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 static void show_thread_status_normal(struct thread_stat *ts,
501 struct group_run_stats *rs)
503 double usr_cpu, sys_cpu;
504 unsigned long runtime;
505 double io_u_dist[FIO_IO_U_MAP_NR];
509 if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
513 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
516 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
517 ts->name, ts->groupid, ts->members,
518 ts->error, (int) ts->pid, time_buf);
520 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
521 ts->name, ts->groupid, ts->members,
522 ts->error, ts->verror, (int) ts->pid,
526 if (strlen(ts->description))
527 log_info(" Description : [%s]\n", ts->description);
529 if (ts->io_bytes[DDIR_READ])
530 show_ddir_status(rs, ts, DDIR_READ);
531 if (ts->io_bytes[DDIR_WRITE])
532 show_ddir_status(rs, ts, DDIR_WRITE);
533 if (ts->io_bytes[DDIR_TRIM])
534 show_ddir_status(rs, ts, DDIR_TRIM);
538 runtime = ts->total_run_time;
540 double runt = (double) runtime;
542 usr_cpu = (double) ts->usr_time * 100 / runt;
543 sys_cpu = (double) ts->sys_time * 100 / runt;
549 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
550 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
551 (unsigned long long) ts->ctx,
552 (unsigned long long) ts->majf,
553 (unsigned long long) ts->minf);
555 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
556 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
557 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
558 io_u_dist[1], io_u_dist[2],
559 io_u_dist[3], io_u_dist[4],
560 io_u_dist[5], io_u_dist[6]);
562 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
563 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
564 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
565 io_u_dist[1], io_u_dist[2],
566 io_u_dist[3], io_u_dist[4],
567 io_u_dist[5], io_u_dist[6]);
568 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
569 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
570 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
571 io_u_dist[1], io_u_dist[2],
572 io_u_dist[3], io_u_dist[4],
573 io_u_dist[5], io_u_dist[6]);
574 log_info(" issued : total=r=%llu/w=%llu/d=%llu,"
575 " short=r=%llu/w=%llu/d=%llu,"
576 " drop=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 (unsigned long long) ts->drop_io_u[0],
584 (unsigned long long) ts->drop_io_u[1],
585 (unsigned long long) ts->drop_io_u[2]);
586 if (ts->continue_on_error) {
587 log_info(" errors : total=%llu, first_error=%d/<%s>\n",
588 (unsigned long long)ts->total_err_count,
590 strerror(ts->first_error));
592 if (ts->latency_depth) {
593 log_info(" latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
594 (unsigned long long)ts->latency_target,
595 (unsigned long long)ts->latency_window,
596 ts->latency_percentile.u.f,
601 static void show_ddir_status_terse(struct thread_stat *ts,
602 struct group_run_stats *rs, int ddir)
604 unsigned long min, max;
605 unsigned long long bw, iops;
606 unsigned int *ovals = NULL;
608 unsigned int len, minv, maxv;
611 assert(ddir_rw(ddir));
614 if (ts->runtime[ddir]) {
615 uint64_t runt = ts->runtime[ddir];
617 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
618 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
621 log_info(";%llu;%llu;%llu;%llu",
622 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
623 (unsigned long long) ts->runtime[ddir]);
625 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
626 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
628 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
630 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
631 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
633 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
635 if (ts->clat_percentiles) {
636 len = calc_clat_percentiles(ts->io_u_plat[ddir],
637 ts->clat_stat[ddir].samples,
638 ts->percentile_list, &ovals, &maxv,
643 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
648 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
651 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
652 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
654 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
659 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
660 double p_of_agg = 100.0;
663 p_of_agg = mean * 100 / (double) rs->agg[ddir];
664 if (p_of_agg > 100.0)
668 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
670 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
673 static void add_ddir_status_json(struct thread_stat *ts,
674 struct group_run_stats *rs, int ddir, struct json_object *parent)
676 unsigned long min, max;
677 unsigned long long bw, iops;
678 unsigned int *ovals = NULL;
680 unsigned int len, minv, maxv;
682 const char *ddirname[] = {"read", "write", "trim"};
683 struct json_object *dir_object, *tmp_object, *percentile_object;
685 double p_of_agg = 100.0;
687 assert(ddir_rw(ddir));
689 if (ts->unified_rw_rep && ddir != DDIR_READ)
692 dir_object = json_create_object();
693 json_object_add_value_object(parent,
694 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
697 if (ts->runtime[ddir]) {
698 uint64_t runt = ts->runtime[ddir];
700 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
701 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
704 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
705 json_object_add_value_int(dir_object, "bw", bw);
706 json_object_add_value_int(dir_object, "iops", iops);
707 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
708 json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
709 json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
710 json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
712 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
716 tmp_object = json_create_object();
717 json_object_add_value_object(dir_object, "slat", tmp_object);
718 json_object_add_value_int(tmp_object, "min", min);
719 json_object_add_value_int(tmp_object, "max", max);
720 json_object_add_value_float(tmp_object, "mean", mean);
721 json_object_add_value_float(tmp_object, "stddev", dev);
723 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
727 tmp_object = json_create_object();
728 json_object_add_value_object(dir_object, "clat", tmp_object);
729 json_object_add_value_int(tmp_object, "min", min);
730 json_object_add_value_int(tmp_object, "max", max);
731 json_object_add_value_float(tmp_object, "mean", mean);
732 json_object_add_value_float(tmp_object, "stddev", dev);
734 if (ts->clat_percentiles) {
735 len = calc_clat_percentiles(ts->io_u_plat[ddir],
736 ts->clat_stat[ddir].samples,
737 ts->percentile_list, &ovals, &maxv,
742 percentile_object = json_create_object();
743 json_object_add_value_object(tmp_object, "percentile", percentile_object);
744 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
746 json_object_add_value_int(percentile_object, "0.00", 0);
749 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
750 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
753 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
757 tmp_object = json_create_object();
758 json_object_add_value_object(dir_object, "lat", tmp_object);
759 json_object_add_value_int(tmp_object, "min", min);
760 json_object_add_value_int(tmp_object, "max", max);
761 json_object_add_value_float(tmp_object, "mean", mean);
762 json_object_add_value_float(tmp_object, "stddev", dev);
766 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
768 p_of_agg = mean * 100 / (double) rs->agg[ddir];
769 if (p_of_agg > 100.0)
774 p_of_agg = mean = dev = 0.0;
776 json_object_add_value_int(dir_object, "bw_min", min);
777 json_object_add_value_int(dir_object, "bw_max", max);
778 json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
779 json_object_add_value_float(dir_object, "bw_mean", mean);
780 json_object_add_value_float(dir_object, "bw_dev", dev);
783 static void show_thread_status_terse_v2(struct thread_stat *ts,
784 struct group_run_stats *rs)
786 double io_u_dist[FIO_IO_U_MAP_NR];
787 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
788 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
789 double usr_cpu, sys_cpu;
793 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
794 /* Log Read Status */
795 show_ddir_status_terse(ts, rs, DDIR_READ);
796 /* Log Write Status */
797 show_ddir_status_terse(ts, rs, DDIR_WRITE);
798 /* Log Trim Status */
799 show_ddir_status_terse(ts, rs, DDIR_TRIM);
802 if (ts->total_run_time) {
803 double runt = (double) ts->total_run_time;
805 usr_cpu = (double) ts->usr_time * 100 / runt;
806 sys_cpu = (double) ts->sys_time * 100 / runt;
812 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
813 (unsigned long long) ts->ctx,
814 (unsigned long long) ts->majf,
815 (unsigned long long) ts->minf);
817 /* Calc % distribution of IO depths, usecond, msecond latency */
818 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
819 stat_calc_lat_u(ts, io_u_lat_u);
820 stat_calc_lat_m(ts, io_u_lat_m);
822 /* Only show fixed 7 I/O depth levels*/
823 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
824 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
825 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
827 /* Microsecond latency */
828 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
829 log_info(";%3.2f%%", io_u_lat_u[i]);
830 /* Millisecond latency */
831 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
832 log_info(";%3.2f%%", io_u_lat_m[i]);
833 /* Additional output if continue_on_error set - default off*/
834 if (ts->continue_on_error)
835 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
838 /* Additional output if description is set */
839 if (strlen(ts->description))
840 log_info(";%s", ts->description);
845 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
846 struct group_run_stats *rs, int ver)
848 double io_u_dist[FIO_IO_U_MAP_NR];
849 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
850 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
851 double usr_cpu, sys_cpu;
855 log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
856 ts->name, ts->groupid, ts->error);
857 /* Log Read Status */
858 show_ddir_status_terse(ts, rs, DDIR_READ);
859 /* Log Write Status */
860 show_ddir_status_terse(ts, rs, DDIR_WRITE);
861 /* Log Trim Status */
863 show_ddir_status_terse(ts, rs, DDIR_TRIM);
866 if (ts->total_run_time) {
867 double runt = (double) ts->total_run_time;
869 usr_cpu = (double) ts->usr_time * 100 / runt;
870 sys_cpu = (double) ts->sys_time * 100 / runt;
876 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
877 (unsigned long long) ts->ctx,
878 (unsigned long long) ts->majf,
879 (unsigned long long) ts->minf);
881 /* Calc % distribution of IO depths, usecond, msecond latency */
882 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
883 stat_calc_lat_u(ts, io_u_lat_u);
884 stat_calc_lat_m(ts, io_u_lat_m);
886 /* Only show fixed 7 I/O depth levels*/
887 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
888 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
889 io_u_dist[4], io_u_dist[5], io_u_dist[6]);
891 /* Microsecond latency */
892 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
893 log_info(";%3.2f%%", io_u_lat_u[i]);
894 /* Millisecond latency */
895 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
896 log_info(";%3.2f%%", io_u_lat_m[i]);
898 /* disk util stats, if any */
900 show_disk_util(1, NULL);
902 /* Additional output if continue_on_error set - default off*/
903 if (ts->continue_on_error)
904 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
906 /* Additional output if description is set */
907 if (strlen(ts->description))
908 log_info(";%s", ts->description);
913 static struct json_object *show_thread_status_json(struct thread_stat *ts,
914 struct group_run_stats *rs)
916 struct json_object *root, *tmp;
917 double io_u_dist[FIO_IO_U_MAP_NR];
918 double io_u_lat_u[FIO_IO_U_LAT_U_NR];
919 double io_u_lat_m[FIO_IO_U_LAT_M_NR];
920 double usr_cpu, sys_cpu;
923 root = json_create_object();
924 json_object_add_value_string(root, "jobname", ts->name);
925 json_object_add_value_int(root, "groupid", ts->groupid);
926 json_object_add_value_int(root, "error", ts->error);
928 add_ddir_status_json(ts, rs, DDIR_READ, root);
929 add_ddir_status_json(ts, rs, DDIR_WRITE, root);
930 add_ddir_status_json(ts, rs, DDIR_TRIM, root);
933 if (ts->total_run_time) {
934 double runt = (double) ts->total_run_time;
936 usr_cpu = (double) ts->usr_time * 100 / runt;
937 sys_cpu = (double) ts->sys_time * 100 / runt;
942 json_object_add_value_float(root, "usr_cpu", usr_cpu);
943 json_object_add_value_float(root, "sys_cpu", sys_cpu);
944 json_object_add_value_int(root, "ctx", ts->ctx);
945 json_object_add_value_int(root, "majf", ts->majf);
946 json_object_add_value_int(root, "minf", ts->minf);
949 /* Calc % distribution of IO depths, usecond, msecond latency */
950 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
951 stat_calc_lat_u(ts, io_u_lat_u);
952 stat_calc_lat_m(ts, io_u_lat_m);
954 tmp = json_create_object();
955 json_object_add_value_object(root, "iodepth_level", tmp);
956 /* Only show fixed 7 I/O depth levels*/
957 for (i = 0; i < 7; i++) {
960 snprintf(name, 20, "%d", 1 << i);
962 snprintf(name, 20, ">=%d", 1 << i);
963 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
966 tmp = json_create_object();
967 json_object_add_value_object(root, "latency_us", tmp);
968 /* Microsecond latency */
969 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
970 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
971 "250", "500", "750", "1000", };
972 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
974 /* Millisecond latency */
975 tmp = json_create_object();
976 json_object_add_value_object(root, "latency_ms", tmp);
977 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
978 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
979 "250", "500", "750", "1000", "2000",
981 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
984 /* Additional output if continue_on_error set - default off*/
985 if (ts->continue_on_error) {
986 json_object_add_value_int(root, "total_err", ts->total_err_count);
987 json_object_add_value_int(root, "first_error", ts->first_error);
990 if (ts->latency_depth) {
991 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
992 json_object_add_value_int(root, "latency_target", ts->latency_target);
993 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
994 json_object_add_value_int(root, "latency_window", ts->latency_window);
997 /* Additional output if description is set */
998 if (strlen(ts->description))
999 json_object_add_value_string(root, "desc", ts->description);
1004 static void show_thread_status_terse(struct thread_stat *ts,
1005 struct group_run_stats *rs)
1007 if (terse_version == 2)
1008 show_thread_status_terse_v2(ts, rs);
1009 else if (terse_version == 3 || terse_version == 4)
1010 show_thread_status_terse_v3_v4(ts, rs, terse_version);
1012 log_err("fio: bad terse version!? %d\n", terse_version);
1015 struct json_object *show_thread_status(struct thread_stat *ts,
1016 struct group_run_stats *rs)
1018 if (output_format == FIO_OUTPUT_TERSE)
1019 show_thread_status_terse(ts, rs);
1020 else if (output_format == FIO_OUTPUT_JSON)
1021 return show_thread_status_json(ts, rs);
1023 show_thread_status_normal(ts, rs);
1027 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1031 if (src->samples == 0)
1034 dst->min_val = min(dst->min_val, src->min_val);
1035 dst->max_val = max(dst->max_val, src->max_val);
1038 * Compute new mean and S after the merge
1039 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1040 * #Parallel_algorithm>
1043 mean = src->mean.u.f;
1046 double delta = src->mean.u.f - dst->mean.u.f;
1048 mean = ((src->mean.u.f * src->samples) +
1049 (dst->mean.u.f * dst->samples)) /
1050 (dst->samples + src->samples);
1052 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1053 (dst->samples * src->samples) /
1054 (dst->samples + src->samples);
1057 dst->samples += src->samples;
1058 dst->mean.u.f = mean;
1062 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1066 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1067 if (dst->max_run[i] < src->max_run[i])
1068 dst->max_run[i] = src->max_run[i];
1069 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1070 dst->min_run[i] = src->min_run[i];
1071 if (dst->max_bw[i] < src->max_bw[i])
1072 dst->max_bw[i] = src->max_bw[i];
1073 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1074 dst->min_bw[i] = src->min_bw[i];
1076 dst->io_kb[i] += src->io_kb[i];
1077 dst->agg[i] += src->agg[i];
1081 dst->kb_base = src->kb_base;
1082 if (!dst->unit_base)
1083 dst->unit_base = src->unit_base;
1086 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1090 for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1091 if (!dst->unified_rw_rep) {
1092 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1093 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1094 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1095 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1097 dst->io_bytes[l] += src->io_bytes[l];
1099 if (dst->runtime[l] < src->runtime[l])
1100 dst->runtime[l] = src->runtime[l];
1102 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1103 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1104 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1105 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1107 dst->io_bytes[0] += src->io_bytes[l];
1109 if (dst->runtime[0] < src->runtime[l])
1110 dst->runtime[0] = src->runtime[l];
1114 dst->usr_time += src->usr_time;
1115 dst->sys_time += src->sys_time;
1116 dst->ctx += src->ctx;
1117 dst->majf += src->majf;
1118 dst->minf += src->minf;
1120 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1121 dst->io_u_map[k] += src->io_u_map[k];
1122 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1123 dst->io_u_submit[k] += src->io_u_submit[k];
1124 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1125 dst->io_u_complete[k] += src->io_u_complete[k];
1126 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1127 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1128 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1129 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1131 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1132 if (!dst->unified_rw_rep) {
1133 dst->total_io_u[k] += src->total_io_u[k];
1134 dst->short_io_u[k] += src->short_io_u[k];
1135 dst->drop_io_u[k] += src->drop_io_u[k];
1137 dst->total_io_u[0] += src->total_io_u[k];
1138 dst->short_io_u[0] += src->short_io_u[k];
1139 dst->drop_io_u[0] += src->drop_io_u[k];
1143 for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1146 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1147 if (!dst->unified_rw_rep)
1148 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1150 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1154 dst->total_run_time += src->total_run_time;
1155 dst->total_submit += src->total_submit;
1156 dst->total_complete += src->total_complete;
1159 void init_group_run_stat(struct group_run_stats *gs)
1162 memset(gs, 0, sizeof(*gs));
1164 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1165 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1168 void init_thread_stat(struct thread_stat *ts)
1172 memset(ts, 0, sizeof(*ts));
1174 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1175 ts->lat_stat[j].min_val = -1UL;
1176 ts->clat_stat[j].min_val = -1UL;
1177 ts->slat_stat[j].min_val = -1UL;
1178 ts->bw_stat[j].min_val = -1UL;
1183 void __show_run_stats(void)
1185 struct group_run_stats *runstats, *rs;
1186 struct thread_data *td;
1187 struct thread_stat *threadstats, *ts;
1188 int i, j, nr_ts, last_ts, idx;
1189 int kb_base_warned = 0;
1190 int unit_base_warned = 0;
1191 struct json_object *root = NULL;
1192 struct json_array *array = NULL;
1194 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1196 for (i = 0; i < groupid + 1; i++)
1197 init_group_run_stat(&runstats[i]);
1200 * find out how many threads stats we need. if group reporting isn't
1201 * enabled, it's one-per-td.
1205 for_each_td(td, i) {
1206 if (!td->o.group_reporting) {
1210 if (last_ts == td->groupid)
1213 last_ts = td->groupid;
1217 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1219 for (i = 0; i < nr_ts; i++)
1220 init_thread_stat(&threadstats[i]);
1225 for_each_td(td, i) {
1226 if (idx && (!td->o.group_reporting ||
1227 (td->o.group_reporting && last_ts != td->groupid))) {
1232 last_ts = td->groupid;
1234 ts = &threadstats[j];
1236 ts->clat_percentiles = td->o.clat_percentiles;
1237 ts->percentile_precision = td->o.percentile_precision;
1238 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1243 if (ts->groupid == -1) {
1245 * These are per-group shared already
1247 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1248 if (td->o.description)
1249 strncpy(ts->description, td->o.description,
1250 FIO_JOBDESC_SIZE - 1);
1252 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1255 * If multiple entries in this group, this is
1258 ts->thread_number = td->thread_number;
1259 ts->groupid = td->groupid;
1262 * first pid in group, not very useful...
1266 ts->kb_base = td->o.kb_base;
1267 ts->unit_base = td->o.unit_base;
1268 ts->unified_rw_rep = td->o.unified_rw_rep;
1269 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1270 log_info("fio: kb_base differs for jobs in group, using"
1271 " %u as the base\n", ts->kb_base);
1273 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1274 log_info("fio: unit_base differs for jobs in group, using"
1275 " %u as the base\n", ts->unit_base);
1276 unit_base_warned = 1;
1279 ts->continue_on_error = td->o.continue_on_error;
1280 ts->total_err_count += td->total_err_count;
1281 ts->first_error = td->first_error;
1283 if (!td->error && td->o.continue_on_error &&
1285 ts->error = td->first_error;
1286 ts->verror[sizeof(ts->verror) - 1] = '\0';
1287 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1288 } else if (td->error) {
1289 ts->error = td->error;
1290 ts->verror[sizeof(ts->verror) - 1] = '\0';
1291 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1295 ts->latency_depth = td->latency_qd;
1296 ts->latency_target = td->o.latency_target;
1297 ts->latency_percentile = td->o.latency_percentile;
1298 ts->latency_window = td->o.latency_window;
1300 sum_thread_stats(ts, &td->ts, idx);
1303 for (i = 0; i < nr_ts; i++) {
1304 unsigned long long bw;
1306 ts = &threadstats[i];
1307 rs = &runstats[ts->groupid];
1308 rs->kb_base = ts->kb_base;
1309 rs->unit_base = ts->unit_base;
1310 rs->unified_rw_rep += ts->unified_rw_rep;
1312 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1313 if (!ts->runtime[j])
1315 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1316 rs->min_run[j] = ts->runtime[j];
1317 if (ts->runtime[j] > rs->max_run[j])
1318 rs->max_run[j] = ts->runtime[j];
1321 if (ts->runtime[j]) {
1322 unsigned long runt = ts->runtime[j];
1323 unsigned long long kb;
1325 kb = ts->io_bytes[j] / rs->kb_base;
1326 bw = kb * 1000 / runt;
1328 if (bw < rs->min_bw[j])
1330 if (bw > rs->max_bw[j])
1333 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1337 for (i = 0; i < groupid + 1; i++) {
1342 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1343 if (rs->max_run[ddir])
1344 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1350 * don't overwrite last signal output
1352 if (output_format == FIO_OUTPUT_NORMAL)
1354 else if (output_format == FIO_OUTPUT_JSON) {
1355 root = json_create_object();
1356 json_object_add_value_string(root, "fio version", fio_version_string);
1357 array = json_create_array();
1358 json_object_add_value_array(root, "jobs", array);
1361 for (i = 0; i < nr_ts; i++) {
1362 ts = &threadstats[i];
1363 rs = &runstats[ts->groupid];
1366 fio_server_send_ts(ts, rs);
1367 else if (output_format == FIO_OUTPUT_TERSE)
1368 show_thread_status_terse(ts, rs);
1369 else if (output_format == FIO_OUTPUT_JSON) {
1370 struct json_object *tmp = show_thread_status_json(ts, rs);
1371 json_array_add_value_object(array, tmp);
1373 show_thread_status_normal(ts, rs);
1375 if (output_format == FIO_OUTPUT_JSON) {
1376 /* disk util stats, if any */
1377 show_disk_util(1, root);
1379 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1381 json_print_object(root);
1383 json_free_object(root);
1386 for (i = 0; i < groupid + 1; i++) {
1391 fio_server_send_gs(rs);
1392 else if (output_format == FIO_OUTPUT_NORMAL)
1393 show_group_stats(rs);
1397 fio_server_send_du();
1398 else if (output_format == FIO_OUTPUT_NORMAL) {
1399 show_disk_util(0, NULL);
1400 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1403 if ( !(output_format == FIO_OUTPUT_TERSE) && append_terse_output) {
1404 log_info("\nAdditional Terse Output:\n");
1406 for (i = 0; i < nr_ts; i++) {
1407 ts = &threadstats[i];
1408 rs = &runstats[ts->groupid];
1409 show_thread_status_terse(ts, rs);
1418 void show_run_stats(void)
1420 fio_mutex_down(stat_mutex);
1422 fio_mutex_up(stat_mutex);
1425 void __show_running_run_stats(void)
1427 struct thread_data *td;
1428 unsigned long long *rt;
1432 fio_mutex_down(stat_mutex);
1434 rt = malloc(thread_number * sizeof(unsigned long long));
1435 fio_gettime(&tv, NULL);
1437 for_each_td(td, i) {
1438 rt[i] = mtime_since(&td->start, &tv);
1439 if (td_read(td) && td->io_bytes[DDIR_READ])
1440 td->ts.runtime[DDIR_READ] += rt[i];
1441 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1442 td->ts.runtime[DDIR_WRITE] += rt[i];
1443 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1444 td->ts.runtime[DDIR_TRIM] += rt[i];
1446 td->update_rusage = 1;
1447 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1448 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1449 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1450 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1453 for_each_td(td, i) {
1454 if (td->runstate >= TD_EXITED)
1456 if (td->rusage_sem) {
1457 td->update_rusage = 1;
1458 fio_mutex_down(td->rusage_sem);
1460 td->update_rusage = 0;
1465 for_each_td(td, i) {
1466 if (td_read(td) && td->io_bytes[DDIR_READ])
1467 td->ts.runtime[DDIR_READ] -= rt[i];
1468 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1469 td->ts.runtime[DDIR_WRITE] -= rt[i];
1470 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1471 td->ts.runtime[DDIR_TRIM] -= rt[i];
1475 fio_mutex_up(stat_mutex);
1478 static int status_interval_init;
1479 static struct timeval status_time;
1480 static int status_file_disabled;
1482 #define FIO_STATUS_FILE "fio-dump-status"
1484 static int check_status_file(void)
1487 const char *temp_dir;
1488 char fio_status_file_path[PATH_MAX];
1490 if (status_file_disabled)
1493 temp_dir = getenv("TMPDIR");
1494 if (temp_dir == NULL) {
1495 temp_dir = getenv("TEMP");
1496 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1499 if (temp_dir == NULL)
1502 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1504 if (stat(fio_status_file_path, &sb))
1507 if (unlink(fio_status_file_path) < 0) {
1508 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1510 log_err("fio: disabling status file updates\n");
1511 status_file_disabled = 1;
1517 void check_for_running_stats(void)
1519 if (status_interval) {
1520 if (!status_interval_init) {
1521 fio_gettime(&status_time, NULL);
1522 status_interval_init = 1;
1523 } else if (mtime_since_now(&status_time) >= status_interval) {
1524 show_running_run_stats();
1525 fio_gettime(&status_time, NULL);
1529 if (check_status_file()) {
1530 show_running_run_stats();
1535 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1540 if (data > is->max_val)
1542 if (data < is->min_val)
1545 delta = val - is->mean.u.f;
1547 is->mean.u.f += delta / (is->samples + 1.0);
1548 is->S.u.f += delta * (val - is->mean.u.f);
1554 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1555 enum fio_ddir ddir, unsigned int bs,
1556 unsigned long t, uint64_t offset)
1558 uint64_t nr_samples = iolog->nr_samples;
1559 struct io_sample *s;
1561 if (iolog->disabled)
1564 if (!iolog->nr_samples)
1565 iolog->avg_last = t;
1567 if (iolog->nr_samples == iolog->max_samples) {
1571 new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
1573 if (iolog->log_gz && (new_size > iolog->log_gz)) {
1574 if (iolog_flush(iolog, 0)) {
1575 log_err("fio: failed flushing iolog! Will stop logging.\n");
1576 iolog->disabled = 1;
1579 nr_samples = iolog->nr_samples;
1581 new_log = realloc(iolog->log, new_size);
1583 log_err("fio: failed extending iolog! Will stop logging.\n");
1584 iolog->disabled = 1;
1587 iolog->log = new_log;
1588 iolog->max_samples <<= 1;
1592 s = get_sample(iolog, nr_samples);
1596 io_sample_set_ddir(iolog, s, ddir);
1599 if (iolog->log_offset) {
1600 struct io_sample_offset *so = (void *) s;
1602 so->offset = offset;
1605 iolog->nr_samples++;
1608 static inline void reset_io_stat(struct io_stat *ios)
1610 ios->max_val = ios->min_val = ios->samples = 0;
1611 ios->mean.u.f = ios->S.u.f = 0;
1614 void reset_io_stats(struct thread_data *td)
1616 struct thread_stat *ts = &td->ts;
1619 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1620 reset_io_stat(&ts->clat_stat[i]);
1621 reset_io_stat(&ts->slat_stat[i]);
1622 reset_io_stat(&ts->lat_stat[i]);
1623 reset_io_stat(&ts->bw_stat[i]);
1624 reset_io_stat(&ts->iops_stat[i]);
1626 ts->io_bytes[i] = 0;
1629 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1630 ts->io_u_plat[i][j] = 0;
1633 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1634 ts->io_u_map[i] = 0;
1635 ts->io_u_submit[i] = 0;
1636 ts->io_u_complete[i] = 0;
1637 ts->io_u_lat_u[i] = 0;
1638 ts->io_u_lat_m[i] = 0;
1639 ts->total_submit = 0;
1640 ts->total_complete = 0;
1643 for (i = 0; i < 3; i++) {
1644 ts->total_io_u[i] = 0;
1645 ts->short_io_u[i] = 0;
1646 ts->drop_io_u[i] = 0;
1650 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
1653 * Note an entry in the log. Use the mean from the logged samples,
1654 * making sure to properly round up. Only write a log entry if we
1655 * had actual samples done.
1657 if (iolog->avg_window[DDIR_READ].samples) {
1660 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1661 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed, 0);
1663 if (iolog->avg_window[DDIR_WRITE].samples) {
1666 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1667 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
1669 if (iolog->avg_window[DDIR_TRIM].samples) {
1672 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1673 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed, 0);
1676 reset_io_stat(&iolog->avg_window[DDIR_READ]);
1677 reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1678 reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1681 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1682 unsigned long val, enum fio_ddir ddir,
1683 unsigned int bs, uint64_t offset)
1685 unsigned long elapsed, this_window;
1690 elapsed = mtime_since_now(&td->epoch);
1693 * If no time averaging, just add the log sample.
1695 if (!iolog->avg_msec) {
1696 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
1701 * Add the sample. If the time period has passed, then
1702 * add that entry to the log and clear.
1704 add_stat_sample(&iolog->avg_window[ddir], val);
1707 * If period hasn't passed, adding the above sample is all we
1710 this_window = elapsed - iolog->avg_last;
1711 if (this_window < iolog->avg_msec)
1714 _add_stat_to_log(iolog, elapsed);
1716 iolog->avg_last = elapsed;
1719 void finalize_logs(struct thread_data *td)
1721 unsigned long elapsed;
1723 elapsed = mtime_since_now(&td->epoch);
1726 _add_stat_to_log(td->clat_log, elapsed);
1728 _add_stat_to_log(td->slat_log, elapsed);
1730 _add_stat_to_log(td->lat_log, elapsed);
1732 _add_stat_to_log(td->bw_log, elapsed);
1734 _add_stat_to_log(td->iops_log, elapsed);
1737 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1739 struct io_log *iolog;
1744 iolog = agg_io_log[ddir];
1745 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
1748 static void add_clat_percentile_sample(struct thread_stat *ts,
1749 unsigned long usec, enum fio_ddir ddir)
1751 unsigned int idx = plat_val_to_idx(usec);
1752 assert(idx < FIO_IO_U_PLAT_NR);
1754 ts->io_u_plat[ddir][idx]++;
1757 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1758 unsigned long usec, unsigned int bs, uint64_t offset)
1760 struct thread_stat *ts = &td->ts;
1765 add_stat_sample(&ts->clat_stat[ddir], usec);
1768 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
1770 if (ts->clat_percentiles)
1771 add_clat_percentile_sample(ts, usec, ddir);
1774 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1775 unsigned long usec, unsigned int bs, uint64_t offset)
1777 struct thread_stat *ts = &td->ts;
1782 add_stat_sample(&ts->slat_stat[ddir], usec);
1785 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
1788 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1789 unsigned long usec, unsigned int bs, uint64_t offset)
1791 struct thread_stat *ts = &td->ts;
1796 add_stat_sample(&ts->lat_stat[ddir], usec);
1799 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
1802 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1805 struct thread_stat *ts = &td->ts;
1806 unsigned long spent, rate;
1811 spent = mtime_since(&td->bw_sample_time, t);
1812 if (spent < td->o.bw_avg_time)
1816 * Compute both read and write rates for the interval.
1818 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1821 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
1823 continue; /* No entries for interval */
1826 rate = delta * 1000 / spent / 1024;
1830 add_stat_sample(&ts->bw_stat[ddir], rate);
1833 add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
1835 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
1838 fio_gettime(&td->bw_sample_time, NULL);
1841 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
1844 struct thread_stat *ts = &td->ts;
1845 unsigned long spent, iops;
1850 spent = mtime_since(&td->iops_sample_time, t);
1851 if (spent < td->o.iops_avg_time)
1855 * Compute both read and write rates for the interval.
1857 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
1860 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
1862 continue; /* No entries for interval */
1865 iops = (delta * 1000) / spent;
1869 add_stat_sample(&ts->iops_stat[ddir], iops);
1872 add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
1874 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
1877 fio_gettime(&td->iops_sample_time, NULL);
1880 void stat_init(void)
1882 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1885 void stat_exit(void)
1888 * When we have the mutex, we know out-of-band access to it
1891 fio_mutex_down(stat_mutex);
1892 fio_mutex_remove(stat_mutex);
1896 * Called from signal handler. Wake up status thread.
1898 void show_running_run_stats(void)
1901 pthread_cond_signal(&helper_cond);