12 #include "lib/ieee754.h"
14 #include "lib/getrusage.h"
17 static struct fio_mutex *stat_mutex;
19 void update_rusage_stat(struct thread_data *td)
21 struct thread_stat *ts = &td->ts;
23 fio_getrusage(&td->ru_end);
24 ts->usr_time += mtime_since(&td->ru_start.ru_utime,
25 &td->ru_end.ru_utime);
26 ts->sys_time += mtime_since(&td->ru_start.ru_stime,
27 &td->ru_end.ru_stime);
28 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
29 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
30 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
31 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
33 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
37 * Given a latency, return the index of the corresponding bucket in
38 * the structure tracking percentiles.
40 * (1) find the group (and error bits) that the value (latency)
41 * belongs to by looking at its MSB. (2) find the bucket number in the
42 * group by looking at the index bits.
45 static unsigned int plat_val_to_idx(unsigned int val)
47 unsigned int msb, error_bits, base, offset, idx;
49 /* Find MSB starting from bit 0 */
53 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
56 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
57 * all bits of the sample as index
59 if (msb <= FIO_IO_U_PLAT_BITS)
62 /* Compute the number of error bits to discard*/
63 error_bits = msb - FIO_IO_U_PLAT_BITS;
65 /* Compute the number of buckets before the group */
66 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
69 * Discard the error bits and apply the mask to find the
70 * index for the buckets in the group
72 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
74 /* Make sure the index does not exceed (array size - 1) */
75 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
76 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
82 * Convert the given index of the bucket array to the value
83 * represented by the bucket
85 static unsigned int plat_idx_to_val(unsigned int idx)
87 unsigned int error_bits, k, base;
89 assert(idx < FIO_IO_U_PLAT_NR);
91 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
92 * all bits of the sample as index */
93 if (idx < (FIO_IO_U_PLAT_VAL << 1))
96 /* Find the group and compute the minimum value of that group */
97 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
98 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
100 /* Find its bucket number of the group */
101 k = idx % FIO_IO_U_PLAT_VAL;
103 /* Return the mean of the range of the bucket */
104 return base + ((k + 0.5) * (1 << error_bits));
107 static int double_cmp(const void *a, const void *b)
109 const fio_fp64_t fa = *(const fio_fp64_t *) a;
110 const fio_fp64_t fb = *(const fio_fp64_t *) b;
115 else if (fa.u.f < fb.u.f)
121 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
122 fio_fp64_t *plist, unsigned int **output,
123 unsigned int *maxv, unsigned int *minv)
125 unsigned long sum = 0;
126 unsigned int len, i, j = 0;
127 unsigned int oval_len = 0;
128 unsigned int *ovals = NULL;
135 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
142 * Sort the percentile list. Note that it may already be sorted if
143 * we are using the default values, but since it's a short list this
144 * isn't a worry. Also note that this does not work for NaN values.
147 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
150 * Calculate bucket values, note down max and min values
153 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
155 while (sum >= (plist[j].u.f / 100.0 * nr)) {
156 assert(plist[j].u.f <= 100.0);
160 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
163 ovals[j] = plat_idx_to_val(i);
164 if (ovals[j] < *minv)
166 if (ovals[j] > *maxv)
169 is_last = (j == len - 1);
182 * Find and display the p-th percentile of clat
184 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
185 fio_fp64_t *plist, unsigned int precision)
187 unsigned int len, j = 0, minv, maxv;
189 int is_last, per_line, scale_down;
192 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
197 * We default to usecs, but if the value range is such that we
198 * should scale down to msecs, do that.
200 if (minv > 2000 && maxv > 99999) {
202 log_info(" clat percentiles (msec):\n |");
205 log_info(" clat percentiles (usec):\n |");
208 snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
209 per_line = (80 - 7) / (precision + 14);
211 for (j = 0; j < len; j++) {
212 char fbuf[16], *ptr = fbuf;
215 if (j != 0 && (j % per_line) == 0)
218 /* end of the list */
219 is_last = (j == len - 1);
221 if (plist[j].u.f < 10.0)
222 ptr += sprintf(fbuf, " ");
224 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
227 ovals[j] = (ovals[j] + 999) / 1000;
229 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
234 if ((j % per_line) == per_line - 1) /* for formatting */
243 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
244 double *mean, double *dev)
246 double n = (double) is->samples;
253 *mean = is->mean.u.f;
256 *dev = sqrt(is->S.u.f / (n - 1.0));
263 void show_group_stats(struct group_run_stats *rs)
265 char *p1, *p2, *p3, *p4;
266 const char *ddir_str[] = { " READ", " WRITE" , " TRIM"};
269 log_info("\nRun status group %d (all jobs):\n", rs->groupid);
271 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
272 const int i2p = is_power_of_2(rs->kb_base);
277 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
278 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
279 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
280 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
282 log_info("%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
283 " mint=%llumsec, maxt=%llumsec\n",
284 rs->unified_rw_rep ? " MIXED" : ddir_str[i],
286 (unsigned long long) rs->min_run[i],
287 (unsigned long long) rs->max_run[i]);
296 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
301 * Do depth distribution calculations
303 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
305 io_u_dist[i] = (double) map[i] / (double) total;
306 io_u_dist[i] *= 100.0;
307 if (io_u_dist[i] < 0.1 && map[i])
314 static void stat_calc_lat(struct thread_stat *ts, double *dst,
315 unsigned int *src, int nr)
317 unsigned long total = ddir_rw_sum(ts->total_io_u);
321 * Do latency distribution calculations
323 for (i = 0; i < nr; i++) {
325 dst[i] = (double) src[i] / (double) total;
327 if (dst[i] < 0.01 && src[i])
334 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
336 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
339 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
341 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
344 static void display_lat(const char *name, unsigned long min, unsigned long max,
345 double mean, double dev)
347 const char *base = "(usec)";
350 if (!usec_to_msec(&min, &max, &mean, &dev))
353 minp = num2str(min, 6, 1, 0, 0);
354 maxp = num2str(max, 6, 1, 0, 0);
356 log_info(" %s %s: min=%s, max=%s, avg=%5.02f,"
357 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
363 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
366 const char *ddir_str[] = { "read ", "write", "trim" };
367 unsigned long min, max, runt;
368 unsigned long long bw, iops;
370 char *io_p, *bw_p, *iops_p;
373 assert(ddir_rw(ddir));
375 if (!ts->runtime[ddir])
378 i2p = is_power_of_2(rs->kb_base);
379 runt = ts->runtime[ddir];
381 bw = (1000 * ts->io_bytes[ddir]) / runt;
382 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
383 bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
385 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
386 iops_p = num2str(iops, 6, 1, 0, 0);
388 log_info(" %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
389 rs->unified_rw_rep ? "mixed" : ddir_str[ddir],
391 (unsigned long long) ts->runtime[ddir]);
397 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
398 display_lat("slat", min, max, mean, dev);
399 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
400 display_lat("clat", min, max, mean, dev);
401 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
402 display_lat(" lat", min, max, mean, dev);
404 if (ts->clat_percentiles) {
405 show_clat_percentiles(ts->io_u_plat[ddir],
406 ts->clat_stat[ddir].samples,
408 ts->percentile_precision);
410 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
411 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
412 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
414 if (rs->unit_base == 1) {
422 p_of_agg = mean * 100 / (double) rs->agg[ddir];
423 if (p_of_agg > 100.0)
427 if (mean > fkb_base * fkb_base) {
432 bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
435 log_info(" bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
436 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
437 p_of_agg, mean, dev);
441 static int show_lat(double *io_u_lat, int nr, const char **ranges,
444 int new_line = 1, i, line = 0, shown = 0;
446 for (i = 0; i < nr; i++) {
447 if (io_u_lat[i] <= 0.0)
453 log_info(" lat (%s) : ", msg);
459 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
471 static void show_lat_u(double *io_u_lat_u)
473 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
474 "250=", "500=", "750=", "1000=", };
476 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
479 static void show_lat_m(double *io_u_lat_m)
481 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
482 "250=", "500=", "750=", "1000=", "2000=",
485 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
488 static void show_latencies(struct thread_stat *ts)
490 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
491 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
493 stat_calc_lat_u(ts, io_u_lat_u);
494 stat_calc_lat_m(ts, io_u_lat_m);
496 show_lat_u(io_u_lat_u);
497 show_lat_m(io_u_lat_m);
500 void show_thread_status_normal(struct thread_stat *ts, struct group_run_stats *rs)
502 double usr_cpu, sys_cpu;
503 unsigned long runtime;
504 double io_u_dist[FIO_IO_U_MAP_NR];
508 if (!(ts->io_bytes[DDIR_READ] + ts->io_bytes[DDIR_WRITE] +
509 ts->io_bytes[DDIR_TRIM]) && !(ts->total_io_u[DDIR_READ] +
510 ts->total_io_u[DDIR_WRITE] + ts->total_io_u[DDIR_TRIM]))
514 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
517 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
518 ts->name, ts->groupid, ts->members,
519 ts->error, (int) ts->pid, time_buf);
521 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
522 ts->name, ts->groupid, ts->members,
523 ts->error, ts->verror, (int) ts->pid,
527 if (strlen(ts->description))
528 log_info(" Description : [%s]\n", ts->description);
530 if (ts->io_bytes[DDIR_READ])
531 show_ddir_status(rs, ts, DDIR_READ);
532 if (ts->io_bytes[DDIR_WRITE])
533 show_ddir_status(rs, ts, DDIR_WRITE);
534 if (ts->io_bytes[DDIR_TRIM])
535 show_ddir_status(rs, ts, DDIR_TRIM);
539 runtime = ts->total_run_time;
541 double runt = (double) runtime;
543 usr_cpu = (double) ts->usr_time * 100 / runt;
544 sys_cpu = (double) ts->sys_time * 100 / runt;
550 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
551 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
552 (unsigned long long) ts->ctx,
553 (unsigned long long) ts->majf,
554 (unsigned long long) ts->minf);
556 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
557 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
558 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
559 io_u_dist[1], io_u_dist[2],
560 io_u_dist[3], io_u_dist[4],
561 io_u_dist[5], io_u_dist[6]);
563 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
564 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
565 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
566 io_u_dist[1], io_u_dist[2],
567 io_u_dist[3], io_u_dist[4],
568 io_u_dist[5], io_u_dist[6]);
569 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
570 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
571 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
572 io_u_dist[1], io_u_dist[2],
573 io_u_dist[3], io_u_dist[4],
574 io_u_dist[5], io_u_dist[6]);
575 log_info(" issued : total=r=%llu/w=%llu/d=%llu,"
576 " short=r=%llu/w=%llu/d=%llu\n",
577 (unsigned long long) ts->total_io_u[0],
578 (unsigned long long) ts->total_io_u[1],
579 (unsigned long long) ts->total_io_u[2],
580 (unsigned long long) ts->short_io_u[0],
581 (unsigned long long) ts->short_io_u[1],
582 (unsigned long long) ts->short_io_u[2]);
583 if (ts->continue_on_error) {
584 log_info(" errors : total=%llu, first_error=%d/<%s>\n",
585 (unsigned long long)ts->total_err_count,
587 strerror(ts->first_error));
589 if (ts->latency_depth) {
590 log_info(" latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
591 (unsigned long long)ts->latency_target,
592 (unsigned long long)ts->latency_window,
593 ts->latency_percentile.u.f,
598 static void show_ddir_status_terse(struct thread_stat *ts,
599 struct group_run_stats *rs, int ddir)
601 unsigned long min, max;
602 unsigned long long bw, iops;
603 unsigned int *ovals = NULL;
605 unsigned int len, minv, maxv;
608 assert(ddir_rw(ddir));
611 if (ts->runtime[ddir]) {
612 uint64_t runt = ts->runtime[ddir];
614 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
615 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
618 log_info(";%llu;%llu;%llu;%llu",
619 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
620 (unsigned long long) ts->runtime[ddir]);
622 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
623 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
625 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
627 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
628 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
630 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
632 if (ts->clat_percentiles) {
633 len = calc_clat_percentiles(ts->io_u_plat[ddir],
634 ts->clat_stat[ddir].samples,
635 ts->percentile_list, &ovals, &maxv,
640 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
645 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
648 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
649 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
651 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
656 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
657 double p_of_agg = 100.0;
660 p_of_agg = mean * 100 / (double) rs->agg[ddir];
661 if (p_of_agg > 100.0)
665 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
667 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
670 static void add_ddir_status_json(struct thread_stat *ts,
671 struct group_run_stats *rs, int ddir, struct json_object *parent)
673 unsigned long min, max;
674 unsigned long long bw, iops;
675 unsigned int *ovals = NULL;
677 unsigned int len, minv, maxv;
679 const char *ddirname[] = {"read", "write", "trim"};
680 struct json_object *dir_object, *tmp_object, *percentile_object;
682 double p_of_agg = 100.0;
684 assert(ddir_rw(ddir));
686 if (ts->unified_rw_rep && ddir != DDIR_READ)
689 dir_object = json_create_object();
690 json_object_add_value_object(parent,
691 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
694 if (ts->runtime[ddir]) {
695 uint64_t runt = ts->runtime[ddir];
697 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
698 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
701 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
702 json_object_add_value_int(dir_object, "bw", bw);
703 json_object_add_value_int(dir_object, "iops", iops);
704 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
706 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
710 tmp_object = json_create_object();
711 json_object_add_value_object(dir_object, "slat", tmp_object);
712 json_object_add_value_int(tmp_object, "min", min);
713 json_object_add_value_int(tmp_object, "max", max);
714 json_object_add_value_float(tmp_object, "mean", mean);
715 json_object_add_value_float(tmp_object, "stddev", dev);
717 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
721 tmp_object = json_create_object();
722 json_object_add_value_object(dir_object, "clat", tmp_object);
723 json_object_add_value_int(tmp_object, "min", min);
724 json_object_add_value_int(tmp_object, "max", max);
725 json_object_add_value_float(tmp_object, "mean", mean);
726 json_object_add_value_float(tmp_object, "stddev", dev);
728 if (ts->clat_percentiles) {
729 len = calc_clat_percentiles(ts->io_u_plat[ddir],
730 ts->clat_stat[ddir].samples,
731 ts->percentile_list, &ovals, &maxv,
736 percentile_object = json_create_object();
737 json_object_add_value_object(tmp_object, "percentile", percentile_object);
738 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
740 json_object_add_value_int(percentile_object, "0.00", 0);
743 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
744 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
747 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
751 tmp_object = json_create_object();
752 json_object_add_value_object(dir_object, "lat", tmp_object);
753 json_object_add_value_int(tmp_object, "min", min);
754 json_object_add_value_int(tmp_object, "max", max);
755 json_object_add_value_float(tmp_object, "mean", mean);
756 json_object_add_value_float(tmp_object, "stddev", dev);
760 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
762 p_of_agg = mean * 100 / (double) rs->agg[ddir];
763 if (p_of_agg > 100.0)
768 p_of_agg = mean = dev = 0.0;
770 json_object_add_value_int(dir_object, "bw_min", min);
771 json_object_add_value_int(dir_object, "bw_max", max);
772 json_object_add_value_float(dir_object, "bw_agg", mean);
773 json_object_add_value_float(dir_object, "bw_mean", mean);
774 json_object_add_value_float(dir_object, "bw_dev", dev);
777 static void show_thread_status_terse_v2(struct thread_stat *ts,
778 struct group_run_stats *rs)
780 double io_u_dist[FIO_IO_U_MAP_NR];
781 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
782 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
783 double usr_cpu, sys_cpu;
787 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
788 /* Log Read Status */
789 show_ddir_status_terse(ts, rs, DDIR_READ);
790 /* Log Write Status */
791 show_ddir_status_terse(ts, rs, DDIR_WRITE);
792 /* Log Trim Status */
793 show_ddir_status_terse(ts, rs, DDIR_TRIM);
796 if (ts->total_run_time) {
797 double runt = (double) ts->total_run_time;
799 usr_cpu = (double) ts->usr_time * 100 / runt;
800 sys_cpu = (double) ts->sys_time * 100 / runt;
806 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
807 (unsigned long long) ts->ctx,
808 (unsigned long long) ts->majf,
809 (unsigned long long) ts->minf);
811 /* Calc % distribution of IO depths, usecond, msecond latency */
812 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
813 stat_calc_lat_u(ts, io_u_lat_u);
814 stat_calc_lat_m(ts, io_u_lat_m);
816 /* Only show fixed 7 I/O depth levels*/
817 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
818 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
819 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
821 /* Microsecond latency */
822 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
823 log_info(";%3.2f%%", io_u_lat_u[i]);
824 /* Millisecond latency */
825 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
826 log_info(";%3.2f%%", io_u_lat_m[i]);
827 /* Additional output if continue_on_error set - default off*/
828 if (ts->continue_on_error)
829 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
832 /* Additional output if description is set */
834 log_info(";%s", ts->description);
839 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
840 struct group_run_stats *rs, int ver)
842 double io_u_dist[FIO_IO_U_MAP_NR];
843 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
844 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
845 double usr_cpu, sys_cpu;
849 log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
850 ts->name, ts->groupid, ts->error);
851 /* Log Read Status */
852 show_ddir_status_terse(ts, rs, DDIR_READ);
853 /* Log Write Status */
854 show_ddir_status_terse(ts, rs, DDIR_WRITE);
855 /* Log Trim Status */
857 show_ddir_status_terse(ts, rs, DDIR_TRIM);
860 if (ts->total_run_time) {
861 double runt = (double) ts->total_run_time;
863 usr_cpu = (double) ts->usr_time * 100 / runt;
864 sys_cpu = (double) ts->sys_time * 100 / runt;
870 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
871 (unsigned long long) ts->ctx,
872 (unsigned long long) ts->majf,
873 (unsigned long long) ts->minf);
875 /* Calc % distribution of IO depths, usecond, msecond latency */
876 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
877 stat_calc_lat_u(ts, io_u_lat_u);
878 stat_calc_lat_m(ts, io_u_lat_m);
880 /* Only show fixed 7 I/O depth levels*/
881 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
882 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
883 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
885 /* Microsecond latency */
886 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
887 log_info(";%3.2f%%", io_u_lat_u[i]);
888 /* Millisecond latency */
889 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
890 log_info(";%3.2f%%", io_u_lat_m[i]);
892 /* disk util stats, if any */
894 show_disk_util(1, NULL);
896 /* Additional output if continue_on_error set - default off*/
897 if (ts->continue_on_error)
898 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
900 /* Additional output if description is set */
901 if (strlen(ts->description))
902 log_info(";%s", ts->description);
907 static struct json_object *show_thread_status_json(struct thread_stat *ts,
908 struct group_run_stats *rs)
910 struct json_object *root, *tmp;
911 double io_u_dist[FIO_IO_U_MAP_NR];
912 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
913 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
914 double usr_cpu, sys_cpu;
917 root = json_create_object();
918 json_object_add_value_string(root, "jobname", ts->name);
919 json_object_add_value_int(root, "groupid", ts->groupid);
920 json_object_add_value_int(root, "error", ts->error);
922 add_ddir_status_json(ts, rs, DDIR_READ, root);
923 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
924 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
927 if (ts->total_run_time) {
928 double runt = (double) ts->total_run_time;
930 usr_cpu = (double) ts->usr_time * 100 / runt;
931 sys_cpu = (double) ts->sys_time * 100 / runt;
936 json_object_add_value_float(root, "usr_cpu", usr_cpu);
937 json_object_add_value_float(root, "sys_cpu", sys_cpu);
938 json_object_add_value_int(root, "ctx", ts->ctx);
939 json_object_add_value_int(root, "majf", ts->majf);
940 json_object_add_value_int(root, "minf", ts->minf);
943 /* Calc % distribution of IO depths, usecond, msecond latency */
944 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
945 stat_calc_lat_u(ts, io_u_lat_u);
946 stat_calc_lat_m(ts, io_u_lat_m);
948 tmp = json_create_object();
949 json_object_add_value_object(root, "iodepth_level", tmp);
950 /* Only show fixed 7 I/O depth levels*/
951 for (i = 0; i < 7; i++) {
954 snprintf(name, 20, "%d", 1 << i);
956 snprintf(name, 20, ">=%d", 1 << i);
957 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
960 tmp = json_create_object();
961 json_object_add_value_object(root, "latency_us", tmp);
962 /* Microsecond latency */
963 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
964 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
965 "250", "500", "750", "1000", };
966 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
968 /* Millisecond latency */
969 tmp = json_create_object();
970 json_object_add_value_object(root, "latency_ms", tmp);
971 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
972 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
973 "250", "500", "750", "1000", "2000",
975 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
978 /* Additional output if continue_on_error set - default off*/
979 if (ts->continue_on_error) {
980 json_object_add_value_int(root, "total_err", ts->total_err_count);
981 json_object_add_value_int(root, "first_error", ts->first_error);
984 if (ts->latency_depth) {
985 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
986 json_object_add_value_int(root, "latency_target", ts->latency_target);
987 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
988 json_object_add_value_int(root, "latency_window", ts->latency_window);
991 /* Additional output if description is set */
992 if (strlen(ts->description))
993 json_object_add_value_string(root, "desc", ts->description);
998 static void show_thread_status_terse(struct thread_stat *ts,
999 struct group_run_stats *rs)
1001 if (terse_version == 2)
1002 show_thread_status_terse_v2(ts, rs);
1003 else if (terse_version == 3 || terse_version == 4)
1004 show_thread_status_terse_v3_v4(ts, rs, terse_version);
1006 log_err("fio: bad terse version!? %d\n", terse_version);
1009 struct json_object *show_thread_status(struct thread_stat *ts,
1010 struct group_run_stats *rs)
1012 if (output_format == FIO_OUTPUT_TERSE)
1013 show_thread_status_terse(ts, rs);
1014 else if (output_format == FIO_OUTPUT_JSON)
1015 return(show_thread_status_json(ts, rs));
1017 show_thread_status_normal(ts, rs);
1021 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1025 if (src->samples == 0)
1028 dst->min_val = min(dst->min_val, src->min_val);
1029 dst->max_val = max(dst->max_val, src->max_val);
1032 * Compute new mean and S after the merge
1033 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1034 * #Parallel_algorithm>
1037 mean = src->mean.u.f;
1040 double delta = src->mean.u.f - dst->mean.u.f;
1042 mean = ((src->mean.u.f * src->samples) +
1043 (dst->mean.u.f * dst->samples)) /
1044 (dst->samples + src->samples);
1046 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1047 (dst->samples * src->samples) /
1048 (dst->samples + src->samples);
1051 dst->samples += src->samples;
1052 dst->mean.u.f = mean;
1056 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1060 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1061 if (dst->max_run[i] < src->max_run[i])
1062 dst->max_run[i] = src->max_run[i];
1063 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1064 dst->min_run[i] = src->min_run[i];
1065 if (dst->max_bw[i] < src->max_bw[i])
1066 dst->max_bw[i] = src->max_bw[i];
1067 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1068 dst->min_bw[i] = src->min_bw[i];
1070 dst->io_kb[i] += src->io_kb[i];
1071 dst->agg[i] += src->agg[i];
1076 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1080 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1081 if (!dst->unified_rw_rep) {
1082 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1083 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1084 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1085 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1087 dst->io_bytes[l] += src->io_bytes[l];
1089 if (dst->runtime[l] < src->runtime[l])
1090 dst->runtime[l] = src->runtime[l];
1092 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1093 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1094 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1095 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1097 dst->io_bytes[0] += src->io_bytes[l];
1099 if (dst->runtime[0] < src->runtime[l])
1100 dst->runtime[0] = src->runtime[l];
1104 dst->usr_time += src->usr_time;
1105 dst->sys_time += src->sys_time;
1106 dst->ctx += src->ctx;
1107 dst->majf += src->majf;
1108 dst->minf += src->minf;
1110 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1111 dst->io_u_map[k] += src->io_u_map[k];
1112 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1113 dst->io_u_submit[k] += src->io_u_submit[k];
1114 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1115 dst->io_u_complete[k] += src->io_u_complete[k];
1116 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1117 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1118 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1119 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1121 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1122 if (!dst->unified_rw_rep) {
1123 dst->total_io_u[k] += src->total_io_u[k];
1124 dst->short_io_u[k] += src->short_io_u[k];
1126 dst->total_io_u[0] += src->total_io_u[k];
1127 dst->short_io_u[0] += src->short_io_u[k];
1131 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1134 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1135 if (!dst->unified_rw_rep)
1136 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1138 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1142 dst->total_run_time += src->total_run_time;
1143 dst->total_submit += src->total_submit;
1144 dst->total_complete += src->total_complete;
1147 void init_group_run_stat(struct group_run_stats *gs)
1150 memset(gs, 0, sizeof(*gs));
1152 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1153 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1156 void init_thread_stat(struct thread_stat *ts)
1160 memset(ts, 0, sizeof(*ts));
1162 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1163 ts->lat_stat[j].min_val = -1UL;
1164 ts->clat_stat[j].min_val = -1UL;
1165 ts->slat_stat[j].min_val = -1UL;
1166 ts->bw_stat[j].min_val = -1UL;
1171 static void __show_run_stats(void)
1173 struct group_run_stats *runstats, *rs;
1174 struct thread_data *td;
1175 struct thread_stat *threadstats, *ts;
1176 int i, j, nr_ts, last_ts, idx;
1177 int kb_base_warned = 0;
1178 int unit_base_warned = 0;
1179 struct json_object *root = NULL;
1180 struct json_array *array = NULL;
1182 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1184 for (i = 0; i < groupid + 1; i++)
1185 init_group_run_stat(&runstats[i]);
1188 * find out how many threads stats we need. if group reporting isn't
1189 * enabled, it's one-per-td.
1193 for_each_td(td, i) {
1194 if (!td->o.group_reporting) {
1198 if (last_ts == td->groupid)
1201 last_ts = td->groupid;
1205 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1207 for (i = 0; i < nr_ts; i++)
1208 init_thread_stat(&threadstats[i]);
1213 for_each_td(td, i) {
1214 if (idx && (!td->o.group_reporting ||
1215 (td->o.group_reporting && last_ts != td->groupid))) {
1220 last_ts = td->groupid;
1222 ts = &threadstats[j];
1224 ts->clat_percentiles = td->o.clat_percentiles;
1225 ts->percentile_precision = td->o.percentile_precision;
1226 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1231 if (ts->groupid == -1) {
1233 * These are per-group shared already
1235 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
1236 if (td->o.description)
1237 strncpy(ts->description, td->o.description,
1240 memset(ts->description, 0, FIO_JOBNAME_SIZE);
1243 * If multiple entries in this group, this is
1246 ts->thread_number = td->thread_number;
1247 ts->groupid = td->groupid;
1250 * first pid in group, not very useful...
1254 ts->kb_base = td->o.kb_base;
1255 ts->unit_base = td->o.unit_base;
1256 ts->unified_rw_rep = td->o.unified_rw_rep;
1257 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1258 log_info("fio: kb_base differs for jobs in group, using"
1259 " %u as the base\n", ts->kb_base);
1261 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1262 log_info("fio: unit_base differs for jobs in group, using"
1263 " %u as the base\n", ts->unit_base);
1264 unit_base_warned = 1;
1267 ts->continue_on_error = td->o.continue_on_error;
1268 ts->total_err_count += td->total_err_count;
1269 ts->first_error = td->first_error;
1271 if (!td->error && td->o.continue_on_error &&
1273 ts->error = td->first_error;
1274 strcpy(ts->verror, td->verror);
1275 } else if (td->error) {
1276 ts->error = td->error;
1277 strcpy(ts->verror, td->verror);
1281 ts->latency_depth = td->latency_qd;
1282 ts->latency_target = td->o.latency_target;
1283 ts->latency_percentile = td->o.latency_percentile;
1284 ts->latency_window = td->o.latency_window;
1286 sum_thread_stats(ts, &td->ts, idx);
1289 for (i = 0; i < nr_ts; i++) {
1290 unsigned long long bw;
1292 ts = &threadstats[i];
1293 rs = &runstats[ts->groupid];
1294 rs->kb_base = ts->kb_base;
1295 rs->unit_base = ts->unit_base;
1296 rs->unified_rw_rep += ts->unified_rw_rep;
1298 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1299 if (!ts->runtime[j])
1301 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1302 rs->min_run[j] = ts->runtime[j];
1303 if (ts->runtime[j] > rs->max_run[j])
1304 rs->max_run[j] = ts->runtime[j];
1307 if (ts->runtime[j]) {
1308 unsigned long runt = ts->runtime[j];
1309 unsigned long long kb;
1311 kb = ts->io_bytes[j] / rs->kb_base;
1312 bw = kb * 1000 / runt;
1314 if (bw < rs->min_bw[j])
1316 if (bw > rs->max_bw[j])
1319 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1323 for (i = 0; i < groupid + 1; i++) {
1328 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1329 if (rs->max_run[ddir])
1330 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1336 * don't overwrite last signal output
1338 if (output_format == FIO_OUTPUT_NORMAL)
1340 else if (output_format == FIO_OUTPUT_JSON) {
1341 root = json_create_object();
1342 json_object_add_value_string(root, "fio version", fio_version_string);
1343 array = json_create_array();
1344 json_object_add_value_array(root, "jobs", array);
1347 for (i = 0; i < nr_ts; i++) {
1348 ts = &threadstats[i];
1349 rs = &runstats[ts->groupid];
1352 fio_server_send_ts(ts, rs);
1353 else if (output_format == FIO_OUTPUT_TERSE)
1354 show_thread_status_terse(ts, rs);
1355 else if (output_format == FIO_OUTPUT_JSON) {
1356 struct json_object *tmp = show_thread_status_json(ts, rs);
1357 json_array_add_value_object(array, tmp);
1359 show_thread_status_normal(ts, rs);
1361 if (output_format == FIO_OUTPUT_JSON) {
1362 /* disk util stats, if any */
1363 show_disk_util(1, root);
1365 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1367 json_print_object(root);
1369 json_free_object(root);
1372 for (i = 0; i < groupid + 1; i++) {
1377 fio_server_send_gs(rs);
1378 else if (output_format == FIO_OUTPUT_NORMAL)
1379 show_group_stats(rs);
1383 fio_server_send_du();
1384 else if (output_format == FIO_OUTPUT_NORMAL) {
1385 show_disk_util(0, NULL);
1386 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1394 void show_run_stats(void)
1396 fio_mutex_down(stat_mutex);
1398 fio_mutex_up(stat_mutex);
1401 static void *__show_running_run_stats(void fio_unused *arg)
1403 struct thread_data *td;
1404 unsigned long long *rt;
1408 rt = malloc(thread_number * sizeof(unsigned long long));
1409 fio_gettime(&tv, NULL);
1411 for_each_td(td, i) {
1412 rt[i] = mtime_since(&td->start, &tv);
1413 if (td_read(td) && td->io_bytes[DDIR_READ])
1414 td->ts.runtime[DDIR_READ] += rt[i];
1415 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1416 td->ts.runtime[DDIR_WRITE] += rt[i];
1417 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1418 td->ts.runtime[DDIR_TRIM] += rt[i];
1420 td->update_rusage = 1;
1421 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1422 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1423 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1424 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1427 for_each_td(td, i) {
1428 if (td->rusage_sem) {
1429 td->update_rusage = 1;
1430 fio_mutex_down(td->rusage_sem);
1432 td->update_rusage = 0;
1437 for_each_td(td, i) {
1438 if (td_read(td) && td->io_bytes[DDIR_READ])
1439 td->ts.runtime[DDIR_READ] -= rt[i];
1440 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1441 td->ts.runtime[DDIR_WRITE] -= rt[i];
1442 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1443 td->ts.runtime[DDIR_TRIM] -= rt[i];
1447 fio_mutex_up(stat_mutex);
1452 * Called from signal handler. It _should_ be safe to just run this inline
1453 * in the sig handler, but we should be disturbing the system less by just
1454 * creating a thread to do it.
1456 void show_running_run_stats(void)
1460 fio_mutex_down(stat_mutex);
1462 if (!pthread_create(&thread, NULL, __show_running_run_stats, NULL)) {
1463 pthread_detach(thread);
1467 fio_mutex_up(stat_mutex);
1470 static int status_interval_init;
1471 static struct timeval status_time;
1473 #define FIO_STATUS_FILE "/tmp/fio-dump-status"
1475 static int check_status_file(void)
1478 const char *temp_dir;
1479 char fio_status_file_path[PATH_MAX];
1481 temp_dir = getenv("TMPDIR");
1482 if (temp_dir == NULL)
1483 temp_dir = getenv("TEMP");
1484 if (temp_dir == NULL)
1487 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1489 if (stat(fio_status_file_path, &sb))
1492 unlink(fio_status_file_path);
1496 void check_for_running_stats(void)
1498 if (status_interval) {
1499 if (!status_interval_init) {
1500 fio_gettime(&status_time, NULL);
1501 status_interval_init = 1;
1502 } else if (mtime_since_now(&status_time) >= status_interval) {
1503 show_running_run_stats();
1504 fio_gettime(&status_time, NULL);
1508 if (check_status_file()) {
1509 show_running_run_stats();
1514 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1519 if (data > is->max_val)
1521 if (data < is->min_val)
1524 delta = val - is->mean.u.f;
1526 is->mean.u.f += delta / (is->samples + 1.0);
1527 is->S.u.f += delta * (val - is->mean.u.f);
1533 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1534 enum fio_ddir ddir, unsigned int bs,
1537 const int nr_samples = iolog->nr_samples;
1539 if (iolog->disabled)
1542 if (!iolog->nr_samples)
1543 iolog->avg_last = t;
1545 if (iolog->nr_samples == iolog->max_samples) {
1546 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
1549 new_log = realloc(iolog->log, new_size);
1551 log_err("fio: failed extending iolog! Will stop logging.\n");
1552 iolog->disabled = 1;
1555 iolog->log = new_log;
1556 iolog->max_samples <<= 1;
1559 iolog->log[nr_samples].val = val;
1560 iolog->log[nr_samples].time = t;
1561 iolog->log[nr_samples].ddir = ddir;
1562 iolog->log[nr_samples].bs = bs;
1563 iolog->nr_samples++;
1566 static inline void reset_io_stat(struct io_stat *ios)
1568 ios->max_val = ios->min_val = ios->samples = 0;
1569 ios->mean.u.f = ios->S.u.f = 0;
1572 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1573 unsigned long val, enum fio_ddir ddir,
1576 unsigned long elapsed, this_window;
1581 elapsed = mtime_since_now(&td->epoch);
1584 * If no time averaging, just add the log sample.
1586 if (!iolog->avg_msec) {
1587 __add_log_sample(iolog, val, ddir, bs, elapsed);
1592 * Add the sample. If the time period has passed, then
1593 * add that entry to the log and clear.
1595 add_stat_sample(&iolog->avg_window[ddir], val);
1598 * If period hasn't passed, adding the above sample is all we
1601 this_window = elapsed - iolog->avg_last;
1602 if (this_window < iolog->avg_msec)
1606 * Note an entry in the log. Use the mean from the logged samples,
1607 * making sure to properly round up. Only write a log entry if we
1608 * had actual samples done.
1610 if (iolog->avg_window[DDIR_READ].samples) {
1613 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1614 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed);
1616 if (iolog->avg_window[DDIR_WRITE].samples) {
1619 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1620 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed);
1622 if (iolog->avg_window[DDIR_TRIM].samples) {
1625 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1626 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed);
1630 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1631 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1632 reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1633 iolog->avg_last = elapsed;
1636 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1638 struct io_log *iolog;
1643 iolog = agg_io_log[ddir];
1644 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
1647 static void add_clat_percentile_sample(struct thread_stat *ts,
1648 unsigned long usec, enum fio_ddir ddir)
1650 unsigned int idx = plat_val_to_idx(usec);
1651 assert(idx < FIO_IO_U_PLAT_NR);
1653 ts->io_u_plat[ddir][idx]++;
1656 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1657 unsigned long usec, unsigned int bs)
1659 struct thread_stat *ts = &td->ts;
1664 add_stat_sample(&ts->clat_stat[ddir], usec);
1667 add_log_sample(td, td->clat_log, usec, ddir, bs);
1669 if (ts->clat_percentiles)
1670 add_clat_percentile_sample(ts, usec, ddir);
1673 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1674 unsigned long usec, unsigned int bs)
1676 struct thread_stat *ts = &td->ts;
1681 add_stat_sample(&ts->slat_stat[ddir], usec);
1684 add_log_sample(td, td->slat_log, usec, ddir, bs);
1687 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1688 unsigned long usec, unsigned int bs)
1690 struct thread_stat *ts = &td->ts;
1695 add_stat_sample(&ts->lat_stat[ddir], usec);
1698 add_log_sample(td, td->lat_log, usec, ddir, bs);
1701 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1704 struct thread_stat *ts = &td->ts;
1705 unsigned long spent, rate;
1710 spent = mtime_since(&td->bw_sample_time, t);
1711 if (spent < td->o.bw_avg_time)
1715 * Compute both read and write rates for the interval.
1717 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1720 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1722 continue; /* No entries for interval */
1724 rate = delta * 1000 / spent / 1024;
1725 add_stat_sample(&ts->bw_stat[ddir], rate);
1728 add_log_sample(td, td->bw_log, rate, ddir, bs);
1730 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1733 fio_gettime(&td->bw_sample_time, NULL);
1736 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1739 struct thread_stat *ts = &td->ts;
1740 unsigned long spent, iops;
1745 spent = mtime_since(&td->iops_sample_time, t);
1746 if (spent < td->o.iops_avg_time)
1750 * Compute both read and write rates for the interval.
1752 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1755 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1757 continue; /* No entries for interval */
1759 iops = (delta * 1000) / spent;
1760 add_stat_sample(&ts->iops_stat[ddir], iops);
1763 add_log_sample(td, td->iops_log, iops, ddir, bs);
1765 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1768 fio_gettime(&td->iops_sample_time, NULL);
1771 void stat_init(void)
1773 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1776 void stat_exit(void)
1779 * When we have the mutex, we know out-of-band access to it
1782 fio_mutex_down(stat_mutex);
1783 fio_mutex_remove(stat_mutex);