9 #include "lib/ieee754.h"
11 #include "lib/getrusage.h"
14 #include "lib/output_buffer.h"
15 #include "helper_thread.h"
18 #include "oslib/asprintf.h"
21 #define LOG_MSEC_SLACK 2
23 #define LOG_MSEC_SLACK 1
26 struct fio_sem *stat_sem;
28 void clear_rusage_stat(struct thread_data *td)
30 struct thread_stat *ts = &td->ts;
32 fio_getrusage(&td->ru_start);
33 ts->usr_time = ts->sys_time = 0;
35 ts->minf = ts->majf = 0;
38 void update_rusage_stat(struct thread_data *td)
40 struct thread_stat *ts = &td->ts;
42 fio_getrusage(&td->ru_end);
43 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
44 &td->ru_end.ru_utime);
45 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
46 &td->ru_end.ru_stime);
47 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
48 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
49 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
50 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
52 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
56 * Given a latency, return the index of the corresponding bucket in
57 * the structure tracking percentiles.
59 * (1) find the group (and error bits) that the value (latency)
60 * belongs to by looking at its MSB. (2) find the bucket number in the
61 * group by looking at the index bits.
64 static unsigned int plat_val_to_idx(unsigned long long val)
66 unsigned int msb, error_bits, base, offset, idx;
68 /* Find MSB starting from bit 0 */
72 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
75 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
76 * all bits of the sample as index
78 if (msb <= FIO_IO_U_PLAT_BITS)
81 /* Compute the number of error bits to discard*/
82 error_bits = msb - FIO_IO_U_PLAT_BITS;
84 /* Compute the number of buckets before the group */
85 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
88 * Discard the error bits and apply the mask to find the
89 * index for the buckets in the group
91 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
93 /* Make sure the index does not exceed (array size - 1) */
94 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
95 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
101 * Convert the given index of the bucket array to the value
102 * represented by the bucket
104 static unsigned long long plat_idx_to_val(unsigned int idx)
106 unsigned int error_bits;
107 unsigned long long k, base;
109 assert(idx < FIO_IO_U_PLAT_NR);
111 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
112 * all bits of the sample as index */
113 if (idx < (FIO_IO_U_PLAT_VAL << 1))
116 /* Find the group and compute the minimum value of that group */
117 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
118 base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
120 /* Find its bucket number of the group */
121 k = idx % FIO_IO_U_PLAT_VAL;
123 /* Return the mean of the range of the bucket */
124 return base + ((k + 0.5) * (1 << error_bits));
127 static int double_cmp(const void *a, const void *b)
129 const fio_fp64_t fa = *(const fio_fp64_t *) a;
130 const fio_fp64_t fb = *(const fio_fp64_t *) b;
135 else if (fa.u.f < fb.u.f)
141 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
142 fio_fp64_t *plist, unsigned long long **output,
143 unsigned long long *maxv, unsigned long long *minv)
145 unsigned long long sum = 0;
146 unsigned int len, i, j = 0;
147 unsigned long long *ovals = NULL;
154 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
161 * Sort the percentile list. Note that it may already be sorted if
162 * we are using the default values, but since it's a short list this
163 * isn't a worry. Also note that this does not work for NaN values.
166 qsort(plist, len, sizeof(plist[0]), double_cmp);
168 ovals = malloc(len * sizeof(*ovals));
173 * Calculate bucket values, note down max and min values
176 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
178 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
179 assert(plist[j].u.f <= 100.0);
181 ovals[j] = plat_idx_to_val(i);
182 if (ovals[j] < *minv)
184 if (ovals[j] > *maxv)
187 is_last = (j == len - 1) != 0;
196 log_err("fio: error calculating latency percentiles\n");
203 * Find and display the p-th percentile of clat
205 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
206 fio_fp64_t *plist, unsigned int precision,
207 const char *pre, struct buf_output *out)
209 unsigned int divisor, len, i, j = 0;
210 unsigned long long minv, maxv;
211 unsigned long long *ovals;
212 int per_line, scale_down, time_width;
216 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
221 * We default to nsecs, but if the value range is such that we
222 * should scale down to usecs or msecs, do that.
224 if (minv > 2000000 && maxv > 99999999ULL) {
227 log_buf(out, " %s percentiles (msec):\n |", pre);
228 } else if (minv > 2000 && maxv > 99999) {
231 log_buf(out, " %s percentiles (usec):\n |", pre);
235 log_buf(out, " %s percentiles (nsec):\n |", pre);
239 time_width = max(5, (int) (log10(maxv / divisor) + 1));
240 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
241 precision, time_width);
242 /* fmt will be something like " %5.2fth=[%4llu]%c" */
243 per_line = (80 - 7) / (precision + 10 + time_width);
245 for (j = 0; j < len; j++) {
247 if (j != 0 && (j % per_line) == 0)
250 /* end of the list */
251 is_last = (j == len - 1) != 0;
253 for (i = 0; i < scale_down; i++)
254 ovals[j] = (ovals[j] + 999) / 1000;
256 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
261 if ((j % per_line) == per_line - 1) /* for formatting */
268 static int get_nr_prios_with_samples(struct thread_stat *ts, enum fio_ddir ddir)
270 int i, nr_prios_with_samples = 0;
272 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
273 if (ts->clat_prio[ddir][i].clat_stat.samples)
274 nr_prios_with_samples++;
277 return nr_prios_with_samples;
280 bool calc_lat(struct io_stat *is, unsigned long long *min,
281 unsigned long long *max, double *mean, double *dev)
283 double n = (double) is->samples;
290 *mean = is->mean.u.f;
293 *dev = sqrt(is->S.u.f / (n - 1.0));
300 void show_mixed_group_stats(struct group_run_stats *rs, struct buf_output *out)
302 char *io, *agg, *min, *max;
303 char *ioalt, *aggalt, *minalt, *maxalt;
304 uint64_t io_mix = 0, agg_mix = 0, min_mix = -1, max_mix = 0;
305 uint64_t min_run = -1, max_run = 0;
306 const int i2p = is_power_of_2(rs->kb_base);
309 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
312 io_mix += rs->iobytes[i];
313 agg_mix += rs->agg[i];
314 min_mix = min_mix < rs->min_bw[i] ? min_mix : rs->min_bw[i];
315 max_mix = max_mix > rs->max_bw[i] ? max_mix : rs->max_bw[i];
316 min_run = min_run < rs->min_run[i] ? min_run : rs->min_run[i];
317 max_run = max_run > rs->max_run[i] ? max_run : rs->max_run[i];
319 io = num2str(io_mix, rs->sig_figs, 1, i2p, N2S_BYTE);
320 ioalt = num2str(io_mix, rs->sig_figs, 1, !i2p, N2S_BYTE);
321 agg = num2str(agg_mix, rs->sig_figs, 1, i2p, rs->unit_base);
322 aggalt = num2str(agg_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
323 min = num2str(min_mix, rs->sig_figs, 1, i2p, rs->unit_base);
324 minalt = num2str(min_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
325 max = num2str(max_mix, rs->sig_figs, 1, i2p, rs->unit_base);
326 maxalt = num2str(max_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
327 log_buf(out, " MIXED: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
328 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
329 (unsigned long long) min_run,
330 (unsigned long long) max_run);
341 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
343 char *io, *agg, *min, *max;
344 char *ioalt, *aggalt, *minalt, *maxalt;
345 const char *str[] = { " READ", " WRITE" , " TRIM"};
348 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
350 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
351 const int i2p = is_power_of_2(rs->kb_base);
356 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
357 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
358 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
359 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
360 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
361 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
362 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
363 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
364 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
365 (rs->unified_rw_rep == UNIFIED_MIXED) ? " MIXED" : str[i],
366 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
367 (unsigned long long) rs->min_run[i],
368 (unsigned long long) rs->max_run[i]);
380 /* Need to aggregate statistics to show mixed values */
381 if (rs->unified_rw_rep == UNIFIED_BOTH)
382 show_mixed_group_stats(rs, out);
385 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
390 * Do depth distribution calculations
392 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
394 io_u_dist[i] = (double) map[i] / (double) total;
395 io_u_dist[i] *= 100.0;
396 if (io_u_dist[i] < 0.1 && map[i])
403 static void stat_calc_lat(struct thread_stat *ts, double *dst,
404 uint64_t *src, int nr)
406 unsigned long total = ddir_rw_sum(ts->total_io_u);
410 * Do latency distribution calculations
412 for (i = 0; i < nr; i++) {
414 dst[i] = (double) src[i] / (double) total;
416 if (dst[i] < 0.01 && src[i])
424 * To keep the terse format unaltered, add all of the ns latency
425 * buckets to the first us latency bucket
427 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
429 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
432 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
434 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
435 ntotal += ts->io_u_lat_n[i];
437 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
440 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
442 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
445 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
447 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
450 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
452 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
455 static void display_lat(const char *name, unsigned long long min,
456 unsigned long long max, double mean, double dev,
457 struct buf_output *out)
459 const char *base = "(nsec)";
462 if (nsec_to_msec(&min, &max, &mean, &dev))
464 else if (nsec_to_usec(&min, &max, &mean, &dev))
467 minp = num2str(min, 6, 1, 0, N2S_NONE);
468 maxp = num2str(max, 6, 1, 0, N2S_NONE);
470 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
471 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
477 static struct thread_stat *gen_mixed_ddir_stats_from_ts(struct thread_stat *ts)
479 struct thread_stat *ts_lcl;
482 * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
485 ts_lcl = malloc(sizeof(struct thread_stat));
487 log_err("fio: failed to allocate local thread stat\n");
491 init_thread_stat(ts_lcl);
493 /* calculate mixed stats */
494 ts_lcl->unified_rw_rep = UNIFIED_MIXED;
495 ts_lcl->lat_percentiles = ts->lat_percentiles;
496 ts_lcl->clat_percentiles = ts->clat_percentiles;
497 ts_lcl->slat_percentiles = ts->slat_percentiles;
498 ts_lcl->percentile_precision = ts->percentile_precision;
499 memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
501 sum_thread_stats(ts_lcl, ts);
506 static double convert_agg_kbytes_percent(struct group_run_stats *rs,
507 enum fio_ddir ddir, int mean)
509 double p_of_agg = 100.0;
510 if (rs && rs->agg[ddir] > 1024) {
511 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
513 if (p_of_agg > 100.0)
519 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
520 enum fio_ddir ddir, struct buf_output *out)
523 unsigned long long min, max, bw, iops;
525 char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
527 const char *clat_type = ts->lat_percentiles ? "lat" : "clat";
529 if (ddir_sync(ddir)) {
530 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
531 log_buf(out, " %s:\n", "fsync/fdatasync/sync_file_range");
532 display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
533 show_clat_percentiles(ts->io_u_sync_plat,
534 ts->sync_stat.samples,
536 ts->percentile_precision,
537 io_ddir_name(ddir), out);
542 assert(ddir_rw(ddir));
544 if (!ts->runtime[ddir])
547 i2p = is_power_of_2(rs->kb_base);
548 runt = ts->runtime[ddir];
550 bw = (1000 * ts->io_bytes[ddir]) / runt;
551 io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
552 bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
553 bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
555 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
556 iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
557 if (ddir == DDIR_WRITE)
558 post_st = zbd_write_status(ts);
559 else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
563 total = ts->cachehit + ts->cachemiss;
564 hit = (double) ts->cachehit / (double) total;
566 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
570 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
571 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
572 iops_p, bw_p, bw_p_alt, io_p,
573 (unsigned long long) ts->runtime[ddir],
582 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
583 display_lat("slat", min, max, mean, dev, out);
584 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
585 display_lat("clat", min, max, mean, dev, out);
586 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
587 display_lat(" lat", min, max, mean, dev, out);
589 /* Only print per prio stats if there are >= 2 prios with samples */
590 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
591 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
592 if (calc_lat(&ts->clat_prio[ddir][i].clat_stat, &min,
593 &max, &mean, &dev)) {
596 snprintf(buf, sizeof(buf),
599 ts->clat_prio[ddir][i].ioprio >> 13,
600 ts->clat_prio[ddir][i].ioprio & 7);
601 display_lat(buf, min, max, mean, dev, out);
606 if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
607 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
608 ts->slat_stat[ddir].samples,
610 ts->percentile_precision, "slat", out);
611 if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
612 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
613 ts->clat_stat[ddir].samples,
615 ts->percentile_precision, "clat", out);
616 if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
617 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
618 ts->lat_stat[ddir].samples,
620 ts->percentile_precision, "lat", out);
622 if (ts->clat_percentiles || ts->lat_percentiles) {
626 if (ts->lat_percentiles)
627 samples = ts->lat_stat[ddir].samples;
629 samples = ts->clat_stat[ddir].samples;
631 /* Only print per prio stats if there are >= 2 prios with samples */
632 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
633 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
634 uint64_t prio_samples = ts->clat_prio[ddir][i].clat_stat.samples;
636 if (prio_samples > 0) {
637 snprintf(prio_name, sizeof(prio_name),
638 "%s prio %u/%u (%.2f%% of IOs)",
640 ts->clat_prio[ddir][i].ioprio >> 13,
641 ts->clat_prio[ddir][i].ioprio & 7,
642 100. * (double) prio_samples / (double) samples);
643 show_clat_percentiles(ts->clat_prio[ddir][i].io_u_plat,
644 prio_samples, ts->percentile_list,
645 ts->percentile_precision,
652 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
653 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
656 if ((rs->unit_base == 1) && i2p)
658 else if (rs->unit_base == 1)
665 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
667 if (rs->unit_base == 1) {
674 if (mean > fkb_base * fkb_base) {
679 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
682 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
683 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
684 bw_str, min, max, p_of_agg, mean, dev,
685 (&ts->bw_stat[ddir])->samples);
687 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
688 log_buf(out, " iops : min=%5llu, max=%5llu, "
689 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
690 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
694 static void show_mixed_ddir_status(struct group_run_stats *rs,
695 struct thread_stat *ts,
696 struct buf_output *out)
698 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
701 show_ddir_status(rs, ts_lcl, DDIR_READ, out);
703 free_clat_prio_stats(ts_lcl);
707 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
708 const char *msg, struct buf_output *out)
710 bool new_line = true, shown = false;
713 for (i = 0; i < nr; i++) {
714 if (io_u_lat[i] <= 0.0)
720 log_buf(out, " lat (%s) : ", msg);
726 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
738 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
740 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
741 "250=", "500=", "750=", "1000=", };
743 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
746 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
748 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
749 "250=", "500=", "750=", "1000=", };
751 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
754 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
756 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
757 "250=", "500=", "750=", "1000=", "2000=",
760 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
763 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
765 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
766 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
767 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
769 stat_calc_lat_n(ts, io_u_lat_n);
770 stat_calc_lat_u(ts, io_u_lat_u);
771 stat_calc_lat_m(ts, io_u_lat_m);
773 show_lat_n(io_u_lat_n, out);
774 show_lat_u(io_u_lat_u, out);
775 show_lat_m(io_u_lat_m, out);
778 static int block_state_category(int block_state)
780 switch (block_state) {
781 case BLOCK_STATE_UNINIT:
783 case BLOCK_STATE_TRIMMED:
784 case BLOCK_STATE_WRITTEN:
786 case BLOCK_STATE_WRITE_FAILURE:
787 case BLOCK_STATE_TRIM_FAILURE:
790 /* Silence compile warning on some BSDs and have a return */
796 static int compare_block_infos(const void *bs1, const void *bs2)
798 uint64_t block1 = *(uint64_t *)bs1;
799 uint64_t block2 = *(uint64_t *)bs2;
800 int state1 = BLOCK_INFO_STATE(block1);
801 int state2 = BLOCK_INFO_STATE(block2);
802 int bscat1 = block_state_category(state1);
803 int bscat2 = block_state_category(state2);
804 int cycles1 = BLOCK_INFO_TRIMS(block1);
805 int cycles2 = BLOCK_INFO_TRIMS(block2);
812 if (cycles1 < cycles2)
814 if (cycles1 > cycles2)
822 assert(block1 == block2);
826 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
827 fio_fp64_t *plist, unsigned int **percentiles,
833 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
835 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
842 * Sort the percentile list. Note that it may already be sorted if
843 * we are using the default values, but since it's a short list this
844 * isn't a worry. Also note that this does not work for NaN values.
847 qsort(plist, len, sizeof(plist[0]), double_cmp);
849 /* Start only after the uninit entries end */
851 nr_uninit < nr_block_infos
852 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
856 if (nr_uninit == nr_block_infos)
859 *percentiles = calloc(len, sizeof(**percentiles));
861 for (i = 0; i < len; i++) {
862 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
864 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
867 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
868 for (i = 0; i < nr_block_infos; i++)
869 types[BLOCK_INFO_STATE(block_infos[i])]++;
874 static const char *block_state_names[] = {
875 [BLOCK_STATE_UNINIT] = "unwritten",
876 [BLOCK_STATE_TRIMMED] = "trimmed",
877 [BLOCK_STATE_WRITTEN] = "written",
878 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
879 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
882 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
883 fio_fp64_t *plist, struct buf_output *out)
886 unsigned int *percentiles = NULL;
887 unsigned int block_state_counts[BLOCK_STATE_COUNT];
889 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
890 &percentiles, block_state_counts);
892 log_buf(out, " block lifetime percentiles :\n |");
894 for (i = 0; i < len; i++) {
895 uint32_t block_info = percentiles[i];
896 #define LINE_LENGTH 75
897 char str[LINE_LENGTH];
898 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
899 plist[i].u.f, block_info,
900 i == len - 1 ? '\n' : ',');
901 assert(strln < LINE_LENGTH);
902 if (pos + strln > LINE_LENGTH) {
904 log_buf(out, "\n |");
906 log_buf(out, "%s", str);
913 log_buf(out, " states :");
914 for (i = 0; i < BLOCK_STATE_COUNT; i++)
915 log_buf(out, " %s=%u%c",
916 block_state_names[i], block_state_counts[i],
917 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
920 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
922 char *p1, *p1alt, *p2;
923 unsigned long long bw_mean, iops_mean;
924 const int i2p = is_power_of_2(ts->kb_base);
929 bw_mean = steadystate_bw_mean(ts);
930 iops_mean = steadystate_iops_mean(ts);
932 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
933 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
934 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
936 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
937 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
939 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
940 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
941 ts->ss_criterion.u.f,
942 ts->ss_state & FIO_SS_PCT ? "%" : "");
949 static void show_agg_stats(struct disk_util_agg *agg, int terse,
950 struct buf_output *out)
952 if (!agg->slavecount)
956 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
957 "aggrticks=%llu/%llu, aggrin_queue=%llu, "
959 (unsigned long long) agg->ios[0] / agg->slavecount,
960 (unsigned long long) agg->ios[1] / agg->slavecount,
961 (unsigned long long) agg->merges[0] / agg->slavecount,
962 (unsigned long long) agg->merges[1] / agg->slavecount,
963 (unsigned long long) agg->ticks[0] / agg->slavecount,
964 (unsigned long long) agg->ticks[1] / agg->slavecount,
965 (unsigned long long) agg->time_in_queue / agg->slavecount,
968 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
969 (unsigned long long) agg->ios[0] / agg->slavecount,
970 (unsigned long long) agg->ios[1] / agg->slavecount,
971 (unsigned long long) agg->merges[0] / agg->slavecount,
972 (unsigned long long) agg->merges[1] / agg->slavecount,
973 (unsigned long long) agg->ticks[0] / agg->slavecount,
974 (unsigned long long) agg->ticks[1] / agg->slavecount,
975 (unsigned long long) agg->time_in_queue / agg->slavecount,
980 static void aggregate_slaves_stats(struct disk_util *masterdu)
982 struct disk_util_agg *agg = &masterdu->agg;
983 struct disk_util_stat *dus;
984 struct flist_head *entry;
985 struct disk_util *slavedu;
988 flist_for_each(entry, &masterdu->slaves) {
989 slavedu = flist_entry(entry, struct disk_util, slavelist);
991 agg->ios[0] += dus->s.ios[0];
992 agg->ios[1] += dus->s.ios[1];
993 agg->merges[0] += dus->s.merges[0];
994 agg->merges[1] += dus->s.merges[1];
995 agg->sectors[0] += dus->s.sectors[0];
996 agg->sectors[1] += dus->s.sectors[1];
997 agg->ticks[0] += dus->s.ticks[0];
998 agg->ticks[1] += dus->s.ticks[1];
999 agg->time_in_queue += dus->s.time_in_queue;
1002 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
1003 /* System utilization is the utilization of the
1004 * component with the highest utilization.
1006 if (util > agg->max_util.u.f)
1007 agg->max_util.u.f = util;
1011 if (agg->max_util.u.f > 100.0)
1012 agg->max_util.u.f = 100.0;
1015 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
1016 int terse, struct buf_output *out)
1021 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1026 if (agg->slavecount)
1029 log_buf(out, " %s: ios=%llu/%llu, merge=%llu/%llu, "
1030 "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
1032 (unsigned long long) dus->s.ios[0],
1033 (unsigned long long) dus->s.ios[1],
1034 (unsigned long long) dus->s.merges[0],
1035 (unsigned long long) dus->s.merges[1],
1036 (unsigned long long) dus->s.ticks[0],
1037 (unsigned long long) dus->s.ticks[1],
1038 (unsigned long long) dus->s.time_in_queue,
1041 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1043 (unsigned long long) dus->s.ios[0],
1044 (unsigned long long) dus->s.ios[1],
1045 (unsigned long long) dus->s.merges[0],
1046 (unsigned long long) dus->s.merges[1],
1047 (unsigned long long) dus->s.ticks[0],
1048 (unsigned long long) dus->s.ticks[1],
1049 (unsigned long long) dus->s.time_in_queue,
1054 * If the device has slaves, aggregate the stats for
1055 * those slave devices also.
1057 show_agg_stats(agg, terse, out);
1063 void json_array_add_disk_util(struct disk_util_stat *dus,
1064 struct disk_util_agg *agg, struct json_array *array)
1066 struct json_object *obj;
1070 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1074 obj = json_create_object();
1075 json_array_add_value_object(array, obj);
1077 json_object_add_value_string(obj, "name", (const char *)dus->name);
1078 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1079 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1080 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1081 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1082 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1083 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1084 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1085 json_object_add_value_float(obj, "util", util);
1088 * If the device has slaves, aggregate the stats for
1089 * those slave devices also.
1091 if (!agg->slavecount)
1093 json_object_add_value_int(obj, "aggr_read_ios",
1094 agg->ios[0] / agg->slavecount);
1095 json_object_add_value_int(obj, "aggr_write_ios",
1096 agg->ios[1] / agg->slavecount);
1097 json_object_add_value_int(obj, "aggr_read_merges",
1098 agg->merges[0] / agg->slavecount);
1099 json_object_add_value_int(obj, "aggr_write_merge",
1100 agg->merges[1] / agg->slavecount);
1101 json_object_add_value_int(obj, "aggr_read_ticks",
1102 agg->ticks[0] / agg->slavecount);
1103 json_object_add_value_int(obj, "aggr_write_ticks",
1104 agg->ticks[1] / agg->slavecount);
1105 json_object_add_value_int(obj, "aggr_in_queue",
1106 agg->time_in_queue / agg->slavecount);
1107 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1110 static void json_object_add_disk_utils(struct json_object *obj,
1111 struct flist_head *head)
1113 struct json_array *array = json_create_array();
1114 struct flist_head *entry;
1115 struct disk_util *du;
1117 json_object_add_value_array(obj, "disk_util", array);
1119 flist_for_each(entry, head) {
1120 du = flist_entry(entry, struct disk_util, list);
1122 aggregate_slaves_stats(du);
1123 json_array_add_disk_util(&du->dus, &du->agg, array);
1127 void show_disk_util(int terse, struct json_object *parent,
1128 struct buf_output *out)
1130 struct flist_head *entry;
1131 struct disk_util *du;
1134 if (!is_running_backend())
1137 if (flist_empty(&disk_list))
1140 if ((output_format & FIO_OUTPUT_JSON) && parent)
1145 if (!terse && !do_json)
1146 log_buf(out, "\nDisk stats (read/write):\n");
1149 json_object_add_disk_utils(parent, &disk_list);
1150 } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1151 flist_for_each(entry, &disk_list) {
1152 du = flist_entry(entry, struct disk_util, list);
1154 aggregate_slaves_stats(du);
1155 print_disk_util(&du->dus, &du->agg, terse, out);
1160 static void show_thread_status_normal(struct thread_stat *ts,
1161 struct group_run_stats *rs,
1162 struct buf_output *out)
1164 double usr_cpu, sys_cpu;
1165 unsigned long runtime;
1166 double io_u_dist[FIO_IO_U_MAP_NR];
1170 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1173 memset(time_buf, 0, sizeof(time_buf));
1176 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1179 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1180 ts->name, ts->groupid, ts->members,
1181 ts->error, (int) ts->pid, time_buf);
1183 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1184 ts->name, ts->groupid, ts->members,
1185 ts->error, ts->verror, (int) ts->pid,
1189 if (strlen(ts->description))
1190 log_buf(out, " Description : [%s]\n", ts->description);
1192 for_each_rw_ddir(ddir) {
1193 if (ts->io_bytes[ddir])
1194 show_ddir_status(rs, ts, ddir, out);
1197 if (ts->unified_rw_rep == UNIFIED_BOTH)
1198 show_mixed_ddir_status(rs, ts, out);
1200 show_latencies(ts, out);
1202 if (ts->sync_stat.samples)
1203 show_ddir_status(rs, ts, DDIR_SYNC, out);
1205 runtime = ts->total_run_time;
1207 double runt = (double) runtime;
1209 usr_cpu = (double) ts->usr_time * 100 / runt;
1210 sys_cpu = (double) ts->sys_time * 100 / runt;
1216 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1217 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1218 (unsigned long long) ts->ctx,
1219 (unsigned long long) ts->majf,
1220 (unsigned long long) ts->minf);
1222 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1223 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1224 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1225 io_u_dist[1], io_u_dist[2],
1226 io_u_dist[3], io_u_dist[4],
1227 io_u_dist[5], io_u_dist[6]);
1229 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1230 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1231 " 32=%3.1f%%, 64=%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]);
1235 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1236 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1237 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1238 io_u_dist[1], io_u_dist[2],
1239 io_u_dist[3], io_u_dist[4],
1240 io_u_dist[5], io_u_dist[6]);
1241 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
1242 " short=%llu,%llu,%llu,0"
1243 " dropped=%llu,%llu,%llu,0\n",
1244 (unsigned long long) ts->total_io_u[0],
1245 (unsigned long long) ts->total_io_u[1],
1246 (unsigned long long) ts->total_io_u[2],
1247 (unsigned long long) ts->total_io_u[3],
1248 (unsigned long long) ts->short_io_u[0],
1249 (unsigned long long) ts->short_io_u[1],
1250 (unsigned long long) ts->short_io_u[2],
1251 (unsigned long long) ts->drop_io_u[0],
1252 (unsigned long long) ts->drop_io_u[1],
1253 (unsigned long long) ts->drop_io_u[2]);
1254 if (ts->continue_on_error) {
1255 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
1256 (unsigned long long)ts->total_err_count,
1258 strerror(ts->first_error));
1260 if (ts->latency_depth) {
1261 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1262 (unsigned long long)ts->latency_target,
1263 (unsigned long long)ts->latency_window,
1264 ts->latency_percentile.u.f,
1268 if (ts->nr_block_infos)
1269 show_block_infos(ts->nr_block_infos, ts->block_infos,
1270 ts->percentile_list, out);
1273 show_ss_normal(ts, out);
1276 static void show_ddir_status_terse(struct thread_stat *ts,
1277 struct group_run_stats *rs,
1278 enum fio_ddir ddir, int ver,
1279 struct buf_output *out)
1281 unsigned long long min, max, minv, maxv, bw, iops;
1282 unsigned long long *ovals = NULL;
1287 assert(ddir_rw(ddir));
1290 if (ts->runtime[ddir]) {
1291 uint64_t runt = ts->runtime[ddir];
1293 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1294 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1297 log_buf(out, ";%llu;%llu;%llu;%llu",
1298 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1299 (unsigned long long) ts->runtime[ddir]);
1301 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1302 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1304 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1306 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1307 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1309 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1311 if (ts->lat_percentiles) {
1312 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1313 ts->lat_stat[ddir].samples,
1314 ts->percentile_list, &ovals, &maxv,
1316 } else if (ts->clat_percentiles) {
1317 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1318 ts->clat_stat[ddir].samples,
1319 ts->percentile_list, &ovals, &maxv,
1325 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1327 log_buf(out, ";0%%=0");
1330 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1333 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1334 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1336 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1340 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1342 double p_of_agg = 100.0;
1344 if (rs->agg[ddir]) {
1345 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1346 if (p_of_agg > 100.0)
1350 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1352 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1357 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1359 log_buf(out, ";%lu", 0UL);
1361 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1362 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1363 mean, dev, (&ts->iops_stat[ddir])->samples);
1365 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1369 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1370 struct group_run_stats *rs,
1371 int ver, struct buf_output *out)
1373 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1376 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1378 free_clat_prio_stats(ts_lcl);
1382 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1383 uint32_t percentiles,
1384 struct io_stat *lat_stat,
1385 uint64_t *io_u_plat)
1389 unsigned int i, len;
1390 struct json_object *lat_object, *percentile_object, *clat_bins_object;
1391 unsigned long long min, max, maxv, minv, *ovals = NULL;
1393 if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1397 lat_object = json_create_object();
1398 json_object_add_value_int(lat_object, "min", min);
1399 json_object_add_value_int(lat_object, "max", max);
1400 json_object_add_value_float(lat_object, "mean", mean);
1401 json_object_add_value_float(lat_object, "stddev", dev);
1402 json_object_add_value_int(lat_object, "N", lat_stat->samples);
1404 if (percentiles && lat_stat->samples) {
1405 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1406 ts->percentile_list, &ovals, &maxv, &minv);
1408 if (len > FIO_IO_U_LIST_MAX_LEN)
1409 len = FIO_IO_U_LIST_MAX_LEN;
1411 percentile_object = json_create_object();
1412 json_object_add_value_object(lat_object, "percentile", percentile_object);
1413 for (i = 0; i < len; i++) {
1414 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1415 json_object_add_value_int(percentile_object, buf, ovals[i]);
1419 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1420 clat_bins_object = json_create_object();
1421 json_object_add_value_object(lat_object, "bins", clat_bins_object);
1423 for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1425 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1426 json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1434 static void add_ddir_status_json(struct thread_stat *ts,
1435 struct group_run_stats *rs, enum fio_ddir ddir,
1436 struct json_object *parent)
1438 unsigned long long min, max;
1439 unsigned long long bw_bytes, bw;
1440 double mean, dev, iops;
1441 struct json_object *dir_object, *tmp_object;
1442 double p_of_agg = 100.0;
1444 assert(ddir_rw(ddir) || ddir_sync(ddir));
1446 if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1449 dir_object = json_create_object();
1450 json_object_add_value_object(parent,
1451 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1453 if (ddir_rw(ddir)) {
1457 if (ts->runtime[ddir]) {
1458 uint64_t runt = ts->runtime[ddir];
1460 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1461 bw = bw_bytes / 1024; /* KiB/s */
1462 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1465 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1466 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1467 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1468 json_object_add_value_int(dir_object, "bw", bw);
1469 json_object_add_value_float(dir_object, "iops", iops);
1470 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1471 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1472 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1473 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1475 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1476 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1477 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1479 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1480 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1481 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1483 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1484 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1485 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1487 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1488 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1489 &ts->sync_stat, ts->io_u_sync_plat);
1490 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1496 /* Only include per prio stats if there are >= 2 prios with samples */
1497 if (get_nr_prios_with_samples(ts, ddir) >= 2) {
1498 struct json_array *array = json_create_array();
1499 const char *obj_name;
1502 if (ts->lat_percentiles)
1503 obj_name = "lat_ns";
1505 obj_name = "clat_ns";
1507 json_object_add_value_array(dir_object, "prios", array);
1509 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
1510 if (ts->clat_prio[ddir][i].clat_stat.samples > 0) {
1511 struct json_object *obj = json_create_object();
1512 unsigned long long class, level;
1514 class = ts->clat_prio[ddir][i].ioprio >> 13;
1515 json_object_add_value_int(obj, "prioclass", class);
1516 level = ts->clat_prio[ddir][i].ioprio & 7;
1517 json_object_add_value_int(obj, "prio", level);
1519 tmp_object = add_ddir_lat_json(ts,
1520 ts->clat_percentiles | ts->lat_percentiles,
1521 &ts->clat_prio[ddir][i].clat_stat,
1522 ts->clat_prio[ddir][i].io_u_plat);
1523 json_object_add_value_object(obj, obj_name, tmp_object);
1524 json_array_add_value_object(array, obj);
1529 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1530 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1533 p_of_agg = mean = dev = 0.0;
1536 json_object_add_value_int(dir_object, "bw_min", min);
1537 json_object_add_value_int(dir_object, "bw_max", max);
1538 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1539 json_object_add_value_float(dir_object, "bw_mean", mean);
1540 json_object_add_value_float(dir_object, "bw_dev", dev);
1541 json_object_add_value_int(dir_object, "bw_samples",
1542 (&ts->bw_stat[ddir])->samples);
1544 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1548 json_object_add_value_int(dir_object, "iops_min", min);
1549 json_object_add_value_int(dir_object, "iops_max", max);
1550 json_object_add_value_float(dir_object, "iops_mean", mean);
1551 json_object_add_value_float(dir_object, "iops_stddev", dev);
1552 json_object_add_value_int(dir_object, "iops_samples",
1553 (&ts->iops_stat[ddir])->samples);
1555 if (ts->cachehit + ts->cachemiss) {
1559 total = ts->cachehit + ts->cachemiss;
1560 hit = (double) ts->cachehit / (double) total;
1562 json_object_add_value_float(dir_object, "cachehit", hit);
1566 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1567 struct group_run_stats *rs, struct json_object *parent)
1569 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1571 /* add the aggregated stats to json parent */
1573 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1575 free_clat_prio_stats(ts_lcl);
1579 static void show_thread_status_terse_all(struct thread_stat *ts,
1580 struct group_run_stats *rs, int ver,
1581 struct buf_output *out)
1583 double io_u_dist[FIO_IO_U_MAP_NR];
1584 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1585 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1586 double usr_cpu, sys_cpu;
1591 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1593 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1594 ts->name, ts->groupid, ts->error);
1596 /* Log Read Status, or mixed if unified_rw_rep = 1 */
1597 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1598 if (ts->unified_rw_rep != UNIFIED_MIXED) {
1599 /* Log Write Status */
1600 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1601 /* Log Trim Status */
1602 if (ver == 2 || ver == 4 || ver == 5)
1603 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1605 if (ts->unified_rw_rep == UNIFIED_BOTH)
1606 show_mixed_ddir_status_terse(ts, rs, ver, out);
1608 if (ts->total_run_time) {
1609 double runt = (double) ts->total_run_time;
1611 usr_cpu = (double) ts->usr_time * 100 / runt;
1612 sys_cpu = (double) ts->sys_time * 100 / runt;
1618 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1619 (unsigned long long) ts->ctx,
1620 (unsigned long long) ts->majf,
1621 (unsigned long long) ts->minf);
1623 /* Calc % distribution of IO depths, usecond, msecond latency */
1624 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1625 stat_calc_lat_nu(ts, io_u_lat_u);
1626 stat_calc_lat_m(ts, io_u_lat_m);
1628 /* Only show fixed 7 I/O depth levels*/
1629 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1630 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1631 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1633 /* Microsecond latency */
1634 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1635 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1636 /* Millisecond latency */
1637 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1638 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1640 /* disk util stats, if any */
1641 if (ver >= 3 && is_running_backend())
1642 show_disk_util(1, NULL, out);
1644 /* Additional output if continue_on_error set - default off*/
1645 if (ts->continue_on_error)
1646 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1648 /* Additional output if description is set */
1649 if (strlen(ts->description)) {
1652 log_buf(out, ";%s", ts->description);
1658 static void json_add_job_opts(struct json_object *root, const char *name,
1659 struct flist_head *opt_list)
1661 struct json_object *dir_object;
1662 struct flist_head *entry;
1663 struct print_option *p;
1665 if (flist_empty(opt_list))
1668 dir_object = json_create_object();
1669 json_object_add_value_object(root, name, dir_object);
1671 flist_for_each(entry, opt_list) {
1672 p = flist_entry(entry, struct print_option, list);
1673 json_object_add_value_string(dir_object, p->name, p->value);
1677 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1678 struct group_run_stats *rs,
1679 struct flist_head *opt_list)
1681 struct json_object *root, *tmp;
1682 struct jobs_eta *je;
1683 double io_u_dist[FIO_IO_U_MAP_NR];
1684 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1685 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1686 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1687 double usr_cpu, sys_cpu;
1691 root = json_create_object();
1692 json_object_add_value_string(root, "jobname", ts->name);
1693 json_object_add_value_int(root, "groupid", ts->groupid);
1694 json_object_add_value_int(root, "error", ts->error);
1697 je = get_jobs_eta(true, &size);
1699 json_object_add_value_int(root, "eta", je->eta_sec);
1700 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1704 json_add_job_opts(root, "job options", opt_list);
1706 add_ddir_status_json(ts, rs, DDIR_READ, root);
1707 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1708 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1709 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1711 if (ts->unified_rw_rep == UNIFIED_BOTH)
1712 add_mixed_ddir_status_json(ts, rs, root);
1715 if (ts->total_run_time) {
1716 double runt = (double) ts->total_run_time;
1718 usr_cpu = (double) ts->usr_time * 100 / runt;
1719 sys_cpu = (double) ts->sys_time * 100 / runt;
1724 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1725 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1726 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1727 json_object_add_value_int(root, "ctx", ts->ctx);
1728 json_object_add_value_int(root, "majf", ts->majf);
1729 json_object_add_value_int(root, "minf", ts->minf);
1731 /* Calc % distribution of IO depths */
1732 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1733 tmp = json_create_object();
1734 json_object_add_value_object(root, "iodepth_level", tmp);
1735 /* Only show fixed 7 I/O depth levels*/
1736 for (i = 0; i < 7; i++) {
1739 snprintf(name, 20, "%d", 1 << i);
1741 snprintf(name, 20, ">=%d", 1 << i);
1742 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1745 /* Calc % distribution of submit IO depths */
1746 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1747 tmp = json_create_object();
1748 json_object_add_value_object(root, "iodepth_submit", tmp);
1749 /* Only show fixed 7 I/O depth levels*/
1750 for (i = 0; i < 7; i++) {
1753 snprintf(name, 20, "0");
1755 snprintf(name, 20, "%d", 1 << (i+1));
1757 snprintf(name, 20, ">=%d", 1 << i);
1758 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1761 /* Calc % distribution of completion IO depths */
1762 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1763 tmp = json_create_object();
1764 json_object_add_value_object(root, "iodepth_complete", tmp);
1765 /* Only show fixed 7 I/O depth levels*/
1766 for (i = 0; i < 7; i++) {
1769 snprintf(name, 20, "0");
1771 snprintf(name, 20, "%d", 1 << (i+1));
1773 snprintf(name, 20, ">=%d", 1 << i);
1774 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1777 /* Calc % distribution of nsecond, usecond, msecond latency */
1778 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1779 stat_calc_lat_n(ts, io_u_lat_n);
1780 stat_calc_lat_u(ts, io_u_lat_u);
1781 stat_calc_lat_m(ts, io_u_lat_m);
1783 /* Nanosecond latency */
1784 tmp = json_create_object();
1785 json_object_add_value_object(root, "latency_ns", tmp);
1786 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1787 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1788 "250", "500", "750", "1000", };
1789 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1791 /* Microsecond latency */
1792 tmp = json_create_object();
1793 json_object_add_value_object(root, "latency_us", tmp);
1794 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1795 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1796 "250", "500", "750", "1000", };
1797 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1799 /* Millisecond latency */
1800 tmp = json_create_object();
1801 json_object_add_value_object(root, "latency_ms", tmp);
1802 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1803 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1804 "250", "500", "750", "1000", "2000",
1806 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1809 /* Additional output if continue_on_error set - default off*/
1810 if (ts->continue_on_error) {
1811 json_object_add_value_int(root, "total_err", ts->total_err_count);
1812 json_object_add_value_int(root, "first_error", ts->first_error);
1815 if (ts->latency_depth) {
1816 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1817 json_object_add_value_int(root, "latency_target", ts->latency_target);
1818 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1819 json_object_add_value_int(root, "latency_window", ts->latency_window);
1822 /* Additional output if description is set */
1823 if (strlen(ts->description))
1824 json_object_add_value_string(root, "desc", ts->description);
1826 if (ts->nr_block_infos) {
1827 /* Block error histogram and types */
1829 unsigned int *percentiles = NULL;
1830 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1832 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1833 ts->percentile_list,
1834 &percentiles, block_state_counts);
1837 struct json_object *block, *percentile_object, *states;
1839 block = json_create_object();
1840 json_object_add_value_object(root, "block", block);
1842 percentile_object = json_create_object();
1843 json_object_add_value_object(block, "percentiles",
1845 for (i = 0; i < len; i++) {
1847 snprintf(buf, sizeof(buf), "%f",
1848 ts->percentile_list[i].u.f);
1849 json_object_add_value_int(percentile_object,
1854 states = json_create_object();
1855 json_object_add_value_object(block, "states", states);
1856 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1857 json_object_add_value_int(states,
1858 block_state_names[state],
1859 block_state_counts[state]);
1866 struct json_object *data;
1867 struct json_array *iops, *bw;
1871 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1872 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1873 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1874 (float) ts->ss_limit.u.f,
1875 ts->ss_state & FIO_SS_PCT ? "%" : "");
1877 tmp = json_create_object();
1878 json_object_add_value_object(root, "steadystate", tmp);
1879 json_object_add_value_string(tmp, "ss", ss_buf);
1880 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1881 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1883 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1884 ts->ss_state & FIO_SS_PCT ? "%" : "");
1885 json_object_add_value_string(tmp, "criterion", ss_buf);
1886 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1887 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1889 data = json_create_object();
1890 json_object_add_value_object(tmp, "data", data);
1891 bw = json_create_array();
1892 iops = json_create_array();
1895 ** if ss was attained or the buffer is not full,
1896 ** ss->head points to the first element in the list.
1897 ** otherwise it actually points to the second element
1900 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1903 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1904 for (l = 0; l < ts->ss_dur; l++) {
1905 k = (j + l) % ts->ss_dur;
1906 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1907 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1909 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1910 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1911 json_object_add_value_array(data, "iops", iops);
1912 json_object_add_value_array(data, "bw", bw);
1918 static void show_thread_status_terse(struct thread_stat *ts,
1919 struct group_run_stats *rs,
1920 struct buf_output *out)
1922 if (terse_version >= 2 && terse_version <= 5)
1923 show_thread_status_terse_all(ts, rs, terse_version, out);
1925 log_err("fio: bad terse version!? %d\n", terse_version);
1928 struct json_object *show_thread_status(struct thread_stat *ts,
1929 struct group_run_stats *rs,
1930 struct flist_head *opt_list,
1931 struct buf_output *out)
1933 struct json_object *ret = NULL;
1935 if (output_format & FIO_OUTPUT_TERSE)
1936 show_thread_status_terse(ts, rs, out);
1937 if (output_format & FIO_OUTPUT_JSON)
1938 ret = show_thread_status_json(ts, rs, opt_list);
1939 if (output_format & FIO_OUTPUT_NORMAL)
1940 show_thread_status_normal(ts, rs, out);
1945 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1949 dst->min_val = min(dst->min_val, src->min_val);
1950 dst->max_val = max(dst->max_val, src->max_val);
1953 * Compute new mean and S after the merge
1954 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1955 * #Parallel_algorithm>
1958 mean = src->mean.u.f;
1961 double delta = src->mean.u.f - dst->mean.u.f;
1963 mean = ((src->mean.u.f * src->samples) +
1964 (dst->mean.u.f * dst->samples)) /
1965 (dst->samples + src->samples);
1967 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1968 (dst->samples * src->samples) /
1969 (dst->samples + src->samples);
1972 dst->samples += src->samples;
1973 dst->mean.u.f = mean;
1979 * We sum two kinds of stats - one that is time based, in which case we
1980 * apply the proper summing technique, and then one that is iops/bw
1981 * numbers. For group_reporting, we should just add those up, not make
1982 * them the mean of everything.
1984 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
1986 bool first = dst->samples == 0;
1988 if (src->samples == 0)
1992 __sum_stat(dst, src, first);
1997 dst->min_val = src->min_val;
1998 dst->max_val = src->max_val;
1999 dst->samples = src->samples;
2000 dst->mean.u.f = src->mean.u.f;
2001 dst->S.u.f = src->S.u.f;
2003 dst->min_val += src->min_val;
2004 dst->max_val += src->max_val;
2005 dst->samples += src->samples;
2006 dst->mean.u.f += src->mean.u.f;
2007 dst->S.u.f += src->S.u.f;
2011 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
2015 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2016 if (dst->max_run[i] < src->max_run[i])
2017 dst->max_run[i] = src->max_run[i];
2018 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
2019 dst->min_run[i] = src->min_run[i];
2020 if (dst->max_bw[i] < src->max_bw[i])
2021 dst->max_bw[i] = src->max_bw[i];
2022 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
2023 dst->min_bw[i] = src->min_bw[i];
2025 dst->iobytes[i] += src->iobytes[i];
2026 dst->agg[i] += src->agg[i];
2030 dst->kb_base = src->kb_base;
2031 if (!dst->unit_base)
2032 dst->unit_base = src->unit_base;
2034 dst->sig_figs = src->sig_figs;
2038 * Free the clat_prio_stat arrays allocated by alloc_clat_prio_stat_ddir().
2040 void free_clat_prio_stats(struct thread_stat *ts)
2047 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2048 sfree(ts->clat_prio[ddir]);
2049 ts->clat_prio[ddir] = NULL;
2050 ts->nr_clat_prio[ddir] = 0;
2055 * Allocate a clat_prio_stat array. The array has to be allocated/freed using
2056 * smalloc/sfree, so that it is accessible by the process/thread summing the
2059 int alloc_clat_prio_stat_ddir(struct thread_stat *ts, enum fio_ddir ddir,
2062 struct clat_prio_stat *clat_prio;
2065 clat_prio = scalloc(nr_prios, sizeof(*ts->clat_prio[ddir]));
2067 log_err("fio: failed to allocate ts clat data\n");
2071 for (i = 0; i < nr_prios; i++)
2072 clat_prio[i].clat_stat.min_val = ULONG_MAX;
2074 ts->clat_prio[ddir] = clat_prio;
2075 ts->nr_clat_prio[ddir] = nr_prios;
2080 static int grow_clat_prio_stat(struct thread_stat *dst, enum fio_ddir ddir)
2082 int curr_len = dst->nr_clat_prio[ddir];
2085 new_arr = scalloc(curr_len + 1, sizeof(*dst->clat_prio[ddir]));
2087 log_err("fio: failed to grow clat prio array\n");
2091 memcpy(new_arr, dst->clat_prio[ddir],
2092 curr_len * sizeof(*dst->clat_prio[ddir]));
2093 sfree(dst->clat_prio[ddir]);
2095 dst->clat_prio[ddir] = new_arr;
2096 dst->clat_prio[ddir][curr_len].clat_stat.min_val = ULONG_MAX;
2097 dst->nr_clat_prio[ddir]++;
2102 static int find_clat_prio_index(struct thread_stat *dst, enum fio_ddir ddir,
2105 int i, nr_prios = dst->nr_clat_prio[ddir];
2107 for (i = 0; i < nr_prios; i++) {
2108 if (dst->clat_prio[ddir][i].ioprio == ioprio)
2115 static int alloc_or_get_clat_prio_index(struct thread_stat *dst,
2116 enum fio_ddir ddir, uint32_t ioprio,
2119 int index = find_clat_prio_index(dst, ddir, ioprio);
2122 index = dst->nr_clat_prio[ddir];
2124 if (grow_clat_prio_stat(dst, ddir))
2127 dst->clat_prio[ddir][index].ioprio = ioprio;
2135 static int clat_prio_stats_copy(struct thread_stat *dst, struct thread_stat *src,
2136 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2138 size_t sz = sizeof(*src->clat_prio[src_ddir]) *
2139 src->nr_clat_prio[src_ddir];
2141 dst->clat_prio[dst_ddir] = smalloc(sz);
2142 if (!dst->clat_prio[dst_ddir]) {
2143 log_err("fio: failed to alloc clat prio array\n");
2147 memcpy(dst->clat_prio[dst_ddir], src->clat_prio[src_ddir], sz);
2148 dst->nr_clat_prio[dst_ddir] = src->nr_clat_prio[src_ddir];
2153 static int clat_prio_stat_add_samples(struct thread_stat *dst,
2154 enum fio_ddir dst_ddir, uint32_t ioprio,
2155 struct io_stat *io_stat,
2156 uint64_t *io_u_plat)
2160 if (!io_stat->samples)
2163 if (alloc_or_get_clat_prio_index(dst, dst_ddir, ioprio, &dst_index))
2166 sum_stat(&dst->clat_prio[dst_ddir][dst_index].clat_stat, io_stat,
2169 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
2170 dst->clat_prio[dst_ddir][dst_index].io_u_plat[i] += io_u_plat[i];
2175 static int sum_clat_prio_stats_src_single_prio(struct thread_stat *dst,
2176 struct thread_stat *src,
2177 enum fio_ddir dst_ddir,
2178 enum fio_ddir src_ddir)
2180 struct io_stat *io_stat;
2181 uint64_t *io_u_plat;
2184 * If src ts has no clat_prio_stat array, then all I/Os were submitted
2185 * using src->ioprio. Thus, the global samples in src->clat_stat (or
2186 * src->lat_stat) can be used as the 'per prio' samples for src->ioprio.
2188 assert(!src->clat_prio[src_ddir]);
2189 assert(src->nr_clat_prio[src_ddir] == 0);
2191 if (src->lat_percentiles) {
2192 io_u_plat = src->io_u_plat[FIO_LAT][src_ddir];
2193 io_stat = &src->lat_stat[src_ddir];
2195 io_u_plat = src->io_u_plat[FIO_CLAT][src_ddir];
2196 io_stat = &src->clat_stat[src_ddir];
2199 return clat_prio_stat_add_samples(dst, dst_ddir, src->ioprio, io_stat,
2203 static int sum_clat_prio_stats_src_multi_prio(struct thread_stat *dst,
2204 struct thread_stat *src,
2205 enum fio_ddir dst_ddir,
2206 enum fio_ddir src_ddir)
2211 * If src ts has a clat_prio_stat array, then there are multiple prios
2212 * in use (i.e. src ts had cmdprio_percentage or cmdprio_bssplit set).
2213 * The samples for the default prio will exist in the src->clat_prio
2214 * array, just like the samples for any other prio.
2216 assert(src->clat_prio[src_ddir]);
2217 assert(src->nr_clat_prio[src_ddir]);
2219 /* If the dst ts doesn't yet have a clat_prio array, simply memcpy. */
2220 if (!dst->clat_prio[dst_ddir])
2221 return clat_prio_stats_copy(dst, src, dst_ddir, src_ddir);
2223 /* The dst ts already has a clat_prio_array, add src stats into it. */
2224 for (i = 0; i < src->nr_clat_prio[src_ddir]; i++) {
2225 struct io_stat *io_stat = &src->clat_prio[src_ddir][i].clat_stat;
2226 uint64_t *io_u_plat = src->clat_prio[src_ddir][i].io_u_plat;
2227 uint32_t ioprio = src->clat_prio[src_ddir][i].ioprio;
2229 if (clat_prio_stat_add_samples(dst, dst_ddir, ioprio, io_stat, io_u_plat))
2236 static int sum_clat_prio_stats(struct thread_stat *dst, struct thread_stat *src,
2237 enum fio_ddir dst_ddir, enum fio_ddir src_ddir)
2239 if (dst->disable_prio_stat)
2242 if (!src->clat_prio[src_ddir])
2243 return sum_clat_prio_stats_src_single_prio(dst, src, dst_ddir,
2246 return sum_clat_prio_stats_src_multi_prio(dst, src, dst_ddir, src_ddir);
2249 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2253 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2254 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2255 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2256 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2257 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2258 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2259 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2260 sum_clat_prio_stats(dst, src, l, l);
2262 dst->io_bytes[l] += src->io_bytes[l];
2264 if (dst->runtime[l] < src->runtime[l])
2265 dst->runtime[l] = src->runtime[l];
2267 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2268 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2269 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2270 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2271 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2272 sum_clat_prio_stats(dst, src, 0, l);
2274 dst->io_bytes[0] += src->io_bytes[l];
2276 if (dst->runtime[0] < src->runtime[l])
2277 dst->runtime[0] = src->runtime[l];
2281 sum_stat(&dst->sync_stat, &src->sync_stat, false);
2282 dst->usr_time += src->usr_time;
2283 dst->sys_time += src->sys_time;
2284 dst->ctx += src->ctx;
2285 dst->majf += src->majf;
2286 dst->minf += src->minf;
2288 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2289 dst->io_u_map[k] += src->io_u_map[k];
2290 dst->io_u_submit[k] += src->io_u_submit[k];
2291 dst->io_u_complete[k] += src->io_u_complete[k];
2294 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2295 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2296 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2297 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2298 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2299 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2301 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2302 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2303 dst->total_io_u[k] += src->total_io_u[k];
2304 dst->short_io_u[k] += src->short_io_u[k];
2305 dst->drop_io_u[k] += src->drop_io_u[k];
2307 dst->total_io_u[0] += src->total_io_u[k];
2308 dst->short_io_u[0] += src->short_io_u[k];
2309 dst->drop_io_u[0] += src->drop_io_u[k];
2313 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2315 for (k = 0; k < FIO_LAT_CNT; k++)
2316 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2317 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2318 if (dst->unified_rw_rep != UNIFIED_MIXED)
2319 dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2321 dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2323 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2324 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2326 dst->total_run_time += src->total_run_time;
2327 dst->total_submit += src->total_submit;
2328 dst->total_complete += src->total_complete;
2329 dst->nr_zone_resets += src->nr_zone_resets;
2330 dst->cachehit += src->cachehit;
2331 dst->cachemiss += src->cachemiss;
2334 void init_group_run_stat(struct group_run_stats *gs)
2337 memset(gs, 0, sizeof(*gs));
2339 for (i = 0; i < DDIR_RWDIR_CNT; i++)
2340 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2343 void init_thread_stat_min_vals(struct thread_stat *ts)
2347 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2348 ts->clat_stat[i].min_val = ULONG_MAX;
2349 ts->slat_stat[i].min_val = ULONG_MAX;
2350 ts->lat_stat[i].min_val = ULONG_MAX;
2351 ts->bw_stat[i].min_val = ULONG_MAX;
2352 ts->iops_stat[i].min_val = ULONG_MAX;
2354 ts->sync_stat.min_val = ULONG_MAX;
2357 void init_thread_stat(struct thread_stat *ts)
2359 memset(ts, 0, sizeof(*ts));
2361 init_thread_stat_min_vals(ts);
2365 static void init_per_prio_stats(struct thread_stat *threadstats, int nr_ts)
2367 struct thread_data *td;
2368 struct thread_stat *ts;
2369 int i, j, last_ts, idx;
2377 * Loop through all tds, if a td requires per prio stats, temporarily
2378 * store a 1 in ts->disable_prio_stat, and then do an additional
2379 * loop at the end where we invert the ts->disable_prio_stat values.
2381 for_each_td(td, i) {
2385 (!td->o.group_reporting ||
2386 (td->o.group_reporting && last_ts != td->groupid))) {
2391 last_ts = td->groupid;
2392 ts = &threadstats[j];
2394 /* idx == 0 means first td in group, or td is not in a group. */
2396 ts->ioprio = td->ioprio;
2397 else if (td->ioprio != ts->ioprio)
2398 ts->disable_prio_stat = 1;
2400 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2401 if (td->ts.clat_prio[ddir]) {
2402 ts->disable_prio_stat = 1;
2410 /* Loop through all dst threadstats and fixup the values. */
2411 for (i = 0; i < nr_ts; i++) {
2412 ts = &threadstats[i];
2413 ts->disable_prio_stat = !ts->disable_prio_stat;
2417 void __show_run_stats(void)
2419 struct group_run_stats *runstats, *rs;
2420 struct thread_data *td;
2421 struct thread_stat *threadstats, *ts;
2422 int i, j, k, nr_ts, last_ts, idx;
2423 bool kb_base_warned = false;
2424 bool unit_base_warned = false;
2425 struct json_object *root = NULL;
2426 struct json_array *array = NULL;
2427 struct buf_output output[FIO_OUTPUT_NR];
2428 struct flist_head **opt_lists;
2430 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2432 for (i = 0; i < groupid + 1; i++)
2433 init_group_run_stat(&runstats[i]);
2436 * find out how many threads stats we need. if group reporting isn't
2437 * enabled, it's one-per-td.
2441 for_each_td(td, i) {
2442 if (!td->o.group_reporting) {
2446 if (last_ts == td->groupid)
2451 last_ts = td->groupid;
2455 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2456 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2458 for (i = 0; i < nr_ts; i++) {
2459 init_thread_stat(&threadstats[i]);
2460 opt_lists[i] = NULL;
2463 init_per_prio_stats(threadstats, nr_ts);
2468 for_each_td(td, i) {
2471 if (idx && (!td->o.group_reporting ||
2472 (td->o.group_reporting && last_ts != td->groupid))) {
2477 last_ts = td->groupid;
2479 ts = &threadstats[j];
2481 ts->clat_percentiles = td->o.clat_percentiles;
2482 ts->lat_percentiles = td->o.lat_percentiles;
2483 ts->slat_percentiles = td->o.slat_percentiles;
2484 ts->percentile_precision = td->o.percentile_precision;
2485 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2486 opt_lists[j] = &td->opt_list;
2490 if (ts->groupid == -1) {
2492 * These are per-group shared already
2494 snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2495 if (td->o.description)
2496 snprintf(ts->description,
2497 sizeof(ts->description), "%s",
2500 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2503 * If multiple entries in this group, this is
2506 ts->thread_number = td->thread_number;
2507 ts->groupid = td->groupid;
2510 * first pid in group, not very useful...
2514 ts->kb_base = td->o.kb_base;
2515 ts->unit_base = td->o.unit_base;
2516 ts->sig_figs = td->o.sig_figs;
2517 ts->unified_rw_rep = td->o.unified_rw_rep;
2518 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2519 log_info("fio: kb_base differs for jobs in group, using"
2520 " %u as the base\n", ts->kb_base);
2521 kb_base_warned = true;
2522 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2523 log_info("fio: unit_base differs for jobs in group, using"
2524 " %u as the base\n", ts->unit_base);
2525 unit_base_warned = true;
2528 ts->continue_on_error = td->o.continue_on_error;
2529 ts->total_err_count += td->total_err_count;
2530 ts->first_error = td->first_error;
2532 if (!td->error && td->o.continue_on_error &&
2534 ts->error = td->first_error;
2535 snprintf(ts->verror, sizeof(ts->verror), "%s",
2537 } else if (td->error) {
2538 ts->error = td->error;
2539 snprintf(ts->verror, sizeof(ts->verror), "%s",
2544 ts->latency_depth = td->latency_qd;
2545 ts->latency_target = td->o.latency_target;
2546 ts->latency_percentile = td->o.latency_percentile;
2547 ts->latency_window = td->o.latency_window;
2549 ts->nr_block_infos = td->ts.nr_block_infos;
2550 for (k = 0; k < ts->nr_block_infos; k++)
2551 ts->block_infos[k] = td->ts.block_infos[k];
2553 sum_thread_stats(ts, &td->ts);
2558 ts->ss_state = td->ss.state;
2559 ts->ss_dur = td->ss.dur;
2560 ts->ss_head = td->ss.head;
2561 ts->ss_bw_data = td->ss.bw_data;
2562 ts->ss_iops_data = td->ss.iops_data;
2563 ts->ss_limit.u.f = td->ss.limit;
2564 ts->ss_slope.u.f = td->ss.slope;
2565 ts->ss_deviation.u.f = td->ss.deviation;
2566 ts->ss_criterion.u.f = td->ss.criterion;
2569 ts->ss_dur = ts->ss_state = 0;
2572 for (i = 0; i < nr_ts; i++) {
2573 unsigned long long bw;
2575 ts = &threadstats[i];
2576 if (ts->groupid == -1)
2578 rs = &runstats[ts->groupid];
2579 rs->kb_base = ts->kb_base;
2580 rs->unit_base = ts->unit_base;
2581 rs->sig_figs = ts->sig_figs;
2582 rs->unified_rw_rep |= ts->unified_rw_rep;
2584 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2585 if (!ts->runtime[j])
2587 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2588 rs->min_run[j] = ts->runtime[j];
2589 if (ts->runtime[j] > rs->max_run[j])
2590 rs->max_run[j] = ts->runtime[j];
2594 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2595 if (bw < rs->min_bw[j])
2597 if (bw > rs->max_bw[j])
2600 rs->iobytes[j] += ts->io_bytes[j];
2604 for (i = 0; i < groupid + 1; i++) {
2609 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2610 if (rs->max_run[ddir])
2611 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2616 for (i = 0; i < FIO_OUTPUT_NR; i++)
2617 buf_output_init(&output[i]);
2620 * don't overwrite last signal output
2622 if (output_format & FIO_OUTPUT_NORMAL)
2623 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2624 if (output_format & FIO_OUTPUT_JSON) {
2625 struct thread_data *global;
2628 unsigned long long ms_since_epoch;
2631 gettimeofday(&now, NULL);
2632 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2633 (unsigned long long)(now.tv_usec) / 1000;
2635 tv_sec = now.tv_sec;
2636 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2637 if (time_buf[strlen(time_buf) - 1] == '\n')
2638 time_buf[strlen(time_buf) - 1] = '\0';
2640 root = json_create_object();
2641 json_object_add_value_string(root, "fio version", fio_version_string);
2642 json_object_add_value_int(root, "timestamp", now.tv_sec);
2643 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2644 json_object_add_value_string(root, "time", time_buf);
2645 global = get_global_options();
2646 json_add_job_opts(root, "global options", &global->opt_list);
2647 array = json_create_array();
2648 json_object_add_value_array(root, "jobs", array);
2652 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2654 for (i = 0; i < nr_ts; i++) {
2655 ts = &threadstats[i];
2656 rs = &runstats[ts->groupid];
2659 fio_server_send_job_options(opt_lists[i], i);
2660 fio_server_send_ts(ts, rs);
2662 if (output_format & FIO_OUTPUT_TERSE)
2663 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2664 if (output_format & FIO_OUTPUT_JSON) {
2665 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2666 json_array_add_value_object(array, tmp);
2668 if (output_format & FIO_OUTPUT_NORMAL)
2669 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2672 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2673 /* disk util stats, if any */
2674 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2676 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2678 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2679 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2680 json_free_object(root);
2683 for (i = 0; i < groupid + 1; i++) {
2688 fio_server_send_gs(rs);
2689 else if (output_format & FIO_OUTPUT_NORMAL)
2690 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2694 fio_server_send_du();
2695 else if (output_format & FIO_OUTPUT_NORMAL) {
2696 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2697 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2700 for (i = 0; i < FIO_OUTPUT_NR; i++) {
2701 struct buf_output *out = &output[i];
2703 log_info_buf(out->buf, out->buflen);
2704 buf_output_free(out);
2707 fio_idle_prof_cleanup();
2712 /* free arrays allocated by sum_thread_stats(), if any */
2713 for (i = 0; i < nr_ts; i++) {
2714 ts = &threadstats[i];
2715 free_clat_prio_stats(ts);
2721 int __show_running_run_stats(void)
2723 struct thread_data *td;
2724 unsigned long long *rt;
2728 fio_sem_down(stat_sem);
2730 rt = malloc(thread_number * sizeof(unsigned long long));
2731 fio_gettime(&ts, NULL);
2733 for_each_td(td, i) {
2734 td->update_rusage = 1;
2735 for_each_rw_ddir(ddir) {
2736 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2738 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2740 rt[i] = mtime_since(&td->start, &ts);
2741 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2742 td->ts.runtime[DDIR_READ] += rt[i];
2743 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2744 td->ts.runtime[DDIR_WRITE] += rt[i];
2745 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2746 td->ts.runtime[DDIR_TRIM] += rt[i];
2749 for_each_td(td, i) {
2750 if (td->runstate >= TD_EXITED)
2752 if (td->rusage_sem) {
2753 td->update_rusage = 1;
2754 fio_sem_down(td->rusage_sem);
2756 td->update_rusage = 0;
2761 for_each_td(td, i) {
2762 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2763 td->ts.runtime[DDIR_READ] -= rt[i];
2764 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2765 td->ts.runtime[DDIR_WRITE] -= rt[i];
2766 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2767 td->ts.runtime[DDIR_TRIM] -= rt[i];
2771 fio_sem_up(stat_sem);
2776 static bool status_file_disabled;
2778 #define FIO_STATUS_FILE "fio-dump-status"
2780 static int check_status_file(void)
2783 const char *temp_dir;
2784 char fio_status_file_path[PATH_MAX];
2786 if (status_file_disabled)
2789 temp_dir = getenv("TMPDIR");
2790 if (temp_dir == NULL) {
2791 temp_dir = getenv("TEMP");
2792 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2795 if (temp_dir == NULL)
2798 __coverity_tainted_data_sanitize__(temp_dir);
2801 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2803 if (stat(fio_status_file_path, &sb))
2806 if (unlink(fio_status_file_path) < 0) {
2807 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2809 log_err("fio: disabling status file updates\n");
2810 status_file_disabled = true;
2816 void check_for_running_stats(void)
2818 if (check_status_file()) {
2819 show_running_run_stats();
2824 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2829 if (data > is->max_val)
2831 if (data < is->min_val)
2834 delta = val - is->mean.u.f;
2836 is->mean.u.f += delta / (is->samples + 1.0);
2837 is->S.u.f += delta * (val - is->mean.u.f);
2843 static inline void add_stat_prio_sample(struct clat_prio_stat *clat_prio,
2844 unsigned short clat_prio_index,
2845 unsigned long long nsec)
2848 add_stat_sample(&clat_prio[clat_prio_index].clat_stat, nsec);
2852 * Return a struct io_logs, which is added to the tail of the log
2855 static struct io_logs *get_new_log(struct io_log *iolog)
2858 struct io_logs *cur_log;
2861 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2864 if (!iolog->cur_log_max) {
2865 new_samples = iolog->td->o.log_entries;
2867 new_samples = iolog->cur_log_max * 2;
2868 if (new_samples > MAX_LOG_ENTRIES)
2869 new_samples = MAX_LOG_ENTRIES;
2872 cur_log = smalloc(sizeof(*cur_log));
2874 INIT_FLIST_HEAD(&cur_log->list);
2875 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2877 cur_log->nr_samples = 0;
2878 cur_log->max_samples = new_samples;
2879 flist_add_tail(&cur_log->list, &iolog->io_logs);
2880 iolog->cur_log_max = new_samples;
2890 * Add and return a new log chunk, or return current log if big enough
2892 static struct io_logs *regrow_log(struct io_log *iolog)
2894 struct io_logs *cur_log;
2897 if (!iolog || iolog->disabled)
2900 cur_log = iolog_cur_log(iolog);
2902 cur_log = get_new_log(iolog);
2907 if (cur_log->nr_samples < cur_log->max_samples)
2911 * No room for a new sample. If we're compressing on the fly, flush
2912 * out the current chunk
2914 if (iolog->log_gz) {
2915 if (iolog_cur_flush(iolog, cur_log)) {
2916 log_err("fio: failed flushing iolog! Will stop logging.\n");
2922 * Get a new log array, and add to our list
2924 cur_log = get_new_log(iolog);
2926 log_err("fio: failed extending iolog! Will stop logging.\n");
2930 if (!iolog->pending || !iolog->pending->nr_samples)
2934 * Flush pending items to new log
2936 for (i = 0; i < iolog->pending->nr_samples; i++) {
2937 struct io_sample *src, *dst;
2939 src = get_sample(iolog, iolog->pending, i);
2940 dst = get_sample(iolog, cur_log, i);
2941 memcpy(dst, src, log_entry_sz(iolog));
2943 cur_log->nr_samples = iolog->pending->nr_samples;
2945 iolog->pending->nr_samples = 0;
2949 iolog->disabled = true;
2953 void regrow_logs(struct thread_data *td)
2955 regrow_log(td->slat_log);
2956 regrow_log(td->clat_log);
2957 regrow_log(td->clat_hist_log);
2958 regrow_log(td->lat_log);
2959 regrow_log(td->bw_log);
2960 regrow_log(td->iops_log);
2961 td->flags &= ~TD_F_REGROW_LOGS;
2964 void regrow_agg_logs(void)
2968 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2969 regrow_log(agg_io_log[ddir]);
2972 static struct io_logs *get_cur_log(struct io_log *iolog)
2974 struct io_logs *cur_log;
2976 cur_log = iolog_cur_log(iolog);
2978 cur_log = get_new_log(iolog);
2983 if (cur_log->nr_samples < cur_log->max_samples)
2987 * Out of space. If we're in IO offload mode, or we're not doing
2988 * per unit logging (hence logging happens outside of the IO thread
2989 * as well), add a new log chunk inline. If we're doing inline
2990 * submissions, flag 'td' as needing a log regrow and we'll take
2991 * care of it on the submission side.
2993 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2994 !per_unit_log(iolog))
2995 return regrow_log(iolog);
2998 iolog->td->flags |= TD_F_REGROW_LOGS;
3000 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
3001 return iolog->pending;
3004 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
3005 enum fio_ddir ddir, unsigned long long bs,
3006 unsigned long t, uint64_t offset,
3007 unsigned int priority)
3009 struct io_logs *cur_log;
3011 if (iolog->disabled)
3013 if (flist_empty(&iolog->io_logs))
3014 iolog->avg_last[ddir] = t;
3016 cur_log = get_cur_log(iolog);
3018 struct io_sample *s;
3020 s = get_sample(iolog, cur_log, cur_log->nr_samples);
3023 s->time = t + (iolog->td ? iolog->td->alternate_epoch : 0);
3024 io_sample_set_ddir(iolog, s, ddir);
3026 s->priority = priority;
3028 if (iolog->log_offset) {
3029 struct io_sample_offset *so = (void *) s;
3031 so->offset = offset;
3034 cur_log->nr_samples++;
3038 iolog->disabled = true;
3041 static inline void reset_io_stat(struct io_stat *ios)
3043 ios->min_val = -1ULL;
3044 ios->max_val = ios->samples = 0;
3045 ios->mean.u.f = ios->S.u.f = 0;
3048 static inline void reset_io_u_plat(uint64_t *io_u_plat)
3052 for (i = 0; i < FIO_IO_U_PLAT_NR; i++)
3056 static inline void reset_clat_prio_stats(struct thread_stat *ts)
3061 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3062 if (!ts->clat_prio[ddir])
3065 for (i = 0; i < ts->nr_clat_prio[ddir]; i++) {
3066 reset_io_stat(&ts->clat_prio[ddir][i].clat_stat);
3067 reset_io_u_plat(ts->clat_prio[ddir][i].io_u_plat);
3072 void reset_io_stats(struct thread_data *td)
3074 struct thread_stat *ts = &td->ts;
3077 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
3078 reset_io_stat(&ts->clat_stat[i]);
3079 reset_io_stat(&ts->slat_stat[i]);
3080 reset_io_stat(&ts->lat_stat[i]);
3081 reset_io_stat(&ts->bw_stat[i]);
3082 reset_io_stat(&ts->iops_stat[i]);
3084 ts->io_bytes[i] = 0;
3086 ts->total_io_u[i] = 0;
3087 ts->short_io_u[i] = 0;
3088 ts->drop_io_u[i] = 0;
3091 for (i = 0; i < FIO_LAT_CNT; i++)
3092 for (j = 0; j < DDIR_RWDIR_CNT; j++)
3093 reset_io_u_plat(ts->io_u_plat[i][j]);
3095 reset_clat_prio_stats(ts);
3097 ts->total_io_u[DDIR_SYNC] = 0;
3098 reset_io_u_plat(ts->io_u_sync_plat);
3100 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
3101 ts->io_u_map[i] = 0;
3102 ts->io_u_submit[i] = 0;
3103 ts->io_u_complete[i] = 0;
3106 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
3107 ts->io_u_lat_n[i] = 0;
3108 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
3109 ts->io_u_lat_u[i] = 0;
3110 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
3111 ts->io_u_lat_m[i] = 0;
3113 ts->total_submit = 0;
3114 ts->total_complete = 0;
3115 ts->nr_zone_resets = 0;
3116 ts->cachehit = ts->cachemiss = 0;
3119 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
3120 unsigned long elapsed, bool log_max)
3123 * Note an entry in the log. Use the mean from the logged samples,
3124 * making sure to properly round up. Only write a log entry if we
3125 * had actual samples done.
3127 if (iolog->avg_window[ddir].samples) {
3128 union io_sample_data data;
3131 data.val = iolog->avg_window[ddir].max_val;
3133 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
3135 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
3138 reset_io_stat(&iolog->avg_window[ddir]);
3141 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
3146 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
3147 __add_stat_to_log(iolog, ddir, elapsed, log_max);
3150 static unsigned long add_log_sample(struct thread_data *td,
3151 struct io_log *iolog,
3152 union io_sample_data data,
3153 enum fio_ddir ddir, unsigned long long bs,
3154 uint64_t offset, unsigned int ioprio)
3156 unsigned long elapsed, this_window;
3161 elapsed = mtime_since_now(&td->epoch);
3164 * If no time averaging, just add the log sample.
3166 if (!iolog->avg_msec) {
3167 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
3173 * Add the sample. If the time period has passed, then
3174 * add that entry to the log and clear.
3176 add_stat_sample(&iolog->avg_window[ddir], data.val);
3179 * If period hasn't passed, adding the above sample is all we
3182 this_window = elapsed - iolog->avg_last[ddir];
3183 if (elapsed < iolog->avg_last[ddir])
3184 return iolog->avg_last[ddir] - elapsed;
3185 else if (this_window < iolog->avg_msec) {
3186 unsigned long diff = iolog->avg_msec - this_window;
3188 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
3192 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
3194 iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
3196 return iolog->avg_msec;
3199 void finalize_logs(struct thread_data *td, bool unit_logs)
3201 unsigned long elapsed;
3203 elapsed = mtime_since_now(&td->epoch);
3205 if (td->clat_log && unit_logs)
3206 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
3207 if (td->slat_log && unit_logs)
3208 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
3209 if (td->lat_log && unit_logs)
3210 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
3211 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
3212 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
3213 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
3214 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
3217 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
3218 unsigned long long bs)
3220 struct io_log *iolog;
3225 iolog = agg_io_log[ddir];
3226 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
3229 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
3231 unsigned int idx = plat_val_to_idx(nsec);
3232 assert(idx < FIO_IO_U_PLAT_NR);
3234 ts->io_u_sync_plat[idx]++;
3235 add_stat_sample(&ts->sync_stat, nsec);
3238 static inline void add_lat_percentile_sample(struct thread_stat *ts,
3239 unsigned long long nsec,
3243 unsigned int idx = plat_val_to_idx(nsec);
3244 assert(idx < FIO_IO_U_PLAT_NR);
3246 ts->io_u_plat[lat][ddir][idx]++;
3250 add_lat_percentile_prio_sample(struct thread_stat *ts, unsigned long long nsec,
3252 unsigned short clat_prio_index)
3254 unsigned int idx = plat_val_to_idx(nsec);
3256 if (ts->clat_prio[ddir])
3257 ts->clat_prio[ddir][clat_prio_index].io_u_plat[idx]++;
3260 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
3261 unsigned long long nsec, unsigned long long bs,
3262 uint64_t offset, unsigned int ioprio,
3263 unsigned short clat_prio_index)
3265 const bool needs_lock = td_async_processing(td);
3266 unsigned long elapsed, this_window;
3267 struct thread_stat *ts = &td->ts;
3268 struct io_log *iolog = td->clat_hist_log;
3273 add_stat_sample(&ts->clat_stat[ddir], nsec);
3276 * When lat_percentiles=1 (default 0), the reported per priority
3277 * percentiles and stats are used for describing total latency values,
3278 * even though the variable names themselves start with clat_.
3280 * Because of the above definition, add a prio stat sample only when
3281 * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3282 * when lat_percentiles=1.
3284 if (!ts->lat_percentiles)
3285 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3289 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3292 if (ts->clat_percentiles) {
3294 * Because of the above definition, add a prio lat percentile
3295 * sample only when lat_percentiles=0. add_lat_sample() will add
3296 * the prio lat percentile sample when lat_percentiles=1.
3298 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3299 if (!ts->lat_percentiles)
3300 add_lat_percentile_prio_sample(ts, nsec, ddir,
3304 if (iolog && iolog->hist_msec) {
3305 struct io_hist *hw = &iolog->hist_window[ddir];
3308 elapsed = mtime_since_now(&td->epoch);
3310 hw->hist_last = elapsed;
3311 this_window = elapsed - hw->hist_last;
3313 if (this_window >= iolog->hist_msec) {
3314 uint64_t *io_u_plat;
3315 struct io_u_plat_entry *dst;
3318 * Make a byte-for-byte copy of the latency histogram
3319 * stored in td->ts.io_u_plat[ddir], recording it in a
3320 * log sample. Note that the matching call to free() is
3321 * located in iolog.c after printing this sample to the
3324 io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3325 dst = malloc(sizeof(struct io_u_plat_entry));
3326 memcpy(&(dst->io_u_plat), io_u_plat,
3327 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3328 flist_add(&dst->list, &hw->list);
3329 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3330 elapsed, offset, ioprio);
3333 * Update the last time we recorded as being now, minus
3334 * any drift in time we encountered before actually
3335 * making the record.
3337 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3343 __td_io_u_unlock(td);
3346 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3347 unsigned long long nsec, unsigned long long bs,
3348 uint64_t offset, unsigned int ioprio)
3350 const bool needs_lock = td_async_processing(td);
3351 struct thread_stat *ts = &td->ts;
3359 add_stat_sample(&ts->slat_stat[ddir], nsec);
3362 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3365 if (ts->slat_percentiles)
3366 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3369 __td_io_u_unlock(td);
3372 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3373 unsigned long long nsec, unsigned long long bs,
3374 uint64_t offset, unsigned int ioprio,
3375 unsigned short clat_prio_index)
3377 const bool needs_lock = td_async_processing(td);
3378 struct thread_stat *ts = &td->ts;
3386 add_stat_sample(&ts->lat_stat[ddir], nsec);
3389 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3393 * When lat_percentiles=1 (default 0), the reported per priority
3394 * percentiles and stats are used for describing total latency values,
3395 * even though the variable names themselves start with clat_.
3397 * Because of the above definition, add a prio stat and prio lat
3398 * percentile sample only when lat_percentiles=1. add_clat_sample() will
3399 * add the prio stat and prio lat percentile sample when
3400 * lat_percentiles=0.
3402 if (ts->lat_percentiles) {
3403 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3404 add_lat_percentile_prio_sample(ts, nsec, ddir, clat_prio_index);
3405 add_stat_prio_sample(ts->clat_prio[ddir], clat_prio_index,
3409 __td_io_u_unlock(td);
3412 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3413 unsigned int bytes, unsigned long long spent)
3415 const bool needs_lock = td_async_processing(td);
3416 struct thread_stat *ts = &td->ts;
3420 rate = (unsigned long) (bytes * 1000000ULL / spent);
3427 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3430 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3431 bytes, io_u->offset, io_u->ioprio);
3433 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3436 __td_io_u_unlock(td);
3439 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3440 struct timespec *t, unsigned int avg_time,
3441 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3442 struct io_stat *stat, struct io_log *log,
3445 const bool needs_lock = td_async_processing(td);
3446 unsigned long spent, rate;
3448 unsigned long next, next_log;
3450 next_log = avg_time;
3452 spent = mtime_since(parent_tv, t);
3453 if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3454 return avg_time - spent;
3460 * Compute both read and write rates for the interval.
3462 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3465 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3467 continue; /* No entries for interval */
3471 rate = delta * 1000 / spent / 1024; /* KiB/s */
3473 rate = (delta * 1000) / spent;
3477 add_stat_sample(&stat[ddir], rate);
3480 unsigned long long bs = 0;
3482 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3483 bs = td->o.min_bs[ddir];
3485 next = add_log_sample(td, log, sample_val(rate), ddir,
3487 next_log = min(next_log, next);
3490 stat_io_bytes[ddir] = this_io_bytes[ddir];
3496 __td_io_u_unlock(td);
3498 if (spent <= avg_time)
3501 next = avg_time - (1 + spent - avg_time);
3503 return min(next, next_log);
3506 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3508 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3509 td->this_io_bytes, td->stat_io_bytes,
3510 td->ts.bw_stat, td->bw_log, true);
3513 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3516 const bool needs_lock = td_async_processing(td);
3517 struct thread_stat *ts = &td->ts;
3522 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3525 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3526 bytes, io_u->offset, io_u->ioprio);
3528 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3531 __td_io_u_unlock(td);
3534 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3536 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3537 td->this_io_blocks, td->stat_io_blocks,
3538 td->ts.iops_stat, td->iops_log, false);
3542 * Returns msecs to next event
3544 int calc_log_samples(void)
3546 struct thread_data *td;
3547 unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3548 struct timespec now;
3550 long elapsed_time = 0;
3552 fio_gettime(&now, NULL);
3554 for_each_td(td, i) {
3555 elapsed_time = mtime_since_now(&td->epoch);
3559 if (in_ramp_time(td) ||
3560 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3561 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3565 (td->bw_log && !per_unit_log(td->bw_log))) {
3566 tmp = add_bw_samples(td, &now);
3569 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3571 if (!td->iops_log ||
3572 (td->iops_log && !per_unit_log(td->iops_log))) {
3573 tmp = add_iops_samples(td, &now);
3576 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3583 /* if log_avg_msec_min has not been changed, set it to 0 */
3584 if (log_avg_msec_min == -1U)
3585 log_avg_msec_min = 0;
3587 if (log_avg_msec_min == 0)
3588 next_mod = elapsed_time;
3590 next_mod = elapsed_time % log_avg_msec_min;
3592 /* correction to keep the time on the log avg msec boundary */
3593 next = min(next, (log_avg_msec_min - next_mod));
3595 return next == ~0U ? 0 : next;
3598 void stat_init(void)
3600 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3603 void stat_exit(void)
3606 * When we have the mutex, we know out-of-band access to it
3609 fio_sem_down(stat_sem);
3610 fio_sem_remove(stat_sem);
3614 * Called from signal handler. Wake up status thread.
3616 void show_running_run_stats(void)
3621 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3623 /* Ignore io_u's which span multiple blocks--they will just get
3624 * inaccurate counts. */
3625 int idx = (io_u->offset - io_u->file->file_offset)
3626 / td->o.bs[DDIR_TRIM];
3627 uint32_t *info = &td->ts.block_infos[idx];
3628 assert(idx < td->ts.nr_block_infos);