_window >= iolog->hist_msec) {
3330 uint64_t *io_u_plat;
3331 struct io_u_plat_entry *dst;
3334 * Make a byte-for-byte copy of the latency histogram
3335 * stored in td->ts.io_u_plat[ddir], recording it in a
3336 * log sample. Note that the matching call to free() is
3337 * located in iolog.c after printing this sample to the
3340 io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3341 dst = malloc(sizeof(struct io_u_plat_entry));
3342 memcpy(&(dst->io_u_plat), io_u_plat,
3343 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3344 flist_add(&dst->list, &hw->list);
3345 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3346 elapsed, offset, ioprio);
3349 * Update the last time we recorded as being now, minus
3350 * any drift in time we encountered before actually
3351 * making the record.
3353 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3359 __td_io_u_unlock(td);
3362 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3363 unsigned long long nsec, unsigned long long bs,
3364 uint64_t offset, unsigned int ioprio)
3366 const bool needs_lock = td_async_processing(td);
3367 struct thread_stat *ts = &td->ts;
3375 add_stat_sample(&ts->slat_stat[ddir], nsec);
3378 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3381 if (ts->slat_percentiles)
3382 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3385 __td_io_u_unlock(td);
3388 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3389 unsigned long long nsec, unsigned long long bs,
3390 uint64_t offset, unsigned int ioprio,
3391 unsigned short clat_prio_index)
3393 const bool needs_lock = td_async_processing(td);
3394 struct thread_stat *ts = &td->ts;
3402 add_stat_sample(&ts->lat_stat[ddir], nsec);
3405 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3409 * When lat_percentiles=1 (default 0), the reported per priority
3410 * percentiles and stats are used for describing total latency values,
3411 * even though the variable names themselves start with clat_.
3413 * Because of the above definition, add a prio stat and prio lat
3414 * percentile sample only when lat_percentiles=1. add_clat_sample() will
3415 * add the prio stat and prio lat percentile sample when
3416 * lat_percentiles=0.
3418 if (ts->lat_percentiles) {
3419 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3420 add_lat_percentile_prio_sample(ts, nsec, ddir, clat_prio_index);
3421 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3425 __td_io_u_unlock(td);
3428 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3429 unsigned int bytes, unsigned long long spent)
3431 const bool needs_lock = td_async_processing(td);
3432 struct thread_stat *ts = &td->ts;
3436 rate = (unsigned long) (bytes * 1000000ULL / spent);
3443 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3446 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3447 bytes, io_u->offset, io_u->ioprio);
3449 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3452 __td_io_u_unlock(td);
3455 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3456 struct timespec *t, unsigned int avg_time,
3457 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3458 struct io_stat *stat, struct io_log *log,
3461 const bool needs_lock = td_async_processing(td);
3462 unsigned long spent, rate;
3464 unsigned long next, next_log;
3466 next_log = avg_time;
3468 spent = mtime_since(parent_tv, t);
3469 if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3470 return avg_time - spent;
3476 * Compute both read and write rates for the interval.
3478 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3481 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3483 continue; /* No entries for interval */
3487 rate = delta * 1000 / spent / 1024; /* KiB/s */
3489 rate = (delta * 1000) / spent;
3493 add_stat_sample(&stat[ddir], rate);
3496 unsigned long long bs = 0;
3498 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3499 bs = td->o.min_bs[ddir];
3501 next = add_log_sample(td, log, sample_val(rate), ddir,
3503 next_log = min(next_log, next);
3506 stat_io_bytes[ddir] = this_io_bytes[ddir];
3512 __td_io_u_unlock(td);
3514 if (spent <= avg_time)
3517 next = avg_time - (1 + spent - avg_time);
3519 return min(next, next_log);
3522 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3524 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3525 td->this_io_bytes, td->stat_io_bytes,
3526 td->ts.bw_stat, td->bw_log, true);
3529 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3532 const bool needs_lock = td_async_processing(td);
3533 struct thread_stat *ts = &td->ts;
3538 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3541 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3542 bytes, io_u->offset, io_u->ioprio);
3544 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3547 __td_io_u_unlock(td);
3550 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3552 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3553 td->this_io_blocks, td->stat_io_blocks,
3554 td->ts.iops_stat, td->iops_log, false);
3558 * Returns msecs to next event
3560 int calc_log_samples(void)
3562 unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3563 struct timespec now;
3564 long elapsed_time = 0;
3566 fio_gettime(&now, NULL);
3569 elapsed_time = mtime_since_now(&td->epoch);
3573 if (in_ramp_time(td) ||
3574 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3575 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3579 (td->bw_log && !per_unit_log(td->bw_log))) {
3580 tmp = add_bw_samples(td, &now);
3583 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3585 if (!td->iops_log ||
3586 (td->iops_log && !per_unit_log(td->iops_log))) {
3587 tmp = add_iops_samples(td, &now);
3590 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3597 /* if log_avg_msec_min has not been changed, set it to 0 */
3598 if (log_avg_msec_min == -1U)
3599 log_avg_msec_min = 0;
3601 if (log_avg_msec_min == 0)
3602 next_mod = elapsed_time;
3604 next_mod = elapsed_time % log_avg_msec_min;
3606 /* correction to keep the time on the log avg msec boundary */
3607 next = min(next, (log_avg_msec_min - next_mod));
3609 return next == ~0U ? 0 : next;
3612 void stat_init(void)
3614 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3617 void stat_exit(void)
3620 * When we have the mutex, we know out-of-band access to it
3623 fio_sem_down(stat_sem);
3624 fio_sem_remove(stat_sem);
3628 * Called from signal handler. Wake up status thread.
3630 void show_running_run_stats(void)
3635 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3637 /* Ignore io_u's which span multiple blocks--they will just get
3638 * inaccurate counts. */
3639 int idx = (io_u->offset - io_u->file->file_offset)
3640 / td->o.bs[DDIR_TRIM];
3641 uint32_t *info = &td->ts.block_infos[idx];
3642 assert(idx < td->ts.nr_block_infos);