12 static struct itimerval itimer;
13 static struct list_head disk_list = LIST_HEAD_INIT(disk_list);
14 static dev_t last_dev;
17 * Cheesy number->string conversion, complete with carry rounding error.
19 static char *num2str(unsigned long num, int maxlen, int base)
22 * could be passed in for 10^3 base, but every caller expects
23 * 2^10 base right now.
25 const unsigned int thousand = 1024;
26 char postfix[] = { 'K', 'M', 'G', 'P', 'E' };
32 for (i = 0; base > 1; i++)
38 len = sprintf(buf, "%'lu", num);
40 buf[len] = postfix[i];
45 if ((num % thousand) >= (thousand / 2))
56 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
63 f = fopen(du->path, "r");
67 p = fgets(line, sizeof(line), f);
73 if (sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0], &dus->merges[0], &dus->sectors[0], &dus->ticks[0], &dus->ios[1], &dus->merges[1], &dus->sectors[1], &dus->ticks[1], &in_flight, &dus->io_ticks, &dus->time_in_queue) != 11) {
82 static void update_io_tick_disk(struct disk_util *du)
84 struct disk_util_stat __dus, *dus, *ldus;
87 if (get_io_ticks(du, &__dus))
93 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
94 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
95 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
96 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
97 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
98 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
99 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
100 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
101 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
102 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
104 fio_gettime(&t, NULL);
105 du->msec += mtime_since(&du->time, &t);
106 memcpy(&du->time, &t, sizeof(t));
107 memcpy(ldus, &__dus, sizeof(__dus));
110 void update_io_ticks(void)
112 struct list_head *entry;
113 struct disk_util *du;
115 list_for_each(entry, &disk_list) {
116 du = list_entry(entry, struct disk_util, list);
117 update_io_tick_disk(du);
121 static int disk_util_exists(dev_t dev)
123 struct list_head *entry;
124 struct disk_util *du;
126 list_for_each(entry, &disk_list) {
127 du = list_entry(entry, struct disk_util, list);
136 static void disk_util_add(dev_t dev, char *path)
138 struct disk_util *du, *__du;
139 struct list_head *entry;
141 du = malloc(sizeof(*du));
142 memset(du, 0, sizeof(*du));
143 INIT_LIST_HEAD(&du->list);
144 sprintf(du->path, "%s/stat", path);
145 du->name = strdup(basename(path));
148 list_for_each(entry, &disk_list) {
149 __du = list_entry(entry, struct disk_util, list);
151 if (!strcmp(du->name, __du->name)) {
158 fio_gettime(&du->time, NULL);
159 get_io_ticks(du, &du->last_dus);
161 list_add_tail(&du->list, &disk_list);
164 static int check_dev_match(dev_t dev, char *path)
166 unsigned int major, minor;
170 f = fopen(path, "r");
176 p = fgets(line, sizeof(line), f);
182 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
187 if (((major << 8) | minor) == dev) {
196 static int find_block_dir(dev_t dev, char *path)
207 while ((dir = readdir(D)) != NULL) {
210 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
213 sprintf(full_path, "%s/%s", path, dir->d_name);
215 if (!strcmp(dir->d_name, "dev")) {
216 if (!check_dev_match(dev, full_path)) {
222 if (lstat(full_path, &st) == -1) {
227 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
230 found = find_block_dir(dev, full_path);
232 strcpy(path, full_path);
241 void init_disk_util(struct thread_data *td)
245 char foo[PATH_MAX], tmp[PATH_MAX];
249 if (!td->do_disk_util || (td->io_ops->flags & FIO_DISKLESSIO))
253 * Just use the same file, they are on the same device.
256 if (!stat(f->file_name, &st)) {
257 if (S_ISBLK(st.st_mode))
263 * must be a file, open "." in that path
265 strncpy(foo, f->file_name, PATH_MAX - 1);
268 perror("disk util stat");
275 if (disk_util_exists(dev))
279 * for an fs without a device, we will repeatedly stat through
280 * sysfs which can take oodles of time for thousands of files. so
281 * cache the last lookup and compare with that before going through
289 sprintf(foo, "/sys/block");
290 if (!find_block_dir(dev, foo))
294 * If there's a ../queue/ directory there, we are inside a partition.
295 * Check if that is the case and jump back. For loop/md/dm etc we
296 * are already in the right spot.
298 sprintf(tmp, "%s/../queue", foo);
299 if (!stat(tmp, &st)) {
301 sprintf(tmp, "%s/queue", p);
302 if (stat(tmp, &st)) {
303 log_err("unknown sysfs layout\n");
306 strncpy(tmp, p, PATH_MAX - 1);
307 sprintf(foo, "%s", tmp);
311 td->sysfs_root = strdup(foo);
313 disk_util_add(dev, foo);
316 void disk_util_timer_arm(void)
318 itimer.it_value.tv_sec = 0;
319 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
320 setitimer(ITIMER_REAL, &itimer, NULL);
323 void update_rusage_stat(struct thread_data *td)
325 struct thread_stat *ts = &td->ts;
327 getrusage(RUSAGE_SELF, &ts->ru_end);
329 ts->usr_time += mtime_since(&ts->ru_start.ru_utime, &ts->ru_end.ru_utime);
330 ts->sys_time += mtime_since(&ts->ru_start.ru_stime, &ts->ru_end.ru_stime);
331 ts->ctx += ts->ru_end.ru_nvcsw + ts->ru_end.ru_nivcsw - (ts->ru_start.ru_nvcsw + ts->ru_start.ru_nivcsw);
333 memcpy(&ts->ru_start, &ts->ru_end, sizeof(ts->ru_end));
336 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
337 double *mean, double *dev)
339 double n = is->samples;
341 if (is->samples == 0)
347 n = (double) is->samples;
351 *dev = sqrt(is->S / (n - 1.0));
358 static void show_group_stats(struct group_run_stats *rs, int id)
360 char *p1, *p2, *p3, *p4;
361 const char *ddir_str[] = { " READ", " WRITE" };
364 fprintf(f_out, "\nRun status group %d (all jobs):\n", id);
366 for (i = 0; i <= DDIR_WRITE; i++) {
370 p1 = num2str(rs->io_kb[i], 6, 1);
371 p2 = num2str(rs->agg[i], 6, 1);
372 p3 = num2str(rs->min_bw[i], 6, 1);
373 p4 = num2str(rs->max_bw[i], 6, 1);
375 fprintf(f_out, "%s: io=%siB, aggrb=%siB/s, minb=%siB/s, maxb=%siB/s, mint=%llumsec, maxt=%llumsec\n", ddir_str[i], p1, p2, p3, p4, rs->min_run[0], rs->max_run[0]);
384 static void show_disk_util(void)
386 struct disk_util_stat *dus;
387 struct list_head *entry, *next;
388 struct disk_util *du;
391 fprintf(f_out, "\nDisk stats (read/write):\n");
393 list_for_each(entry, &disk_list) {
394 du = list_entry(entry, struct disk_util, list);
397 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
401 fprintf(f_out, " %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, in_queue=%u, util=%3.2f%%\n", du->name, dus->ios[0], dus->ios[1], dus->merges[0], dus->merges[1], dus->ticks[0], dus->ticks[1], dus->time_in_queue, util);
407 list_for_each_safe(entry, next, &disk_list) {
409 du = list_entry(entry, struct disk_util, list);
415 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
418 const char *ddir_str[] = { "read ", "write" };
419 unsigned long min, max;
420 unsigned long long bw;
424 if (!ts->runtime[ddir])
427 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
428 io_p = num2str(ts->io_bytes[ddir] >> 10, 6, 1);
429 bw_p = num2str(bw, 6, 1);
431 fprintf(f_out, " %s: io=%siB, bw=%siB/s, runt=%6lumsec\n", ddir_str[ddir], io_p, bw_p, ts->runtime[ddir]);
436 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
437 fprintf(f_out, " slat (msec): min=%5lu, max=%5lu, avg=%5.02f, stdev=%5.02f\n", min, max, mean, dev);
439 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
440 fprintf(f_out, " clat (msec): min=%5lu, max=%5lu, avg=%5.02f, stdev=%5.02f\n", min, max, mean, dev);
442 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
445 p_of_agg = mean * 100 / (double) rs->agg[ddir];
446 fprintf(f_out, " bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg, mean, dev);
450 static void show_thread_status(struct thread_stat *ts,
451 struct group_run_stats *rs)
453 double usr_cpu, sys_cpu;
454 unsigned long runtime;
455 double io_u_dist[FIO_IO_U_MAP_NR];
456 double io_u_lat[FIO_IO_U_LAT_NR];
459 if (!(ts->io_bytes[0] + ts->io_bytes[1]))
463 fprintf(f_out, "%s: (groupid=%d): err=%2d: pid=%d\n", ts->name, ts->groupid, ts->error, ts->pid);
465 fprintf(f_out, "%s: (groupid=%d): err=%2d (%s): pid=%d\n", ts->name, ts->groupid, ts->error, ts->verror, ts->pid);
467 if (ts->io_bytes[DDIR_READ])
468 show_ddir_status(rs, ts, DDIR_READ);
469 if (ts->io_bytes[DDIR_WRITE])
470 show_ddir_status(rs, ts, DDIR_WRITE);
472 runtime = ts->total_run_time;
474 double runt = (double) runtime;
476 usr_cpu = (double) ts->usr_time * 100 / runt;
477 sys_cpu = (double) ts->sys_time * 100 / runt;
483 fprintf(f_out, " cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, ts->ctx);
486 * Do depth distribution calculations
488 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
489 io_u_dist[i] = (double) ts->io_u_map[i] / (double) ts->total_io_u;
490 io_u_dist[i] *= 100.0;
493 fprintf(f_out, " IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], io_u_dist[4], io_u_dist[5], io_u_dist[6]);
496 * Do latency distribution calculations
498 for (i = 0; i < FIO_IO_U_LAT_NR; i++) {
499 io_u_lat[i] = (double) ts->io_u_lat[i] / (double) ts->total_io_u;
500 io_u_lat[i] *= 100.0;
503 fprintf(f_out, " lat (msec): 2=%3.1f%%, 4=%3.1f%%, 10=%3.1f%%, 20=%3.1f%%, 50=%3.1f%%, 100=%3.1f%%\n", io_u_lat[0], io_u_lat[1], io_u_lat[2], io_u_lat[3], io_u_lat[4], io_u_lat[5]);
504 fprintf(f_out, " lat (msec): 250=%3.1f%%, 500=%3.1f%%, 750=%3.1f%%, 1000=%3.1f%%, >=2000=%3.1f%%\n", io_u_lat[6], io_u_lat[7], io_u_lat[8], io_u_lat[9], io_u_lat[10]);
507 fprintf(f_out, "%s\n", ts->description);
510 static void show_ddir_status_terse(struct thread_stat *ts,
511 struct group_run_stats *rs, int ddir)
513 unsigned long min, max;
514 unsigned long long bw;
518 if (ts->runtime[ddir])
519 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
521 fprintf(f_out, ",%llu,%llu,%lu", ts->io_bytes[ddir] >> 10, bw, ts->runtime[ddir]);
523 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
524 fprintf(f_out, ",%lu,%lu,%f,%f", min, max, mean, dev);
526 fprintf(f_out, ",%lu,%lu,%f,%f", 0UL, 0UL, 0.0, 0.0);
528 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
529 fprintf(f_out, ",%lu,%lu,%f,%f", min, max, mean, dev);
531 fprintf(f_out, ",%lu,%lu,%f,%f", 0UL, 0UL, 0.0, 0.0);
533 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
536 p_of_agg = mean * 100 / (double) rs->agg[ddir];
537 fprintf(f_out, ",%lu,%lu,%f%%,%f,%f", min, max, p_of_agg, mean, dev);
539 fprintf(f_out, ",%lu,%lu,%f%%,%f,%f", 0UL, 0UL, 0.0, 0.0, 0.0);
543 static void show_thread_status_terse(struct thread_stat *ts,
544 struct group_run_stats *rs)
546 double usr_cpu, sys_cpu;
548 fprintf(f_out, "%s,%d,%d", ts->name, ts->groupid, ts->error);
550 show_ddir_status_terse(ts, rs, 0);
551 show_ddir_status_terse(ts, rs, 1);
553 if (ts->total_run_time) {
554 double runt = (double) ts->total_run_time;
556 usr_cpu = (double) ts->usr_time * 100 / runt;
557 sys_cpu = (double) ts->sys_time * 100 / runt;
563 fprintf(f_out, ",%f%%,%f%%,%lu\n", usr_cpu, sys_cpu, ts->ctx);
566 static void __sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
570 dst->min_val = min(dst->min_val, src->min_val);
571 dst->max_val = max(dst->max_val, src->max_val);
572 dst->samples += src->samples;
575 * Needs a new method for calculating stddev, we cannot just
576 * average them we do below for nr > 1
582 mean = ((src->mean * (double) (nr - 1)) + dst->mean) / ((double) nr);
583 S = ((src->S * (double) (nr - 1)) + dst->S) / ((double) nr);
590 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
592 __sum_stat(&dst[DDIR_READ], &src[DDIR_READ], nr);
593 __sum_stat(&dst[DDIR_WRITE], &src[DDIR_WRITE], nr);
596 void show_run_stats(void)
598 struct group_run_stats *runstats, *rs;
599 struct thread_data *td;
600 struct thread_stat *threadstats, *ts;
601 int i, j, k, nr_ts, last_ts, members;
603 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
605 for (i = 0; i < groupid + 1; i++) {
608 memset(rs, 0, sizeof(*rs));
609 rs->min_bw[0] = rs->min_run[0] = ~0UL;
610 rs->min_bw[1] = rs->min_run[1] = ~0UL;
614 * find out how many threads stats we need. if group reporting isn't
615 * enabled, it's one-per-td.
620 if (!td->group_reporting) {
624 if (last_ts == td->groupid)
627 last_ts = td->groupid;
631 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
633 for (i = 0; i < nr_ts; i++) {
634 ts = &threadstats[i];
636 memset(ts, 0, sizeof(*ts));
637 ts->clat_stat[0].min_val = -1UL;
638 ts->clat_stat[1].min_val = -1UL;
639 ts->slat_stat[0].min_val = -1UL;
640 ts->slat_stat[1].min_val = -1UL;
641 ts->bw_stat[0].min_val = -1UL;
642 ts->bw_stat[1].min_val = -1UL;
649 ts = &threadstats[j];
655 * These are per-group shared already
658 ts->description = td->description;
659 ts->groupid = td->groupid;
662 * first pid in group, not very useful...
667 if (td->error && !ts->error) {
668 ts->error = td->error;
669 ts->verror = td->verror;
672 sum_stat(ts->clat_stat, td->ts.clat_stat, members);
673 sum_stat(ts->slat_stat, td->ts.slat_stat, members);
674 sum_stat(ts->bw_stat, td->ts.bw_stat, members);
676 ts->stat_io_bytes[0] += td->ts.stat_io_bytes[0];
677 ts->stat_io_bytes[1] += td->ts.stat_io_bytes[1];
679 ts->usr_time += td->ts.usr_time;
680 ts->sys_time += td->ts.sys_time;
681 ts->ctx += td->ts.ctx;
683 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
684 ts->io_u_map[k] += td->ts.io_u_map[k];
685 for (k = 0; k < FIO_IO_U_LAT_NR; k++)
686 ts->io_u_lat[k] += td->ts.io_u_lat[k];
688 ts->total_io_u += td->ts.total_io_u;
689 ts->io_bytes[0] += td->ts.io_bytes[0];
690 ts->io_bytes[1] += td->ts.io_bytes[1];
692 if (ts->runtime[0] < td->ts.runtime[0])
693 ts->runtime[0] = td->ts.runtime[0];
694 if (ts->runtime[1] < td->ts.runtime[1])
695 ts->runtime[1] = td->ts.runtime[1];
697 ts->total_run_time += td->ts.total_run_time;
699 if (!td->group_reporting) {
704 if (last_ts == td->groupid)
712 last_ts = td->groupid;
715 for (i = 0; i < nr_ts; i++) {
716 unsigned long long rbw, wbw;
718 ts = &threadstats[i];
719 rs = &runstats[ts->groupid];
721 if (ts->runtime[0] < rs->min_run[0] || !rs->min_run[0])
722 rs->min_run[0] = ts->runtime[0];
723 if (ts->runtime[0] > rs->max_run[0])
724 rs->max_run[0] = ts->runtime[0];
725 if (ts->runtime[1] < rs->min_run[1] || !rs->min_run[1])
726 rs->min_run[1] = ts->runtime[1];
727 if (ts->runtime[1] > rs->max_run[1])
728 rs->max_run[1] = ts->runtime[1];
732 rbw = ts->io_bytes[0] / (unsigned long long) ts->runtime[0];
734 wbw = ts->io_bytes[1] / (unsigned long long) ts->runtime[1];
736 if (rbw < rs->min_bw[0])
738 if (wbw < rs->min_bw[1])
740 if (rbw > rs->max_bw[0])
742 if (wbw > rs->max_bw[1])
745 rs->io_kb[0] += ts->io_bytes[0] >> 10;
746 rs->io_kb[1] += ts->io_bytes[1] >> 10;
749 for (i = 0; i < groupid + 1; i++) {
753 rs->agg[0] = (rs->io_kb[0]*1024) / rs->max_run[0];
755 rs->agg[1] = (rs->io_kb[1]*1024) / rs->max_run[1];
759 * don't overwrite last signal output
764 for (i = 0; i < nr_ts; i++) {
765 ts = &threadstats[i];
766 rs = &runstats[ts->groupid];
769 show_thread_status_terse(ts, rs);
771 show_thread_status(ts, rs);
775 for (i = 0; i < groupid + 1; i++)
776 show_group_stats(&runstats[i], i);
785 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
790 if (data > is->max_val)
792 if (data < is->min_val)
795 delta = val - is->mean;
796 is->mean += delta / (is->samples + 1.0);
797 is->S += delta * (val - is->mean);
802 static void __add_log_sample(struct io_log *iolog, unsigned long val,
803 enum fio_ddir ddir, unsigned long time)
805 if (iolog->nr_samples == iolog->max_samples) {
806 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
808 iolog->log = realloc(iolog->log, new_size);
809 iolog->max_samples <<= 1;
812 iolog->log[iolog->nr_samples].val = val;
813 iolog->log[iolog->nr_samples].time = time;
814 iolog->log[iolog->nr_samples].ddir = ddir;
818 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
819 unsigned long val, enum fio_ddir ddir)
821 __add_log_sample(iolog, val, ddir, mtime_since_now(&td->epoch));
824 void add_agg_sample(unsigned long val, enum fio_ddir ddir)
826 struct io_log *iolog = agg_io_log[ddir];
828 __add_log_sample(iolog, val, ddir, mtime_since_genesis());
831 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
834 struct thread_stat *ts = &td->ts;
836 add_stat_sample(&ts->clat_stat[ddir], msec);
839 add_log_sample(td, ts->clat_log, msec, ddir);
842 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
845 struct thread_stat *ts = &td->ts;
847 add_stat_sample(&ts->slat_stat[ddir], msec);
850 add_log_sample(td, ts->slat_log, msec, ddir);
853 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir,
856 struct thread_stat *ts = &td->ts;
857 unsigned long spent = mtime_since(&ts->stat_sample_time[ddir], t);
860 if (spent < td->bw_avg_time)
863 rate = (td->this_io_bytes[ddir] - ts->stat_io_bytes[ddir]) / spent;
864 add_stat_sample(&ts->bw_stat[ddir], rate);
867 add_log_sample(td, ts->bw_log, rate, ddir);
869 fio_gettime(&ts->stat_sample_time[ddir], NULL);
870 ts->stat_io_bytes[ddir] = td->this_io_bytes[ddir];