12 #include "lib/ieee754.h"
14 #include "lib/getrusage.h"
17 #include "lib/output_buffer.h"
18 #include "helper_thread.h"
21 #define LOG_MSEC_SLACK 10
23 struct fio_mutex *stat_mutex;
25 void clear_rusage_stat(struct thread_data *td)
27 struct thread_stat *ts = &td->ts;
29 fio_getrusage(&td->ru_start);
30 ts->usr_time = ts->sys_time = 0;
32 ts->minf = ts->majf = 0;
35 void update_rusage_stat(struct thread_data *td)
37 struct thread_stat *ts = &td->ts;
39 fio_getrusage(&td->ru_end);
40 ts->usr_time += mtime_since_tv(&td->ru_start.ru_utime,
41 &td->ru_end.ru_utime);
42 ts->sys_time += mtime_since_tv(&td->ru_start.ru_stime,
43 &td->ru_end.ru_stime);
44 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
45 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
46 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
47 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
49 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
53 * Given a latency, return the index of the corresponding bucket in
54 * the structure tracking percentiles.
56 * (1) find the group (and error bits) that the value (latency)
57 * belongs to by looking at its MSB. (2) find the bucket number in the
58 * group by looking at the index bits.
61 static unsigned int plat_val_to_idx(unsigned long long val)
63 unsigned int msb, error_bits, base, offset, idx;
65 /* Find MSB starting from bit 0 */
69 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
72 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
73 * all bits of the sample as index
75 if (msb <= FIO_IO_U_PLAT_BITS)
78 /* Compute the number of error bits to discard*/
79 error_bits = msb - FIO_IO_U_PLAT_BITS;
81 /* Compute the number of buckets before the group */
82 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
85 * Discard the error bits and apply the mask to find the
86 * index for the buckets in the group
88 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
90 /* Make sure the index does not exceed (array size - 1) */
91 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
92 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
98 * Convert the given index of the bucket array to the value
99 * represented by the bucket
101 static unsigned long long plat_idx_to_val(unsigned int idx)
103 unsigned int error_bits;
104 unsigned long long k, base;
106 assert(idx < FIO_IO_U_PLAT_NR);
108 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
109 * all bits of the sample as index */
110 if (idx < (FIO_IO_U_PLAT_VAL << 1))
113 /* Find the group and compute the minimum value of that group */
114 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
115 base = ((unsigned long long) 1) << (error_bits + FIO_IO_U_PLAT_BITS);
117 /* Find its bucket number of the group */
118 k = idx % FIO_IO_U_PLAT_VAL;
120 /* Return the mean of the range of the bucket */
121 return base + ((k + 0.5) * (1 << error_bits));
124 static int double_cmp(const void *a, const void *b)
126 const fio_fp64_t fa = *(const fio_fp64_t *) a;
127 const fio_fp64_t fb = *(const fio_fp64_t *) b;
132 else if (fa.u.f < fb.u.f)
138 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long long nr,
139 fio_fp64_t *plist, unsigned long long **output,
140 unsigned long long *maxv, unsigned long long *minv)
142 unsigned long long sum = 0;
143 unsigned int len, i, j = 0;
144 unsigned int oval_len = 0;
145 unsigned long long *ovals = NULL;
152 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
159 * Sort the percentile list. Note that it may already be sorted if
160 * we are using the default values, but since it's a short list this
161 * isn't a worry. Also note that this does not work for NaN values.
164 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
167 * Calculate bucket values, note down max and min values
170 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
172 while (sum >= (plist[j].u.f / 100.0 * nr)) {
173 assert(plist[j].u.f <= 100.0);
177 ovals = realloc(ovals, oval_len * sizeof(*ovals));
180 ovals[j] = plat_idx_to_val(i);
181 if (ovals[j] < *minv)
183 if (ovals[j] > *maxv)
186 is_last = (j == len - 1) != 0;
199 * Find and display the p-th percentile of clat
201 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long long nr,
202 fio_fp64_t *plist, unsigned int precision,
203 bool is_clat, struct buf_output *out)
205 unsigned int divisor, len, i, j = 0;
206 unsigned long long minv, maxv;
207 unsigned long long *ovals;
208 int per_line, scale_down, time_width;
209 const char *pre = is_clat ? "clat" : " lat";
213 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
218 * We default to nsecs, but if the value range is such that we
219 * should scale down to usecs or msecs, do that.
221 if (minv > 2000000 && maxv > 99999999ULL) {
224 log_buf(out, " %s percentiles (msec):\n |", pre);
225 } else if (minv > 2000 && maxv > 99999) {
228 log_buf(out, " %s percentiles (usec):\n |", pre);
232 log_buf(out, " %s percentiles (nsec):\n |", pre);
236 time_width = max(5, (int) (log10(maxv / divisor) + 1));
237 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
238 precision, time_width);
239 /* fmt will be something like " %5.2fth=[%4llu]%c" */
240 per_line = (80 - 7) / (precision + 10 + time_width);
242 for (j = 0; j < len; j++) {
244 if (j != 0 && (j % per_line) == 0)
247 /* end of the list */
248 is_last = (j == len - 1) != 0;
250 for (i = 0; i < scale_down; i++)
251 ovals[j] = (ovals[j] + 999) / 1000;
253 log_buf(out, fmt, plist[j].u.f, ovals[j], is_last ? '\n' : ',');
258 if ((j % per_line) == per_line - 1) /* for formatting */
267 bool calc_lat(struct io_stat *is, unsigned long long *min,
268 unsigned long long *max, double *mean, double *dev)
270 double n = (double) is->samples;
277 *mean = is->mean.u.f;
280 *dev = sqrt(is->S.u.f / (n - 1.0));
287 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
289 char *io, *agg, *min, *max;
290 char *ioalt, *aggalt, *minalt, *maxalt;
291 const char *str[] = { " READ", " WRITE" , " TRIM"};
294 log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
296 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
297 const int i2p = is_power_of_2(rs->kb_base);
302 io = num2str(rs->iobytes[i], rs->sig_figs, 1, i2p, N2S_BYTE);
303 ioalt = num2str(rs->iobytes[i], rs->sig_figs, 1, !i2p, N2S_BYTE);
304 agg = num2str(rs->agg[i], rs->sig_figs, 1, i2p, rs->unit_base);
305 aggalt = num2str(rs->agg[i], rs->sig_figs, 1, !i2p, rs->unit_base);
306 min = num2str(rs->min_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
307 minalt = num2str(rs->min_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
308 max = num2str(rs->max_bw[i], rs->sig_figs, 1, i2p, rs->unit_base);
309 maxalt = num2str(rs->max_bw[i], rs->sig_figs, 1, !i2p, rs->unit_base);
310 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
311 rs->unified_rw_rep ? " MIXED" : str[i],
312 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
313 (unsigned long long) rs->min_run[i],
314 (unsigned long long) rs->max_run[i]);
327 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
332 * Do depth distribution calculations
334 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
336 io_u_dist[i] = (double) map[i] / (double) total;
337 io_u_dist[i] *= 100.0;
338 if (io_u_dist[i] < 0.1 && map[i])
345 static void stat_calc_lat(struct thread_stat *ts, double *dst,
346 unsigned int *src, int nr)
348 unsigned long total = ddir_rw_sum(ts->total_io_u);
352 * Do latency distribution calculations
354 for (i = 0; i < nr; i++) {
356 dst[i] = (double) src[i] / (double) total;
358 if (dst[i] < 0.01 && src[i])
366 * To keep the terse format unaltered, add all of the ns latency
367 * buckets to the first us latency bucket
369 void stat_calc_lat_nu(struct thread_stat *ts, double *io_u_lat_u)
371 unsigned long ntotal = 0, total = ddir_rw_sum(ts->total_io_u);
374 stat_calc_lat(ts, io_u_lat_u, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
376 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
377 ntotal += ts->io_u_lat_n[i];
379 io_u_lat_u[0] += 100.0 * (double) ntotal / (double) total;
382 void stat_calc_lat_n(struct thread_stat *ts, double *io_u_lat)
384 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_n, FIO_IO_U_LAT_N_NR);
387 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
389 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
392 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
394 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
397 static void display_lat(const char *name, unsigned long long min,
398 unsigned long long max, double mean, double dev,
399 struct buf_output *out)
401 const char *base = "(nsec)";
404 if (nsec_to_msec(&min, &max, &mean, &dev))
406 else if (nsec_to_usec(&min, &max, &mean, &dev))
409 minp = num2str(min, 6, 1, 0, N2S_NONE);
410 maxp = num2str(max, 6, 1, 0, N2S_NONE);
412 log_buf(out, " %s %s: min=%s, max=%s, avg=%5.02f,"
413 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
419 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
420 int ddir, struct buf_output *out)
422 const char *str[] = { " read", "write", " trim" };
424 unsigned long long min, max, bw, iops;
426 char *io_p, *bw_p, *bw_p_alt, *iops_p;
429 assert(ddir_rw(ddir));
431 if (!ts->runtime[ddir])
434 i2p = is_power_of_2(rs->kb_base);
435 runt = ts->runtime[ddir];
437 bw = (1000 * ts->io_bytes[ddir]) / runt;
438 io_p = num2str(ts->io_bytes[ddir], ts->sig_figs, 1, i2p, N2S_BYTE);
439 bw_p = num2str(bw, ts->sig_figs, 1, i2p, ts->unit_base);
440 bw_p_alt = num2str(bw, ts->sig_figs, 1, !i2p, ts->unit_base);
442 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
443 iops_p = num2str(iops, ts->sig_figs, 1, 0, N2S_NONE);
445 log_buf(out, " %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
446 rs->unified_rw_rep ? "mixed" : str[ddir],
447 iops_p, bw_p, bw_p_alt, io_p,
448 (unsigned long long) ts->runtime[ddir]);
455 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
456 display_lat("slat", min, max, mean, dev, out);
457 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
458 display_lat("clat", min, max, mean, dev, out);
459 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
460 display_lat(" lat", min, max, mean, dev, out);
462 if (ts->clat_percentiles || ts->lat_percentiles) {
463 show_clat_percentiles(ts->io_u_plat[ddir],
464 ts->clat_stat[ddir].samples,
466 ts->percentile_precision,
467 ts->clat_percentiles, out);
469 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
470 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
473 if ((rs->unit_base == 1) && i2p)
475 else if (rs->unit_base == 1)
483 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
484 if (p_of_agg > 100.0)
488 if (rs->unit_base == 1) {
495 if (mean > fkb_base * fkb_base) {
500 bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
503 log_buf(out, " bw (%5s/s): min=%5llu, max=%5llu, per=%3.2f%%, "
504 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
505 bw_str, min, max, p_of_agg, mean, dev,
506 (&ts->bw_stat[ddir])->samples);
508 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
509 log_buf(out, " iops : min=%5llu, max=%5llu, "
510 "avg=%5.02f, stdev=%5.02f, samples=%" PRIu64 "\n",
511 min, max, mean, dev, (&ts->iops_stat[ddir])->samples);
515 static bool show_lat(double *io_u_lat, int nr, const char **ranges,
516 const char *msg, struct buf_output *out)
518 bool new_line = true, shown = false;
521 for (i = 0; i < nr; i++) {
522 if (io_u_lat[i] <= 0.0)
528 log_buf(out, " lat (%s) : ", msg);
534 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
546 static void show_lat_n(double *io_u_lat_n, struct buf_output *out)
548 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
549 "250=", "500=", "750=", "1000=", };
551 show_lat(io_u_lat_n, FIO_IO_U_LAT_N_NR, ranges, "nsec", out);
554 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
556 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
557 "250=", "500=", "750=", "1000=", };
559 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
562 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
564 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
565 "250=", "500=", "750=", "1000=", "2000=",
568 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
571 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
573 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
574 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
575 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
577 stat_calc_lat_n(ts, io_u_lat_n);
578 stat_calc_lat_u(ts, io_u_lat_u);
579 stat_calc_lat_m(ts, io_u_lat_m);
581 show_lat_n(io_u_lat_n, out);
582 show_lat_u(io_u_lat_u, out);
583 show_lat_m(io_u_lat_m, out);
586 static int block_state_category(int block_state)
588 switch (block_state) {
589 case BLOCK_STATE_UNINIT:
591 case BLOCK_STATE_TRIMMED:
592 case BLOCK_STATE_WRITTEN:
594 case BLOCK_STATE_WRITE_FAILURE:
595 case BLOCK_STATE_TRIM_FAILURE:
598 /* Silence compile warning on some BSDs and have a return */
604 static int compare_block_infos(const void *bs1, const void *bs2)
606 uint32_t block1 = *(uint32_t *)bs1;
607 uint32_t block2 = *(uint32_t *)bs2;
608 int state1 = BLOCK_INFO_STATE(block1);
609 int state2 = BLOCK_INFO_STATE(block2);
610 int bscat1 = block_state_category(state1);
611 int bscat2 = block_state_category(state2);
612 int cycles1 = BLOCK_INFO_TRIMS(block1);
613 int cycles2 = BLOCK_INFO_TRIMS(block2);
620 if (cycles1 < cycles2)
622 if (cycles1 > cycles2)
630 assert(block1 == block2);
634 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
635 fio_fp64_t *plist, unsigned int **percentiles,
641 qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
643 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
650 * Sort the percentile list. Note that it may already be sorted if
651 * we are using the default values, but since it's a short list this
652 * isn't a worry. Also note that this does not work for NaN values.
655 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
658 /* Start only after the uninit entries end */
660 nr_uninit < nr_block_infos
661 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
665 if (nr_uninit == nr_block_infos)
668 *percentiles = calloc(len, sizeof(**percentiles));
670 for (i = 0; i < len; i++) {
671 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
673 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
676 memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
677 for (i = 0; i < nr_block_infos; i++)
678 types[BLOCK_INFO_STATE(block_infos[i])]++;
683 static const char *block_state_names[] = {
684 [BLOCK_STATE_UNINIT] = "unwritten",
685 [BLOCK_STATE_TRIMMED] = "trimmed",
686 [BLOCK_STATE_WRITTEN] = "written",
687 [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
688 [BLOCK_STATE_WRITE_FAILURE] = "write failure",
691 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
692 fio_fp64_t *plist, struct buf_output *out)
695 unsigned int *percentiles = NULL;
696 unsigned int block_state_counts[BLOCK_STATE_COUNT];
698 len = calc_block_percentiles(nr_block_infos, block_infos, plist,
699 &percentiles, block_state_counts);
701 log_buf(out, " block lifetime percentiles :\n |");
703 for (i = 0; i < len; i++) {
704 uint32_t block_info = percentiles[i];
705 #define LINE_LENGTH 75
706 char str[LINE_LENGTH];
707 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
708 plist[i].u.f, block_info,
709 i == len - 1 ? '\n' : ',');
710 assert(strln < LINE_LENGTH);
711 if (pos + strln > LINE_LENGTH) {
713 log_buf(out, "\n |");
715 log_buf(out, "%s", str);
722 log_buf(out, " states :");
723 for (i = 0; i < BLOCK_STATE_COUNT; i++)
724 log_buf(out, " %s=%u%c",
725 block_state_names[i], block_state_counts[i],
726 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
729 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
731 char *p1, *p1alt, *p2;
732 unsigned long long bw_mean, iops_mean;
733 const int i2p = is_power_of_2(ts->kb_base);
738 bw_mean = steadystate_bw_mean(ts);
739 iops_mean = steadystate_iops_mean(ts);
741 p1 = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, i2p, ts->unit_base);
742 p1alt = num2str(bw_mean / ts->kb_base, ts->sig_figs, ts->kb_base, !i2p, ts->unit_base);
743 p2 = num2str(iops_mean, ts->sig_figs, 1, 0, N2S_NONE);
745 log_buf(out, " steadystate : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
746 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
748 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
749 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
750 ts->ss_criterion.u.f,
751 ts->ss_state & __FIO_SS_PCT ? "%" : "");
758 static void show_thread_status_normal(struct thread_stat *ts,
759 struct group_run_stats *rs,
760 struct buf_output *out)
762 double usr_cpu, sys_cpu;
763 unsigned long runtime;
764 double io_u_dist[FIO_IO_U_MAP_NR];
768 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
771 memset(time_buf, 0, sizeof(time_buf));
774 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
777 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
778 ts->name, ts->groupid, ts->members,
779 ts->error, (int) ts->pid, time_buf);
781 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
782 ts->name, ts->groupid, ts->members,
783 ts->error, ts->verror, (int) ts->pid,
787 if (strlen(ts->description))
788 log_buf(out, " Description : [%s]\n", ts->description);
790 if (ts->io_bytes[DDIR_READ])
791 show_ddir_status(rs, ts, DDIR_READ, out);
792 if (ts->io_bytes[DDIR_WRITE])
793 show_ddir_status(rs, ts, DDIR_WRITE, out);
794 if (ts->io_bytes[DDIR_TRIM])
795 show_ddir_status(rs, ts, DDIR_TRIM, out);
797 show_latencies(ts, out);
799 runtime = ts->total_run_time;
801 double runt = (double) runtime;
803 usr_cpu = (double) ts->usr_time * 100 / runt;
804 sys_cpu = (double) ts->sys_time * 100 / runt;
810 log_buf(out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
811 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
812 (unsigned long long) ts->ctx,
813 (unsigned long long) ts->majf,
814 (unsigned long long) ts->minf);
816 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
817 log_buf(out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
818 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
819 io_u_dist[1], io_u_dist[2],
820 io_u_dist[3], io_u_dist[4],
821 io_u_dist[5], io_u_dist[6]);
823 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
824 log_buf(out, " submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
825 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
826 io_u_dist[1], io_u_dist[2],
827 io_u_dist[3], io_u_dist[4],
828 io_u_dist[5], io_u_dist[6]);
829 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
830 log_buf(out, " complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
831 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
832 io_u_dist[1], io_u_dist[2],
833 io_u_dist[3], io_u_dist[4],
834 io_u_dist[5], io_u_dist[6]);
835 log_buf(out, " issued rwt: total=%llu,%llu,%llu,"
836 " short=%llu,%llu,%llu,"
837 " dropped=%llu,%llu,%llu\n",
838 (unsigned long long) ts->total_io_u[0],
839 (unsigned long long) ts->total_io_u[1],
840 (unsigned long long) ts->total_io_u[2],
841 (unsigned long long) ts->short_io_u[0],
842 (unsigned long long) ts->short_io_u[1],
843 (unsigned long long) ts->short_io_u[2],
844 (unsigned long long) ts->drop_io_u[0],
845 (unsigned long long) ts->drop_io_u[1],
846 (unsigned long long) ts->drop_io_u[2]);
847 if (ts->continue_on_error) {
848 log_buf(out, " errors : total=%llu, first_error=%d/<%s>\n",
849 (unsigned long long)ts->total_err_count,
851 strerror(ts->first_error));
853 if (ts->latency_depth) {
854 log_buf(out, " latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
855 (unsigned long long)ts->latency_target,
856 (unsigned long long)ts->latency_window,
857 ts->latency_percentile.u.f,
861 if (ts->nr_block_infos)
862 show_block_infos(ts->nr_block_infos, ts->block_infos,
863 ts->percentile_list, out);
866 show_ss_normal(ts, out);
869 static void show_ddir_status_terse(struct thread_stat *ts,
870 struct group_run_stats *rs, int ddir,
871 int ver, struct buf_output *out)
873 unsigned long long min, max, minv, maxv, bw, iops;
874 unsigned long long *ovals = NULL;
879 assert(ddir_rw(ddir));
882 if (ts->runtime[ddir]) {
883 uint64_t runt = ts->runtime[ddir];
885 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
886 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
889 log_buf(out, ";%llu;%llu;%llu;%llu",
890 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
891 (unsigned long long) ts->runtime[ddir]);
893 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
894 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
896 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
898 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
899 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
901 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
903 if (ts->clat_percentiles || ts->lat_percentiles) {
904 len = calc_clat_percentiles(ts->io_u_plat[ddir],
905 ts->clat_stat[ddir].samples,
906 ts->percentile_list, &ovals, &maxv,
911 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
913 log_buf(out, ";0%%=0");
916 log_buf(out, ";%f%%=%llu", ts->percentile_list[i].u.f, ovals[i]/1000);
919 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
920 log_buf(out, ";%llu;%llu;%f;%f", min/1000, max/1000, mean/1000, dev/1000);
922 log_buf(out, ";%llu;%llu;%f;%f", 0ULL, 0ULL, 0.0, 0.0);
927 bw_stat = calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev);
929 double p_of_agg = 100.0;
932 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
933 if (p_of_agg > 100.0)
937 log_buf(out, ";%llu;%llu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
939 log_buf(out, ";%llu;%llu;%f%%;%f;%f", 0ULL, 0ULL, 0.0, 0.0, 0.0);
943 log_buf(out, ";%" PRIu64, (&ts->bw_stat[ddir])->samples);
945 log_buf(out, ";%lu", 0UL);
947 if (calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev))
948 log_buf(out, ";%llu;%llu;%f;%f;%" PRIu64, min, max,
949 mean, dev, (&ts->iops_stat[ddir])->samples);
951 log_buf(out, ";%llu;%llu;%f;%f;%lu", 0ULL, 0ULL, 0.0, 0.0, 0UL);
955 static void add_ddir_status_json(struct thread_stat *ts,
956 struct group_run_stats *rs, int ddir, struct json_object *parent)
958 unsigned long long min, max, minv, maxv;
959 unsigned long long bw_bytes, bw;
960 unsigned long long *ovals = NULL;
961 double mean, dev, iops;
964 const char *ddirname[] = {"read", "write", "trim"};
965 struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object = NULL;
967 double p_of_agg = 100.0;
969 assert(ddir_rw(ddir));
971 if (ts->unified_rw_rep && ddir != DDIR_READ)
974 dir_object = json_create_object();
975 json_object_add_value_object(parent,
976 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
981 if (ts->runtime[ddir]) {
982 uint64_t runt = ts->runtime[ddir];
984 bw_bytes = ((1000 * ts->io_bytes[ddir]) / runt); /* Bytes/s */
985 bw = bw_bytes / 1024; /* KiB/s */
986 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
989 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir]);
990 json_object_add_value_int(dir_object, "io_kbytes", ts->io_bytes[ddir] >> 10);
991 json_object_add_value_int(dir_object, "bw_bytes", bw_bytes);
992 json_object_add_value_int(dir_object, "bw", bw);
993 json_object_add_value_float(dir_object, "iops", iops);
994 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
995 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
996 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
997 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
999 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
1003 tmp_object = json_create_object();
1004 json_object_add_value_object(dir_object, "slat_ns", tmp_object);
1005 json_object_add_value_int(tmp_object, "min", min);
1006 json_object_add_value_int(tmp_object, "max", max);
1007 json_object_add_value_float(tmp_object, "mean", mean);
1008 json_object_add_value_float(tmp_object, "stddev", dev);
1010 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
1014 tmp_object = json_create_object();
1015 json_object_add_value_object(dir_object, "clat_ns", tmp_object);
1016 json_object_add_value_int(tmp_object, "min", min);
1017 json_object_add_value_int(tmp_object, "max", max);
1018 json_object_add_value_float(tmp_object, "mean", mean);
1019 json_object_add_value_float(tmp_object, "stddev", dev);
1021 if (ts->clat_percentiles || ts->lat_percentiles) {
1022 len = calc_clat_percentiles(ts->io_u_plat[ddir],
1023 ts->clat_stat[ddir].samples,
1024 ts->percentile_list, &ovals, &maxv,
1029 percentile_object = json_create_object();
1030 json_object_add_value_object(tmp_object, "percentile", percentile_object);
1031 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1033 json_object_add_value_int(percentile_object, "0.00", 0);
1036 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
1037 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
1040 if (output_format & FIO_OUTPUT_JSON_PLUS) {
1041 clat_bins_object = json_create_object();
1042 if (ts->clat_percentiles)
1043 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1045 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
1046 if (ts->io_u_plat[ddir][i]) {
1047 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
1048 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
1053 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
1057 tmp_object = json_create_object();
1058 json_object_add_value_object(dir_object, "lat_ns", tmp_object);
1059 json_object_add_value_int(tmp_object, "min", min);
1060 json_object_add_value_int(tmp_object, "max", max);
1061 json_object_add_value_float(tmp_object, "mean", mean);
1062 json_object_add_value_float(tmp_object, "stddev", dev);
1063 if (output_format & FIO_OUTPUT_JSON_PLUS && ts->lat_percentiles)
1064 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
1069 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
1070 if (rs->agg[ddir]) {
1071 p_of_agg = mean * 100 / (double) (rs->agg[ddir] / 1024);
1072 if (p_of_agg > 100.0)
1077 p_of_agg = mean = dev = 0.0;
1079 json_object_add_value_int(dir_object, "bw_min", min);
1080 json_object_add_value_int(dir_object, "bw_max", max);
1081 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1082 json_object_add_value_float(dir_object, "bw_mean", mean);
1083 json_object_add_value_float(dir_object, "bw_dev", dev);
1084 json_object_add_value_int(dir_object, "bw_samples",
1085 (&ts->bw_stat[ddir])->samples);
1087 if (!calc_lat(&ts->iops_stat[ddir], &min, &max, &mean, &dev)) {
1091 json_object_add_value_int(dir_object, "iops_min", min);
1092 json_object_add_value_int(dir_object, "iops_max", max);
1093 json_object_add_value_float(dir_object, "iops_mean", mean);
1094 json_object_add_value_float(dir_object, "iops_stddev", dev);
1095 json_object_add_value_int(dir_object, "iops_samples",
1096 (&ts->iops_stat[ddir])->samples);
1099 static void show_thread_status_terse_all(struct thread_stat *ts,
1100 struct group_run_stats *rs, int ver,
1101 struct buf_output *out)
1103 double io_u_dist[FIO_IO_U_MAP_NR];
1104 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1105 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1106 double usr_cpu, sys_cpu;
1111 log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1113 log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1114 ts->name, ts->groupid, ts->error);
1116 /* Log Read Status */
1117 show_ddir_status_terse(ts, rs, DDIR_READ, ver, out);
1118 /* Log Write Status */
1119 show_ddir_status_terse(ts, rs, DDIR_WRITE, ver, out);
1120 /* Log Trim Status */
1121 if (ver == 2 || ver == 4 || ver == 5)
1122 show_ddir_status_terse(ts, rs, DDIR_TRIM, ver, out);
1125 if (ts->total_run_time) {
1126 double runt = (double) ts->total_run_time;
1128 usr_cpu = (double) ts->usr_time * 100 / runt;
1129 sys_cpu = (double) ts->sys_time * 100 / runt;
1135 log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1136 (unsigned long long) ts->ctx,
1137 (unsigned long long) ts->majf,
1138 (unsigned long long) ts->minf);
1140 /* Calc % distribution of IO depths, usecond, msecond latency */
1141 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1142 stat_calc_lat_nu(ts, io_u_lat_u);
1143 stat_calc_lat_m(ts, io_u_lat_m);
1145 /* Only show fixed 7 I/O depth levels*/
1146 log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1147 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1148 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1150 /* Microsecond latency */
1151 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1152 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1153 /* Millisecond latency */
1154 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1155 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1157 /* disk util stats, if any */
1159 show_disk_util(1, NULL, out);
1161 /* Additional output if continue_on_error set - default off*/
1162 if (ts->continue_on_error)
1163 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1167 /* Additional output if description is set */
1168 if (strlen(ts->description))
1169 log_buf(out, ";%s", ts->description);
1174 static void json_add_job_opts(struct json_object *root, const char *name,
1175 struct flist_head *opt_list, bool num_jobs)
1177 struct json_object *dir_object;
1178 struct flist_head *entry;
1179 struct print_option *p;
1181 if (flist_empty(opt_list))
1184 dir_object = json_create_object();
1185 json_object_add_value_object(root, name, dir_object);
1187 flist_for_each(entry, opt_list) {
1188 const char *pos = "";
1190 p = flist_entry(entry, struct print_option, list);
1191 if (!num_jobs && !strcmp(p->name, "numjobs"))
1195 json_object_add_value_string(dir_object, p->name, pos);
1199 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1200 struct group_run_stats *rs,
1201 struct flist_head *opt_list)
1203 struct json_object *root, *tmp;
1204 struct jobs_eta *je;
1205 double io_u_dist[FIO_IO_U_MAP_NR];
1206 double io_u_lat_n[FIO_IO_U_LAT_N_NR];
1207 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1208 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1209 double usr_cpu, sys_cpu;
1213 root = json_create_object();
1214 json_object_add_value_string(root, "jobname", ts->name);
1215 json_object_add_value_int(root, "groupid", ts->groupid);
1216 json_object_add_value_int(root, "error", ts->error);
1219 je = get_jobs_eta(true, &size);
1221 json_object_add_value_int(root, "eta", je->eta_sec);
1222 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1226 json_add_job_opts(root, "job options", opt_list, true);
1228 add_ddir_status_json(ts, rs, DDIR_READ, root);
1229 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1230 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1233 if (ts->total_run_time) {
1234 double runt = (double) ts->total_run_time;
1236 usr_cpu = (double) ts->usr_time * 100 / runt;
1237 sys_cpu = (double) ts->sys_time * 100 / runt;
1242 json_object_add_value_float(root, "usr_cpu", usr_cpu);
1243 json_object_add_value_float(root, "sys_cpu", sys_cpu);
1244 json_object_add_value_int(root, "ctx", ts->ctx);
1245 json_object_add_value_int(root, "majf", ts->majf);
1246 json_object_add_value_int(root, "minf", ts->minf);
1249 /* Calc % distribution of IO depths, usecond, msecond latency */
1250 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1251 stat_calc_lat_n(ts, io_u_lat_n);
1252 stat_calc_lat_u(ts, io_u_lat_u);
1253 stat_calc_lat_m(ts, io_u_lat_m);
1255 tmp = json_create_object();
1256 json_object_add_value_object(root, "iodepth_level", tmp);
1257 /* Only show fixed 7 I/O depth levels*/
1258 for (i = 0; i < 7; i++) {
1261 snprintf(name, 20, "%d", 1 << i);
1263 snprintf(name, 20, ">=%d", 1 << i);
1264 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1267 /* Nanosecond latency */
1268 tmp = json_create_object();
1269 json_object_add_value_object(root, "latency_ns", tmp);
1270 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++) {
1271 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1272 "250", "500", "750", "1000", };
1273 json_object_add_value_float(tmp, ranges[i], io_u_lat_n[i]);
1275 /* Microsecond latency */
1276 tmp = json_create_object();
1277 json_object_add_value_object(root, "latency_us", tmp);
1278 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1279 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1280 "250", "500", "750", "1000", };
1281 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1283 /* Millisecond latency */
1284 tmp = json_create_object();
1285 json_object_add_value_object(root, "latency_ms", tmp);
1286 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1287 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1288 "250", "500", "750", "1000", "2000",
1290 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1293 /* Additional output if continue_on_error set - default off*/
1294 if (ts->continue_on_error) {
1295 json_object_add_value_int(root, "total_err", ts->total_err_count);
1296 json_object_add_value_int(root, "first_error", ts->first_error);
1299 if (ts->latency_depth) {
1300 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1301 json_object_add_value_int(root, "latency_target", ts->latency_target);
1302 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1303 json_object_add_value_int(root, "latency_window", ts->latency_window);
1306 /* Additional output if description is set */
1307 if (strlen(ts->description))
1308 json_object_add_value_string(root, "desc", ts->description);
1310 if (ts->nr_block_infos) {
1311 /* Block error histogram and types */
1313 unsigned int *percentiles = NULL;
1314 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1316 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1317 ts->percentile_list,
1318 &percentiles, block_state_counts);
1321 struct json_object *block, *percentile_object, *states;
1323 block = json_create_object();
1324 json_object_add_value_object(root, "block", block);
1326 percentile_object = json_create_object();
1327 json_object_add_value_object(block, "percentiles",
1329 for (i = 0; i < len; i++) {
1331 snprintf(buf, sizeof(buf), "%f",
1332 ts->percentile_list[i].u.f);
1333 json_object_add_value_int(percentile_object,
1338 states = json_create_object();
1339 json_object_add_value_object(block, "states", states);
1340 for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1341 json_object_add_value_int(states,
1342 block_state_names[state],
1343 block_state_counts[state]);
1350 struct json_object *data;
1351 struct json_array *iops, *bw;
1355 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1356 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1357 ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1358 (float) ts->ss_limit.u.f,
1359 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1361 tmp = json_create_object();
1362 json_object_add_value_object(root, "steadystate", tmp);
1363 json_object_add_value_string(tmp, "ss", ss_buf);
1364 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1365 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1367 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1368 ts->ss_state & __FIO_SS_PCT ? "%" : "");
1369 json_object_add_value_string(tmp, "criterion", ss_buf);
1370 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1371 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1373 data = json_create_object();
1374 json_object_add_value_object(tmp, "data", data);
1375 bw = json_create_array();
1376 iops = json_create_array();
1379 ** if ss was attained or the buffer is not full,
1380 ** ss->head points to the first element in the list.
1381 ** otherwise it actually points to the second element
1384 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1387 j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1388 for (i = 0; i < ts->ss_dur; i++) {
1389 k = (j + i) % ts->ss_dur;
1390 json_array_add_value_int(bw, ts->ss_bw_data[k]);
1391 json_array_add_value_int(iops, ts->ss_iops_data[k]);
1393 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1394 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1395 json_object_add_value_array(data, "iops", iops);
1396 json_object_add_value_array(data, "bw", bw);
1402 static void show_thread_status_terse(struct thread_stat *ts,
1403 struct group_run_stats *rs,
1404 struct buf_output *out)
1406 if (terse_version >= 2 && terse_version <= 5)
1407 show_thread_status_terse_all(ts, rs, terse_version, out);
1409 log_err("fio: bad terse version!? %d\n", terse_version);
1412 struct json_object *show_thread_status(struct thread_stat *ts,
1413 struct group_run_stats *rs,
1414 struct flist_head *opt_list,
1415 struct buf_output *out)
1417 struct json_object *ret = NULL;
1419 if (output_format & FIO_OUTPUT_TERSE)
1420 show_thread_status_terse(ts, rs, out);
1421 if (output_format & FIO_OUTPUT_JSON)
1422 ret = show_thread_status_json(ts, rs, opt_list);
1423 if (output_format & FIO_OUTPUT_NORMAL)
1424 show_thread_status_normal(ts, rs, out);
1429 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1433 if (src->samples == 0)
1436 dst->min_val = min(dst->min_val, src->min_val);
1437 dst->max_val = max(dst->max_val, src->max_val);
1440 * Compute new mean and S after the merge
1441 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1442 * #Parallel_algorithm>
1445 mean = src->mean.u.f;
1448 double delta = src->mean.u.f - dst->mean.u.f;
1450 mean = ((src->mean.u.f * src->samples) +
1451 (dst->mean.u.f * dst->samples)) /
1452 (dst->samples + src->samples);
1454 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1455 (dst->samples * src->samples) /
1456 (dst->samples + src->samples);
1459 dst->samples += src->samples;
1460 dst->mean.u.f = mean;
1464 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1468 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1469 if (dst->max_run[i] < src->max_run[i])
1470 dst->max_run[i] = src->max_run[i];
1471 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1472 dst->min_run[i] = src->min_run[i];
1473 if (dst->max_bw[i] < src->max_bw[i])
1474 dst->max_bw[i] = src->max_bw[i];
1475 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1476 dst->min_bw[i] = src->min_bw[i];
1478 dst->iobytes[i] += src->iobytes[i];
1479 dst->agg[i] += src->agg[i];
1483 dst->kb_base = src->kb_base;
1484 if (!dst->unit_base)
1485 dst->unit_base = src->unit_base;
1488 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1493 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1494 if (!dst->unified_rw_rep) {
1495 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1496 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1497 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1498 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1499 sum_stat(&dst->iops_stat[l], &src->iops_stat[l], first);
1501 dst->io_bytes[l] += src->io_bytes[l];
1503 if (dst->runtime[l] < src->runtime[l])
1504 dst->runtime[l] = src->runtime[l];
1506 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1507 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1508 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1509 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1510 sum_stat(&dst->iops_stat[0], &src->iops_stat[l], first);
1512 dst->io_bytes[0] += src->io_bytes[l];
1514 if (dst->runtime[0] < src->runtime[l])
1515 dst->runtime[0] = src->runtime[l];
1518 * We're summing to the same destination, so override
1519 * 'first' after the first iteration of the loop
1525 dst->usr_time += src->usr_time;
1526 dst->sys_time += src->sys_time;
1527 dst->ctx += src->ctx;
1528 dst->majf += src->majf;
1529 dst->minf += src->minf;
1531 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1532 dst->io_u_map[k] += src->io_u_map[k];
1533 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1534 dst->io_u_submit[k] += src->io_u_submit[k];
1535 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1536 dst->io_u_complete[k] += src->io_u_complete[k];
1537 for (k = 0; k < FIO_IO_U_LAT_N_NR; k++)
1538 dst->io_u_lat_n[k] += src->io_u_lat_n[k];
1539 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1540 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1541 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1542 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1544 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1545 if (!dst->unified_rw_rep) {
1546 dst->total_io_u[k] += src->total_io_u[k];
1547 dst->short_io_u[k] += src->short_io_u[k];
1548 dst->drop_io_u[k] += src->drop_io_u[k];
1550 dst->total_io_u[0] += src->total_io_u[k];
1551 dst->short_io_u[0] += src->short_io_u[k];
1552 dst->drop_io_u[0] += src->drop_io_u[k];
1556 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1559 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1560 if (!dst->unified_rw_rep)
1561 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1563 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1567 dst->total_run_time += src->total_run_time;
1568 dst->total_submit += src->total_submit;
1569 dst->total_complete += src->total_complete;
1572 void init_group_run_stat(struct group_run_stats *gs)
1575 memset(gs, 0, sizeof(*gs));
1577 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1578 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1581 void init_thread_stat(struct thread_stat *ts)
1585 memset(ts, 0, sizeof(*ts));
1587 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1588 ts->lat_stat[j].min_val = -1UL;
1589 ts->clat_stat[j].min_val = -1UL;
1590 ts->slat_stat[j].min_val = -1UL;
1591 ts->bw_stat[j].min_val = -1UL;
1592 ts->iops_stat[j].min_val = -1UL;
1597 void __show_run_stats(void)
1599 struct group_run_stats *runstats, *rs;
1600 struct thread_data *td;
1601 struct thread_stat *threadstats, *ts;
1602 int i, j, k, nr_ts, last_ts, idx;
1603 bool kb_base_warned = false;
1604 bool unit_base_warned = false;
1605 struct json_object *root = NULL;
1606 struct json_array *array = NULL;
1607 struct buf_output output[FIO_OUTPUT_NR];
1608 struct flist_head **opt_lists;
1610 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1612 for (i = 0; i < groupid + 1; i++)
1613 init_group_run_stat(&runstats[i]);
1616 * find out how many threads stats we need. if group reporting isn't
1617 * enabled, it's one-per-td.
1621 for_each_td(td, i) {
1622 if (!td->o.group_reporting) {
1626 if (last_ts == td->groupid)
1631 last_ts = td->groupid;
1635 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1636 opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1638 for (i = 0; i < nr_ts; i++) {
1639 init_thread_stat(&threadstats[i]);
1640 opt_lists[i] = NULL;
1646 for_each_td(td, i) {
1649 if (idx && (!td->o.group_reporting ||
1650 (td->o.group_reporting && last_ts != td->groupid))) {
1655 last_ts = td->groupid;
1657 ts = &threadstats[j];
1659 ts->clat_percentiles = td->o.clat_percentiles;
1660 ts->lat_percentiles = td->o.lat_percentiles;
1661 ts->percentile_precision = td->o.percentile_precision;
1662 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1663 opt_lists[j] = &td->opt_list;
1668 if (ts->groupid == -1) {
1670 * These are per-group shared already
1672 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1673 if (td->o.description)
1674 strncpy(ts->description, td->o.description,
1675 FIO_JOBDESC_SIZE - 1);
1677 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1680 * If multiple entries in this group, this is
1683 ts->thread_number = td->thread_number;
1684 ts->groupid = td->groupid;
1687 * first pid in group, not very useful...
1691 ts->kb_base = td->o.kb_base;
1692 ts->unit_base = td->o.unit_base;
1693 ts->sig_figs = td->o.sig_figs;
1694 ts->unified_rw_rep = td->o.unified_rw_rep;
1695 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1696 log_info("fio: kb_base differs for jobs in group, using"
1697 " %u as the base\n", ts->kb_base);
1698 kb_base_warned = true;
1699 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1700 log_info("fio: unit_base differs for jobs in group, using"
1701 " %u as the base\n", ts->unit_base);
1702 unit_base_warned = true;
1705 ts->continue_on_error = td->o.continue_on_error;
1706 ts->total_err_count += td->total_err_count;
1707 ts->first_error = td->first_error;
1709 if (!td->error && td->o.continue_on_error &&
1711 ts->error = td->first_error;
1712 ts->verror[sizeof(ts->verror) - 1] = '\0';
1713 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1714 } else if (td->error) {
1715 ts->error = td->error;
1716 ts->verror[sizeof(ts->verror) - 1] = '\0';
1717 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1721 ts->latency_depth = td->latency_qd;
1722 ts->latency_target = td->o.latency_target;
1723 ts->latency_percentile = td->o.latency_percentile;
1724 ts->latency_window = td->o.latency_window;
1726 ts->nr_block_infos = td->ts.nr_block_infos;
1727 for (k = 0; k < ts->nr_block_infos; k++)
1728 ts->block_infos[k] = td->ts.block_infos[k];
1730 sum_thread_stats(ts, &td->ts, idx == 1);
1733 ts->ss_state = td->ss.state;
1734 ts->ss_dur = td->ss.dur;
1735 ts->ss_head = td->ss.head;
1736 ts->ss_bw_data = td->ss.bw_data;
1737 ts->ss_iops_data = td->ss.iops_data;
1738 ts->ss_limit.u.f = td->ss.limit;
1739 ts->ss_slope.u.f = td->ss.slope;
1740 ts->ss_deviation.u.f = td->ss.deviation;
1741 ts->ss_criterion.u.f = td->ss.criterion;
1744 ts->ss_dur = ts->ss_state = 0;
1747 for (i = 0; i < nr_ts; i++) {
1748 unsigned long long bw;
1750 ts = &threadstats[i];
1751 if (ts->groupid == -1)
1753 rs = &runstats[ts->groupid];
1754 rs->kb_base = ts->kb_base;
1755 rs->unit_base = ts->unit_base;
1756 rs->sig_figs = ts->sig_figs;
1757 rs->unified_rw_rep += ts->unified_rw_rep;
1759 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1760 if (!ts->runtime[j])
1762 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1763 rs->min_run[j] = ts->runtime[j];
1764 if (ts->runtime[j] > rs->max_run[j])
1765 rs->max_run[j] = ts->runtime[j];
1769 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1770 if (bw < rs->min_bw[j])
1772 if (bw > rs->max_bw[j])
1775 rs->iobytes[j] += ts->io_bytes[j];
1779 for (i = 0; i < groupid + 1; i++) {
1784 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1785 if (rs->max_run[ddir])
1786 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1791 for (i = 0; i < FIO_OUTPUT_NR; i++)
1792 buf_output_init(&output[i]);
1795 * don't overwrite last signal output
1797 if (output_format & FIO_OUTPUT_NORMAL)
1798 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1799 if (output_format & FIO_OUTPUT_JSON) {
1800 struct thread_data *global;
1803 unsigned long long ms_since_epoch;
1805 gettimeofday(&now, NULL);
1806 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1807 (unsigned long long)(now.tv_usec) / 1000;
1809 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1811 if (time_buf[strlen(time_buf) - 1] == '\n')
1812 time_buf[strlen(time_buf) - 1] = '\0';
1814 root = json_create_object();
1815 json_object_add_value_string(root, "fio version", fio_version_string);
1816 json_object_add_value_int(root, "timestamp", now.tv_sec);
1817 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1818 json_object_add_value_string(root, "time", time_buf);
1819 global = get_global_options();
1820 json_add_job_opts(root, "global options", &global->opt_list, false);
1821 array = json_create_array();
1822 json_object_add_value_array(root, "jobs", array);
1826 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1828 for (i = 0; i < nr_ts; i++) {
1829 ts = &threadstats[i];
1830 rs = &runstats[ts->groupid];
1833 fio_server_send_job_options(opt_lists[i], i);
1834 fio_server_send_ts(ts, rs);
1836 if (output_format & FIO_OUTPUT_TERSE)
1837 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1838 if (output_format & FIO_OUTPUT_JSON) {
1839 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1840 json_array_add_value_object(array, tmp);
1842 if (output_format & FIO_OUTPUT_NORMAL)
1843 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1846 if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1847 /* disk util stats, if any */
1848 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1850 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1852 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1853 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1854 json_free_object(root);
1857 for (i = 0; i < groupid + 1; i++) {
1862 fio_server_send_gs(rs);
1863 else if (output_format & FIO_OUTPUT_NORMAL)
1864 show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1868 fio_server_send_du();
1869 else if (output_format & FIO_OUTPUT_NORMAL) {
1870 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1871 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1874 for (i = 0; i < FIO_OUTPUT_NR; i++) {
1875 struct buf_output *out = &output[i];
1877 log_info_buf(out->buf, out->buflen);
1878 buf_output_free(out);
1887 void show_run_stats(void)
1889 fio_mutex_down(stat_mutex);
1891 fio_mutex_up(stat_mutex);
1894 void __show_running_run_stats(void)
1896 struct thread_data *td;
1897 unsigned long long *rt;
1901 fio_mutex_down(stat_mutex);
1903 rt = malloc(thread_number * sizeof(unsigned long long));
1904 fio_gettime(&ts, NULL);
1906 for_each_td(td, i) {
1907 td->update_rusage = 1;
1908 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1909 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1910 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1911 td->ts.total_run_time = mtime_since(&td->epoch, &ts);
1913 rt[i] = mtime_since(&td->start, &ts);
1914 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1915 td->ts.runtime[DDIR_READ] += rt[i];
1916 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1917 td->ts.runtime[DDIR_WRITE] += rt[i];
1918 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1919 td->ts.runtime[DDIR_TRIM] += rt[i];
1922 for_each_td(td, i) {
1923 if (td->runstate >= TD_EXITED)
1925 if (td->rusage_sem) {
1926 td->update_rusage = 1;
1927 fio_mutex_down(td->rusage_sem);
1929 td->update_rusage = 0;
1934 for_each_td(td, i) {
1935 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1936 td->ts.runtime[DDIR_READ] -= rt[i];
1937 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1938 td->ts.runtime[DDIR_WRITE] -= rt[i];
1939 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1940 td->ts.runtime[DDIR_TRIM] -= rt[i];
1944 fio_mutex_up(stat_mutex);
1947 static bool status_interval_init;
1948 static struct timespec status_time;
1949 static bool status_file_disabled;
1951 #define FIO_STATUS_FILE "fio-dump-status"
1953 static int check_status_file(void)
1956 const char *temp_dir;
1957 char fio_status_file_path[PATH_MAX];
1959 if (status_file_disabled)
1962 temp_dir = getenv("TMPDIR");
1963 if (temp_dir == NULL) {
1964 temp_dir = getenv("TEMP");
1965 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1968 if (temp_dir == NULL)
1971 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1973 if (stat(fio_status_file_path, &sb))
1976 if (unlink(fio_status_file_path) < 0) {
1977 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1979 log_err("fio: disabling status file updates\n");
1980 status_file_disabled = true;
1986 void check_for_running_stats(void)
1988 if (status_interval) {
1989 if (!status_interval_init) {
1990 fio_gettime(&status_time, NULL);
1991 status_interval_init = true;
1992 } else if (mtime_since_now(&status_time) >= status_interval) {
1993 show_running_run_stats();
1994 fio_gettime(&status_time, NULL);
1998 if (check_status_file()) {
1999 show_running_run_stats();
2004 static inline void add_stat_sample(struct io_stat *is, unsigned long long data)
2009 if (data > is->max_val)
2011 if (data < is->min_val)
2014 delta = val - is->mean.u.f;
2016 is->mean.u.f += delta / (is->samples + 1.0);
2017 is->S.u.f += delta * (val - is->mean.u.f);
2024 * Return a struct io_logs, which is added to the tail of the log
2027 static struct io_logs *get_new_log(struct io_log *iolog)
2029 size_t new_size, new_samples;
2030 struct io_logs *cur_log;
2033 * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
2036 if (!iolog->cur_log_max)
2037 new_samples = DEF_LOG_ENTRIES;
2039 new_samples = iolog->cur_log_max * 2;
2040 if (new_samples > MAX_LOG_ENTRIES)
2041 new_samples = MAX_LOG_ENTRIES;
2044 new_size = new_samples * log_entry_sz(iolog);
2046 cur_log = smalloc(sizeof(*cur_log));
2048 INIT_FLIST_HEAD(&cur_log->list);
2049 cur_log->log = malloc(new_size);
2051 cur_log->nr_samples = 0;
2052 cur_log->max_samples = new_samples;
2053 flist_add_tail(&cur_log->list, &iolog->io_logs);
2054 iolog->cur_log_max = new_samples;
2064 * Add and return a new log chunk, or return current log if big enough
2066 static struct io_logs *regrow_log(struct io_log *iolog)
2068 struct io_logs *cur_log;
2071 if (!iolog || iolog->disabled)
2074 cur_log = iolog_cur_log(iolog);
2076 cur_log = get_new_log(iolog);
2081 if (cur_log->nr_samples < cur_log->max_samples)
2085 * No room for a new sample. If we're compressing on the fly, flush
2086 * out the current chunk
2088 if (iolog->log_gz) {
2089 if (iolog_cur_flush(iolog, cur_log)) {
2090 log_err("fio: failed flushing iolog! Will stop logging.\n");
2096 * Get a new log array, and add to our list
2098 cur_log = get_new_log(iolog);
2100 log_err("fio: failed extending iolog! Will stop logging.\n");
2104 if (!iolog->pending || !iolog->pending->nr_samples)
2108 * Flush pending items to new log
2110 for (i = 0; i < iolog->pending->nr_samples; i++) {
2111 struct io_sample *src, *dst;
2113 src = get_sample(iolog, iolog->pending, i);
2114 dst = get_sample(iolog, cur_log, i);
2115 memcpy(dst, src, log_entry_sz(iolog));
2117 cur_log->nr_samples = iolog->pending->nr_samples;
2119 iolog->pending->nr_samples = 0;
2123 iolog->disabled = true;
2127 void regrow_logs(struct thread_data *td)
2129 regrow_log(td->slat_log);
2130 regrow_log(td->clat_log);
2131 regrow_log(td->clat_hist_log);
2132 regrow_log(td->lat_log);
2133 regrow_log(td->bw_log);
2134 regrow_log(td->iops_log);
2135 td->flags &= ~TD_F_REGROW_LOGS;
2138 static struct io_logs *get_cur_log(struct io_log *iolog)
2140 struct io_logs *cur_log;
2142 cur_log = iolog_cur_log(iolog);
2144 cur_log = get_new_log(iolog);
2149 if (cur_log->nr_samples < cur_log->max_samples)
2153 * Out of space. If we're in IO offload mode, or we're not doing
2154 * per unit logging (hence logging happens outside of the IO thread
2155 * as well), add a new log chunk inline. If we're doing inline
2156 * submissions, flag 'td' as needing a log regrow and we'll take
2157 * care of it on the submission side.
2159 if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2160 !per_unit_log(iolog))
2161 return regrow_log(iolog);
2163 iolog->td->flags |= TD_F_REGROW_LOGS;
2164 assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2165 return iolog->pending;
2168 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2169 enum fio_ddir ddir, unsigned int bs,
2170 unsigned long t, uint64_t offset)
2172 struct io_logs *cur_log;
2174 if (iolog->disabled)
2176 if (flist_empty(&iolog->io_logs))
2177 iolog->avg_last[ddir] = t;
2179 cur_log = get_cur_log(iolog);
2181 struct io_sample *s;
2183 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2186 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2187 io_sample_set_ddir(iolog, s, ddir);
2190 if (iolog->log_offset) {
2191 struct io_sample_offset *so = (void *) s;
2193 so->offset = offset;
2196 cur_log->nr_samples++;
2200 iolog->disabled = true;
2203 static inline void reset_io_stat(struct io_stat *ios)
2205 ios->max_val = ios->min_val = ios->samples = 0;
2206 ios->mean.u.f = ios->S.u.f = 0;
2209 void reset_io_stats(struct thread_data *td)
2211 struct thread_stat *ts = &td->ts;
2214 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2215 reset_io_stat(&ts->clat_stat[i]);
2216 reset_io_stat(&ts->slat_stat[i]);
2217 reset_io_stat(&ts->lat_stat[i]);
2218 reset_io_stat(&ts->bw_stat[i]);
2219 reset_io_stat(&ts->iops_stat[i]);
2221 ts->io_bytes[i] = 0;
2223 ts->total_io_u[i] = 0;
2224 ts->short_io_u[i] = 0;
2225 ts->drop_io_u[i] = 0;
2227 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2228 ts->io_u_plat[i][j] = 0;
2231 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2232 ts->io_u_map[i] = 0;
2233 ts->io_u_submit[i] = 0;
2234 ts->io_u_complete[i] = 0;
2237 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
2238 ts->io_u_lat_n[i] = 0;
2239 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2240 ts->io_u_lat_u[i] = 0;
2241 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2242 ts->io_u_lat_m[i] = 0;
2244 ts->total_submit = 0;
2245 ts->total_complete = 0;
2248 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2249 unsigned long elapsed, bool log_max)
2252 * Note an entry in the log. Use the mean from the logged samples,
2253 * making sure to properly round up. Only write a log entry if we
2254 * had actual samples done.
2256 if (iolog->avg_window[ddir].samples) {
2257 union io_sample_data data;
2260 data.val = iolog->avg_window[ddir].max_val;
2262 data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2264 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2267 reset_io_stat(&iolog->avg_window[ddir]);
2270 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2275 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2276 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2279 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2280 union io_sample_data data, enum fio_ddir ddir,
2281 unsigned int bs, uint64_t offset)
2283 unsigned long elapsed, this_window;
2288 elapsed = mtime_since_now(&td->epoch);
2291 * If no time averaging, just add the log sample.
2293 if (!iolog->avg_msec) {
2294 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2299 * Add the sample. If the time period has passed, then
2300 * add that entry to the log and clear.
2302 add_stat_sample(&iolog->avg_window[ddir], data.val);
2305 * If period hasn't passed, adding the above sample is all we
2308 this_window = elapsed - iolog->avg_last[ddir];
2309 if (elapsed < iolog->avg_last[ddir])
2310 return iolog->avg_last[ddir] - elapsed;
2311 else if (this_window < iolog->avg_msec) {
2312 int diff = iolog->avg_msec - this_window;
2314 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2318 __add_stat_to_log(iolog, ddir, elapsed, td->o.log_max != 0);
2320 iolog->avg_last[ddir] = elapsed - (this_window - iolog->avg_msec);
2321 return iolog->avg_msec;
2324 void finalize_logs(struct thread_data *td, bool unit_logs)
2326 unsigned long elapsed;
2328 elapsed = mtime_since_now(&td->epoch);
2330 if (td->clat_log && unit_logs)
2331 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2332 if (td->slat_log && unit_logs)
2333 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2334 if (td->lat_log && unit_logs)
2335 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2336 if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2337 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2338 if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2339 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2342 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2344 struct io_log *iolog;
2349 iolog = agg_io_log[ddir];
2350 __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2353 static void add_clat_percentile_sample(struct thread_stat *ts,
2354 unsigned long long nsec, enum fio_ddir ddir)
2356 unsigned int idx = plat_val_to_idx(nsec);
2357 assert(idx < FIO_IO_U_PLAT_NR);
2359 ts->io_u_plat[ddir][idx]++;
2362 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2363 unsigned long long nsec, unsigned int bs, uint64_t offset)
2365 unsigned long elapsed, this_window;
2366 struct thread_stat *ts = &td->ts;
2367 struct io_log *iolog = td->clat_hist_log;
2371 add_stat_sample(&ts->clat_stat[ddir], nsec);
2374 add_log_sample(td, td->clat_log, sample_val(nsec), ddir, bs,
2377 if (ts->clat_percentiles)
2378 add_clat_percentile_sample(ts, nsec, ddir);
2380 if (iolog && iolog->hist_msec) {
2381 struct io_hist *hw = &iolog->hist_window[ddir];
2384 elapsed = mtime_since_now(&td->epoch);
2386 hw->hist_last = elapsed;
2387 this_window = elapsed - hw->hist_last;
2389 if (this_window >= iolog->hist_msec) {
2390 unsigned int *io_u_plat;
2391 struct io_u_plat_entry *dst;
2394 * Make a byte-for-byte copy of the latency histogram
2395 * stored in td->ts.io_u_plat[ddir], recording it in a
2396 * log sample. Note that the matching call to free() is
2397 * located in iolog.c after printing this sample to the
2400 io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2401 dst = malloc(sizeof(struct io_u_plat_entry));
2402 memcpy(&(dst->io_u_plat), io_u_plat,
2403 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2404 flist_add(&dst->list, &hw->list);
2405 __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2409 * Update the last time we recorded as being now, minus
2410 * any drift in time we encountered before actually
2411 * making the record.
2413 hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2421 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2422 unsigned long usec, unsigned int bs, uint64_t offset)
2424 struct thread_stat *ts = &td->ts;
2431 add_stat_sample(&ts->slat_stat[ddir], usec);
2434 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2439 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2440 unsigned long long nsec, unsigned int bs, uint64_t offset)
2442 struct thread_stat *ts = &td->ts;
2449 add_stat_sample(&ts->lat_stat[ddir], nsec);
2452 add_log_sample(td, td->lat_log, sample_val(nsec), ddir, bs,
2455 if (ts->lat_percentiles)
2456 add_clat_percentile_sample(ts, nsec, ddir);
2461 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2462 unsigned int bytes, unsigned long long spent)
2464 struct thread_stat *ts = &td->ts;
2468 rate = (unsigned long) (bytes * 1000000ULL / spent);
2474 add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2477 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2478 bytes, io_u->offset);
2480 td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2484 static int __add_samples(struct thread_data *td, struct timespec *parent_tv,
2485 struct timespec *t, unsigned int avg_time,
2486 uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2487 struct io_stat *stat, struct io_log *log,
2490 unsigned long spent, rate;
2492 unsigned int next, next_log;
2494 next_log = avg_time;
2496 spent = mtime_since(parent_tv, t);
2497 if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2498 return avg_time - spent;
2503 * Compute both read and write rates for the interval.
2505 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2508 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2510 continue; /* No entries for interval */
2514 rate = delta * 1000 / spent / 1024; /* KiB/s */
2516 rate = (delta * 1000) / spent;
2520 add_stat_sample(&stat[ddir], rate);
2523 unsigned int bs = 0;
2525 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2526 bs = td->o.min_bs[ddir];
2528 next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2529 next_log = min(next_log, next);
2532 stat_io_bytes[ddir] = this_io_bytes[ddir];
2535 timespec_add_msec(parent_tv, avg_time);
2539 if (spent <= avg_time)
2542 next = avg_time - (1 + spent - avg_time);
2544 return min(next, next_log);
2547 static int add_bw_samples(struct thread_data *td, struct timespec *t)
2549 return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2550 td->this_io_bytes, td->stat_io_bytes,
2551 td->ts.bw_stat, td->bw_log, true);
2554 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2557 struct thread_stat *ts = &td->ts;
2561 add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2564 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2565 bytes, io_u->offset);
2567 td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2571 static int add_iops_samples(struct thread_data *td, struct timespec *t)
2573 return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2574 td->this_io_blocks, td->stat_io_blocks,
2575 td->ts.iops_stat, td->iops_log, false);
2579 * Returns msecs to next event
2581 int calc_log_samples(void)
2583 struct thread_data *td;
2584 unsigned int next = ~0U, tmp;
2585 struct timespec now;
2588 fio_gettime(&now, NULL);
2590 for_each_td(td, i) {
2593 if (in_ramp_time(td) ||
2594 !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2595 next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2599 (td->bw_log && !per_unit_log(td->bw_log))) {
2600 tmp = add_bw_samples(td, &now);
2604 if (!td->iops_log ||
2605 (td->iops_log && !per_unit_log(td->iops_log))) {
2606 tmp = add_iops_samples(td, &now);
2612 return next == ~0U ? 0 : next;
2615 void stat_init(void)
2617 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2620 void stat_exit(void)
2623 * When we have the mutex, we know out-of-band access to it
2626 fio_mutex_down(stat_mutex);
2627 fio_mutex_remove(stat_mutex);
2631 * Called from signal handler. Wake up status thread.
2633 void show_running_run_stats(void)
2638 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2640 /* Ignore io_u's which span multiple blocks--they will just get
2641 * inaccurate counts. */
2642 int idx = (io_u->offset - io_u->file->file_offset)
2643 / td->o.bs[DDIR_TRIM];
2644 uint32_t *info = &td->ts.block_infos[idx];
2645 assert(idx < td->ts.nr_block_infos);