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 double convert_agg_kbytes_percent(struct group_run_stats *rs, int ddir, int mean)
467 double p_of_agg = 100.0;
468 if (rs && rs->agg[ddir] > 1024) {
469 p_of_agg = mean * 100.0 / (double) (rs->agg[ddir] / 1024.0);
471 if (p_of_agg > 100.0)
477 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
478 int ddir, struct buf_output *out)
481 unsigned long long min, max, bw, iops;
483 char *io_p, *bw_p, *bw_p_alt, *iops_p, *post_st = NULL;
486 if (ddir_sync(ddir)) {
487 if (calc_lat(&ts->sync_stat, &min, &max, &mean, &dev)) {
488 log_buf(out, " %s:\n", "fsync/fdatasync/sync_file_range");
489 display_lat(io_ddir_name(ddir), min, max, mean, dev, out);
490 show_clat_percentiles(ts->io_u_sync_plat,
491 ts->sync_stat.samples,
493 ts->percentile_precision,
494 io_ddir_name(ddir), out);
499 assert(ddir_rw(ddir));
501 if (!ts->runtime[ddir])
504 i2p = is_power_of_2(rs->kb_base);
505 runt = ts->runtime[ddir];
507 bw = (1000 * ts->io_bytes[ddir]) / runt;
508 io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
509 bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
510 bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
512 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
513 iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
514 if (ddir == DDIR_WRITE)
515 post_st = zbd_write_status(ts);
516 else if (ddir == DDIR_READ && ts->cachehit && ts->cachemiss) {
520 total = ts->cachehit + ts->cachemiss;
521 hit = (double) ts->cachehit / (double) total;
523 if (asprintf(&post_st, "; Cachehit=%0.2f%%", hit) < 0)
527 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)%s\n",
528 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir),
529 iops_p, bw_p, bw_p_alt, io_p,
530 (unsigned long long) ts->runtime[ddir],
539 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
540 display_lat("slat", min, max, mean, dev, out);
541 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
542 display_lat("clat", min, max, mean, dev, out);
543 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
544 display_lat(" lat", min, max, mean, dev, out);
545 if (calc_lat(&ts->clat_high_prio_stat[ddir], &min, &max, &mean, &dev)) {
546 display_lat(ts->lat_percentiles ? "high prio_lat" : "high prio_clat",
547 min, max, mean, dev, out);
548 if (calc_lat(&ts->clat_low_prio_stat[ddir], &min, &max, &mean, &dev))
549 display_lat(ts->lat_percentiles ? "low prio_lat" : "low prio_clat",
550 min, max, mean, dev, out);
553 if (ts->slat_percentiles && ts->slat_stat[ddir].samples > 0)
554 show_clat_percentiles(ts->io_u_plat[FIO_SLAT][ddir],
555 ts->slat_stat[ddir].samples,
557 ts->percentile_precision, "slat", out);
558 if (ts->clat_percentiles && ts->clat_stat[ddir].samples > 0)
559 show_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
560 ts->clat_stat[ddir].samples,
562 ts->percentile_precision, "clat", out);
563 if (ts->lat_percentiles && ts->lat_stat[ddir].samples > 0)
564 show_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
565 ts->lat_stat[ddir].samples,
567 ts->percentile_precision, "lat", out);
569 if (ts->clat_percentiles || ts->lat_percentiles) {
570 const char *name = ts->lat_percentiles ? "lat" : "clat";
574 if (ts->lat_percentiles)
575 samples = ts->lat_stat[ddir].samples;
577 samples = ts->clat_stat[ddir].samples;
579 /* Only print this if some high and low priority stats were collected */
580 if (ts->clat_high_prio_stat[ddir].samples > 0 &&
581 ts->clat_low_prio_stat[ddir].samples > 0)
583 sprintf(prio_name, "high prio (%.2f%%) %s",
584 100. * (double) ts->clat_high_prio_stat[ddir].samples / (double) samples,
586 show_clat_percentiles(ts->io_u_plat_high_prio[ddir],
587 ts->clat_high_prio_stat[ddir].samples,
589 ts->percentile_precision, prio_name, out);
591 sprintf(prio_name, "low prio (%.2f%%) %s",
592 100. * (double) ts->clat_low_prio_stat[ddir].samples / (double) samples,
594 show_clat_percentiles(ts->io_u_plat_low_prio[ddir],
595 ts->clat_low_prio_stat[ddir].samples,
597 ts->percentile_precision, prio_name, out);
601 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
602 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
605 if ((rs->unit_base == 1) && i2p)
607 else if (rs->unit_base == 1)
614 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
616 if (rs->unit_base == 1) {
623 if (mean > fkb_base * fkb_base) {
628 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
631 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
632 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
633 bw_str, min, max, p_of_agg, mean, dev,
634 (&ts->bw_stat[ddir])->samples);
636 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
637 log_buf(out, " iops : min=%5llu, max=%5llu, "
638 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
639 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
643 static void show_mixed_ddir_status(struct group_run_stats *rs,
644 struct thread_stat *ts,
645 struct buf_output *out)
647 struct thread_stat *ts_lcl;
650 * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
653 ts_lcl = malloc(sizeof(struct thread_stat));
654 memset((void *)ts_lcl, 0, sizeof(struct thread_stat));
655 /* calculate mixed stats */
656 ts_lcl->unified_rw_rep = UNIFIED_MIXED;
657 init_thread_stat_min_vals(ts_lcl);
658 ts_lcl->lat_percentiles = ts->lat_percentiles;
659 ts_lcl->clat_percentiles = ts->clat_percentiles;
660 ts_lcl->slat_percentiles = ts->slat_percentiles;
661 ts_lcl->percentile_precision = ts->percentile_precision;
662 memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
664 sum_thread_stats(ts_lcl, ts);
666 show_ddir_status(rs, ts_lcl, DDIR_READ, out);
670 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
671 const char *msg, struct buf_output *out)
673 bool new_line = true, shown = false;
676 for (i = 0; i < nr; i++) {
677 if (io_u_lat[i] <= 0.0)
683 log_buf(out, " lat (%s) : ", msg);
689 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
701 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
703 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
704 "250=", "500=", "750=", "1000=", };
706 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
709 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
711 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
712 "250=", "500=", "750=", "1000=", };
714 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
717 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
719 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
720 "250=", "500=", "750=", "1000=", "2000=",
723 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
726 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
728 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
729 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
730 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
732 stat_calc_lat_n(ts, io_u_lat_n);
733 stat_calc_lat_u(ts, io_u_lat_u);
734 stat_calc_lat_m(ts, io_u_lat_m);
736 show_lat_n(io_u_lat_n, out);
737 show_lat_u(io_u_lat_u, out);
738 show_lat_m(io_u_lat_m, out);
741 static int block_state_category(int block_state)
743 switch (block_state) {
744 case BLOCK_STATE_UNINIT:
746 case BLOCK_STATE_TRIMMED:
747 case BLOCK_STATE_WRITTEN:
749 case BLOCK_STATE_WRITE_FAILURE:
750 case BLOCK_STATE_TRIM_FAILURE:
753 /* Silence compile warning on some BSDs and have a return */
759 static int compare_block_infos(const void *bs1, const void *bs2)
761 uint64_t block1 = *(uint64_t *)bs1;
762 uint64_t block2 = *(uint64_t *)bs2;
763 int state1 = BLOCK_INFO_STATE(block1);
764 int state2 = BLOCK_INFO_STATE(block2);
765 int bscat1 = block_state_category(state1);
766 int bscat2 = block_state_category(state2);
767 int cycles1 = BLOCK_INFO_TRIMS(block1);
768 int cycles2 = BLOCK_INFO_TRIMS(block2);
775 if (cycles1 < cycles2)
777 if (cycles1 > cycles2)
785 assert(block1 == block2);
789 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
790 fio_fp64_t *plist, unsigned int **percentiles,
796 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
798 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
805 * Sort the percentile list. Note that it may already be sorted if
806 * we are using the default values, but since it's a short list this
807 * isn't a worry. Also note that this does not work for NaN values.
810 qsort(plist, len, sizeof(plist[0]), double_cmp);
812 /* Start only after the uninit entries end */
814 nr_uninit < nr_block_infos
815 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
819 if (nr_uninit == nr_block_infos)
822 *percentiles = calloc(len, sizeof(**percentiles));
824 for (i = 0; i < len; i++) {
825 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
827 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
830 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
831 for (i = 0; i < nr_block_infos; i++)
832 types[BLOCK_INFO_STATE(block_infos[i])]++;
837 static const char *block_state_names[] = {
838 [BLOCK_STATE_UNINIT] = "unwritten",
839 [BLOCK_STATE_TRIMMED] = "trimmed",
840 [BLOCK_STATE_WRITTEN] = "written",
841 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
842 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
845 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
846 fio_fp64_t *plist, struct buf_output *out)
849 unsigned int *percentiles = NULL;
850 unsigned int block_state_counts[BLOCK_STATE_COUNT];
852 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
853 &percentiles, block_state_counts);
855 log_buf(out, " block lifetime percentiles :\n |");
857 for (i = 0; i < len; i++) {
858 uint32_t block_info = percentiles[i];
859 #define LINE_LENGTH 75
860 char str[LINE_LENGTH];
861 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
862 plist[i].u.f, block_info,
863 i == len - 1 ? '\n' : ',');
864 assert(strln < LINE_LENGTH);
865 if (pos + strln > LINE_LENGTH) {
867 log_buf(out, "\n |");
869 log_buf(out, "%s", str);
876 log_buf(out, " states :");
877 for (i = 0; i < BLOCK_STATE_COUNT; i++)
878 log_buf(out, " %s=%u%c",
879 block_state_names[i], block_state_counts[i],
880 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
883 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
885 char *p1, *p1alt, *p2;
886 unsigned long long bw_mean, iops_mean;
887 const int i2p = is_power_of_2(ts->kb_base);
892 bw_mean = steadystate_bw_mean(ts);
893 iops_mean = steadystate_iops_mean(ts);
895 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
896 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
897 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
899 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
900 ts->ss_state & FIO_SS_ATTAINED ? "yes" : "no",
902 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
903 ts->ss_state & FIO_SS_SLOPE ? " slope": " mean dev",
904 ts->ss_criterion.u.f,
905 ts->ss_state & FIO_SS_PCT ? "%" : "");
912 static void show_agg_stats(struct disk_util_agg *agg, int terse,
913 struct buf_output *out)
915 if (!agg->slavecount)
919 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
920 "aggrticks=%llu/%llu, aggrin_queue=%llu, "
922 (unsigned long long) agg->ios[0] / agg->slavecount,
923 (unsigned long long) agg->ios[1] / agg->slavecount,
924 (unsigned long long) agg->merges[0] / agg->slavecount,
925 (unsigned long long) agg->merges[1] / agg->slavecount,
926 (unsigned long long) agg->ticks[0] / agg->slavecount,
927 (unsigned long long) agg->ticks[1] / agg->slavecount,
928 (unsigned long long) agg->time_in_queue / agg->slavecount,
931 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
932 (unsigned long long) agg->ios[0] / agg->slavecount,
933 (unsigned long long) agg->ios[1] / agg->slavecount,
934 (unsigned long long) agg->merges[0] / agg->slavecount,
935 (unsigned long long) agg->merges[1] / agg->slavecount,
936 (unsigned long long) agg->ticks[0] / agg->slavecount,
937 (unsigned long long) agg->ticks[1] / agg->slavecount,
938 (unsigned long long) agg->time_in_queue / agg->slavecount,
943 static void aggregate_slaves_stats(struct disk_util *masterdu)
945 struct disk_util_agg *agg = &masterdu->agg;
946 struct disk_util_stat *dus;
947 struct flist_head *entry;
948 struct disk_util *slavedu;
951 flist_for_each(entry, &masterdu->slaves) {
952 slavedu = flist_entry(entry, struct disk_util, slavelist);
954 agg->ios[0] += dus->s.ios[0];
955 agg->ios[1] += dus->s.ios[1];
956 agg->merges[0] += dus->s.merges[0];
957 agg->merges[1] += dus->s.merges[1];
958 agg->sectors[0] += dus->s.sectors[0];
959 agg->sectors[1] += dus->s.sectors[1];
960 agg->ticks[0] += dus->s.ticks[0];
961 agg->ticks[1] += dus->s.ticks[1];
962 agg->time_in_queue += dus->s.time_in_queue;
965 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
966 /* System utilization is the utilization of the
967 * component with the highest utilization.
969 if (util > agg->max_util.u.f)
970 agg->max_util.u.f = util;
974 if (agg->max_util.u.f > 100.0)
975 agg->max_util.u.f = 100.0;
978 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
979 int terse, struct buf_output *out)
984 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
992 log_buf(out, " %s: ios=%llu/%llu, merge=%llu/%llu, "
993 "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
995 (unsigned long long) dus->s.ios[0],
996 (unsigned long long) dus->s.ios[1],
997 (unsigned long long) dus->s.merges[0],
998 (unsigned long long) dus->s.merges[1],
999 (unsigned long long) dus->s.ticks[0],
1000 (unsigned long long) dus->s.ticks[1],
1001 (unsigned long long) dus->s.time_in_queue,
1004 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
1006 (unsigned long long) dus->s.ios[0],
1007 (unsigned long long) dus->s.ios[1],
1008 (unsigned long long) dus->s.merges[0],
1009 (unsigned long long) dus->s.merges[1],
1010 (unsigned long long) dus->s.ticks[0],
1011 (unsigned long long) dus->s.ticks[1],
1012 (unsigned long long) dus->s.time_in_queue,
1017 * If the device has slaves, aggregate the stats for
1018 * those slave devices also.
1020 show_agg_stats(agg, terse, out);
1026 void json_array_add_disk_util(struct disk_util_stat *dus,
1027 struct disk_util_agg *agg, struct json_array *array)
1029 struct json_object *obj;
1033 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
1037 obj = json_create_object();
1038 json_array_add_value_object(array, obj);
1040 json_object_add_value_string(obj, "name", (const char *)dus->name);
1041 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
1042 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
1043 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
1044 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
1045 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
1046 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
1047 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
1048 json_object_add_value_float(obj, "util", util);
1051 * If the device has slaves, aggregate the stats for
1052 * those slave devices also.
1054 if (!agg->slavecount)
1056 json_object_add_value_int(obj, "aggr_read_ios",
1057 agg->ios[0] / agg->slavecount);
1058 json_object_add_value_int(obj, "aggr_write_ios",
1059 agg->ios[1] / agg->slavecount);
1060 json_object_add_value_int(obj, "aggr_read_merges",
1061 agg->merges[0] / agg->slavecount);
1062 json_object_add_value_int(obj, "aggr_write_merge",
1063 agg->merges[1] / agg->slavecount);
1064 json_object_add_value_int(obj, "aggr_read_ticks",
1065 agg->ticks[0] / agg->slavecount);
1066 json_object_add_value_int(obj, "aggr_write_ticks",
1067 agg->ticks[1] / agg->slavecount);
1068 json_object_add_value_int(obj, "aggr_in_queue",
1069 agg->time_in_queue / agg->slavecount);
1070 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
1073 static void json_object_add_disk_utils(struct json_object *obj,
1074 struct flist_head *head)
1076 struct json_array *array = json_create_array();
1077 struct flist_head *entry;
1078 struct disk_util *du;
1080 json_object_add_value_array(obj, "disk_util", array);
1082 flist_for_each(entry, head) {
1083 du = flist_entry(entry, struct disk_util, list);
1085 aggregate_slaves_stats(du);
1086 json_array_add_disk_util(&du->dus, &du->agg, array);
1090 void show_disk_util(int terse, struct json_object *parent,
1091 struct buf_output *out)
1093 struct flist_head *entry;
1094 struct disk_util *du;
1097 if (!is_running_backend())
1100 if (flist_empty(&disk_list))
1103 if ((output_format & FIO_OUTPUT_JSON) && parent)
1108 if (!terse && !do_json)
1109 log_buf(out, "\nDisk stats (read/write):\n");
1112 json_object_add_disk_utils(parent, &disk_list);
1113 } else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
1114 flist_for_each(entry, &disk_list) {
1115 du = flist_entry(entry, struct disk_util, list);
1117 aggregate_slaves_stats(du);
1118 print_disk_util(&du->dus, &du->agg, terse, out);
1123 static void show_thread_status_normal(struct thread_stat *ts,
1124 struct group_run_stats *rs,
1125 struct buf_output *out)
1127 double usr_cpu, sys_cpu;
1128 unsigned long runtime;
1129 double io_u_dist[FIO_IO_U_MAP_NR];
1133 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
1136 memset(time_buf, 0, sizeof(time_buf));
1139 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
1142 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
1143 ts->name, ts->groupid, ts->members,
1144 ts->error, (int) ts->pid, time_buf);
1146 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
1147 ts->name, ts->groupid, ts->members,
1148 ts->error, ts->verror, (int) ts->pid,
1152 if (strlen(ts->description))
1153 log_buf(out, " Description : [%s]\n", ts->description);
1155 for_each_rw_ddir(ddir) {
1156 if (ts->io_bytes[ddir])
1157 show_ddir_status(rs, ts, ddir, out);
1160 if (ts->unified_rw_rep == UNIFIED_BOTH)
1161 show_mixed_ddir_status(rs, ts, out);
1163 show_latencies(ts, out);
1165 if (ts->sync_stat.samples)
1166 show_ddir_status(rs, ts, DDIR_SYNC, out);
1168 runtime = ts->total_run_time;
1170 double runt = (double) runtime;
1172 usr_cpu = (double) ts->usr_time * 100 / runt;
1173 sys_cpu = (double) ts->sys_time * 100 / runt;
1179 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
1180 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
1181 (unsigned long long) ts->ctx,
1182 (unsigned long long) ts->majf,
1183 (unsigned long long) ts->minf);
1185 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1186 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
1187 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1188 io_u_dist[1], io_u_dist[2],
1189 io_u_dist[3], io_u_dist[4],
1190 io_u_dist[5], io_u_dist[6]);
1192 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1193 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1194 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1195 io_u_dist[1], io_u_dist[2],
1196 io_u_dist[3], io_u_dist[4],
1197 io_u_dist[5], io_u_dist[6]);
1198 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1199 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
1200 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
1201 io_u_dist[1], io_u_dist[2],
1202 io_u_dist[3], io_u_dist[4],
1203 io_u_dist[5], io_u_dist[6]);
1204 log_buf(out, " issued rwts: total=%llu,%llu,%llu,%llu"
1205 " short=%llu,%llu,%llu,0"
1206 " dropped=%llu,%llu,%llu,0\n",
1207 (unsigned long long) ts->total_io_u[0],
1208 (unsigned long long) ts->total_io_u[1],
1209 (unsigned long long) ts->total_io_u[2],
1210 (unsigned long long) ts->total_io_u[3],
1211 (unsigned long long) ts->short_io_u[0],
1212 (unsigned long long) ts->short_io_u[1],
1213 (unsigned long long) ts->short_io_u[2],
1214 (unsigned long long) ts->drop_io_u[0],
1215 (unsigned long long) ts->drop_io_u[1],
1216 (unsigned long long) ts->drop_io_u[2]);
1217 if (ts->continue_on_error) {
1218 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
1219 (unsigned long long)ts->total_err_count,
1221 strerror(ts->first_error));
1223 if (ts->latency_depth) {
1224 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
1225 (unsigned long long)ts->latency_target,
1226 (unsigned long long)ts->latency_window,
1227 ts->latency_percentile.u.f,
1231 if (ts->nr_block_infos)
1232 show_block_infos(ts->nr_block_infos, ts->block_infos,
1233 ts->percentile_list, out);
1236 show_ss_normal(ts, out);
1239 static void show_ddir_status_terse(struct thread_stat *ts,
1240 struct group_run_stats *rs, int ddir,
1241 int ver, struct buf_output *out)
1243 unsigned long long min, max, minv, maxv, bw, iops;
1244 unsigned long long *ovals = NULL;
1249 assert(ddir_rw(ddir));
1252 if (ts->runtime[ddir]) {
1253 uint64_t runt = ts->runtime[ddir];
1255 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
1256 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
1259 log_buf(out, ";%llu;%llu;%llu;%llu",
1260 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
1261 (unsigned long long) ts->runtime[ddir]);
1263 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
1264 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1266 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1268 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
1269 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1271 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1273 if (ts->lat_percentiles) {
1274 len = calc_clat_percentiles(ts->io_u_plat[FIO_LAT][ddir],
1275 ts->lat_stat[ddir].samples,
1276 ts->percentile_list, &ovals, &maxv,
1278 } else if (ts->clat_percentiles) {
1279 len = calc_clat_percentiles(ts->io_u_plat[FIO_CLAT][ddir],
1280 ts->clat_stat[ddir].samples,
1281 ts->percentile_list, &ovals, &maxv,
1287 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1289 log_buf(out, ";0%%=0");
1292 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
1295 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
1296 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
1298 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
1302 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
1304 double p_of_agg = 100.0;
1306 if (rs->agg[ddir]) {
1307 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1308 if (p_of_agg > 100.0)
1312 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
1314 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
1319 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
1321 log_buf(out, ";%lu", 0UL);
1323 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
1324 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
1325 mean, dev, (&ts->iops_stat[ddir])->samples);
1327 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
1331 static void show_mixed_ddir_status_terse(struct thread_stat *ts,
1332 struct group_run_stats *rs,
1333 int ver, struct buf_output *out)
1335 struct thread_stat *ts_lcl;
1338 * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
1341 ts_lcl = malloc(sizeof(struct thread_stat));
1342 memset((void *)ts_lcl, 0, sizeof(struct thread_stat));
1343 /* calculate mixed stats */
1344 ts_lcl->unified_rw_rep = UNIFIED_MIXED;
1345 init_thread_stat_min_vals(ts_lcl);
1346 ts_lcl->lat_percentiles = ts->lat_percentiles;
1347 ts_lcl->clat_percentiles = ts->clat_percentiles;
1348 ts_lcl->slat_percentiles = ts->slat_percentiles;
1349 ts_lcl->percentile_precision = ts->percentile_precision;
1350 memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
1352 sum_thread_stats(ts_lcl, ts);
1354 /* add the aggregated stats to json parent */
1355 show_ddir_status_terse(ts_lcl, rs, DDIR_READ, ver, out);
1359 static struct json_object *add_ddir_lat_json(struct thread_stat *ts,
1360 uint32_t percentiles,
1361 struct io_stat *lat_stat,
1362 uint64_t *io_u_plat)
1366 unsigned int i, len;
1367 struct json_object *lat_object, *percentile_object, *clat_bins_object;
1368 unsigned long long min, max, maxv, minv, *ovals = NULL;
1370 if (!calc_lat(lat_stat, &min, &max, &mean, &dev)) {
1374 lat_object = json_create_object();
1375 json_object_add_value_int(lat_object, "min", min);
1376 json_object_add_value_int(lat_object, "max", max);
1377 json_object_add_value_float(lat_object, "mean", mean);
1378 json_object_add_value_float(lat_object, "stddev", dev);
1379 json_object_add_value_int(lat_object, "N", lat_stat->samples);
1381 if (percentiles && lat_stat->samples) {
1382 len = calc_clat_percentiles(io_u_plat, lat_stat->samples,
1383 ts->percentile_list, &ovals, &maxv, &minv);
1385 if (len > FIO_IO_U_LIST_MAX_LEN)
1386 len = FIO_IO_U_LIST_MAX_LEN;
1388 percentile_object = json_create_object();
1389 json_object_add_value_object(lat_object, "percentile", percentile_object);
1390 for (i = 0; i < len; i++) {
1391 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1392 json_object_add_value_int(percentile_object, buf, ovals[i]);
1396 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1397 clat_bins_object = json_create_object();
1398 json_object_add_value_object(lat_object, "bins", clat_bins_object);
1400 for(i = 0; i < FIO_IO_U_PLAT_NR; i++)
1402 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1403 json_object_add_value_int(clat_bins_object, buf, io_u_plat[i]);
1411 static void add_ddir_status_json(struct thread_stat *ts,
1412 struct group_run_stats *rs, int ddir, struct json_object *parent)
1414 unsigned long long min, max;
1415 unsigned long long bw_bytes, bw;
1416 double mean, dev, iops;
1417 struct json_object *dir_object, *tmp_object;
1418 double p_of_agg = 100.0;
1420 assert(ddir_rw(ddir) || ddir_sync(ddir));
1422 if ((ts->unified_rw_rep == UNIFIED_MIXED) && ddir != DDIR_READ)
1425 dir_object = json_create_object();
1426 json_object_add_value_object(parent,
1427 (ts->unified_rw_rep == UNIFIED_MIXED) ? "mixed" : io_ddir_name(ddir), dir_object);
1429 if (ddir_rw(ddir)) {
1433 if (ts->runtime[ddir]) {
1434 uint64_t runt = ts->runtime[ddir];
1436 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
1437 bw = bw_bytes / 1024; /* KiB/s */
1438 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
1441 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
1442 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
1443 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
1444 json_object_add_value_int(dir_object, "bw", bw);
1445 json_object_add_value_float(dir_object, "iops", iops);
1446 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
1447 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
1448 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
1449 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
1451 tmp_object = add_ddir_lat_json(ts, ts->slat_percentiles,
1452 &ts->slat_stat[ddir], ts->io_u_plat[FIO_SLAT][ddir]);
1453 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1455 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles,
1456 &ts->clat_stat[ddir], ts->io_u_plat[FIO_CLAT][ddir]);
1457 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1459 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles,
1460 &ts->lat_stat[ddir], ts->io_u_plat[FIO_LAT][ddir]);
1461 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1463 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[DDIR_SYNC]);
1464 tmp_object = add_ddir_lat_json(ts, ts->lat_percentiles | ts->clat_percentiles,
1465 &ts->sync_stat, ts->io_u_sync_plat);
1466 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1472 /* Only print PRIO latencies if some high priority samples were gathered */
1473 if (ts->clat_high_prio_stat[ddir].samples > 0) {
1474 const char *high, *low;
1476 if (ts->lat_percentiles) {
1477 high = "lat_high_prio";
1478 low = "lat_low_prio";
1480 high = "clat_high_prio";
1481 low = "clat_low_prio";
1484 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1485 &ts->clat_high_prio_stat[ddir], ts->io_u_plat_high_prio[ddir]);
1486 json_object_add_value_object(dir_object, high, tmp_object);
1488 tmp_object = add_ddir_lat_json(ts, ts->clat_percentiles | ts->lat_percentiles,
1489 &ts->clat_low_prio_stat[ddir], ts->io_u_plat_low_prio[ddir]);
1490 json_object_add_value_object(dir_object, low, tmp_object);
1493 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1494 p_of_agg = convert_agg_kbytes_percent(rs, ddir, mean);
1497 p_of_agg = mean = dev = 0.0;
1500 json_object_add_value_int(dir_object, "bw_min", min);
1501 json_object_add_value_int(dir_object, "bw_max", max);
1502 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1503 json_object_add_value_float(dir_object, "bw_mean", mean);
1504 json_object_add_value_float(dir_object, "bw_dev", dev);
1505 json_object_add_value_int(dir_object, "bw_samples",
1506 (&ts->bw_stat[ddir])->samples);
1508 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1512 json_object_add_value_int(dir_object, "iops_min", min);
1513 json_object_add_value_int(dir_object, "iops_max", max);
1514 json_object_add_value_float(dir_object, "iops_mean", mean);
1515 json_object_add_value_float(dir_object, "iops_stddev", dev);
1516 json_object_add_value_int(dir_object, "iops_samples",
1517 (&ts->iops_stat[ddir])->samples);
1519 if (ts->cachehit + ts->cachemiss) {
1523 total = ts->cachehit + ts->cachemiss;
1524 hit = (double) ts->cachehit / (double) total;
1526 json_object_add_value_float(dir_object, "cachehit", hit);
1530 static void add_mixed_ddir_status_json(struct thread_stat *ts,
1531 struct group_run_stats *rs, struct json_object *parent)
1533 struct thread_stat *ts_lcl;
1536 * Handle aggregation of Reads (ddir = 0), Writes (ddir = 1), and
1539 ts_lcl = malloc(sizeof(struct thread_stat));
1540 memset((void *)ts_lcl, 0, sizeof(struct thread_stat));
1541 /* calculate mixed stats */
1542 ts_lcl->unified_rw_rep = UNIFIED_MIXED;
1543 init_thread_stat_min_vals(ts_lcl);
1544 ts_lcl->lat_percentiles = ts->lat_percentiles;
1545 ts_lcl->clat_percentiles = ts->clat_percentiles;
1546 ts_lcl->slat_percentiles = ts->slat_percentiles;
1547 ts_lcl->percentile_precision = ts->percentile_precision;
1548 memcpy(ts_lcl->percentile_list, ts->percentile_list, sizeof(ts->percentile_list));
1550 sum_thread_stats(ts_lcl, ts);
1552 /* add the aggregated stats to json parent */
1553 add_ddir_status_json(ts_lcl, rs, DDIR_READ, parent);
1557 static void show_thread_status_terse_all(struct thread_stat *ts,
1558 struct group_run_stats *rs, int ver,
1559 struct buf_output *out)
1561 double io_u_dist[FIO_IO_U_MAP_NR];
1562 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1563 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1564 double usr_cpu, sys_cpu;
1569 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1571 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1572 ts->name, ts->groupid, ts->error);
1574 /* Log Read Status, or mixed if unified_rw_rep = 1 */
1575 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1576 if (ts->unified_rw_rep != UNIFIED_MIXED) {
1577 /* Log Write Status */
1578 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1579 /* Log Trim Status */
1580 if (ver == 2 || ver == 4 || ver == 5)
1581 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1583 if (ts->unified_rw_rep == UNIFIED_BOTH)
1584 show_mixed_ddir_status_terse(ts, rs, ver, out);
1586 if (ts->total_run_time) {
1587 double runt = (double) ts->total_run_time;
1589 usr_cpu = (double) ts->usr_time * 100 / runt;
1590 sys_cpu = (double) ts->sys_time * 100 / runt;
1596 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1597 (unsigned long long) ts->ctx,
1598 (unsigned long long) ts->majf,
1599 (unsigned long long) ts->minf);
1601 /* Calc % distribution of IO depths, usecond, msecond latency */
1602 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1603 stat_calc_lat_nu(ts, io_u_lat_u);
1604 stat_calc_lat_m(ts, io_u_lat_m);
1606 /* Only show fixed 7 I/O depth levels*/
1607 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1608 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1609 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1611 /* Microsecond latency */
1612 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1613 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1614 /* Millisecond latency */
1615 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1616 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1618 /* disk util stats, if any */
1619 if (ver >= 3 && is_running_backend())
1620 show_disk_util(1, NULL, out);
1622 /* Additional output if continue_on_error set - default off*/
1623 if (ts->continue_on_error)
1624 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1626 /* Additional output if description is set */
1627 if (strlen(ts->description)) {
1630 log_buf(out, ";%s", ts->description);
1636 static void json_add_job_opts(struct json_object *root, const char *name,
1637 struct flist_head *opt_list)
1639 struct json_object *dir_object;
1640 struct flist_head *entry;
1641 struct print_option *p;
1643 if (flist_empty(opt_list))
1646 dir_object = json_create_object();
1647 json_object_add_value_object(root, name, dir_object);
1649 flist_for_each(entry, opt_list) {
1650 p = flist_entry(entry, struct print_option, list);
1651 json_object_add_value_string(dir_object, p->name, p->value);
1655 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1656 struct group_run_stats *rs,
1657 struct flist_head *opt_list)
1659 struct json_object *root, *tmp;
1660 struct jobs_eta *je;
1661 double io_u_dist[FIO_IO_U_MAP_NR];
1662 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1663 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1664 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1665 double usr_cpu, sys_cpu;
1669 root = json_create_object();
1670 json_object_add_value_string(root, "jobname", ts->name);
1671 json_object_add_value_int(root, "groupid", ts->groupid);
1672 json_object_add_value_int(root, "error", ts->error);
1675 je = get_jobs_eta(true, &size);
1677 json_object_add_value_int(root, "eta", je->eta_sec);
1678 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1682 json_add_job_opts(root, "job options", opt_list);
1684 add_ddir_status_json(ts, rs, DDIR_READ, root);
1685 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1686 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1687 add_ddir_status_json(ts, rs, DDIR_SYNC, root);
1689 if (ts->unified_rw_rep == UNIFIED_BOTH)
1690 add_mixed_ddir_status_json(ts, rs, root);
1693 if (ts->total_run_time) {
1694 double runt = (double) ts->total_run_time;
1696 usr_cpu = (double) ts->usr_time * 100 / runt;
1697 sys_cpu = (double) ts->sys_time * 100 / runt;
1702 json_object_add_value_int(root, "job_runtime", ts->total_run_time);
1703 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1704 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1705 json_object_add_value_int(root, "ctx", ts->ctx);
1706 json_object_add_value_int(root, "majf", ts->majf);
1707 json_object_add_value_int(root, "minf", ts->minf);
1709 /* Calc % distribution of IO depths */
1710 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1711 tmp = json_create_object();
1712 json_object_add_value_object(root, "iodepth_level", tmp);
1713 /* Only show fixed 7 I/O depth levels*/
1714 for (i = 0; i < 7; i++) {
1717 snprintf(name, 20, "%d", 1 << i);
1719 snprintf(name, 20, ">=%d", 1 << i);
1720 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1723 /* Calc % distribution of submit IO depths */
1724 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
1725 tmp = json_create_object();
1726 json_object_add_value_object(root, "iodepth_submit", tmp);
1727 /* Only show fixed 7 I/O depth levels*/
1728 for (i = 0; i < 7; i++) {
1731 snprintf(name, 20, "0");
1733 snprintf(name, 20, "%d", 1 << (i+1));
1735 snprintf(name, 20, ">=%d", 1 << i);
1736 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1739 /* Calc % distribution of completion IO depths */
1740 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
1741 tmp = json_create_object();
1742 json_object_add_value_object(root, "iodepth_complete", tmp);
1743 /* Only show fixed 7 I/O depth levels*/
1744 for (i = 0; i < 7; i++) {
1747 snprintf(name, 20, "0");
1749 snprintf(name, 20, "%d", 1 << (i+1));
1751 snprintf(name, 20, ">=%d", 1 << i);
1752 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1755 /* Calc % distribution of nsecond, usecond, msecond latency */
1756 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1757 stat_calc_lat_n(ts, io_u_lat_n);
1758 stat_calc_lat_u(ts, io_u_lat_u);
1759 stat_calc_lat_m(ts, io_u_lat_m);
1761 /* Nanosecond latency */
1762 tmp = json_create_object();
1763 json_object_add_value_object(root, "latency_ns", tmp);
1764 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1765 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1766 "250", "500", "750", "1000", };
1767 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1769 /* Microsecond latency */
1770 tmp = json_create_object();
1771 json_object_add_value_object(root, "latency_us", tmp);
1772 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1773 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1774 "250", "500", "750", "1000", };
1775 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1777 /* Millisecond latency */
1778 tmp = json_create_object();
1779 json_object_add_value_object(root, "latency_ms", tmp);
1780 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1781 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1782 "250", "500", "750", "1000", "2000",
1784 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1787 /* Additional output if continue_on_error set - default off*/
1788 if (ts->continue_on_error) {
1789 json_object_add_value_int(root, "total_err", ts->total_err_count);
1790 json_object_add_value_int(root, "first_error", ts->first_error);
1793 if (ts->latency_depth) {
1794 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1795 json_object_add_value_int(root, "latency_target", ts->latency_target);
1796 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1797 json_object_add_value_int(root, "latency_window", ts->latency_window);
1800 /* Additional output if description is set */
1801 if (strlen(ts->description))
1802 json_object_add_value_string(root, "desc", ts->description);
1804 if (ts->nr_block_infos) {
1805 /* Block error histogram and types */
1807 unsigned int *percentiles = NULL;
1808 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1810 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1811 ts->percentile_list,
1812 &percentiles, block_state_counts);
1815 struct json_object *block, *percentile_object, *states;
1817 block = json_create_object();
1818 json_object_add_value_object(root, "block", block);
1820 percentile_object = json_create_object();
1821 json_object_add_value_object(block, "percentiles",
1823 for (i = 0; i < len; i++) {
1825 snprintf(buf, sizeof(buf), "%f",
1826 ts->percentile_list[i].u.f);
1827 json_object_add_value_int(percentile_object,
1832 states = json_create_object();
1833 json_object_add_value_object(block, "states", states);
1834 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1835 json_object_add_value_int(states,
1836 block_state_names[state],
1837 block_state_counts[state]);
1844 struct json_object *data;
1845 struct json_array *iops, *bw;
1849 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1850 ts->ss_state & FIO_SS_IOPS ? "iops" : "bw",
1851 ts->ss_state & FIO_SS_SLOPE ? "_slope" : "",
1852 (float) ts->ss_limit.u.f,
1853 ts->ss_state & FIO_SS_PCT ? "%" : "");
1855 tmp = json_create_object();
1856 json_object_add_value_object(root, "steadystate", tmp);
1857 json_object_add_value_string(tmp, "ss", ss_buf);
1858 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1859 json_object_add_value_int(tmp, "attained", (ts->ss_state & FIO_SS_ATTAINED) > 0);
1861 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1862 ts->ss_state & FIO_SS_PCT ? "%" : "");
1863 json_object_add_value_string(tmp, "criterion", ss_buf);
1864 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1865 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1867 data = json_create_object();
1868 json_object_add_value_object(tmp, "data", data);
1869 bw = json_create_array();
1870 iops = json_create_array();
1873 ** if ss was attained or the buffer is not full,
1874 ** ss->head points to the first element in the list.
1875 ** otherwise it actually points to the second element
1878 if ((ts->ss_state & FIO_SS_ATTAINED) || !(ts->ss_state & FIO_SS_BUFFER_FULL))
1881 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1882 for (l = 0; l < ts->ss_dur; l++) {
1883 k = (j + l) % ts->ss_dur;
1884 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1885 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1887 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1888 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1889 json_object_add_value_array(data, "iops", iops);
1890 json_object_add_value_array(data, "bw", bw);
1896 static void show_thread_status_terse(struct thread_stat *ts,
1897 struct group_run_stats *rs,
1898 struct buf_output *out)
1900 if (terse_version >= 2 && terse_version <= 5)
1901 show_thread_status_terse_all(ts, rs, terse_version, out);
1903 log_err("fio: bad terse version!? %d\n", terse_version);
1906 struct json_object *show_thread_status(struct thread_stat *ts,
1907 struct group_run_stats *rs,
1908 struct flist_head *opt_list,
1909 struct buf_output *out)
1911 struct json_object *ret = NULL;
1913 if (output_format & FIO_OUTPUT_TERSE)
1914 show_thread_status_terse(ts, rs, out);
1915 if (output_format & FIO_OUTPUT_JSON)
1916 ret = show_thread_status_json(ts, rs, opt_list);
1917 if (output_format & FIO_OUTPUT_NORMAL)
1918 show_thread_status_normal(ts, rs, out);
1923 static void __sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1927 dst->min_val = min(dst->min_val, src->min_val);
1928 dst->max_val = max(dst->max_val, src->max_val);
1931 * Compute new mean and S after the merge
1932 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1933 * #Parallel_algorithm>
1936 mean = src->mean.u.f;
1939 double delta = src->mean.u.f - dst->mean.u.f;
1941 mean = ((src->mean.u.f * src->samples) +
1942 (dst->mean.u.f * dst->samples)) /
1943 (dst->samples + src->samples);
1945 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1946 (dst->samples * src->samples) /
1947 (dst->samples + src->samples);
1950 dst->samples += src->samples;
1951 dst->mean.u.f = mean;
1957 * We sum two kinds of stats - one that is time based, in which case we
1958 * apply the proper summing technique, and then one that is iops/bw
1959 * numbers. For group_reporting, we should just add those up, not make
1960 * them the mean of everything.
1962 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool pure_sum)
1964 bool first = dst->samples == 0;
1966 if (src->samples == 0)
1970 __sum_stat(dst, src, first);
1975 dst->min_val = src->min_val;
1976 dst->max_val = src->max_val;
1977 dst->samples = src->samples;
1978 dst->mean.u.f = src->mean.u.f;
1979 dst->S.u.f = src->S.u.f;
1981 dst->min_val += src->min_val;
1982 dst->max_val += src->max_val;
1983 dst->samples += src->samples;
1984 dst->mean.u.f += src->mean.u.f;
1985 dst->S.u.f += src->S.u.f;
1989 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1993 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1994 if (dst->max_run[i] < src->max_run[i])
1995 dst->max_run[i] = src->max_run[i];
1996 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1997 dst->min_run[i] = src->min_run[i];
1998 if (dst->max_bw[i] < src->max_bw[i])
1999 dst->max_bw[i] = src->max_bw[i];
2000 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
2001 dst->min_bw[i] = src->min_bw[i];
2003 dst->iobytes[i] += src->iobytes[i];
2004 dst->agg[i] += src->agg[i];
2008 dst->kb_base = src->kb_base;
2009 if (!dst->unit_base)
2010 dst->unit_base = src->unit_base;
2012 dst->sig_figs = src->sig_figs;
2015 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src)
2019 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
2020 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2021 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], false);
2022 sum_stat(&dst->clat_high_prio_stat[l], &src->clat_high_prio_stat[l], false);
2023 sum_stat(&dst->clat_low_prio_stat[l], &src->clat_low_prio_stat[l], false);
2024 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], false);
2025 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], false);
2026 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], true);
2027 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], true);
2029 dst->io_bytes[l] += src->io_bytes[l];
2031 if (dst->runtime[l] < src->runtime[l])
2032 dst->runtime[l] = src->runtime[l];
2034 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], false);
2035 sum_stat(&dst->clat_high_prio_stat[0], &src->clat_high_prio_stat[l], false);
2036 sum_stat(&dst->clat_low_prio_stat[0], &src->clat_low_prio_stat[l], false);
2037 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], false);
2038 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], false);
2039 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], true);
2040 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], true);
2042 dst->io_bytes[0] += src->io_bytes[l];
2044 if (dst->runtime[0] < src->runtime[l])
2045 dst->runtime[0] = src->runtime[l];
2049 sum_stat(&dst->sync_stat, &src->sync_stat, false);
2050 dst->usr_time += src->usr_time;
2051 dst->sys_time += src->sys_time;
2052 dst->ctx += src->ctx;
2053 dst->majf += src->majf;
2054 dst->minf += src->minf;
2056 for (k = 0; k < FIO_IO_U_MAP_NR; k++) {
2057 dst->io_u_map[k] += src->io_u_map[k];
2058 dst->io_u_submit[k] += src->io_u_submit[k];
2059 dst->io_u_complete[k] += src->io_u_complete[k];
2062 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
2063 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
2064 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
2065 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
2066 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
2067 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
2069 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2070 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2071 dst->total_io_u[k] += src->total_io_u[k];
2072 dst->short_io_u[k] += src->short_io_u[k];
2073 dst->drop_io_u[k] += src->drop_io_u[k];
2075 dst->total_io_u[0] += src->total_io_u[k];
2076 dst->short_io_u[0] += src->short_io_u[k];
2077 dst->drop_io_u[0] += src->drop_io_u[k];
2081 dst->total_io_u[DDIR_SYNC] += src->total_io_u[DDIR_SYNC];
2083 for (k = 0; k < FIO_LAT_CNT; k++)
2084 for (l = 0; l < DDIR_RWDIR_CNT; l++)
2085 for (m = 0; m < FIO_IO_U_PLAT_NR; m++)
2086 if (dst->unified_rw_rep != UNIFIED_MIXED)
2087 dst->io_u_plat[k][l][m] += src->io_u_plat[k][l][m];
2089 dst->io_u_plat[k][0][m] += src->io_u_plat[k][l][m];
2091 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2092 dst->io_u_sync_plat[k] += src->io_u_sync_plat[k];
2094 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
2095 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
2096 if (dst->unified_rw_rep != UNIFIED_MIXED) {
2097 dst->io_u_plat_high_prio[k][m] += src->io_u_plat_high_prio[k][m];
2098 dst->io_u_plat_low_prio[k][m] += src->io_u_plat_low_prio[k][m];
2100 dst->io_u_plat_high_prio[0][m] += src->io_u_plat_high_prio[k][m];
2101 dst->io_u_plat_low_prio[0][m] += src->io_u_plat_low_prio[k][m];
2107 dst->total_run_time += src->total_run_time;
2108 dst->total_submit += src->total_submit;
2109 dst->total_complete += src->total_complete;
2110 dst->nr_zone_resets += src->nr_zone_resets;
2111 dst->cachehit += src->cachehit;
2112 dst->cachemiss += src->cachemiss;
2115 void init_group_run_stat(struct group_run_stats *gs)
2118 memset(gs, 0, sizeof(*gs));
2120 for (i = 0; i < DDIR_RWDIR_CNT; i++)
2121 gs->min_bw[i] = gs->min_run[i] = ~0UL;
2124 void init_thread_stat_min_vals(struct thread_stat *ts)
2128 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2129 ts->clat_stat[i].min_val = ULONG_MAX;
2130 ts->slat_stat[i].min_val = ULONG_MAX;
2131 ts->lat_stat[i].min_val = ULONG_MAX;
2132 ts->bw_stat[i].min_val = ULONG_MAX;
2133 ts->iops_stat[i].min_val = ULONG_MAX;
2134 ts->clat_high_prio_stat[i].min_val = ULONG_MAX;
2135 ts->clat_low_prio_stat[i].min_val = ULONG_MAX;
2137 ts->sync_stat.min_val = ULONG_MAX;
2140 void init_thread_stat(struct thread_stat *ts)
2142 memset(ts, 0, sizeof(*ts));
2144 init_thread_stat_min_vals(ts);
2148 void __show_run_stats(void)
2150 struct group_run_stats *runstats, *rs;
2151 struct thread_data *td;
2152 struct thread_stat *threadstats, *ts;
2153 int i, j, k, nr_ts, last_ts, idx;
2154 bool kb_base_warned = false;
2155 bool unit_base_warned = false;
2156 struct json_object *root = NULL;
2157 struct json_array *array = NULL;
2158 struct buf_output output[FIO_OUTPUT_NR];
2159 struct flist_head **opt_lists;
2161 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2163 for (i = 0; i < groupid + 1; i++)
2164 init_group_run_stat(&runstats[i]);
2167 * find out how many threads stats we need. if group reporting isn't
2168 * enabled, it's one-per-td.
2172 for_each_td(td, i) {
2173 if (!td->o.group_reporting) {
2177 if (last_ts == td->groupid)
2182 last_ts = td->groupid;
2186 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
2187 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
2189 for (i = 0; i < nr_ts; i++) {
2190 init_thread_stat(&threadstats[i]);
2191 opt_lists[i] = NULL;
2197 for_each_td(td, i) {
2200 if (idx && (!td->o.group_reporting ||
2201 (td->o.group_reporting && last_ts != td->groupid))) {
2206 last_ts = td->groupid;
2208 ts = &threadstats[j];
2210 ts->clat_percentiles = td->o.clat_percentiles;
2211 ts->lat_percentiles = td->o.lat_percentiles;
2212 ts->slat_percentiles = td->o.slat_percentiles;
2213 ts->percentile_precision = td->o.percentile_precision;
2214 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
2215 opt_lists[j] = &td->opt_list;
2220 if (ts->groupid == -1) {
2222 * These are per-group shared already
2224 snprintf(ts->name, sizeof(ts->name), "%s", td->o.name);
2225 if (td->o.description)
2226 snprintf(ts->description,
2227 sizeof(ts->description), "%s",
2230 memset(ts->description, 0, FIO_JOBDESC_SIZE);
2233 * If multiple entries in this group, this is
2236 ts->thread_number = td->thread_number;
2237 ts->groupid = td->groupid;
2240 * first pid in group, not very useful...
2244 ts->kb_base = td->o.kb_base;
2245 ts->unit_base = td->o.unit_base;
2246 ts->sig_figs = td->o.sig_figs;
2247 ts->unified_rw_rep = td->o.unified_rw_rep;
2248 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
2249 log_info("fio: kb_base differs for jobs in group, using"
2250 " %u as the base\n", ts->kb_base);
2251 kb_base_warned = true;
2252 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
2253 log_info("fio: unit_base differs for jobs in group, using"
2254 " %u as the base\n", ts->unit_base);
2255 unit_base_warned = true;
2258 ts->continue_on_error = td->o.continue_on_error;
2259 ts->total_err_count += td->total_err_count;
2260 ts->first_error = td->first_error;
2262 if (!td->error && td->o.continue_on_error &&
2264 ts->error = td->first_error;
2265 snprintf(ts->verror, sizeof(ts->verror), "%s",
2267 } else if (td->error) {
2268 ts->error = td->error;
2269 snprintf(ts->verror, sizeof(ts->verror), "%s",
2274 ts->latency_depth = td->latency_qd;
2275 ts->latency_target = td->o.latency_target;
2276 ts->latency_percentile = td->o.latency_percentile;
2277 ts->latency_window = td->o.latency_window;
2279 ts->nr_block_infos = td->ts.nr_block_infos;
2280 for (k = 0; k < ts->nr_block_infos; k++)
2281 ts->block_infos[k] = td->ts.block_infos[k];
2283 sum_thread_stats(ts, &td->ts);
2286 ts->ss_state = td->ss.state;
2287 ts->ss_dur = td->ss.dur;
2288 ts->ss_head = td->ss.head;
2289 ts->ss_bw_data = td->ss.bw_data;
2290 ts->ss_iops_data = td->ss.iops_data;
2291 ts->ss_limit.u.f = td->ss.limit;
2292 ts->ss_slope.u.f = td->ss.slope;
2293 ts->ss_deviation.u.f = td->ss.deviation;
2294 ts->ss_criterion.u.f = td->ss.criterion;
2297 ts->ss_dur = ts->ss_state = 0;
2300 for (i = 0; i < nr_ts; i++) {
2301 unsigned long long bw;
2303 ts = &threadstats[i];
2304 if (ts->groupid == -1)
2306 rs = &runstats[ts->groupid];
2307 rs->kb_base = ts->kb_base;
2308 rs->unit_base = ts->unit_base;
2309 rs->sig_figs = ts->sig_figs;
2310 rs->unified_rw_rep |= ts->unified_rw_rep;
2312 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
2313 if (!ts->runtime[j])
2315 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
2316 rs->min_run[j] = ts->runtime[j];
2317 if (ts->runtime[j] > rs->max_run[j])
2318 rs->max_run[j] = ts->runtime[j];
2322 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
2323 if (bw < rs->min_bw[j])
2325 if (bw > rs->max_bw[j])
2328 rs->iobytes[j] += ts->io_bytes[j];
2332 for (i = 0; i < groupid + 1; i++) {
2337 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2338 if (rs->max_run[ddir])
2339 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
2344 for (i = 0; i < FIO_OUTPUT_NR; i++)
2345 buf_output_init(&output[i]);
2348 * don't overwrite last signal output
2350 if (output_format & FIO_OUTPUT_NORMAL)
2351 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
2352 if (output_format & FIO_OUTPUT_JSON) {
2353 struct thread_data *global;
2356 unsigned long long ms_since_epoch;
2359 gettimeofday(&now, NULL);
2360 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
2361 (unsigned long long)(now.tv_usec) / 1000;
2363 tv_sec = now.tv_sec;
2364 os_ctime_r(&tv_sec, time_buf, sizeof(time_buf));
2365 if (time_buf[strlen(time_buf) - 1] == '\n')
2366 time_buf[strlen(time_buf) - 1] = '\0';
2368 root = json_create_object();
2369 json_object_add_value_string(root, "fio version", fio_version_string);
2370 json_object_add_value_int(root, "timestamp", now.tv_sec);
2371 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
2372 json_object_add_value_string(root, "time", time_buf);
2373 global = get_global_options();
2374 json_add_job_opts(root, "global options", &global->opt_list);
2375 array = json_create_array();
2376 json_object_add_value_array(root, "jobs", array);
2380 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
2382 for (i = 0; i < nr_ts; i++) {
2383 ts = &threadstats[i];
2384 rs = &runstats[ts->groupid];
2387 fio_server_send_job_options(opt_lists[i], i);
2388 fio_server_send_ts(ts, rs);
2390 if (output_format & FIO_OUTPUT_TERSE)
2391 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
2392 if (output_format & FIO_OUTPUT_JSON) {
2393 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
2394 json_array_add_value_object(array, tmp);
2396 if (output_format & FIO_OUTPUT_NORMAL)
2397 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
2400 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
2401 /* disk util stats, if any */
2402 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
2404 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
2406 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
2407 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
2408 json_free_object(root);
2411 for (i = 0; i < groupid + 1; i++) {
2416 fio_server_send_gs(rs);
2417 else if (output_format & FIO_OUTPUT_NORMAL)
2418 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
2422 fio_server_send_du();
2423 else if (output_format & FIO_OUTPUT_NORMAL) {
2424 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
2425 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
2428 for (i = 0; i < FIO_OUTPUT_NR; i++) {
2429 struct buf_output *out = &output[i];
2431 log_info_buf(out->buf, out->buflen);
2432 buf_output_free(out);
2435 fio_idle_prof_cleanup();
2443 int __show_running_run_stats(void)
2445 struct thread_data *td;
2446 unsigned long long *rt;
2450 fio_sem_down(stat_sem);
2452 rt = malloc(thread_number * sizeof(unsigned long long));
2453 fio_gettime(&ts, NULL);
2455 for_each_td(td, i) {
2456 td->update_rusage = 1;
2457 for_each_rw_ddir(ddir) {
2458 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2460 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
2462 rt[i] = mtime_since(&td->start, &ts);
2463 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2464 td->ts.runtime[DDIR_READ] += rt[i];
2465 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2466 td->ts.runtime[DDIR_WRITE] += rt[i];
2467 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2468 td->ts.runtime[DDIR_TRIM] += rt[i];
2471 for_each_td(td, i) {
2472 if (td->runstate >= TD_EXITED)
2474 if (td->rusage_sem) {
2475 td->update_rusage = 1;
2476 fio_sem_down(td->rusage_sem);
2478 td->update_rusage = 0;
2483 for_each_td(td, i) {
2484 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
2485 td->ts.runtime[DDIR_READ] -= rt[i];
2486 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
2487 td->ts.runtime[DDIR_WRITE] -= rt[i];
2488 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
2489 td->ts.runtime[DDIR_TRIM] -= rt[i];
2493 fio_sem_up(stat_sem);
2498 static bool status_file_disabled;
2500 #define FIO_STATUS_FILE "fio-dump-status"
2502 static int check_status_file(void)
2505 const char *temp_dir;
2506 char fio_status_file_path[PATH_MAX];
2508 if (status_file_disabled)
2511 temp_dir = getenv("TMPDIR");
2512 if (temp_dir == NULL) {
2513 temp_dir = getenv("TEMP");
2514 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
2517 if (temp_dir == NULL)
2520 __coverity_tainted_data_sanitize__(temp_dir);
2523 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
2525 if (stat(fio_status_file_path, &sb))
2528 if (unlink(fio_status_file_path) < 0) {
2529 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
2531 log_err("fio: disabling status file updates\n");
2532 status_file_disabled = true;
2538 void check_for_running_stats(void)
2540 if (check_status_file()) {
2541 show_running_run_stats();
2546 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2551 if (data > is->max_val)
2553 if (data < is->min_val)
2556 delta = val - is->mean.u.f;
2558 is->mean.u.f += delta / (is->samples + 1.0);
2559 is->S.u.f += delta * (val - is->mean.u.f);
2566 * Return a struct io_logs, which is added to the tail of the log
2569 static struct io_logs *get_new_log(struct io_log *iolog)
2572 struct io_logs *cur_log;
2575 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2578 if (!iolog->cur_log_max) {
2579 new_samples = iolog->td->o.log_entries;
2581 new_samples = iolog->cur_log_max * 2;
2582 if (new_samples > MAX_LOG_ENTRIES)
2583 new_samples = MAX_LOG_ENTRIES;
2586 cur_log = smalloc(sizeof(*cur_log));
2588 INIT_FLIST_HEAD(&cur_log->list);
2589 cur_log->log = calloc(new_samples, log_entry_sz(iolog));
2591 cur_log->nr_samples = 0;
2592 cur_log->max_samples = new_samples;
2593 flist_add_tail(&cur_log->list, &iolog->io_logs);
2594 iolog->cur_log_max = new_samples;
2604 * Add and return a new log chunk, or return current log if big enough
2606 static struct io_logs *regrow_log(struct io_log *iolog)
2608 struct io_logs *cur_log;
2611 if (!iolog || iolog->disabled)
2614 cur_log = iolog_cur_log(iolog);
2616 cur_log = get_new_log(iolog);
2621 if (cur_log->nr_samples < cur_log->max_samples)
2625 * No room for a new sample. If we're compressing on the fly, flush
2626 * out the current chunk
2628 if (iolog->log_gz) {
2629 if (iolog_cur_flush(iolog, cur_log)) {
2630 log_err("fio: failed flushing iolog! Will stop logging.\n");
2636 * Get a new log array, and add to our list
2638 cur_log = get_new_log(iolog);
2640 log_err("fio: failed extending iolog! Will stop logging.\n");
2644 if (!iolog->pending || !iolog->pending->nr_samples)
2648 * Flush pending items to new log
2650 for (i = 0; i < iolog->pending->nr_samples; i++) {
2651 struct io_sample *src, *dst;
2653 src = get_sample(iolog, iolog->pending, i);
2654 dst = get_sample(iolog, cur_log, i);
2655 memcpy(dst, src, log_entry_sz(iolog));
2657 cur_log->nr_samples = iolog->pending->nr_samples;
2659 iolog->pending->nr_samples = 0;
2663 iolog->disabled = true;
2667 void regrow_logs(struct thread_data *td)
2669 regrow_log(td->slat_log);
2670 regrow_log(td->clat_log);
2671 regrow_log(td->clat_hist_log);
2672 regrow_log(td->lat_log);
2673 regrow_log(td->bw_log);
2674 regrow_log(td->iops_log);
2675 td->flags &= ~TD_F_REGROW_LOGS;
2678 void regrow_agg_logs(void)
2682 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2683 regrow_log(agg_io_log[ddir]);
2686 static struct io_logs *get_cur_log(struct io_log *iolog)
2688 struct io_logs *cur_log;
2690 cur_log = iolog_cur_log(iolog);
2692 cur_log = get_new_log(iolog);
2697 if (cur_log->nr_samples < cur_log->max_samples)
2701 * Out of space. If we're in IO offload mode, or we're not doing
2702 * per unit logging (hence logging happens outside of the IO thread
2703 * as well), add a new log chunk inline. If we're doing inline
2704 * submissions, flag 'td' as needing a log regrow and we'll take
2705 * care of it on the submission side.
2707 if ((iolog->td && iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD) ||
2708 !per_unit_log(iolog))
2709 return regrow_log(iolog);
2712 iolog->td->flags |= TD_F_REGROW_LOGS;
2714 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2715 return iolog->pending;
2718 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2719 enum fio_ddir ddir, unsigned long long bs,
2720 unsigned long t, uint64_t offset,
2721 unsigned int priority)
2723 struct io_logs *cur_log;
2725 if (iolog->disabled)
2727 if (flist_empty(&iolog->io_logs))
2728 iolog->avg_last[ddir] = t;
2730 cur_log = get_cur_log(iolog);
2732 struct io_sample *s;
2734 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2737 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2738 io_sample_set_ddir(iolog, s, ddir);
2740 s->priority = priority;
2742 if (iolog->log_offset) {
2743 struct io_sample_offset *so = (void *) s;
2745 so->offset = offset;
2748 cur_log->nr_samples++;
2752 iolog->disabled = true;
2755 static inline void reset_io_stat(struct io_stat *ios)
2757 ios->min_val = -1ULL;
2758 ios->max_val = ios->samples = 0;
2759 ios->mean.u.f = ios->S.u.f = 0;
2762 void reset_io_stats(struct thread_data *td)
2764 struct thread_stat *ts = &td->ts;
2767 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2768 reset_io_stat(&ts->clat_high_prio_stat[i]);
2769 reset_io_stat(&ts->clat_low_prio_stat[i]);
2770 reset_io_stat(&ts->clat_stat[i]);
2771 reset_io_stat(&ts->slat_stat[i]);
2772 reset_io_stat(&ts->lat_stat[i]);
2773 reset_io_stat(&ts->bw_stat[i]);
2774 reset_io_stat(&ts->iops_stat[i]);
2776 ts->io_bytes[i] = 0;
2778 ts->total_io_u[i] = 0;
2779 ts->short_io_u[i] = 0;
2780 ts->drop_io_u[i] = 0;
2782 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2783 ts->io_u_plat_high_prio[i][j] = 0;
2784 ts->io_u_plat_low_prio[i][j] = 0;
2786 ts->io_u_sync_plat[j] = 0;
2790 for (i = 0; i < FIO_LAT_CNT; i++)
2791 for (j = 0; j < DDIR_RWDIR_CNT; j++)
2792 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
2793 ts->io_u_plat[i][j][k] = 0;
2795 ts->total_io_u[DDIR_SYNC] = 0;
2797 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2798 ts->io_u_map[i] = 0;
2799 ts->io_u_submit[i] = 0;
2800 ts->io_u_complete[i] = 0;
2803 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2804 ts->io_u_lat_n[i] = 0;
2805 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2806 ts->io_u_lat_u[i] = 0;
2807 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2808 ts->io_u_lat_m[i] = 0;
2810 ts->total_submit = 0;
2811 ts->total_complete = 0;
2812 ts->nr_zone_resets = 0;
2813 ts->cachehit = ts->cachemiss = 0;
2816 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2817 unsigned long elapsed, bool log_max)
2820 * Note an entry in the log. Use the mean from the logged samples,
2821 * making sure to properly round up. Only write a log entry if we
2822 * had actual samples done.
2824 if (iolog->avg_window[ddir].samples) {
2825 union io_sample_data data;
2828 data.val = iolog->avg_window[ddir].max_val;
2830 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2832 __add_log_sample(iolog, data, ddir, 0, elapsed, 0, 0);
2835 reset_io_stat(&iolog->avg_window[ddir]);
2838 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2843 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2844 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2847 static unsigned long add_log_sample(struct thread_data *td,
2848 struct io_log *iolog,
2849 union io_sample_data data,
2850 enum fio_ddir ddir, unsigned long long bs,
2851 uint64_t offset, unsigned int ioprio)
2853 unsigned long elapsed, this_window;
2858 elapsed = mtime_since_now(&td->epoch);
2861 * If no time averaging, just add the log sample.
2863 if (!iolog->avg_msec) {
2864 __add_log_sample(iolog, data, ddir, bs, elapsed, offset,
2870 * Add the sample. If the time period has passed, then
2871 * add that entry to the log and clear.
2873 add_stat_sample(&iolog->avg_window[ddir], data.val);
2876 * If period hasn't passed, adding the above sample is all we
2879 this_window = elapsed - iolog->avg_last[ddir];
2880 if (elapsed < iolog->avg_last[ddir])
2881 return iolog->avg_last[ddir] - elapsed;
2882 else if (this_window < iolog->avg_msec) {
2883 unsigned long diff = iolog->avg_msec - this_window;
2885 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2889 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2891 iolog->avg_last[ddir] = elapsed - (elapsed % iolog->avg_msec);
2893 return iolog->avg_msec;
2896 void finalize_logs(struct thread_data *td, bool unit_logs)
2898 unsigned long elapsed;
2900 elapsed = mtime_since_now(&td->epoch);
2902 if (td->clat_log && unit_logs)
2903 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2904 if (td->slat_log && unit_logs)
2905 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2906 if (td->lat_log && unit_logs)
2907 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2908 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2909 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2910 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2911 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2914 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir,
2915 unsigned long long bs)
2917 struct io_log *iolog;
2922 iolog = agg_io_log[ddir];
2923 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0, 0);
2926 void add_sync_clat_sample(struct thread_stat *ts, unsigned long long nsec)
2928 unsigned int idx = plat_val_to_idx(nsec);
2929 assert(idx < FIO_IO_U_PLAT_NR);
2931 ts->io_u_sync_plat[idx]++;
2932 add_stat_sample(&ts->sync_stat, nsec);
2935 static inline void add_lat_percentile_sample(struct thread_stat *ts,
2936 unsigned long long nsec,
2940 unsigned int idx = plat_val_to_idx(nsec);
2941 assert(idx < FIO_IO_U_PLAT_NR);
2943 ts->io_u_plat[lat][ddir][idx]++;
2946 static inline void add_lat_percentile_prio_sample(struct thread_stat *ts,
2947 unsigned long long nsec,
2951 unsigned int idx = plat_val_to_idx(nsec);
2954 ts->io_u_plat_low_prio[ddir][idx]++;
2956 ts->io_u_plat_high_prio[ddir][idx]++;
2959 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2960 unsigned long long nsec, unsigned long long bs,
2961 uint64_t offset, unsigned int ioprio, bool high_prio)
2963 const bool needs_lock = td_async_processing(td);
2964 unsigned long elapsed, this_window;
2965 struct thread_stat *ts = &td->ts;
2966 struct io_log *iolog = td->clat_hist_log;
2971 add_stat_sample(&ts->clat_stat[ddir], nsec);
2974 * When lat_percentiles=1 (default 0), the reported high/low priority
2975 * percentiles and stats are used for describing total latency values,
2976 * even though the variable names themselves start with clat_.
2978 * Because of the above definition, add a prio stat sample only when
2979 * lat_percentiles=0. add_lat_sample() will add the prio stat sample
2980 * when lat_percentiles=1.
2982 if (!ts->lat_percentiles) {
2984 add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
2986 add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
2990 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2993 if (ts->clat_percentiles) {
2995 * Because of the above definition, add a prio lat percentile
2996 * sample only when lat_percentiles=0. add_lat_sample() will add
2997 * the prio lat percentile sample when lat_percentiles=1.
2999 add_lat_percentile_sample(ts, nsec, ddir, FIO_CLAT);
3000 if (!ts->lat_percentiles)
3001 add_lat_percentile_prio_sample(ts, nsec, ddir,
3005 if (iolog && iolog->hist_msec) {
3006 struct io_hist *hw = &iolog->hist_window[ddir];
3009 elapsed = mtime_since_now(&td->epoch);
3011 hw->hist_last = elapsed;
3012 this_window = elapsed - hw->hist_last;
3014 if (this_window >= iolog->hist_msec) {
3015 uint64_t *io_u_plat;
3016 struct io_u_plat_entry *dst;
3019 * Make a byte-for-byte copy of the latency histogram
3020 * stored in td->ts.io_u_plat[ddir], recording it in a
3021 * log sample. Note that the matching call to free() is
3022 * located in iolog.c after printing this sample to the
3025 io_u_plat = (uint64_t *) td->ts.io_u_plat[FIO_CLAT][ddir];
3026 dst = malloc(sizeof(struct io_u_plat_entry));
3027 memcpy(&(dst->io_u_plat), io_u_plat,
3028 FIO_IO_U_PLAT_NR * sizeof(uint64_t));
3029 flist_add(&dst->list, &hw->list);
3030 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
3031 elapsed, offset, ioprio);
3034 * Update the last time we recorded as being now, minus
3035 * any drift in time we encountered before actually
3036 * making the record.
3038 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
3044 __td_io_u_unlock(td);
3047 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
3048 unsigned long long nsec, unsigned long long bs,
3049 uint64_t offset, unsigned int ioprio)
3051 const bool needs_lock = td_async_processing(td);
3052 struct thread_stat *ts = &td->ts;
3060 add_stat_sample(&ts->slat_stat[ddir], nsec);
3063 add_log_sample(td, td->slat_log, sample_val(nsec), ddir, bs,
3066 if (ts->slat_percentiles)
3067 add_lat_percentile_sample(ts, nsec, ddir, FIO_SLAT);
3070 __td_io_u_unlock(td);
3073 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
3074 unsigned long long nsec, unsigned long long bs,
3075 uint64_t offset, unsigned int ioprio, bool high_prio)
3077 const bool needs_lock = td_async_processing(td);
3078 struct thread_stat *ts = &td->ts;
3086 add_stat_sample(&ts->lat_stat[ddir], nsec);
3089 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
3093 * When lat_percentiles=1 (default 0), the reported high/low priority
3094 * percentiles and stats are used for describing total latency values,
3095 * even though the variable names themselves start with clat_.
3097 * Because of the above definition, add a prio stat and prio lat
3098 * percentile sample only when lat_percentiles=1. add_clat_sample() will
3099 * add the prio stat and prio lat percentile sample when
3100 * lat_percentiles=0.
3102 if (ts->lat_percentiles) {
3103 add_lat_percentile_sample(ts, nsec, ddir, FIO_LAT);
3104 add_lat_percentile_prio_sample(ts, nsec, ddir, high_prio);
3106 add_stat_sample(&ts->clat_high_prio_stat[ddir], nsec);
3108 add_stat_sample(&ts->clat_low_prio_stat[ddir], nsec);
3112 __td_io_u_unlock(td);
3115 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
3116 unsigned int bytes, unsigned long long spent)
3118 const bool needs_lock = td_async_processing(td);
3119 struct thread_stat *ts = &td->ts;
3123 rate = (unsigned long) (bytes * 1000000ULL / spent);
3130 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
3133 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
3134 bytes, io_u->offset, io_u->ioprio);
3136 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
3139 __td_io_u_unlock(td);
3142 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
3143 struct timespec *t, unsigned int avg_time,
3144 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
3145 struct io_stat *stat, struct io_log *log,
3148 const bool needs_lock = td_async_processing(td);
3149 unsigned long spent, rate;
3151 unsigned long next, next_log;
3153 next_log = avg_time;
3155 spent = mtime_since(parent_tv, t);
3156 if (spent < avg_time && avg_time - spent > LOG_MSEC_SLACK)
3157 return avg_time - spent;
3163 * Compute both read and write rates for the interval.
3165 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
3168 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
3170 continue; /* No entries for interval */
3174 rate = delta * 1000 / spent / 1024; /* KiB/s */
3176 rate = (delta * 1000) / spent;
3180 add_stat_sample(&stat[ddir], rate);
3183 unsigned long long bs = 0;
3185 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
3186 bs = td->o.min_bs[ddir];
3188 next = add_log_sample(td, log, sample_val(rate), ddir,
3190 next_log = min(next_log, next);
3193 stat_io_bytes[ddir] = this_io_bytes[ddir];
3199 __td_io_u_unlock(td);
3201 if (spent <= avg_time)
3204 next = avg_time - (1 + spent - avg_time);
3206 return min(next, next_log);
3209 static int add_bw_samples(struct thread_data *td, struct timespec *t)
3211 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
3212 td->this_io_bytes, td->stat_io_bytes,
3213 td->ts.bw_stat, td->bw_log, true);
3216 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
3219 const bool needs_lock = td_async_processing(td);
3220 struct thread_stat *ts = &td->ts;
3225 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
3228 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
3229 bytes, io_u->offset, io_u->ioprio);
3231 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
3234 __td_io_u_unlock(td);
3237 static int add_iops_samples(struct thread_data *td, struct timespec *t)
3239 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
3240 td->this_io_blocks, td->stat_io_blocks,
3241 td->ts.iops_stat, td->iops_log, false);
3245 * Returns msecs to next event
3247 int calc_log_samples(void)
3249 struct thread_data *td;
3250 unsigned int next = ~0U, tmp = 0, next_mod = 0, log_avg_msec_min = -1U;
3251 struct timespec now;
3253 long elapsed_time = 0;
3255 fio_gettime(&now, NULL);
3257 for_each_td(td, i) {
3258 elapsed_time = mtime_since_now(&td->epoch);
3262 if (in_ramp_time(td) ||
3263 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
3264 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
3268 (td->bw_log && !per_unit_log(td->bw_log))) {
3269 tmp = add_bw_samples(td, &now);
3272 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->bw_log->avg_msec);
3274 if (!td->iops_log ||
3275 (td->iops_log && !per_unit_log(td->iops_log))) {
3276 tmp = add_iops_samples(td, &now);
3279 log_avg_msec_min = min(log_avg_msec_min, (unsigned int)td->iops_log->avg_msec);
3286 /* if log_avg_msec_min has not been changed, set it to 0 */
3287 if (log_avg_msec_min == -1U)
3288 log_avg_msec_min = 0;
3290 if (log_avg_msec_min == 0)
3291 next_mod = elapsed_time;
3293 next_mod = elapsed_time % log_avg_msec_min;
3295 /* correction to keep the time on the log avg msec boundary */
3296 next = min(next, (log_avg_msec_min - next_mod));
3298 return next == ~0U ? 0 : next;
3301 void stat_init(void)
3303 stat_sem = fio_sem_init(FIO_SEM_UNLOCKED);
3306 void stat_exit(void)
3309 * When we have the mutex, we know out-of-band access to it
3312 fio_sem_down(stat_sem);
3313 fio_sem_remove(stat_sem);
3317 * Called from signal handler. Wake up status thread.
3319 void show_running_run_stats(void)
3324 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
3326 /* Ignore io_u's which span multiple blocks--they will just get
3327 * inaccurate counts. */
3328 int idx = (io_u->offset - io_u->file->file_offset)
3329 / td->o.bs[DDIR_TRIM];
3330 uint32_t *info = &td->ts.block_infos[idx];
3331 assert(idx < td->ts.nr_block_infos);