9 #include "lib/ieee754.h"
11 #include "lib/getrusage.h"
14 #include "lib/output_buffer.h"
15 #include "helper_thread.h"
18 #include "oslib/asprintf.h"
21 #define LOG_MSEC_SLACK 2
23 #define LOG_MSEC_SLACK 1
26 struct fio_sem *stat_sem;
28 void clear_rusage_stat(struct thread_data *td)
30 struct thread_stat *ts = &td->ts;
32 fio_getrusage(&td->ru_start);
33 ts->usr_time = ts->sys_time = 0;
35 ts->minf = ts->majf = 0;
38 void update_rusage_stat(struct thread_data *td)
40 struct thread_stat *ts = &td->ts;
42 fio_getrusage(&td->ru_end);
43 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
44 &td->ru_end.ru_utime);
45 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
46 &td->ru_end.ru_stime);
47 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
48 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
49 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
50 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
52 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
56 * Given a latency, return the index of the corresponding bucket in
57 * the structure tracking percentiles.
59 * (1) find the group (and error bits) that the value (latency)
60 * belongs to by looking at its MSB. (2) find the bucket number in the
61 * group by looking at the index bits.
64 static unsigned int plat_val_to_idx(unsigned long long val)
66 unsigned int msb, error_bits, base, offset, idx;
68 /* Find MSB starting from bit 0 */
72 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
75 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
76 * all bits of the sample as index
78 if (msb <= FIO_IO_U_PLAT_BITS)
81 /* Compute the number of error bits to discard*/
82 error_bits = msb - FIO_IO_U_PLAT_BITS;
84 /* Compute the number of buckets before the group */
85 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
88 * Discard the error bits and apply the mask to find the
89 * index for the buckets in the group
91 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
93 /* Make sure the index does not exceed (array size - 1) */
94 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
95 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
101 * Convert the given index of the bucket array to the value
102 * represented by the bucket
104 static unsigned long long plat_idx_to_val(unsigned int idx)
106 unsigned int error_bits;
107 unsigned long long k, base;
109 assert(idx < FIO_IO_U_PLAT_NR);
111 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
112 * all bits of the sample as index */
113 if (idx < (FIO_IO_U_PLAT_VAL << 1))
116 /* Find the group and compute the minimum value of that group */
117 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
118 base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
120 /* Find its bucket number of the group */
121 k = idx % FIO_IO_U_PLAT_VAL;
123 /* Return the mean of the range of the bucket */
124 return base + ((k + 0.5) * (1 << error_bits));
127 static int double_cmp(const void *a, const void *b)
129 const fio_fp64_t fa = *(const fio_fp64_t *) a;
130 const fio_fp64_t fb = *(const fio_fp64_t *) b;
135 else if (fa.u.f < fb.u.f)
141 unsigned int calc_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
142 fio_fp64_t *plist, unsigned long long **output,
143 unsigned long long *maxv, unsigned long long *minv)
145 unsigned long long sum = 0;
146 unsigned int len, i, j = 0;
147 unsigned long long *ovals = NULL;
154 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
161 * Sort the percentile list. Note that it may already be sorted if
162 * we are using the default values, but since it's a short list this
163 * isn't a worry. Also note that this does not work for NaN values.
166 qsort(plist, len, sizeof(plist[0]), double_cmp);
168 ovals = malloc(len * sizeof(*ovals));
173 * Calculate bucket values, note down max and min values
176 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
178 while (sum >= ((long double) plist[j].u.f / 100.0 * nr)) {
179 assert(plist[j].u.f <= 100.0);
181 ovals[j] = plat_idx_to_val(i);
182 if (ovals[j] < *minv)
184 if (ovals[j] > *maxv)
187 is_last = (j == len - 1) != 0;
196 log_err("fio: error calculating latency percentiles\n");
203 * Find and display the p-th percentile of clat
205 static void show_clat_percentiles(uint64_t *io_u_plat, unsigned long long nr,
206 fio_fp64_t *plist, unsigned int precision,
207 const char *pre, struct buf_output *out)
209 unsigned int divisor, len, i, j = 0;
210 unsigned long long minv, maxv;
211 unsigned long long *ovals;
212 int per_line, scale_down, time_width;
216 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
221 * We default to nsecs, but if the value range is such that we
222 * should scale down to usecs or msecs, do that.
224 if (minv > 2000000 && maxv > 99999999ULL) {
227 log_buf(out, " %s percentiles (msec):\n |", pre);
228 } else if (minv > 2000 && maxv > 99999) {
231 log_buf(out, " %s percentiles (usec):\n |", pre);
235 log_buf(out, " %s percentiles (nsec):\n |", pre);
239 time_width = max(5, (int) (log10(maxv / divisor) + 1));
240 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
241 precision, time_width);
242 /* fmt will be something like " %5.2fth=[%4llu]%c" */
243 per_line = (80 - 7) / (precision + 10 + time_width);
245 for (j = 0; j < len; j++) {
247 if (j != 0 && (j % per_line) == 0)
250 /* end of the list */
251 is_last = (j == len - 1) != 0;
253 for (i = 0; i < scale_down; i++)
254 ovals[j] = (ovals[j] + 999) / 1000;
256 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
261 if ((j % per_line) == per_line - 1) /* for formatting */
268 bool calc_lat(struct io_stat *is, unsigned long long *min,
269 unsigned long long *max, double *mean, double *dev)
271 double n = (double) is->samples;
278 *mean = is->mean.u.f;
281 *dev = sqrt(is->S.u.f / (n - 1.0));
288 void show_mixed_group_stats(struct group_run_stats *rs, struct buf_output *out)
290 char *io, *agg, *min, *max;
291 char *ioalt, *aggalt, *minalt, *maxalt;
292 uint64_t io_mix = 0, agg_mix = 0, min_mix = -1, max_mix = 0;
293 uint64_t min_run = -1, max_run = 0;
294 const int i2p = is_power_of_2(rs->kb_base);
297 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
300 io_mix += rs->iobytes[i];
301 agg_mix += rs->agg[i];
302 min_mix = min_mix < rs->min_bw[i] ? min_mix : rs->min_bw[i];
303 max_mix = max_mix > rs->max_bw[i] ? max_mix : rs->max_bw[i];
304 min_run = min_run < rs->min_run[i] ? min_run : rs->min_run[i];
305 max_run = max_run > rs->max_run[i] ? max_run : rs->max_run[i];
307 io = num2str(io_mix, rs->sig_figs, 1, i2p, N2S_BYTE);
308 ioalt = num2str(io_mix, rs->sig_figs, 1, !i2p, N2S_BYTE);
309 agg = num2str(agg_mix, rs->sig_figs, 1, i2p, rs->unit_base);
310 aggalt = num2str(agg_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
311 min = num2str(min_mix, rs->sig_figs, 1, i2p, rs->unit_base);
312 minalt = num2str(min_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
313 max = num2str(max_mix, rs->sig_figs, 1, i2p, rs->unit_base);
314 maxalt = num2str(max_mix, rs->sig_figs, 1, !i2p, rs->unit_base);
315 log_buf(out, " MIXED: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
316 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
317 (unsigned long long) min_run,
318 (unsigned long long) max_run);
329 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
331 char *io, *agg, *min, *max;
332 char *ioalt, *aggalt, *minalt, *maxalt;
333 const char *str[] = { " READ", " WRITE" , " TRIM"};
336 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
338 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
339 const int i2p = is_power_of_2(rs->kb_base);
344 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
345 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
346 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
347 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
348 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
349 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
350 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
351 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
352 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
353 (rs->unified_rw_rep == UNIFIED_MIXED) ? " MIXED" : str[i],
354 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
355 (unsigned long long) rs->min_run[i],
356 (unsigned long long) rs->max_run[i]);
368 /* Need to aggregate statisitics to show mixed values */
369 if (rs->unified_rw_rep == UNIFIED_BOTH)
370 show_mixed_group_stats(rs, out);
373 void stat_calc_dist(uint64_t *map, unsigned long total, double *io_u_dist)
378 * Do depth distribution calculations
380 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
382 io_u_dist[i] = (double) map[i] / (double) total;
383 io_u_dist[i] *= 100.0;
384 if (io_u_dist[i] < 0.1 && map[i])
391 static void stat_calc_lat(struct thread_stat *ts, double *dst,
392 uint64_t *src, int nr)
394 unsigned long total = ddir_rw_sum(ts->total_io_u);
398 * Do latency distribution calculations
400 for (i = 0; i < nr; i++) {
402 dst[i] = (double) src[i] / (double) total;
404 if (dst[i] < 0.01 && src[i])
412 * To keep the terse format unaltered, add all of the ns latency
413 * buckets to the first us latency bucket
415 static void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
417 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
420 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
422 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
423 ntotal += ts->io_u_lat_n[i];
425 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
428 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
430 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
433 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
435 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
438 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
440 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
443 static void display_lat(const char *name, unsigned long long min,
444 unsigned long long max, double mean, double dev,
445 struct buf_output *out)
447 const char *base = "(nsec)";
450 if (nsec_to_msec(&min, &max, &mean, &dev))
452 else if (nsec_to_usec(&min, &max, &mean, &dev))
455 minp = num2str(min, 6, 1, 0, N2S_NONE);
456 maxp = num2str(max, 6, 1, 0, N2S_NONE);
458 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
459 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
465 static struct thread_stat *gen_mixed_ddir_stats_from_ts(struct thread_stat *ts)
467 struct thread_stat *ts_lcl;
470 * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
473 ts_lcl = malloc(sizeof(struct thread_stat));
475 log_err("fio: failed to allocate local thread stat\n");
479 init_thread_stat(ts_lcl);
481 /* calculate mixed stats */
482 ts_lcl->unified_rw_rep = UNIFIED_MIXED;
483 ts_lcl->lat_percentiles = ts->lat_percentiles;
484 ts_lcl->clat_percentiles = ts->clat_percentiles;
485 ts_lcl->slat_percentiles = ts->slat_percentiles;
486 ts_lcl->percentile_precision = ts->percentile_precision;
487 memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
489 sum_thread_stats(ts_lcl, ts);
494 static double convert_agg_kbytes_percent(struct group_run_stats *rs, int ddir, int mean)
496 double p_of_agg = 100.0;
497 if (rs && rs->agg[ddir] > 1024) {
498 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
500 if (p_of_agg > 100.0)
506 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
507 int ddir, struct buf_output *out)
510 unsigned long long min, max, bw, iops;
512 char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
515 if (ddir_sync(ddir)) {
516 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
517 log_buf(out, " %s:\n", "fsync/fdatasync/sync_file_range");
518 display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
519 show_clat_percentiles(ts->io_u_sync_plat,
520 ts->sync_stat.samples,
522 ts->percentile_precision,
523 io_ddir_name(ddir), out);
528 assert(ddir_rw(ddir));
530 if (!ts->runtime[ddir])
533 i2p = is_power_of_2(rs->kb_base);
534 runt = ts->runtime[ddir];
536 bw = (1000 * ts->io_bytes[ddir]) / runt;
537 io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
538 bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
539 bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
541 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
542 iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
543 if (ddir == DDIR_WRITE)
544 post_st = zbd_write_status(ts);
545 else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
549 total = ts->cachehit + ts->cachemiss;
550 hit = (double) ts->cachehit / (double) total;
552 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
556 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
557 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
558 iops_p, bw_p, bw_p_alt, io_p,
559 (unsigned long long) ts->runtime[ddir],
568 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
569 display_lat("slat", min, max, mean, dev, out);
570 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
571 display_lat("clat", min, max, mean, dev, out);
572 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
573 display_lat(" lat", min, max, mean, dev, out);
574 if (calc_lat(&ts->clat_high_prio_stat[ddir], &min, &max, &mean, &dev)) {
575 display_lat(ts->lat_percentiles ? "high prio_lat" : "high prio_clat",
576 min, max, mean, dev, out);
577 if (calc_lat(&ts->clat_low_prio_stat[ddir], &min, &max, &mean, &dev))
578 display_lat(ts->lat_percentiles ? "low prio_lat" : "low prio_clat",
579 min, max, mean, dev, out);
582 if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
583 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
584 ts->slat_stat[ddir].samples,
586 ts->percentile_precision, "slat", out);
587 if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
588 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
589 ts->clat_stat[ddir].samples,
591 ts->percentile_precision, "clat", out);
592 if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
593 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
594 ts->lat_stat[ddir].samples,
596 ts->percentile_precision, "lat", out);
598 if (ts->clat_percentiles || ts->lat_percentiles) {
599 const char *name = ts->lat_percentiles ? "lat" : "clat";
603 if (ts->lat_percentiles)
604 samples = ts->lat_stat[ddir].samples;
606 samples = ts->clat_stat[ddir].samples;
608 /* Only print this if some high and low priority stats were collected */
609 if (ts->clat_high_prio_stat[ddir].samples > 0 &&
610 ts->clat_low_prio_stat[ddir].samples > 0)
612 sprintf(prio_name, "high prio (%.2f%%) %s",
613 100. * (double) ts->clat_high_prio_stat[ddir].samples / (double) samples,
615 show_clat_percentiles(ts->io_u_plat_high_prio[ddir],
616 ts->clat_high_prio_stat[ddir].samples,
618 ts->percentile_precision, prio_name, out);
620 sprintf(prio_name, "low prio (%.2f%%) %s",
621 100. * (double) ts->clat_low_prio_stat[ddir].samples / (double) samples,
623 show_clat_percentiles(ts->io_u_plat_low_prio[ddir],
624 ts->clat_low_prio_stat[ddir].samples,
626 ts->percentile_precision, prio_name, out);
630 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
631 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
634 if ((rs->unit_base == 1) && i2p)
636 else if (rs->unit_base == 1)
643 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
645 if (rs->unit_base == 1) {
652 if (mean > fkb_base * fkb_base) {
657 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
660 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
661 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
662 bw_str, min, max, p_of_agg, mean, dev,
663 (&ts->bw_stat[ddir])->samples);
665 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
666 log_buf(out, " iops : min=%5llu, max=%5llu, "
667 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
668 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
672 static void show_mixed_ddir_status(struct group_run_stats *rs,
673 struct thread_stat *ts,
674 struct buf_output *out)
676 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
679 show_ddir_status(rs, ts_lcl, DDIR_READ, out);
684 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
685 const char *msg, struct buf_output *out)
687 bool new_line = true, shown = false;
690 for (i = 0; i < nr; i++) {
691 if (io_u_lat[i] <= 0.0)
697 log_buf(out, " lat (%s) : ", msg);
703 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
715 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
717 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
718 "250=", "500=", "750=", "1000=", };
720 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
723 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
725 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
726 "250=", "500=", "750=", "1000=", };
728 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
731 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
733 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
734 "250=", "500=", "750=", "1000=", "2000=",
737 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
740 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
742 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
743 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
744 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
746 stat_calc_lat_n(ts, io_u_lat_n);
747 stat_calc_lat_u(ts, io_u_lat_u);
748 stat_calc_lat_m(ts, io_u_lat_m);
750 show_lat_n(io_u_lat_n, out);
751 show_lat_u(io_u_lat_u, out);
752 show_lat_m(io_u_lat_m, out);
755 static int block_state_category(int block_state)
757 switch (block_state) {
758 case BLOCK_STATE_UNINIT:
760 case BLOCK_STATE_TRIMMED:
761 case BLOCK_STATE_WRITTEN:
763 case BLOCK_STATE_WRITE_FAILURE:
764 case BLOCK_STATE_TRIM_FAILURE:
767 /* Silence compile warning on some BSDs and have a return */
773 static int compare_block_infos(const void *bs1, const void *bs2)
775 uint64_t block1 = *(uint64_t *)bs1;
776 uint64_t block2 = *(uint64_t *)bs2;
777 int state1 = BLOCK_INFO_STATE(block1);
778 int state2 = BLOCK_INFO_STATE(block2);
779 int bscat1 = block_state_category(state1);
780 int bscat2 = block_state_category(state2);
781 int cycles1 = BLOCK_INFO_TRIMS(block1);
782 int cycles2 = BLOCK_INFO_TRIMS(block2);
789 if (cycles1 < cycles2)
791 if (cycles1 > cycles2)
799 assert(block1 == block2);
803 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
804 fio_fp64_t *plist, unsigned int **percentiles,
810 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
812 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
819 * Sort the percentile list. Note that it may already be sorted if
820 * we are using the default values, but since it's a short list this
821 * isn't a worry. Also note that this does not work for NaN values.
824 qsort(plist, len, sizeof(plist[0]), double_cmp);
826 /* Start only after the uninit entries end */
828 nr_uninit < nr_block_infos
829 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
833 if (nr_uninit == nr_block_infos)
836 *percentiles = calloc(len, sizeof(**percentiles));
838 for (i = 0; i < len; i++) {
839 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
841 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
844 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
845 for (i = 0; i < nr_block_infos; i++)
846 types[BLOCK_INFO_STATE(block_infos[i])]++;
851 static const char *block_state_names[] = {
852 [BLOCK_STATE_UNINIT] = "unwritten",
853 [BLOCK_STATE_TRIMMED] = "trimmed",
854 [BLOCK_STATE_WRITTEN] = "written",
855 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
856 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
859 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
860 fio_fp64_t *plist, struct buf_output *out)
863 unsigned int *percentiles = NULL;
864 unsigned int block_state_counts[BLOCK_STATE_COUNT];
866 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
867 &percentiles, block_state_counts);
869 log_buf(out, " block lifetime percentiles :\n |");
871 for (i = 0; i < len; i++) {
872 uint32_t block_info = percentiles[i];
873 #define LINE_LENGTH 75
874 char str[LINE_LENGTH];
875 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
876 plist[i].u.f, block_info,
877 i == len - 1 ? '\n' : ',');
878 assert(strln < LINE_LENGTH);
879 if (pos + strln > LINE_LENGTH) {
881 log_buf(out, "\n |");
883 log_buf(out, "%s", str);
890 log_buf(out, " states :");
891 for (i = 0; i < BLOCK_STATE_COUNT; i++)
892 log_buf(out, " %s=%u%c",
893 block_state_names[i], block_state_counts[i],
894 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
897 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
899 char *p1, *p1alt, *p2;
900 unsigned long long bw_mean, iops_mean;
901 const int i2p = is_power_of_2(ts->kb_base);
906 bw_mean = steadystate_bw_mean(ts);
907 iops_mean = steadystate_iops_mean(ts);
909 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
910 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
911 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
913 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
914 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
916 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
917 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
918 ts->ss_criterion.u.f,
919 ts->ss_state & FIO_SS_PCT ? "%" : "");
926 static void show_agg_stats(struct disk_util_agg *agg, int terse,
927 struct buf_output *out)
929 if (!agg->slavecount)
933 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
934 "aggrticks=%llu/%llu, aggrin_queue=%llu, "
936 (unsigned long long) agg->ios[0] / agg->slavecount,
937 (unsigned long long) agg->ios[1] / agg->slavecount,
938 (unsigned long long) agg->merges[0] / agg->slavecount,
939 (unsigned long long) agg->merges[1] / agg->slavecount,
940 (unsigned long long) agg->ticks[0] / agg->slavecount,
941 (unsigned long long) agg->ticks[1] / agg->slavecount,
942 (unsigned long long) agg->time_in_queue / agg->slavecount,
945 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
946 (unsigned long long) agg->ios[0] / agg->slavecount,
947 (unsigned long long) agg->ios[1] / agg->slavecount,
948 (unsigned long long) agg->merges[0] / agg->slavecount,
949 (unsigned long long) agg->merges[1] / agg->slavecount,
950 (unsigned long long) agg->ticks[0] / agg->slavecount,
951 (unsigned long long) agg->ticks[1] / agg->slavecount,
952 (unsigned long long) agg->time_in_queue / agg->slavecount,
957 static void aggregate_slaves_stats(struct disk_util *masterdu)
959 struct disk_util_agg *agg = &masterdu->agg;
960 struct disk_util_stat *dus;
961 struct flist_head *entry;
962 struct disk_util *slavedu;
965 flist_for_each(entry, &masterdu->slaves) {
966 slavedu = flist_entry(entry, struct disk_util, slavelist);
968 agg->ios[0] += dus->s.ios[0];
969 agg->ios[1] += dus->s.ios[1];
970 agg->merges[0] += dus->s.merges[0];
971 agg->merges[1] += dus->s.merges[1];
972 agg->sectors[0] += dus->s.sectors[0];
973 agg->sectors[1] += dus->s.sectors[1];
974 agg->ticks[0] += dus->s.ticks[0];
975 agg->ticks[1] += dus->s.ticks[1];
976 agg->time_in_queue += dus->s.time_in_queue;
979 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
980 /* System utilization is the utilization of the
981 * component with the highest utilization.
983 if (util > agg->max_util.u.f)
984 agg->max_util.u.f = util;
988 if (agg->max_util.u.f > 100.0)
989 agg->max_util.u.f = 100.0;
992 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
993 int terse, struct buf_output *out)
998 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1003 if (agg->slavecount)
1006 log_buf(out, " %s: ios=%llu/%llu, merge=%llu/%llu, "
1007 "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
1009 (unsigned long long) dus->s.ios[0],
1010 (unsigned long long) dus->s.ios[1],
1011 (unsigned long long) dus->s.merges[0],
1012 (unsigned long long) dus->s.merges[1],
1013 (unsigned long long) dus->s.ticks[0],
1014 (unsigned long long) dus->s.ticks[1],
1015 (unsigned long long) dus->s.time_in_queue,
1018 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1020 (unsigned long long) dus->s.ios[0],
1021 (unsigned long long) dus->s.ios[1],
1022 (unsigned long long) dus->s.merges[0],
1023 (unsigned long long) dus->s.merges[1],
1024 (unsigned long long) dus->s.ticks[0],
1025 (unsigned long long) dus->s.ticks[1],
1026 (unsigned long long) dus->s.time_in_queue,
1031 * If the device has slaves, aggregate the stats for
1032 * those slave devices also.
1034 show_agg_stats(agg, terse, out);
1040 void json_array_add_disk_util(struct disk_util_stat *dus,
1041 struct disk_util_agg *agg, struct json_array *array)
1043 struct json_object *obj;
1047 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1051 obj = json_create_object();
1052 json_array_add_value_object(array, obj);
1054 json_object_add_value_string(obj, "name", (const char *)dus->name);
1055 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1056 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1057 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1058 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1059 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1060 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1061 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1062 json_object_add_value_float(obj, "util", util);
1065 * If the device has slaves, aggregate the stats for
1066 * those slave devices also.
1068 if (!agg->slavecount)
1070 json_object_add_value_int(obj, "aggr_read_ios",
1071 agg->ios[0] / agg->slavecount);
1072 json_object_add_value_int(obj, "aggr_write_ios",
1073 agg->ios[1] / agg->slavecount);
1074 json_object_add_value_int(obj, "aggr_read_merges",
1075 agg->merges[0] / agg->slavecount);
1076 json_object_add_value_int(obj, "aggr_write_merge",
1077 agg->merges[1] / agg->slavecount);
1078 json_object_add_value_int(obj, "aggr_read_ticks",
1079 agg->ticks[0] / agg->slavecount);
1080 json_object_add_value_int(obj, "aggr_write_ticks",
1081 agg->ticks[1] / agg->slavecount);
1082 json_object_add_value_int(obj, "aggr_in_queue",
1083 agg->time_in_queue / agg->slavecount);
1084 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1087 static void json_object_add_disk_utils(struct json_object *obj,
1088 struct flist_head *head)
1090 struct json_array *array = json_create_array();
1091 struct flist_head *entry;
1092 struct disk_util *du;
1094 json_object_add_value_array(obj, "disk_util", array);
1096 flist_for_each(entry, head) {
1097 du = flist_entry(entry, struct disk_util, list);
1099 aggregate_slaves_stats(du);
1100 json_array_add_disk_util(&du->dus, &du->agg, array);
1104 void show_disk_util(int terse, struct json_object *parent,
1105 struct buf_output *out)
1107 struct flist_head *entry;
1108 struct disk_util *du;
1111 if (!is_running_backend())
1114 if (flist_empty(&disk_list))
1117 if ((output_format & FIO_OUTPUT_JSON) && parent)
1122 if (!terse && !do_json)
1123 log_buf(out, "\nDisk stats (read/write):\n");
1126 json_object_add_disk_utils(parent, &disk_list);
1127 } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1128 flist_for_each(entry, &disk_list) {
1129 du = flist_entry(entry, struct disk_util, list);
1131 aggregate_slaves_stats(du);
1132 print_disk_util(&du->dus, &du->agg, terse, out);
1137 static void show_thread_status_normal(struct thread_stat *ts,
1138 struct group_run_stats *rs,
1139 struct buf_output *out)
1141 double usr_cpu, sys_cpu;
1142 unsigned long runtime;
1143 double io_u_dist[FIO_IO_U_MAP_NR];
1147 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1150 memset(time_buf, 0, sizeof(time_buf));
1153 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1156 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1157 ts->name, ts->groupid, ts->members,
1158 ts->error, (int) ts->pid, time_buf);
1160 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1161 ts->name, ts->groupid, ts->members,
1162 ts->error, ts->verror, (int) ts->pid,
1166 if (strlen(ts->description))
1167 log_buf(out, " Description : [%s]\n", ts->description);
1169 for_each_rw_ddir(ddir) {
1170 if (ts->io_bytes[ddir])
1171 show_ddir_status(rs, ts, ddir, out);
1174 if (ts->unified_rw_rep == UNIFIED_BOTH)
1175 show_mixed_ddir_status(rs, ts, out);
1177 show_latencies(ts, out);
1179 if (ts->sync_stat.samples)
1180 show_ddir_status(rs, ts, DDIR_SYNC, out);
1182 runtime = ts->total_run_time;
1184 double runt = (double) runtime;
1186 usr_cpu = (double) ts->usr_time * 100 / runt;
1187 sys_cpu = (double) ts->sys_time * 100 / runt;
1193 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1194 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1195 (unsigned long long) ts->ctx,
1196 (unsigned long long) ts->majf,
1197 (unsigned long long) ts->minf);
1199 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1200 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1201 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1202 io_u_dist[1], io_u_dist[2],
1203 io_u_dist[3], io_u_dist[4],
1204 io_u_dist[5], io_u_dist[6]);
1206 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1207 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1208 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1209 io_u_dist[1], io_u_dist[2],
1210 io_u_dist[3], io_u_dist[4],
1211 io_u_dist[5], io_u_dist[6]);
1212 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1213 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1214 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1215 io_u_dist[1], io_u_dist[2],
1216 io_u_dist[3], io_u_dist[4],
1217 io_u_dist[5], io_u_dist[6]);
1218 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
1219 " short=%llu,%llu,%llu,0"
1220 " dropped=%llu,%llu,%llu,0\n",
1221 (unsigned long long) ts->total_io_u[0],
1222 (unsigned long long) ts->total_io_u[1],
1223 (unsigned long long) ts->total_io_u[2],
1224 (unsigned long long) ts->total_io_u[3],
1225 (unsigned long long) ts->short_io_u[0],
1226 (unsigned long long) ts->short_io_u[1],
1227 (unsigned long long) ts->short_io_u[2],
1228 (unsigned long long) ts->drop_io_u[0],
1229 (unsigned long long) ts->drop_io_u[1],
1230 (unsigned long long) ts->drop_io_u[2]);
1231 if (ts->continue_on_error) {
1232 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
1233 (unsigned long long)ts->total_err_count,
1235 strerror(ts->first_error));
1237 if (ts->latency_depth) {
1238 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1239 (unsigned long long)ts->latency_target,
1240 (unsigned long long)ts->latency_window,
1241 ts->latency_percentile.u.f,
1245 if (ts->nr_block_infos)
1246 show_block_infos(ts->nr_block_infos, ts->block_infos,
1247 ts->percentile_list, out);
1250 show_ss_normal(ts, out);
1253 static void show_ddir_status_terse(struct thread_stat *ts,
1254 struct group_run_stats *rs, int ddir,
1255 int ver, struct buf_output *out)
1257 unsigned long long min, max, minv, maxv, bw, iops;
1258 unsigned long long *ovals = NULL;
1263 assert(ddir_rw(ddir));
1266 if (ts->runtime[ddir]) {
1267 uint64_t runt = ts->runtime[ddir];
1269 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1270 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1273 log_buf(out, ";%llu;%llu;%llu;%llu",
1274 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1275 (unsigned long long) ts->runtime[ddir]);
1277 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1278 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1280 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1282 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1283 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1285 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1287 if (ts->lat_percentiles) {
1288 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1289 ts->lat_stat[ddir].samples,
1290 ts->percentile_list, &ovals, &maxv,
1292 } else if (ts->clat_percentiles) {
1293 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1294 ts->clat_stat[ddir].samples,
1295 ts->percentile_list, &ovals, &maxv,
1301 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1303 log_buf(out, ";0%%=0");
1306 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1309 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1310 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1312 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1316 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1318 double p_of_agg = 100.0;
1320 if (rs->agg[ddir]) {
1321 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1322 if (p_of_agg > 100.0)
1326 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1328 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1333 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1335 log_buf(out, ";%lu", 0UL);
1337 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1338 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1339 mean, dev, (&ts->iops_stat[ddir])->samples);
1341 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1345 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1346 struct group_run_stats *rs,
1347 int ver, struct buf_output *out)
1349 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1352 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1357 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1358 uint32_t percentiles,
1359 struct io_stat *lat_stat,
1360 uint64_t *io_u_plat)
1364 unsigned int i, len;
1365 struct json_object *lat_object, *percentile_object, *clat_bins_object;
1366 unsigned long long min, max, maxv, minv, *ovals = NULL;
1368 if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1372 lat_object = json_create_object();
1373 json_object_add_value_int(lat_object, "min", min);
1374 json_object_add_value_int(lat_object, "max", max);
1375 json_object_add_value_float(lat_object, "mean", mean);
1376 json_object_add_value_float(lat_object, "stddev", dev);
1377 json_object_add_value_int(lat_object, "N", lat_stat->samples);
1379 if (percentiles && lat_stat->samples) {
1380 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1381 ts->percentile_list, &ovals, &maxv, &minv);
1383 if (len > FIO_IO_U_LIST_MAX_LEN)
1384 len = FIO_IO_U_LIST_MAX_LEN;
1386 percentile_object = json_create_object();
1387 json_object_add_value_object(lat_object, "percentile", percentile_object);
1388 for (i = 0; i < len; i++) {
1389 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1390 json_object_add_value_int(percentile_object, buf, ovals[i]);
1394 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1395 clat_bins_object = json_create_object();
1396 json_object_add_value_object(lat_object, "bins", clat_bins_object);
1398 for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1400 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1401 json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1409 static void add_ddir_status_json(struct thread_stat *ts,
1410 struct group_run_stats *rs, int ddir, struct json_object *parent)
1412 unsigned long long min, max;
1413 unsigned long long bw_bytes, bw;
1414 double mean, dev, iops;
1415 struct json_object *dir_object, *tmp_object;
1416 double p_of_agg = 100.0;
1418 assert(ddir_rw(ddir) || ddir_sync(ddir));
1420 if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1423 dir_object = json_create_object();
1424 json_object_add_value_object(parent,
1425 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1427 if (ddir_rw(ddir)) {
1431 if (ts->runtime[ddir]) {
1432 uint64_t runt = ts->runtime[ddir];
1434 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1435 bw = bw_bytes / 1024; /* KiB/s */
1436 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1439 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1440 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1441 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1442 json_object_add_value_int(dir_object, "bw", bw);
1443 json_object_add_value_float(dir_object, "iops", iops);
1444 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1445 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1446 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1447 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1449 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1450 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1451 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1453 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1454 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1455 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1457 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1458 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1459 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1461 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1462 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1463 &ts->sync_stat, ts->io_u_sync_plat);
1464 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1470 /* Only print PRIO latencies if some high priority samples were gathered */
1471 if (ts->clat_high_prio_stat[ddir].samples > 0) {
1472 const char *high, *low;
1474 if (ts->lat_percentiles) {
1475 high = "lat_high_prio";
1476 low = "lat_low_prio";
1478 high = "clat_high_prio";
1479 low = "clat_low_prio";
1482 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1483 &ts->clat_high_prio_stat[ddir], ts->io_u_plat_high_prio[ddir]);
1484 json_object_add_value_object(dir_object, high, tmp_object);
1486 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1487 &ts->clat_low_prio_stat[ddir], ts->io_u_plat_low_prio[ddir]);
1488 json_object_add_value_object(dir_object, low, tmp_object);
1491 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1492 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1495 p_of_agg = mean = dev = 0.0;
1498 json_object_add_value_int(dir_object, "bw_min", min);
1499 json_object_add_value_int(dir_object, "bw_max", max);
1500 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1501 json_object_add_value_float(dir_object, "bw_mean", mean);
1502 json_object_add_value_float(dir_object, "bw_dev", dev);
1503 json_object_add_value_int(dir_object, "bw_samples",
1504 (&ts->bw_stat[ddir])->samples);
1506 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1510 json_object_add_value_int(dir_object, "iops_min", min);
1511 json_object_add_value_int(dir_object, "iops_max", max);
1512 json_object_add_value_float(dir_object, "iops_mean", mean);
1513 json_object_add_value_float(dir_object, "iops_stddev", dev);
1514 json_object_add_value_int(dir_object, "iops_samples",
1515 (&ts->iops_stat[ddir])->samples);
1517 if (ts->cachehit + ts->cachemiss) {
1521 total = ts->cachehit + ts->cachemiss;
1522 hit = (double) ts->cachehit / (double) total;
1524 json_object_add_value_float(dir_object, "cachehit", hit);
1528 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1529 struct group_run_stats *rs, struct json_object *parent)
1531 struct thread_stat *ts_lcl = gen_mixed_ddir_stats_from_ts(ts);
1533 /* add the aggregated stats to json parent */
1535 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1540 static void show_thread_status_terse_all(struct thread_stat *ts,
1541 struct group_run_stats *rs, int ver,
1542 struct buf_output *out)
1544 double io_u_dist[FIO_IO_U_MAP_NR];
1545 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1546 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1547 double usr_cpu, sys_cpu;
1552 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1554 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1555 ts->name, ts->groupid, ts->error);
1557 /* Log Read Status, or mixed if unified_rw_rep = 1 */
1558 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1559 if (ts->unified_rw_rep != UNIFIED_MIXED) {
1560 /* Log Write Status */
1561 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1562 /* Log Trim Status */
1563 if (ver == 2 || ver == 4 || ver == 5)
1564 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1566 if (ts->unified_rw_rep == UNIFIED_BOTH)
1567 show_mixed_ddir_status_terse(ts, rs, ver, out);
1569 if (ts->total_run_time) {
1570 double runt = (double) ts->total_run_time;
1572 usr_cpu = (double) ts->usr_time * 100 / runt;
1573 sys_cpu = (double) ts->sys_time * 100 / runt;
1579 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1580 (unsigned long long) ts->ctx,
1581 (unsigned long long) ts->majf,
1582 (unsigned long long) ts->minf);
1584 /* Calc % distribution of IO depths, usecond, msecond latency */
1585 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1586 stat_calc_lat_nu(ts, io_u_lat_u);
1587 stat_calc_lat_m(ts, io_u_lat_m);
1589 /* Only show fixed 7 I/O depth levels*/
1590 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1591 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1592 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1594 /* Microsecond latency */
1595 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1596 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1597 /* Millisecond latency */
1598 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1599 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1601 /* disk util stats, if any */
1602 if (ver >= 3 && is_running_backend())
1603 show_disk_util(1, NULL, out);
1605 /* Additional output if continue_on_error set - default off*/
1606 if (ts->continue_on_error)
1607 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1609 /* Additional output if description is set */
1610 if (strlen(ts->description)) {
1613 log_buf(out, ";%s", ts->description);
1619 static void json_add_job_opts(struct json_object *root, const char *name,
1620 struct flist_head *opt_list)
1622 struct json_object *dir_object;
1623 struct flist_head *entry;
1624 struct print_option *p;
1626 if (flist_empty(opt_list))
1629 dir_object = json_create_object();
1630 json_object_add_value_object(root, name, dir_object);
1632 flist_for_each(entry, opt_list) {
1633 p = flist_entry(entry, struct print_option, list);
1634 json_object_add_value_string(dir_object, p->name, p->value);
1638 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1639 struct group_run_stats *rs,
1640 struct flist_head *opt_list)
1642 struct json_object *root, *tmp;
1643 struct jobs_eta *je;
1644 double io_u_dist[FIO_IO_U_MAP_NR];
1645 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1646 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1647 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1648 double usr_cpu, sys_cpu;
1652 root = json_create_object();
1653 json_object_add_value_string(root, "jobname", ts->name);
1654 json_object_add_value_int(root, "groupid", ts->groupid);
1655 json_object_add_value_int(root, "error", ts->error);
1658 je = get_jobs_eta(true, &size);
1660 json_object_add_value_int(root, "eta", je->eta_sec);
1661 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1665 json_add_job_opts(root, "job options", opt_list);
1667 add_ddir_status_json(ts, rs, DDIR_READ, root);
1668 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1669 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1670 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1672 if (ts->unified_rw_rep == UNIFIED_BOTH)
1673 add_mixed_ddir_status_json(ts, rs, root);
1676 if (ts->total_run_time) {
1677 double runt = (double) ts->total_run_time;
1679 usr_cpu = (double) ts->usr_time * 100 / runt;
1680 sys_cpu = (double) ts->sys_time * 100 / runt;
1685 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1686 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1687 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1688 json_object_add_value_int(root, "ctx", ts->ctx);
1689 json_object_add_value_int(root, "majf", ts->majf);
1690 json_object_add_value_int(root, "minf", ts->minf);
1692 /* Calc % distribution of IO depths */
1693 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1694 tmp = json_create_object();
1695 json_object_add_value_object(root, "iodepth_level", tmp);
1696 /* Only show fixed 7 I/O depth levels*/
1697 for (i = 0; i < 7; i++) {
1700 snprintf(name, 20, "%d", 1 << i);
1702 snprintf(name, 20, ">=%d", 1 << i);
1703 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1706 /* Calc % distribution of submit IO depths */
1707 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1708 tmp = json_create_object();
1709 json_object_add_value_object(root, "iodepth_submit", tmp);
1710 /* Only show fixed 7 I/O depth levels*/
1711 for (i = 0; i < 7; i++) {
1714 snprintf(name, 20, "0");
1716 snprintf(name, 20, "%d", 1 << (i+1));
1718 snprintf(name, 20, ">=%d", 1 << i);
1719 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1722 /* Calc % distribution of completion IO depths */
1723 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1724 tmp = json_create_object();
1725 json_object_add_value_object(root, "iodepth_complete", tmp);
1726 /* Only show fixed 7 I/O depth levels*/
1727 for (i = 0; i < 7; i++) {
1730 snprintf(name, 20, "0");
1732 snprintf(name, 20, "%d", 1 << (i+1));
1734 snprintf(name, 20, ">=%d", 1 << i);
1735 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1738 /* Calc % distribution of nsecond, usecond, msecond latency */
1739 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1740 stat_calc_lat_n(ts, io_u_lat_n);
1741 stat_calc_lat_u(ts, io_u_lat_u);
1742 stat_calc_lat_m(ts, io_u_lat_m);
1744 /* Nanosecond latency */
1745 tmp = json_create_object();
1746 json_object_add_value_object(root, "latency_ns", tmp);
1747 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1748 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1749 "250", "500", "750", "1000", };
1750 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1752 /* Microsecond latency */
1753 tmp = json_create_object();
1754 json_object_add_value_object(root, "latency_us", tmp);
1755 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1756 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1757 "250", "500", "750", "1000", };
1758 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1760 /* Millisecond latency */
1761 tmp = json_create_object();
1762 json_object_add_value_object(root, "latency_ms", tmp);
1763 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1764 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1765 "250", "500", "750", "1000", "2000",
1767 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1770 /* Additional output if continue_on_error set - default off*/
1771 if (ts->continue_on_error) {
1772 json_object_add_value_int(root, "total_err", ts->total_err_count);
1773 json_object_add_value_int(root, "first_error", ts->first_error);
1776 if (ts->latency_depth) {
1777 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1778 json_object_add_value_int(root, "latency_target", ts->latency_target);
1779 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1780 json_object_add_value_int(root, "latency_window", ts->latency_window);
1783 /* Additional output if description is set */
1784 if (strlen(ts->description))
1785 json_object_add_value_string(root, "desc", ts->description);
1787 if (ts->nr_block_infos) {
1788 /* Block error histogram and types */
1790 unsigned int *percentiles = NULL;
1791 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1793 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1794 ts->percentile_list,
1795 &percentiles, block_state_counts);
1798 struct json_object *block, *percentile_object, *states;
1800 block = json_create_object();
1801 json_object_add_value_object(root, "block", block);
1803 percentile_object = json_create_object();
1804 json_object_add_value_object(block, "percentiles",
1806 for (i = 0; i < len; i++) {
1808 snprintf(buf, sizeof(buf), "%f",
1809 ts->percentile_list[i].u.f);
1810 json_object_add_value_int(percentile_object,
1815 states = json_create_object();
1816 json_object_add_value_object(block, "states", states);
1817 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1818 json_object_add_value_int(states,
1819 block_state_names[state],
1820 block_state_counts[state]);
1827 struct json_object *data;
1828 struct json_array *iops, *bw;
1832 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1833 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1834 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1835 (float) ts->ss_limit.u.f,
1836 ts->ss_state & FIO_SS_PCT ? "%" : "");
1838 tmp = json_create_object();
1839 json_object_add_value_object(root, "steadystate", tmp);
1840 json_object_add_value_string(tmp, "ss", ss_buf);
1841 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1842 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1844 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1845 ts->ss_state & FIO_SS_PCT ? "%" : "");
1846 json_object_add_value_string(tmp, "criterion", ss_buf);
1847 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1848 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1850 data = json_create_object();
1851 json_object_add_value_object(tmp, "data", data);
1852 bw = json_create_array();
1853 iops = json_create_array();
1856 ** if ss was attained or the buffer is not full,
1857 ** ss->head points to the first element in the list.
1858 ** otherwise it actually points to the second element
1861 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1864 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1865 for (l = 0; l < ts->ss_dur; l++) {
1866 k = (j + l) % ts->ss_dur;
1867 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1868 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1870 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1871 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1872 json_object_add_value_array(data, "iops", iops);
1873 json_object_add_value_array(data, "bw", bw);
1879 static void show_thread_status_terse(struct thread_stat *ts,
1880 struct group_run_stats *rs,
1881 struct buf_output *out)
1883 if (terse_version >= 2 && terse_version <= 5)
1884 show_thread_status_terse_all(ts, rs, terse_version, out);
1886 log_err("fio: bad terse version!? %d\n", terse_version);
1889 struct json_object *show_thread_status(struct thread_stat *ts,
1890 struct group_run_stats *rs,
1891 struct flist_head *opt_list,
1892 struct buf_output *out)
1894 struct json_object *ret = NULL;
1896 if (output_format & FIO_OUTPUT_TERSE)
1897 show_thread_status_terse(ts, rs, out);
1898 if (output_format & FIO_OUTPUT_JSON)
1899 ret = show_thread_status_json(ts, rs, opt_list);
1900 if (output_format & FIO_OUTPUT_NORMAL)
1901 show_thread_status_normal(ts, rs, out);
1906 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1910 dst->min_val = min(dst->min_val, src->min_val);
1911 dst->max_val = max(dst->max_val, src->max_val);
1914 * Compute new mean and S after the merge
1915 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1916 * #Parallel_algorithm>
1919 mean = src->mean.u.f;
1922 double delta = src->mean.u.f - dst->mean.u.f;
1924 mean = ((src->mean.u.f * src->samples) +
1925 (dst->mean.u.f * dst->samples)) /
1926 (dst->samples + src->samples);
1928 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1929 (dst->samples * src->samples) /
1930 (dst->samples + src->samples);
1933 dst->samples += src->samples;
1934 dst->mean.u.f = mean;
1940 * We sum two kinds of stats - one that is time based, in which case we
1941 * apply the proper summing technique, and then one that is iops/bw
1942 * numbers. For group_reporting, we should just add those up, not make
1943 * them the mean of everything.
1945 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
1947 bool first = dst->samples == 0;
1949 if (src->samples == 0)
1953 __sum_stat(dst, src, first);
1958 dst->min_val = src->min_val;
1959 dst->max_val = src->max_val;
1960 dst->samples = src->samples;
1961 dst->mean.u.f = src->mean.u.f;
1962 dst->S.u.f = src->S.u.f;
1964 dst->min_val += src->min_val;
1965 dst->max_val += src->max_val;
1966 dst->samples += src->samples;
1967 dst->mean.u.f += src->mean.u.f;
1968 dst->S.u.f += src->S.u.f;
1972 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1976 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1977 if (dst->max_run[i] < src->max_run[i])
1978 dst->max_run[i] = src->max_run[i];
1979 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1980 dst->min_run[i] = src->min_run[i];
1981 if (dst->max_bw[i] < src->max_bw[i])
1982 dst->max_bw[i] = src->max_bw[i];
1983 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1984 dst->min_bw[i] = src->min_bw[i];
1986 dst->iobytes[i] += src->iobytes[i];
1987 dst->agg[i] += src->agg[i];
1991 dst->kb_base = src->kb_base;
1992 if (!dst->unit_base)
1993 dst->unit_base = src->unit_base;
1995 dst->sig_figs = src->sig_figs;
1999 * Free the clat_prio_stat arrays allocated by alloc_clat_prio_stat_ddir().
2001 void free_clat_prio_stats(struct thread_stat *ts)
2005 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2006 sfree(ts->clat_prio[ddir]);
2007 ts->clat_prio[ddir] = NULL;
2008 ts->nr_clat_prio[ddir] = 0;
2013 * Allocate a clat_prio_stat array. The array has to be allocated/freed using
2014 * smalloc/sfree, so that it is accessible by the process/thread summing the
2017 int alloc_clat_prio_stat_ddir(struct thread_stat *ts, enum fio_ddir ddir,
2020 struct clat_prio_stat *clat_prio;
2023 clat_prio = scalloc(nr_prios, sizeof(*ts->clat_prio[ddir]));
2025 log_err("fio: failed to allocate ts clat data\n");
2029 for (i = 0; i < nr_prios; i++)
2030 clat_prio[i].clat_stat.min_val = ULONG_MAX;
2032 ts->clat_prio[ddir] = clat_prio;
2033 ts->nr_clat_prio[ddir] = nr_prios;
2038 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2042 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2043 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2044 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2045 sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], false);
2046 sum_stat(&dst->clat_low_prio_stat[l], &src->clat_low_prio_stat[l], false);
2047 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2048 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2049 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2050 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2052 dst->io_bytes[l] += src->io_bytes[l];
2054 if (dst->runtime[l] < src->runtime[l])
2055 dst->runtime[l] = src->runtime[l];
2057 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2058 sum_stat(&dst->clat_high_prio_stat[0], &src->clat_high_prio_stat[l], false);
2059 sum_stat(&dst->clat_low_prio_stat[0], &src->clat_low_prio_stat[l], false);
2060 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2061 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2062 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2063 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2065 dst->io_bytes[0] += src->io_bytes[l];
2067 if (dst->runtime[0] < src->runtime[l])
2068 dst->runtime[0] = src->runtime[l];
2072 sum_stat(&dst->sync_stat, &src->sync_stat, false);
2073 dst->usr_time += src->usr_time;
2074 dst->sys_time += src->sys_time;
2075 dst->ctx += src->ctx;
2076 dst->majf += src->majf;
2077 dst->minf += src->minf;
2079 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2080 dst->io_u_map[k] += src->io_u_map[k];
2081 dst->io_u_submit[k] += src->io_u_submit[k];
2082 dst->io_u_complete[k] += src->io_u_complete[k];
2085 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2086 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2087 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2088 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2089 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2090 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2092 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2093 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2094 dst->total_io_u[k] += src->total_io_u[k];
2095 dst->short_io_u[k] += src->short_io_u[k];
2096 dst->drop_io_u[k] += src->drop_io_u[k];
2098 dst->total_io_u[0] += src->total_io_u[k];
2099 dst->short_io_u[0] += src->short_io_u[k];
2100 dst->drop_io_u[0] += src->drop_io_u[k];
2104 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2106 for (k = 0; k < FIO_LAT_CNT; k++)
2107 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2108 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2109 if (dst->unified_rw_rep != UNIFIED_MIXED)
2110 dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2112 dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2114 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2115 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2117 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2118 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
2119 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2120 dst->io_u_plat_high_prio[k][m] += src->io_u_plat_high_prio[k][m];
2121 dst->io_u_plat_low_prio[k][m] += src->io_u_plat_low_prio[k][m];
2123 dst->io_u_plat_high_prio[0][m] += src->io_u_plat_high_prio[k][m];
2124 dst->io_u_plat_low_prio[0][m] += src->io_u_plat_low_prio[k][m];
2130 dst->total_run_time += src->total_run_time;
2131 dst->total_submit += src->total_submit;
2132 dst->total_complete += src->total_complete;
2133 dst->nr_zone_resets += src->nr_zone_resets;
2134 dst->cachehit += src->cachehit;
2135 dst->cachemiss += src->cachemiss;
2138 void init_group_run_stat(struct group_run_stats *gs)
2141 memset(gs, 0, sizeof(*gs));
2143 for (i = 0; i < DDIR_RWDIR_CNT; i++)
2144 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2147 void init_thread_stat_min_vals(struct thread_stat *ts)
2151 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2152 ts->clat_stat[i].min_val = ULONG_MAX;
2153 ts->slat_stat[i].min_val = ULONG_MAX;
2154 ts->lat_stat[i].min_val = ULONG_MAX;
2155 ts->bw_stat[i].min_val = ULONG_MAX;
2156 ts->iops_stat[i].min_val = ULONG_MAX;
2157 ts->clat_high_prio_stat[i].min_val = ULONG_MAX;
2158 ts->clat_low_prio_stat[i].min_val = ULONG_MAX;
2160 ts->sync_stat.min_val = ULONG_MAX;
2163 void init_thread_stat(struct thread_stat *ts)
2165 memset(ts, 0, sizeof(*ts));
2167 init_thread_stat_min_vals(ts);
2171 void __show_run_stats(void)
2173 struct group_run_stats *runstats, *rs;
2174 struct thread_data *td;
2175 struct thread_stat *threadstats, *ts;
2176 int i, j, k, nr_ts, last_ts, idx;
2177 bool kb_base_warned = false;
2178 bool unit_base_warned = false;
2179 struct json_object *root = NULL;
2180 struct json_array *array = NULL;
2181 struct buf_output output[FIO_OUTPUT_NR];
2182 struct flist_head **opt_lists;
2184 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2186 for (i = 0; i < groupid + 1; i++)
2187 init_group_run_stat(&runstats[i]);
2190 * find out how many threads stats we need. if group reporting isn't
2191 * enabled, it's one-per-td.
2195 for_each_td(td, i) {
2196 if (!td->o.group_reporting) {
2200 if (last_ts == td->groupid)
2205 last_ts = td->groupid;
2209 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2210 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2212 for (i = 0; i < nr_ts; i++) {
2213 init_thread_stat(&threadstats[i]);
2214 opt_lists[i] = NULL;
2220 for_each_td(td, i) {
2223 if (idx && (!td->o.group_reporting ||
2224 (td->o.group_reporting && last_ts != td->groupid))) {
2229 last_ts = td->groupid;
2231 ts = &threadstats[j];
2233 ts->clat_percentiles = td->o.clat_percentiles;
2234 ts->lat_percentiles = td->o.lat_percentiles;
2235 ts->slat_percentiles = td->o.slat_percentiles;
2236 ts->percentile_precision = td->o.percentile_precision;
2237 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2238 opt_lists[j] = &td->opt_list;
2243 if (ts->groupid == -1) {
2245 * These are per-group shared already
2247 snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2248 if (td->o.description)
2249 snprintf(ts->description,
2250 sizeof(ts->description), "%s",
2253 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2256 * If multiple entries in this group, this is
2259 ts->thread_number = td->thread_number;
2260 ts->groupid = td->groupid;
2263 * first pid in group, not very useful...
2267 ts->kb_base = td->o.kb_base;
2268 ts->unit_base = td->o.unit_base;
2269 ts->sig_figs = td->o.sig_figs;
2270 ts->unified_rw_rep = td->o.unified_rw_rep;
2271 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2272 log_info("fio: kb_base differs for jobs in group, using"
2273 " %u as the base\n", ts->kb_base);
2274 kb_base_warned = true;
2275 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2276 log_info("fio: unit_base differs for jobs in group, using"
2277 " %u as the base\n", ts->unit_base);
2278 unit_base_warned = true;
2281 ts->continue_on_error = td->o.continue_on_error;
2282 ts->total_err_count += td->total_err_count;
2283 ts->first_error = td->first_error;
2285 if (!td->error && td->o.continue_on_error &&
2287 ts->error = td->first_error;
2288 snprintf(ts->verror, sizeof(ts->verror), "%s",
2290 } else if (td->error) {
2291 ts->error = td->error;
2292 snprintf(ts->verror, sizeof(ts->verror), "%s",
2297 ts->latency_depth = td->latency_qd;
2298 ts->latency_target = td->o.latency_target;
2299 ts->latency_percentile = td->o.latency_percentile;
2300 ts->latency_window = td->o.latency_window;
2302 ts->nr_block_infos = td->ts.nr_block_infos;
2303 for (k = 0; k < ts->nr_block_infos; k++)
2304 ts->block_infos[k] = td->ts.block_infos[k];
2306 sum_thread_stats(ts, &td->ts);
2309 ts->ss_state = td->ss.state;
2310 ts->ss_dur = td->ss.dur;
2311 ts->ss_head = td->ss.head;
2312 ts->ss_bw_data = td->ss.bw_data;
2313 ts->ss_iops_data = td->ss.iops_data;
2314 ts->ss_limit.u.f = td->ss.limit;
2315 ts->ss_slope.u.f = td->ss.slope;
2316 ts->ss_deviation.u.f = td->ss.deviation;
2317 ts->ss_criterion.u.f = td->ss.criterion;
2320 ts->ss_dur = ts->ss_state = 0;
2323 for (i = 0; i < nr_ts; i++) {
2324 unsigned long long bw;
2326 ts = &threadstats[i];
2327 if (ts->groupid == -1)
2329 rs = &runstats[ts->groupid];
2330 rs->kb_base = ts->kb_base;
2331 rs->unit_base = ts->unit_base;
2332 rs->sig_figs = ts->sig_figs;
2333 rs->unified_rw_rep |= ts->unified_rw_rep;
2335 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2336 if (!ts->runtime[j])
2338 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2339 rs->min_run[j] = ts->runtime[j];
2340 if (ts->runtime[j] > rs->max_run[j])
2341 rs->max_run[j] = ts->runtime[j];
2345 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2346 if (bw < rs->min_bw[j])
2348 if (bw > rs->max_bw[j])
2351 rs->iobytes[j] += ts->io_bytes[j];
2355 for (i = 0; i < groupid + 1; i++) {
2360 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2361 if (rs->max_run[ddir])
2362 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2367 for (i = 0; i < FIO_OUTPUT_NR; i++)
2368 buf_output_init(&output[i]);
2371 * don't overwrite last signal output
2373 if (output_format & FIO_OUTPUT_NORMAL)
2374 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2375 if (output_format & FIO_OUTPUT_JSON) {
2376 struct thread_data *global;
2379 unsigned long long ms_since_epoch;
2382 gettimeofday(&now, NULL);
2383 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2384 (unsigned long long)(now.tv_usec) / 1000;
2386 tv_sec = now.tv_sec;
2387 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2388 if (time_buf[strlen(time_buf) - 1] == '\n')
2389 time_buf[strlen(time_buf) - 1] = '\0';
2391 root = json_create_object();
2392 json_object_add_value_string(root, "fio version", fio_version_string);
2393 json_object_add_value_int(root, "timestamp", now.tv_sec);
2394 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2395 json_object_add_value_string(root, "time", time_buf);
2396 global = get_global_options();
2397 json_add_job_opts(root, "global options", &global->opt_list);
2398 array = json_create_array();
2399 json_object_add_value_array(root, "jobs", array);
2403 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2405 for (i = 0; i < nr_ts; i++) {
2406 ts = &threadstats[i];
2407 rs = &runstats[ts->groupid];
2410 fio_server_send_job_options(opt_lists[i], i);
2411 fio_server_send_ts(ts, rs);
2413 if (output_format & FIO_OUTPUT_TERSE)
2414 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2415 if (output_format & FIO_OUTPUT_JSON) {
2416 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2417 json_array_add_value_object(array, tmp);
2419 if (output_format & FIO_OUTPUT_NORMAL)
2420 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2423 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2424 /* disk util stats, if any */
2425 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2427 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2429 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2430 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2431 json_free_object(root);
2434 for (i = 0; i < groupid + 1; i++) {
2439 fio_server_send_gs(rs);
2440 else if (output_format & FIO_OUTPUT_NORMAL)
2441 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2445 fio_server_send_du();
2446 else if (output_format & FIO_OUTPUT_NORMAL) {
2447 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2448 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2451 for (i = 0; i < FIO_OUTPUT_NR; i++) {
2452 struct buf_output *out = &output[i];
2454 log_info_buf(out->buf, out->buflen);
2455 buf_output_free(out);
2458 fio_idle_prof_cleanup();
2466 int __show_running_run_stats(void)
2468 struct thread_data *td;
2469 unsigned long long *rt;
2473 fio_sem_down(stat_sem);
2475 rt = malloc(thread_number * sizeof(unsigned long long));
2476 fio_gettime(&ts, NULL);
2478 for_each_td(td, i) {
2479 td->update_rusage = 1;
2480 for_each_rw_ddir(ddir) {
2481 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2483 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2485 rt[i] = mtime_since(&td->start, &ts);
2486 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2487 td->ts.runtime[DDIR_READ] += rt[i];
2488 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2489 td->ts.runtime[DDIR_WRITE] += rt[i];
2490 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2491 td->ts.runtime[DDIR_TRIM] += rt[i];
2494 for_each_td(td, i) {
2495 if (td->runstate >= TD_EXITED)
2497 if (td->rusage_sem) {
2498 td->update_rusage = 1;
2499 fio_sem_down(td->rusage_sem);
2501 td->update_rusage = 0;
2506 for_each_td(td, i) {
2507 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2508 td->ts.runtime[DDIR_READ] -= rt[i];
2509 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2510 td->ts.runtime[DDIR_WRITE] -= rt[i];
2511 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2512 td->ts.runtime[DDIR_TRIM] -= rt[i];
2516 fio_sem_up(stat_sem);
2521 static bool status_file_disabled;
2523 #define FIO_STATUS_FILE "fio-dump-status"
2525 static int check_status_file(void)
2528 const char *temp_dir;
2529 char fio_status_file_path[PATH_MAX];
2531 if (status_file_disabled)
2534 temp_dir = getenv("TMPDIR");
2535 if (temp_dir == NULL) {
2536 temp_dir = getenv("TEMP");
2537 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2540 if (temp_dir == NULL)
2543 __coverity_tainted_data_sanitize__(temp_dir);
2546 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2548 if (stat(fio_status_file_path, &sb))
2551 if (unlink(fio_status_file_path) < 0) {
2552 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2554 log_err("fio: disabling status file updates\n");
2555 status_file_disabled = true;
2561 void check_for_running_stats(void)
2563 if (check_status_file()) {
2564 show_running_run_stats();
2569 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2574 if (data > is->max_val)
2576 if (data < is->min_val)
2579 delta = val - is->mean.u.f;
2581 is->mean.u.f += delta / (is->samples + 1.0);
2582 is->S.u.f += delta * (val - is->mean.u.f);
2589 * Return a struct io_logs, which is added to the tail of the log
2592 static struct io_logs *get_new_log(struct io_log *iolog)
2595 struct io_logs *cur_log;
2598 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2601 if (!iolog->cur_log_max) {
2602 new_samples = iolog->td->o.log_entries;
2604 new_samples = iolog->cur_log_max * 2;
2605 if (new_samples > MAX_LOG_ENTRIES)
2606 new_samples = MAX_LOG_ENTRIES;
2609 cur_log = smalloc(sizeof(*cur_log));
2611 INIT_FLIST_HEAD(&cur_log->list);
2612 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2614 cur_log->nr_samples = 0;
2615 cur_log->max_samples = new_samples;
2616 flist_add_tail(&cur_log->list, &iolog->io_logs);
2617 iolog->cur_log_max = new_samples;
2627 * Add and return a new log chunk, or return current log if big enough
2629 static struct io_logs *regrow_log(struct io_log *iolog)
2631 struct io_logs *cur_log;
2634 if (!iolog || iolog->disabled)
2637 cur_log = iolog_cur_log(iolog);
2639 cur_log = get_new_log(iolog);
2644 if (cur_log->nr_samples < cur_log->max_samples)
2648 * No room for a new sample. If we're compressing on the fly, flush
2649 * out the current chunk
2651 if (iolog->log_gz) {
2652 if (iolog_cur_flush(iolog, cur_log)) {
2653 log_err("fio: failed flushing iolog! Will stop logging.\n");
2659 * Get a new log array, and add to our list
2661 cur_log = get_new_log(iolog);
2663 log_err("fio: failed extending iolog! Will stop logging.\n");
2667 if (!iolog->pending || !iolog->pending->nr_samples)
2671 * Flush pending items to new log
2673 for (i = 0; i < iolog->pending->nr_samples; i++) {
2674 struct io_sample *src, *dst;
2676 src = get_sample(iolog, iolog->pending, i);
2677 dst = get_sample(iolog, cur_log, i);
2678 memcpy(dst, src, log_entry_sz(iolog));
2680 cur_log->nr_samples = iolog->pending->nr_samples;
2682 iolog->pending->nr_samples = 0;
2686 iolog->disabled = true;
2690 void regrow_logs(struct thread_data *td)
2692 regrow_log(td->slat_log);
2693 regrow_log(td->clat_log);
2694 regrow_log(td->clat_hist_log);
2695 regrow_log(td->lat_log);
2696 regrow_log(td->bw_log);
2697 regrow_log(td->iops_log);
2698 td->flags &= ~TD_F_REGROW_LOGS;
2701 void regrow_agg_logs(void)
2705 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2706 regrow_log(agg_io_log[ddir]);
2709 static struct io_logs *get_cur_log(struct io_log *iolog)
2711 struct io_logs *cur_log;
2713 cur_log = iolog_cur_log(iolog);
2715 cur_log = get_new_log(iolog);
2720 if (cur_log->nr_samples < cur_log->max_samples)
2724 * Out of space. If we're in IO offload mode, or we're not doing
2725 * per unit logging (hence logging happens outside of the IO thread
2726 * as well), add a new log chunk inline. If we're doing inline
2727 * submissions, flag 'td' as needing a log regrow and we'll take
2728 * care of it on the submission side.
2730 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2731 !per_unit_log(iolog))
2732 return regrow_log(iolog);
2735 iolog->td->flags |= TD_F_REGROW_LOGS;
2737 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2738 return iolog->pending;
2741 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2742 enum fio_ddir ddir, unsigned long long bs,
2743 unsigned long t, uint64_t offset,
2744 unsigned int priority)
2746 struct io_logs *cur_log;
2748 if (iolog->disabled)
2750 if (flist_empty(&iolog->io_logs))
2751 iolog->avg_last[ddir] = t;
2753 cur_log = get_cur_log(iolog);
2755 struct io_sample *s;
2757 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2760 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2761 io_sample_set_ddir(iolog, s, ddir);
2763 s->priority = priority;
2765 if (iolog->log_offset) {
2766 struct io_sample_offset *so = (void *) s;
2768 so->offset = offset;
2771 cur_log->nr_samples++;
2775 iolog->disabled = true;
2778 static inline void reset_io_stat(struct io_stat *ios)
2780 ios->min_val = -1ULL;
2781 ios->max_val = ios->samples = 0;
2782 ios->mean.u.f = ios->S.u.f = 0;
2785 void reset_io_stats(struct thread_data *td)
2787 struct thread_stat *ts = &td->ts;
2790 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2791 reset_io_stat(&ts->clat_high_prio_stat[i]);
2792 reset_io_stat(&ts->clat_low_prio_stat[i]);
2793 reset_io_stat(&ts->clat_stat[i]);
2794 reset_io_stat(&ts->slat_stat[i]);
2795 reset_io_stat(&ts->lat_stat[i]);
2796 reset_io_stat(&ts->bw_stat[i]);
2797 reset_io_stat(&ts->iops_stat[i]);
2799 ts->io_bytes[i] = 0;
2801 ts->total_io_u[i] = 0;
2802 ts->short_io_u[i] = 0;
2803 ts->drop_io_u[i] = 0;
2805 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2806 ts->io_u_plat_high_prio[i][j] = 0;
2807 ts->io_u_plat_low_prio[i][j] = 0;
2809 ts->io_u_sync_plat[j] = 0;
2813 for (i = 0; i < FIO_LAT_CNT; i++)
2814 for (j = 0; j < DDIR_RWDIR_CNT; j++)
2815 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2816 ts->io_u_plat[i][j][k] = 0;
2818 ts->total_io_u[DDIR_SYNC] = 0;
2820 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2821 ts->io_u_map[i] = 0;
2822 ts->io_u_submit[i] = 0;
2823 ts->io_u_complete[i] = 0;
2826 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2827 ts->io_u_lat_n[i] = 0;
2828 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2829 ts->io_u_lat_u[i] = 0;
2830 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2831 ts->io_u_lat_m[i] = 0;
2833 ts->total_submit = 0;
2834 ts->total_complete = 0;
2835 ts->nr_zone_resets = 0;
2836 ts->cachehit = ts->cachemiss = 0;
2839 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2840 unsigned long elapsed, bool log_max)
2843 * Note an entry in the log. Use the mean from the logged samples,
2844 * making sure to properly round up. Only write a log entry if we
2845 * had actual samples done.
2847 if (iolog->avg_window[ddir].samples) {
2848 union io_sample_data data;
2851 data.val = iolog->avg_window[ddir].max_val;
2853 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2855 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
2858 reset_io_stat(&iolog->avg_window[ddir]);
2861 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2866 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2867 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2870 static unsigned long add_log_sample(struct thread_data *td,
2871 struct io_log *iolog,
2872 union io_sample_data data,
2873 enum fio_ddir ddir, unsigned long long bs,
2874 uint64_t offset, unsigned int ioprio)
2876 unsigned long elapsed, this_window;
2881 elapsed = mtime_since_now(&td->epoch);
2884 * If no time averaging, just add the log sample.
2886 if (!iolog->avg_msec) {
2887 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
2893 * Add the sample. If the time period has passed, then
2894 * add that entry to the log and clear.
2896 add_stat_sample(&iolog->avg_window[ddir], data.val);
2899 * If period hasn't passed, adding the above sample is all we
2902 this_window = elapsed - iolog->avg_last[ddir];
2903 if (elapsed < iolog->avg_last[ddir])
2904 return iolog->avg_last[ddir] - elapsed;
2905 else if (this_window < iolog->avg_msec) {
2906 unsigned long diff = iolog->avg_msec - this_window;
2908 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2912 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2914 iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
2916 return iolog->avg_msec;
2919 void finalize_logs(struct thread_data *td, bool unit_logs)
2921 unsigned long elapsed;
2923 elapsed = mtime_since_now(&td->epoch);
2925 if (td->clat_log && unit_logs)
2926 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2927 if (td->slat_log && unit_logs)
2928 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2929 if (td->lat_log && unit_logs)
2930 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2931 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2932 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2933 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2934 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2937 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
2938 unsigned long long bs)
2940 struct io_log *iolog;
2945 iolog = agg_io_log[ddir];
2946 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
2949 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2951 unsigned int idx = plat_val_to_idx(nsec);
2952 assert(idx < FIO_IO_U_PLAT_NR);
2954 ts->io_u_sync_plat[idx]++;
2955 add_stat_sample(&ts->sync_stat, nsec);
2958 static inline void add_lat_percentile_sample(struct thread_stat *ts,
2959 unsigned long long nsec,
2963 unsigned int idx = plat_val_to_idx(nsec);
2964 assert(idx < FIO_IO_U_PLAT_NR);
2966 ts->io_u_plat[lat][ddir][idx]++;
2969 static inline void add_lat_percentile_prio_sample(struct thread_stat *ts,
2970 unsigned long long nsec,
2974 unsigned int idx = plat_val_to_idx(nsec);
2977 ts->io_u_plat_low_prio[ddir][idx]++;
2979 ts->io_u_plat_high_prio[ddir][idx]++;
2982 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2983 unsigned long long nsec, unsigned long long bs,
2984 uint64_t offset, unsigned int ioprio, bool high_prio)
2986 const bool needs_lock = td_async_processing(td);
2987 unsigned long elapsed, this_window;
2988 struct thread_stat *ts = &td->ts;
2989 struct io_log *iolog = td->clat_hist_log;
2994 add_stat_sample(&ts->clat_stat[ddir], nsec);
2997 * When lat_percentiles=1 (default 0), the reported high/low priority
2998 * percentiles and stats are used for describing total latency values,
2999 * even though the variable names themselves start with clat_.
3001 * Because of the above definition, add a prio stat sample only when
3002 * lat_percentiles=0. add_lat_sample() will add the prio stat sample
3003 * when lat_percentiles=1.
3005 if (!ts->lat_percentiles) {
3007 add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
3009 add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
3013 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
3016 if (ts->clat_percentiles) {
3018 * Because of the above definition, add a prio lat percentile
3019 * sample only when lat_percentiles=0. add_lat_sample() will add
3020 * the prio lat percentile sample when lat_percentiles=1.
3022 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3023 if (!ts->lat_percentiles)
3024 add_lat_percentile_prio_sample(ts, nsec, ddir,
3028 if (iolog && iolog->hist_msec) {
3029 struct io_hist *hw = &iolog->hist_window[ddir];
3032 elapsed = mtime_since_now(&td->epoch);
3034 hw->hist_last = elapsed;
3035 this_window = elapsed - hw->hist_last;
3037 if (this_window >= iolog->hist_msec) {
3038 uint64_t *io_u_plat;
3039 struct io_u_plat_entry *dst;
3042 * Make a byte-for-byte copy of the latency histogram
3043 * stored in td->ts.io_u_plat[ddir], recording it in a
3044 * log sample. Note that the matching call to free() is
3045 * located in iolog.c after printing this sample to the
3048 io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3049 dst = malloc(sizeof(struct io_u_plat_entry));
3050 memcpy(&(dst->io_u_plat), io_u_plat,
3051 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3052 flist_add(&dst->list, &hw->list);
3053 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3054 elapsed, offset, ioprio);
3057 * Update the last time we recorded as being now, minus
3058 * any drift in time we encountered before actually
3059 * making the record.
3061 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3067 __td_io_u_unlock(td);
3070 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3071 unsigned long long nsec, unsigned long long bs,
3072 uint64_t offset, unsigned int ioprio)
3074 const bool needs_lock = td_async_processing(td);
3075 struct thread_stat *ts = &td->ts;
3083 add_stat_sample(&ts->slat_stat[ddir], nsec);
3086 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3089 if (ts->slat_percentiles)
3090 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3093 __td_io_u_unlock(td);
3096 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3097 unsigned long long nsec, unsigned long long bs,
3098 uint64_t offset, unsigned int ioprio, bool high_prio)
3100 const bool needs_lock = td_async_processing(td);
3101 struct thread_stat *ts = &td->ts;
3109 add_stat_sample(&ts->lat_stat[ddir], nsec);
3112 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3116 * When lat_percentiles=1 (default 0), the reported high/low priority
3117 * percentiles and stats are used for describing total latency values,
3118 * even though the variable names themselves start with clat_.
3120 * Because of the above definition, add a prio stat and prio lat
3121 * percentile sample only when lat_percentiles=1. add_clat_sample() will
3122 * add the prio stat and prio lat percentile sample when
3123 * lat_percentiles=0.
3125 if (ts->lat_percentiles) {
3126 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3127 add_lat_percentile_prio_sample(ts, nsec, ddir, high_prio);
3129 add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
3131 add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
3135 __td_io_u_unlock(td);
3138 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3139 unsigned int bytes, unsigned long long spent)
3141 const bool needs_lock = td_async_processing(td);
3142 struct thread_stat *ts = &td->ts;
3146 rate = (unsigned long) (bytes * 1000000ULL / spent);
3153 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3156 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3157 bytes, io_u->offset, io_u->ioprio);
3159 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3162 __td_io_u_unlock(td);
3165 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3166 struct timespec *t, unsigned int avg_time,
3167 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3168 struct io_stat *stat, struct io_log *log,
3171 const bool needs_lock = td_async_processing(td);
3172 unsigned long spent, rate;
3174 unsigned long next, next_log;
3176 next_log = avg_time;
3178 spent = mtime_since(parent_tv, t);
3179 if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3180 return avg_time - spent;
3186 * Compute both read and write rates for the interval.
3188 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3191 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3193 continue; /* No entries for interval */
3197 rate = delta * 1000 / spent / 1024; /* KiB/s */
3199 rate = (delta * 1000) / spent;
3203 add_stat_sample(&stat[ddir], rate);
3206 unsigned long long bs = 0;
3208 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3209 bs = td->o.min_bs[ddir];
3211 next = add_log_sample(td, log, sample_val(rate), ddir,
3213 next_log = min(next_log, next);
3216 stat_io_bytes[ddir] = this_io_bytes[ddir];
3222 __td_io_u_unlock(td);
3224 if (spent <= avg_time)
3227 next = avg_time - (1 + spent - avg_time);
3229 return min(next, next_log);
3232 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3234 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3235 td->this_io_bytes, td->stat_io_bytes,
3236 td->ts.bw_stat, td->bw_log, true);
3239 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3242 const bool needs_lock = td_async_processing(td);
3243 struct thread_stat *ts = &td->ts;
3248 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3251 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3252 bytes, io_u->offset, io_u->ioprio);
3254 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3257 __td_io_u_unlock(td);
3260 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3262 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3263 td->this_io_blocks, td->stat_io_blocks,
3264 td->ts.iops_stat, td->iops_log, false);
3268 * Returns msecs to next event
3270 int calc_log_samples(void)
3272 struct thread_data *td;
3273 unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3274 struct timespec now;
3276 long elapsed_time = 0;
3278 fio_gettime(&now, NULL);
3280 for_each_td(td, i) {
3281 elapsed_time = mtime_since_now(&td->epoch);
3285 if (in_ramp_time(td) ||
3286 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3287 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3291 (td->bw_log && !per_unit_log(td->bw_log))) {
3292 tmp = add_bw_samples(td, &now);
3295 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3297 if (!td->iops_log ||
3298 (td->iops_log && !per_unit_log(td->iops_log))) {
3299 tmp = add_iops_samples(td, &now);
3302 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3309 /* if log_avg_msec_min has not been changed, set it to 0 */
3310 if (log_avg_msec_min == -1U)
3311 log_avg_msec_min = 0;
3313 if (log_avg_msec_min == 0)
3314 next_mod = elapsed_time;
3316 next_mod = elapsed_time % log_avg_msec_min;
3318 /* correction to keep the time on the log avg msec boundary */
3319 next = min(next, (log_avg_msec_min - next_mod));
3321 return next == ~0U ? 0 : next;
3324 void stat_init(void)
3326 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3329 void stat_exit(void)
3332 * When we have the mutex, we know out-of-band access to it
3335 fio_sem_down(stat_sem);
3336 fio_sem_remove(stat_sem);
3340 * Called from signal handler. Wake up status thread.
3342 void show_running_run_stats(void)
3347 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3349 /* Ignore io_u's which span multiple blocks--they will just get
3350 * inaccurate counts. */
3351 int idx = (io_u->offset - io_u->file->file_offset)
3352 / td->o.bs[DDIR_TRIM];
3353 uint32_t *info = &td->ts.block_infos[idx];
3354 assert(idx < td->ts.nr_block_infos);