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, int pow2)
21 char postfix[] = { ' ', 'K', 'M', 'G', 'P', 'E' };
22 unsigned int thousand;
33 for (i = 0; base > 1; i++)
39 len = sprintf(buf, "%'lu", num);
42 buf[len] = postfix[i];
48 if ((num % thousand) >= (thousand / 2))
59 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
66 f = fopen(du->path, "r");
70 p = fgets(line, sizeof(line), f);
76 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) {
85 static void update_io_tick_disk(struct disk_util *du)
87 struct disk_util_stat __dus, *dus, *ldus;
90 if (get_io_ticks(du, &__dus))
96 dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
97 dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
98 dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
99 dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
100 dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
101 dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
102 dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
103 dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
104 dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
105 dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
107 fio_gettime(&t, NULL);
108 du->msec += mtime_since(&du->time, &t);
109 memcpy(&du->time, &t, sizeof(t));
110 memcpy(ldus, &__dus, sizeof(__dus));
113 void update_io_ticks(void)
115 struct list_head *entry;
116 struct disk_util *du;
118 list_for_each(entry, &disk_list) {
119 du = list_entry(entry, struct disk_util, list);
120 update_io_tick_disk(du);
124 static int disk_util_exists(dev_t dev)
126 struct list_head *entry;
127 struct disk_util *du;
129 list_for_each(entry, &disk_list) {
130 du = list_entry(entry, struct disk_util, list);
139 static void disk_util_add(dev_t dev, char *path)
141 struct disk_util *du, *__du;
142 struct list_head *entry;
144 du = malloc(sizeof(*du));
145 memset(du, 0, sizeof(*du));
146 INIT_LIST_HEAD(&du->list);
147 sprintf(du->path, "%s/stat", path);
148 du->name = strdup(basename(path));
151 list_for_each(entry, &disk_list) {
152 __du = list_entry(entry, struct disk_util, list);
154 if (!strcmp(du->name, __du->name)) {
161 fio_gettime(&du->time, NULL);
162 get_io_ticks(du, &du->last_dus);
164 list_add_tail(&du->list, &disk_list);
167 static int check_dev_match(dev_t dev, char *path)
169 unsigned int major, minor;
173 f = fopen(path, "r");
179 p = fgets(line, sizeof(line), f);
185 if (sscanf(p, "%u:%u", &major, &minor) != 2) {
190 if (((major << 8) | minor) == dev) {
199 static int find_block_dir(dev_t dev, char *path)
210 while ((dir = readdir(D)) != NULL) {
213 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
216 sprintf(full_path, "%s/%s", path, dir->d_name);
218 if (!strcmp(dir->d_name, "dev")) {
219 if (!check_dev_match(dev, full_path)) {
225 if (lstat(full_path, &st) == -1) {
230 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
233 found = find_block_dir(dev, full_path);
235 strcpy(path, full_path);
244 static void __init_disk_util(struct thread_data *td, struct fio_file *f)
247 char foo[PATH_MAX], tmp[PATH_MAX];
251 if (!stat(f->file_name, &st)) {
252 if (S_ISBLK(st.st_mode))
258 * must be a file, open "." in that path
260 strncpy(foo, f->file_name, PATH_MAX - 1);
263 perror("disk util stat");
270 if (disk_util_exists(dev))
274 * for an fs without a device, we will repeatedly stat through
275 * sysfs which can take oodles of time for thousands of files. so
276 * cache the last lookup and compare with that before going through
284 sprintf(foo, "/sys/block");
285 if (!find_block_dir(dev, foo))
289 * If there's a ../queue/ directory there, we are inside a partition.
290 * Check if that is the case and jump back. For loop/md/dm etc we
291 * are already in the right spot.
293 sprintf(tmp, "%s/../queue", foo);
294 if (!stat(tmp, &st)) {
296 sprintf(tmp, "%s/queue", p);
297 if (stat(tmp, &st)) {
298 log_err("unknown sysfs layout\n");
301 strncpy(tmp, p, PATH_MAX - 1);
302 sprintf(foo, "%s", tmp);
305 if (td->o.ioscheduler && !td->sysfs_root)
306 td->sysfs_root = strdup(foo);
308 disk_util_add(dev, foo);
311 void init_disk_util(struct thread_data *td)
316 if (!td->o.do_disk_util ||
317 (td->io_ops->flags & (FIO_DISKLESSIO | FIO_NODISKUTIL)))
320 for_each_file(td, f, i)
321 __init_disk_util(td, f);
324 void disk_util_timer_arm(void)
326 itimer.it_value.tv_sec = 0;
327 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
328 setitimer(ITIMER_REAL, &itimer, NULL);
331 void update_rusage_stat(struct thread_data *td)
333 struct thread_stat *ts = &td->ts;
335 getrusage(RUSAGE_SELF, &ts->ru_end);
337 ts->usr_time += mtime_since(&ts->ru_start.ru_utime, &ts->ru_end.ru_utime);
338 ts->sys_time += mtime_since(&ts->ru_start.ru_stime, &ts->ru_end.ru_stime);
339 ts->ctx += ts->ru_end.ru_nvcsw + ts->ru_end.ru_nivcsw - (ts->ru_start.ru_nvcsw + ts->ru_start.ru_nivcsw);
341 memcpy(&ts->ru_start, &ts->ru_end, sizeof(ts->ru_end));
344 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
345 double *mean, double *dev)
347 double n = is->samples;
349 if (is->samples == 0)
355 n = (double) is->samples;
359 *dev = sqrt(is->S / (n - 1.0));
366 static void show_group_stats(struct group_run_stats *rs, int id)
368 char *p1, *p2, *p3, *p4;
369 const char *ddir_str[] = { " READ", " WRITE" };
372 log_info("\nRun status group %d (all jobs):\n", id);
374 for (i = 0; i <= DDIR_WRITE; i++) {
378 p1 = num2str(rs->io_kb[i], 6, 1000, 1);
379 p2 = num2str(rs->agg[i], 6, 1000, 1);
380 p3 = num2str(rs->min_bw[i], 6, 1000, 1);
381 p4 = num2str(rs->max_bw[i], 6, 1000, 1);
383 log_info("%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[i], rs->max_run[i]);
392 static void show_disk_util(void)
394 struct disk_util_stat *dus;
395 struct list_head *entry, *next;
396 struct disk_util *du;
399 log_info("\nDisk stats (read/write):\n");
401 list_for_each(entry, &disk_list) {
402 du = list_entry(entry, struct disk_util, list);
405 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
409 log_info(" %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);
415 list_for_each_safe(entry, next, &disk_list) {
417 du = list_entry(entry, struct disk_util, list);
423 #define ts_total_io_u(ts) \
424 ((ts)->total_io_u[0] + (ts)->total_io_u[1])
426 static void stat_calc_dist(struct thread_stat *ts, double *io_u_dist)
431 * Do depth distribution calculations
433 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
434 io_u_dist[i] = (double) ts->io_u_map[i] / (double) ts_total_io_u(ts);
435 io_u_dist[i] *= 100.0;
439 static void stat_calc_lat(struct thread_stat *ts, double *io_u_lat)
444 * Do latency distribution calculations
446 for (i = 0; i < FIO_IO_U_LAT_NR; i++) {
447 io_u_lat[i] = (double) ts->io_u_lat[i] / (double) ts_total_io_u(ts);
448 io_u_lat[i] *= 100.0;
452 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
455 const char *ddir_str[] = { "read ", "write" };
456 unsigned long min, max;
457 unsigned long long bw, iops;
459 char *io_p, *bw_p, *iops_p;
461 if (!ts->runtime[ddir])
464 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
465 iops = (1000 * ts->total_io_u[ddir]) / ts->runtime[ddir];
466 io_p = num2str(ts->io_bytes[ddir] >> 10, 6, 1000, 1);
467 bw_p = num2str(bw, 6, 1000, 1);
468 iops_p = num2str(iops, 6, 1, 0);
470 log_info(" %s: io=%siB, bw=%siB/s, iops=%s, runt=%6lumsec\n", ddir_str[ddir], io_p, bw_p, iops_p, ts->runtime[ddir]);
476 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
477 log_info(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, stdev=%5.02f\n", min, max, mean, dev);
479 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
480 log_info(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, stdev=%5.02f\n", min, max, mean, dev);
482 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
485 p_of_agg = mean * 100 / (double) rs->agg[ddir];
486 log_info(" bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, stdev=%5.02f\n", min, max, p_of_agg, mean, dev);
490 static void show_thread_status(struct thread_stat *ts,
491 struct group_run_stats *rs)
493 double usr_cpu, sys_cpu;
494 unsigned long runtime;
495 double io_u_dist[FIO_IO_U_MAP_NR];
496 double io_u_lat[FIO_IO_U_LAT_NR];
498 if (!(ts->io_bytes[0] + ts->io_bytes[1]))
502 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d\n", ts->name, ts->groupid, ts->members, ts->error, ts->pid);
504 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d\n", ts->name, ts->groupid, ts->members, ts->error, ts->verror, ts->pid);
507 log_info(" Description : [%s]\n", ts->description);
509 if (ts->io_bytes[DDIR_READ])
510 show_ddir_status(rs, ts, DDIR_READ);
511 if (ts->io_bytes[DDIR_WRITE])
512 show_ddir_status(rs, ts, DDIR_WRITE);
514 runtime = ts->total_run_time;
516 double runt = (double) runtime;
518 usr_cpu = (double) ts->usr_time * 100 / runt;
519 sys_cpu = (double) ts->sys_time * 100 / runt;
525 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, ts->ctx);
527 stat_calc_dist(ts, io_u_dist);
528 stat_calc_lat(ts, io_u_lat);
530 log_info(" 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]);
532 log_info(" 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]);
533 log_info(" 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]);
536 static void show_ddir_status_terse(struct thread_stat *ts,
537 struct group_run_stats *rs, int ddir)
539 unsigned long min, max;
540 unsigned long long bw;
544 if (ts->runtime[ddir])
545 bw = ts->io_bytes[ddir] / ts->runtime[ddir];
547 log_info(";%llu;%llu;%lu", ts->io_bytes[ddir] >> 10, bw, ts->runtime[ddir]);
549 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
550 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
552 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
554 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
555 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
557 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
559 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
562 p_of_agg = mean * 100 / (double) rs->agg[ddir];
563 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
565 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
569 static void show_thread_status_terse(struct thread_stat *ts,
570 struct group_run_stats *rs)
572 double io_u_dist[FIO_IO_U_MAP_NR];
573 double io_u_lat[FIO_IO_U_LAT_NR];
574 double usr_cpu, sys_cpu;
576 log_info("%s;%d;%d", ts->name, ts->groupid, ts->error);
578 show_ddir_status_terse(ts, rs, 0);
579 show_ddir_status_terse(ts, rs, 1);
581 if (ts->total_run_time) {
582 double runt = (double) ts->total_run_time;
584 usr_cpu = (double) ts->usr_time * 100 / runt;
585 sys_cpu = (double) ts->sys_time * 100 / runt;
591 log_info(";%f%%;%f%%;%lu", usr_cpu, sys_cpu, ts->ctx);
593 stat_calc_dist(ts, io_u_dist);
594 stat_calc_lat(ts, io_u_lat);
596 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 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]);
598 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%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]);
599 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", io_u_lat[6], io_u_lat[7], io_u_lat[8], io_u_lat[9], io_u_lat[10]);
602 log_info(";%s", ts->description);
607 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
611 dst->min_val = min(dst->min_val, src->min_val);
612 dst->max_val = max(dst->max_val, src->max_val);
613 dst->samples += src->samples;
616 * Needs a new method for calculating stddev, we cannot just
617 * average them we do below for nr > 1
623 mean = ((src->mean * (double) (nr - 1)) + dst->mean) / ((double) nr);
624 S = ((src->S * (double) (nr - 1)) + dst->S) / ((double) nr);
631 void show_run_stats(void)
633 struct group_run_stats *runstats, *rs;
634 struct thread_data *td;
635 struct thread_stat *threadstats, *ts;
636 int i, j, k, l, nr_ts, last_ts, idx;
638 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
640 for (i = 0; i < groupid + 1; i++) {
643 memset(rs, 0, sizeof(*rs));
644 rs->min_bw[0] = rs->min_run[0] = ~0UL;
645 rs->min_bw[1] = rs->min_run[1] = ~0UL;
649 * find out how many threads stats we need. if group reporting isn't
650 * enabled, it's one-per-td.
655 if (!td->o.group_reporting) {
659 if (last_ts == td->groupid)
662 last_ts = td->groupid;
666 threadstats = malloc(nr_ts * sizeof(struct thread_stat));
668 for (i = 0; i < nr_ts; i++) {
669 ts = &threadstats[i];
671 memset(ts, 0, sizeof(*ts));
672 for (j = 0; j <= DDIR_WRITE; j++) {
673 ts->clat_stat[j].min_val = -1UL;
674 ts->slat_stat[j].min_val = -1UL;
675 ts->bw_stat[j].min_val = -1UL;
684 if (idx && (!td->o.group_reporting ||
685 (td->o.group_reporting && last_ts != td->groupid))) {
690 last_ts = td->groupid;
692 ts = &threadstats[j];
697 if (ts->groupid == -1) {
699 * These are per-group shared already
701 ts->name = td->o.name;
702 ts->description = td->o.description;
703 ts->groupid = td->groupid;
706 * first pid in group, not very useful...
711 if (td->error && !ts->error) {
712 ts->error = td->error;
713 ts->verror = td->verror;
716 for (l = 0; l <= DDIR_WRITE; l++) {
717 sum_stat(&ts->clat_stat[l], &td->ts.clat_stat[l], idx);
718 sum_stat(&ts->slat_stat[l], &td->ts.slat_stat[l], idx);
719 sum_stat(&ts->bw_stat[l], &td->ts.bw_stat[l], idx);
721 ts->stat_io_bytes[l] += td->ts.stat_io_bytes[l];
722 ts->io_bytes[l] += td->ts.io_bytes[l];
724 if (ts->runtime[l] < td->ts.runtime[l])
725 ts->runtime[l] = td->ts.runtime[l];
728 ts->usr_time += td->ts.usr_time;
729 ts->sys_time += td->ts.sys_time;
730 ts->ctx += td->ts.ctx;
732 for (k = 0; k < FIO_IO_U_MAP_NR; k++)
733 ts->io_u_map[k] += td->ts.io_u_map[k];
734 for (k = 0; k < FIO_IO_U_LAT_NR; k++)
735 ts->io_u_lat[k] += td->ts.io_u_lat[k];
737 for (k = 0; k <= DDIR_WRITE; k++)
738 ts->total_io_u[k] += td->ts.total_io_u[k];
740 ts->total_run_time += td->ts.total_run_time;
743 for (i = 0; i < nr_ts; i++) {
744 unsigned long long bw;
746 ts = &threadstats[i];
747 rs = &runstats[ts->groupid];
749 for (j = 0; j <= DDIR_WRITE; j++) {
752 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
753 rs->min_run[j] = ts->runtime[j];
754 if (ts->runtime[j] > rs->max_run[j])
755 rs->max_run[j] = ts->runtime[j];
759 bw = ts->io_bytes[j] / (unsigned long long) ts->runtime[j];
760 if (bw < rs->min_bw[j])
762 if (bw > rs->max_bw[j])
765 rs->io_kb[j] += ts->io_bytes[j] >> 10;
769 for (i = 0; i < groupid + 1; i++) {
773 rs->agg[0] = (rs->io_kb[0]*1024) / rs->max_run[0];
775 rs->agg[1] = (rs->io_kb[1]*1024) / rs->max_run[1];
779 * don't overwrite last signal output
784 for (i = 0; i < nr_ts; i++) {
785 ts = &threadstats[i];
786 rs = &runstats[ts->groupid];
789 show_thread_status_terse(ts, rs);
791 show_thread_status(ts, rs);
795 for (i = 0; i < groupid + 1; i++)
796 show_group_stats(&runstats[i], i);
805 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
810 if (data > is->max_val)
812 if (data < is->min_val)
815 delta = val - is->mean;
816 is->mean += delta / (is->samples + 1.0);
817 is->S += delta * (val - is->mean);
822 static void __add_log_sample(struct io_log *iolog, unsigned long val,
823 enum fio_ddir ddir, unsigned long time)
825 if (iolog->nr_samples == iolog->max_samples) {
826 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
828 iolog->log = realloc(iolog->log, new_size);
829 iolog->max_samples <<= 1;
832 iolog->log[iolog->nr_samples].val = val;
833 iolog->log[iolog->nr_samples].time = time;
834 iolog->log[iolog->nr_samples].ddir = ddir;
838 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
839 unsigned long val, enum fio_ddir ddir)
841 __add_log_sample(iolog, val, ddir, mtime_since_now(&td->epoch));
844 void add_agg_sample(unsigned long val, enum fio_ddir ddir)
846 struct io_log *iolog = agg_io_log[ddir];
848 __add_log_sample(iolog, val, ddir, mtime_since_genesis());
851 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
854 struct thread_stat *ts = &td->ts;
856 add_stat_sample(&ts->clat_stat[ddir], msec);
859 add_log_sample(td, ts->clat_log, msec, ddir);
862 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
865 struct thread_stat *ts = &td->ts;
867 add_stat_sample(&ts->slat_stat[ddir], msec);
870 add_log_sample(td, ts->slat_log, msec, ddir);
873 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir,
876 struct thread_stat *ts = &td->ts;
877 unsigned long spent = mtime_since(&ts->stat_sample_time[ddir], t);
880 if (spent < td->o.bw_avg_time)
883 rate = (td->this_io_bytes[ddir] - ts->stat_io_bytes[ddir]) / spent;
884 add_stat_sample(&ts->bw_stat[ddir], rate);
887 add_log_sample(td, ts->bw_log, rate, ddir);
889 fio_gettime(&ts->stat_sample_time[ddir], NULL);
890 ts->stat_io_bytes[ddir] = td->this_io_bytes[ddir];