14 #include "helper_thread.h"
16 static int last_majdev, last_mindev;
17 static struct disk_util *last_du;
19 static struct fio_mutex *disk_util_mutex;
21 FLIST_HEAD(disk_list);
23 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
24 int majdev, int mindev, char *path);
26 static void disk_util_free(struct disk_util *du)
31 while (!flist_empty(&du->slaves)) {
32 struct disk_util *slave;
34 slave = flist_first_entry(&du->slaves, struct disk_util, slavelist);
35 flist_del(&slave->slavelist);
39 fio_mutex_remove(du->lock);
44 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
47 unsigned long long sectors[2];
53 dprint(FD_DISKUTIL, "open stat file: %s\n", du->path);
55 f = fopen(du->path, "r");
59 p = fgets(line, sizeof(line), f);
65 dprint(FD_DISKUTIL, "%s: %s", du->path, p);
67 ret = sscanf(p, "%llu %llu %llu %llu %llu %llu %llu %llu %u %llu %llu\n",
68 (unsigned long long *) &dus->s.ios[0],
69 (unsigned long long *) &dus->s.merges[0],
71 (unsigned long long *) &dus->s.ticks[0],
72 (unsigned long long *) &dus->s.ios[1],
73 (unsigned long long *) &dus->s.merges[1],
75 (unsigned long long *) &dus->s.ticks[1],
77 (unsigned long long *) &dus->s.io_ticks,
78 (unsigned long long *) &dus->s.time_in_queue);
80 dprint(FD_DISKUTIL, "%s: stat read ok? %d\n", du->path, ret == 1);
81 dus->s.sectors[0] = sectors[0];
82 dus->s.sectors[1] = sectors[1];
86 static void update_io_tick_disk(struct disk_util *du)
88 struct disk_util_stat __dus, *dus, *ldus;
93 if (get_io_ticks(du, &__dus))
99 dus->s.sectors[0] += (__dus.s.sectors[0] - ldus->s.sectors[0]);
100 dus->s.sectors[1] += (__dus.s.sectors[1] - ldus->s.sectors[1]);
101 dus->s.ios[0] += (__dus.s.ios[0] - ldus->s.ios[0]);
102 dus->s.ios[1] += (__dus.s.ios[1] - ldus->s.ios[1]);
103 dus->s.merges[0] += (__dus.s.merges[0] - ldus->s.merges[0]);
104 dus->s.merges[1] += (__dus.s.merges[1] - ldus->s.merges[1]);
105 dus->s.ticks[0] += (__dus.s.ticks[0] - ldus->s.ticks[0]);
106 dus->s.ticks[1] += (__dus.s.ticks[1] - ldus->s.ticks[1]);
107 dus->s.io_ticks += (__dus.s.io_ticks - ldus->s.io_ticks);
108 dus->s.time_in_queue += (__dus.s.time_in_queue - ldus->s.time_in_queue);
110 fio_gettime(&t, NULL);
111 dus->s.msec += mtime_since(&du->time, &t);
112 memcpy(&du->time, &t, sizeof(t));
113 memcpy(&ldus->s, &__dus.s, sizeof(__dus.s));
116 int update_io_ticks(void)
118 struct flist_head *entry;
119 struct disk_util *du;
122 dprint(FD_DISKUTIL, "update io ticks\n");
124 fio_mutex_down(disk_util_mutex);
126 if (!helper_should_exit()) {
127 flist_for_each(entry, &disk_list) {
128 du = flist_entry(entry, struct disk_util, list);
129 update_io_tick_disk(du);
134 fio_mutex_up(disk_util_mutex);
138 static struct disk_util *disk_util_exists(int major, int minor)
140 struct flist_head *entry;
141 struct disk_util *du;
143 fio_mutex_down(disk_util_mutex);
145 flist_for_each(entry, &disk_list) {
146 du = flist_entry(entry, struct disk_util, list);
148 if (major == du->major && minor == du->minor) {
149 fio_mutex_up(disk_util_mutex);
154 fio_mutex_up(disk_util_mutex);
158 static int get_device_numbers(char *file_name, int *maj, int *min)
162 char tempname[PATH_MAX], *p;
164 if (!lstat(file_name, &st)) {
165 if (S_ISBLK(st.st_mode)) {
166 majdev = major(st.st_rdev);
167 mindev = minor(st.st_rdev);
168 } else if (S_ISCHR(st.st_mode)) {
169 majdev = major(st.st_rdev);
170 mindev = minor(st.st_rdev);
171 if (fio_lookup_raw(st.st_rdev, &majdev, &mindev))
173 } else if (S_ISFIFO(st.st_mode))
176 majdev = major(st.st_dev);
177 mindev = minor(st.st_dev);
181 * must be a file, open "." in that path
183 tempname[PATH_MAX - 1] = '\0';
184 strncpy(tempname, file_name, PATH_MAX - 1);
185 p = dirname(tempname);
187 perror("disk util stat");
191 majdev = major(st.st_dev);
192 mindev = minor(st.st_dev);
201 static int read_block_dev_entry(char *path, int *maj, int *min)
206 f = fopen(path, "r");
212 p = fgets(line, sizeof(line), f);
218 if (sscanf(p, "%u:%u", maj, min) != 2)
224 static void find_add_disk_slaves(struct thread_data *td, char *path,
225 struct disk_util *masterdu)
227 DIR *dirhandle = NULL;
228 struct dirent *dirent = NULL;
229 char slavesdir[PATH_MAX], temppath[PATH_MAX], slavepath[PATH_MAX];
230 struct disk_util *slavedu = NULL;
234 sprintf(slavesdir, "%s/%s", path, "slaves");
235 dirhandle = opendir(slavesdir);
239 while ((dirent = readdir(dirhandle)) != NULL) {
240 if (!strcmp(dirent->d_name, ".") ||
241 !strcmp(dirent->d_name, ".."))
244 sprintf(temppath, "%s/%s", slavesdir, dirent->d_name);
245 /* Can we always assume that the slaves device entries
246 * are links to the real directories for the slave
249 linklen = readlink(temppath, slavepath, PATH_MAX - 1);
251 perror("readlink() for slave device.");
255 slavepath[linklen] = '\0';
257 sprintf(temppath, "%s/%s/dev", slavesdir, slavepath);
258 if (read_block_dev_entry(temppath, &majdev, &mindev)) {
259 perror("Error getting slave device numbers.");
265 * See if this maj,min already exists
267 slavedu = disk_util_exists(majdev, mindev);
271 sprintf(temppath, "%s/%s", slavesdir, slavepath);
272 __init_per_file_disk_util(td, majdev, mindev, temppath);
273 slavedu = disk_util_exists(majdev, mindev);
275 /* Should probably use an assert here. slavedu should
276 * always be present at this point. */
279 flist_add_tail(&slavedu->slavelist, &masterdu->slaves);
286 static struct disk_util *disk_util_add(struct thread_data *td, int majdev,
287 int mindev, char *path)
289 struct disk_util *du, *__du;
290 struct flist_head *entry;
293 dprint(FD_DISKUTIL, "add maj/min %d/%d: %s\n", majdev, mindev, path);
295 du = smalloc(sizeof(*du));
299 memset(du, 0, sizeof(*du));
300 INIT_FLIST_HEAD(&du->list);
301 l = snprintf(du->path, sizeof(du->path), "%s/stat", path);
302 if (l < 0 || l >= sizeof(du->path)) {
303 log_err("constructed path \"%.100s[...]/stat\" larger than buffer (%zu bytes)\n",
304 path, sizeof(du->path) - 1);
308 strncpy((char *) du->dus.name, basename(path), FIO_DU_NAME_SZ - 1);
309 du->sysfs_root = strdup(path);
312 INIT_FLIST_HEAD(&du->slavelist);
313 INIT_FLIST_HEAD(&du->slaves);
314 du->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
317 fio_mutex_down(disk_util_mutex);
319 flist_for_each(entry, &disk_list) {
320 __du = flist_entry(entry, struct disk_util, list);
322 dprint(FD_DISKUTIL, "found %s in list\n", __du->dus.name);
324 if (!strcmp((char *) du->dus.name, (char *) __du->dus.name)) {
326 fio_mutex_up(disk_util_mutex);
331 dprint(FD_DISKUTIL, "add %s to list\n", du->dus.name);
333 fio_gettime(&du->time, NULL);
334 get_io_ticks(du, &du->last_dus);
336 flist_add_tail(&du->list, &disk_list);
337 fio_mutex_up(disk_util_mutex);
339 find_add_disk_slaves(td, path, du);
343 static int check_dev_match(int majdev, int mindev, char *path)
347 if (read_block_dev_entry(path, &major, &minor))
350 if (majdev == major && mindev == minor)
356 static int find_block_dir(int majdev, int mindev, char *path, int link_ok)
367 while ((dir = readdir(D)) != NULL) {
370 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
373 sprintf(full_path, "%s/%s", path, dir->d_name);
375 if (!strcmp(dir->d_name, "dev")) {
376 if (!check_dev_match(majdev, mindev, full_path)) {
383 if (stat(full_path, &st) == -1) {
388 if (lstat(full_path, &st) == -1) {
394 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
397 found = find_block_dir(majdev, mindev, full_path, 0);
399 strcpy(path, full_path);
408 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
409 int majdev, int mindev,
417 * If there's a ../queue/ directory there, we are inside a partition.
418 * Check if that is the case and jump back. For loop/md/dm etc we
419 * are already in the right spot.
421 sprintf(tmp, "%s/../queue", path);
422 if (!stat(tmp, &st)) {
424 sprintf(tmp, "%s/queue", p);
425 if (stat(tmp, &st)) {
426 log_err("unknown sysfs layout\n");
429 tmp[PATH_MAX - 1] = '\0';
430 strncpy(tmp, p, PATH_MAX - 1);
431 sprintf(path, "%s", tmp);
434 return disk_util_add(td, majdev, mindev, path);
437 static struct disk_util *init_per_file_disk_util(struct thread_data *td,
442 struct disk_util *du;
445 if (get_device_numbers(filename, &majdev, &mindev))
448 dprint(FD_DISKUTIL, "%s belongs to maj/min %d/%d\n", filename, majdev,
451 du = disk_util_exists(majdev, mindev);
456 * for an fs without a device, we will repeatedly stat through
457 * sysfs which can take oodles of time for thousands of files. so
458 * cache the last lookup and compare with that before going through
461 if (mindev == last_mindev && majdev == last_majdev)
464 last_mindev = mindev;
465 last_majdev = majdev;
467 sprintf(foo, "/sys/block");
468 if (!find_block_dir(majdev, mindev, foo, 1))
471 return __init_per_file_disk_util(td, majdev, mindev, foo);
474 static struct disk_util *__init_disk_util(struct thread_data *td,
477 return init_per_file_disk_util(td, f->file_name);
480 void init_disk_util(struct thread_data *td)
485 if (!td->o.do_disk_util ||
486 td_ioengine_flagged(td, FIO_DISKLESSIO | FIO_NODISKUTIL))
489 for_each_file(td, f, i)
490 f->du = __init_disk_util(td, f);
493 static void show_agg_stats(struct disk_util_agg *agg, int terse,
494 struct buf_output *out)
496 if (!agg->slavecount)
500 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
501 "aggrticks=%llu/%llu, aggrin_queue=%llu, "
503 (unsigned long long) agg->ios[0] / agg->slavecount,
504 (unsigned long long) agg->ios[1] / agg->slavecount,
505 (unsigned long long) agg->merges[0] / agg->slavecount,
506 (unsigned long long) agg->merges[1] / agg->slavecount,
507 (unsigned long long) agg->ticks[0] / agg->slavecount,
508 (unsigned long long) agg->ticks[1] / agg->slavecount,
509 (unsigned long long) agg->time_in_queue / agg->slavecount,
512 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
513 (unsigned long long) agg->ios[0] / agg->slavecount,
514 (unsigned long long) agg->ios[1] / agg->slavecount,
515 (unsigned long long) agg->merges[0] / agg->slavecount,
516 (unsigned long long) agg->merges[1] / agg->slavecount,
517 (unsigned long long) agg->ticks[0] / agg->slavecount,
518 (unsigned long long) agg->ticks[1] / agg->slavecount,
519 (unsigned long long) agg->time_in_queue / agg->slavecount,
524 static void aggregate_slaves_stats(struct disk_util *masterdu)
526 struct disk_util_agg *agg = &masterdu->agg;
527 struct disk_util_stat *dus;
528 struct flist_head *entry;
529 struct disk_util *slavedu;
532 flist_for_each(entry, &masterdu->slaves) {
533 slavedu = flist_entry(entry, struct disk_util, slavelist);
535 agg->ios[0] += dus->s.ios[0];
536 agg->ios[1] += dus->s.ios[1];
537 agg->merges[0] += dus->s.merges[0];
538 agg->merges[1] += dus->s.merges[1];
539 agg->sectors[0] += dus->s.sectors[0];
540 agg->sectors[1] += dus->s.sectors[1];
541 agg->ticks[0] += dus->s.ticks[0];
542 agg->ticks[1] += dus->s.ticks[1];
543 agg->time_in_queue += dus->s.time_in_queue;
546 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
547 /* System utilization is the utilization of the
548 * component with the highest utilization.
550 if (util > agg->max_util.u.f)
551 agg->max_util.u.f = util;
555 if (agg->max_util.u.f > 100.0)
556 agg->max_util.u.f = 100.0;
559 void disk_util_prune_entries(void)
561 fio_mutex_down(disk_util_mutex);
563 while (!flist_empty(&disk_list)) {
564 struct disk_util *du;
566 du = flist_first_entry(&disk_list, struct disk_util, list);
567 flist_del(&du->list);
571 last_majdev = last_mindev = -1;
572 fio_mutex_up(disk_util_mutex);
573 fio_mutex_remove(disk_util_mutex);
576 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
577 int terse, struct buf_output *out)
582 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
590 log_buf(out, " %s: ios=%llu/%llu, merge=%llu/%llu, "
591 "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
593 (unsigned long long) dus->s.ios[0],
594 (unsigned long long) dus->s.ios[1],
595 (unsigned long long) dus->s.merges[0],
596 (unsigned long long) dus->s.merges[1],
597 (unsigned long long) dus->s.ticks[0],
598 (unsigned long long) dus->s.ticks[1],
599 (unsigned long long) dus->s.time_in_queue,
602 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
604 (unsigned long long) dus->s.ios[0],
605 (unsigned long long) dus->s.ios[1],
606 (unsigned long long) dus->s.merges[0],
607 (unsigned long long) dus->s.merges[1],
608 (unsigned long long) dus->s.ticks[0],
609 (unsigned long long) dus->s.ticks[1],
610 (unsigned long long) dus->s.time_in_queue,
615 * If the device has slaves, aggregate the stats for
616 * those slave devices also.
618 show_agg_stats(agg, terse, out);
624 void json_array_add_disk_util(struct disk_util_stat *dus,
625 struct disk_util_agg *agg, struct json_array *array)
627 struct json_object *obj;
631 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
635 obj = json_create_object();
636 json_array_add_value_object(array, obj);
638 json_object_add_value_string(obj, "name", dus->name);
639 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
640 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
641 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
642 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
643 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
644 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
645 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
646 json_object_add_value_float(obj, "util", util);
649 * If the device has slaves, aggregate the stats for
650 * those slave devices also.
652 if (!agg->slavecount)
654 json_object_add_value_int(obj, "aggr_read_ios",
655 agg->ios[0] / agg->slavecount);
656 json_object_add_value_int(obj, "aggr_write_ios",
657 agg->ios[1] / agg->slavecount);
658 json_object_add_value_int(obj, "aggr_read_merges",
659 agg->merges[0] / agg->slavecount);
660 json_object_add_value_int(obj, "aggr_write_merge",
661 agg->merges[1] / agg->slavecount);
662 json_object_add_value_int(obj, "aggr_read_ticks",
663 agg->ticks[0] / agg->slavecount);
664 json_object_add_value_int(obj, "aggr_write_ticks",
665 agg->ticks[1] / agg->slavecount);
666 json_object_add_value_int(obj, "aggr_in_queue",
667 agg->time_in_queue / agg->slavecount);
668 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
671 static void json_object_add_disk_utils(struct json_object *obj,
672 struct flist_head *head)
674 struct json_array *array = json_create_array();
675 struct flist_head *entry;
676 struct disk_util *du;
678 json_object_add_value_array(obj, "disk_util", array);
680 flist_for_each(entry, head) {
681 du = flist_entry(entry, struct disk_util, list);
683 aggregate_slaves_stats(du);
684 json_array_add_disk_util(&du->dus, &du->agg, array);
688 void show_disk_util(int terse, struct json_object *parent,
689 struct buf_output *out)
691 struct flist_head *entry;
692 struct disk_util *du;
695 if (!disk_util_mutex)
698 fio_mutex_down(disk_util_mutex);
700 if (flist_empty(&disk_list)) {
701 fio_mutex_up(disk_util_mutex);
705 if ((output_format & FIO_OUTPUT_JSON) && parent)
710 if (!terse && !do_json)
711 log_buf(out, "\nDisk stats (read/write):\n");
714 json_object_add_disk_utils(parent, &disk_list);
715 else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
716 flist_for_each(entry, &disk_list) {
717 du = flist_entry(entry, struct disk_util, list);
719 aggregate_slaves_stats(du);
720 print_disk_util(&du->dus, &du->agg, terse, out);
724 fio_mutex_up(disk_util_mutex);
727 void setup_disk_util(void)
729 disk_util_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);