5 #include <sys/sysmacros.h>
8 #ifdef CONFIG_VALGRIND_DEV
9 #include <valgrind/drd.h>
11 #define DRD_IGNORE_VAR(x) do { } while (0)
17 #include "helper_thread.h"
19 static int last_majdev, last_mindev;
20 static struct disk_util *last_du;
22 static struct fio_sem *disk_util_sem;
24 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
25 int majdev, int mindev, char *path);
27 static void disk_util_free(struct disk_util *du)
32 while (!flist_empty(&du->slaves)) {
33 struct disk_util *slave;
35 slave = flist_first_entry(&du->slaves, struct disk_util, slavelist);
36 flist_del(&slave->slavelist);
40 fio_sem_remove(du->lock);
45 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
48 unsigned long long sectors[2];
54 dprint(FD_DISKUTIL, "open stat file: %s\n", du->path);
56 f = fopen(du->path, "r");
60 p = fgets(line, sizeof(line), f);
66 dprint(FD_DISKUTIL, "%s: %s", du->path, p);
68 ret = sscanf(p, "%llu %llu %llu %llu %llu %llu %llu %llu %u %llu %llu\n",
69 (unsigned long long *) &dus->s.ios[0],
70 (unsigned long long *) &dus->s.merges[0],
72 (unsigned long long *) &dus->s.ticks[0],
73 (unsigned long long *) &dus->s.ios[1],
74 (unsigned long long *) &dus->s.merges[1],
76 (unsigned long long *) &dus->s.ticks[1],
78 (unsigned long long *) &dus->s.io_ticks,
79 (unsigned long long *) &dus->s.time_in_queue);
81 dprint(FD_DISKUTIL, "%s: stat read ok? %d\n", du->path, ret == 1);
82 dus->s.sectors[0] = sectors[0];
83 dus->s.sectors[1] = sectors[1];
87 static void update_io_tick_disk(struct disk_util *du)
89 struct disk_util_stat __dus, *dus, *ldus;
94 if (get_io_ticks(du, &__dus))
100 dus->s.sectors[0] += (__dus.s.sectors[0] - ldus->s.sectors[0]);
101 dus->s.sectors[1] += (__dus.s.sectors[1] - ldus->s.sectors[1]);
102 dus->s.ios[0] += (__dus.s.ios[0] - ldus->s.ios[0]);
103 dus->s.ios[1] += (__dus.s.ios[1] - ldus->s.ios[1]);
104 dus->s.merges[0] += (__dus.s.merges[0] - ldus->s.merges[0]);
105 dus->s.merges[1] += (__dus.s.merges[1] - ldus->s.merges[1]);
106 dus->s.ticks[0] += (__dus.s.ticks[0] - ldus->s.ticks[0]);
107 dus->s.ticks[1] += (__dus.s.ticks[1] - ldus->s.ticks[1]);
108 dus->s.io_ticks += (__dus.s.io_ticks - ldus->s.io_ticks);
109 dus->s.time_in_queue += (__dus.s.time_in_queue - ldus->s.time_in_queue);
111 fio_gettime(&t, NULL);
112 dus->s.msec += mtime_since(&du->time, &t);
113 memcpy(&du->time, &t, sizeof(t));
114 memcpy(&ldus->s, &__dus.s, sizeof(__dus.s));
117 int update_io_ticks(void)
119 struct flist_head *entry;
120 struct disk_util *du;
123 dprint(FD_DISKUTIL, "update io ticks\n");
125 fio_sem_down(disk_util_sem);
127 if (!helper_should_exit()) {
128 flist_for_each(entry, &disk_list) {
129 du = flist_entry(entry, struct disk_util, list);
130 update_io_tick_disk(du);
135 fio_sem_up(disk_util_sem);
139 static struct disk_util *disk_util_exists(int major, int minor)
141 struct flist_head *entry;
142 struct disk_util *du;
144 fio_sem_down(disk_util_sem);
146 flist_for_each(entry, &disk_list) {
147 du = flist_entry(entry, struct disk_util, list);
149 if (major == du->major && minor == du->minor) {
150 fio_sem_up(disk_util_sem);
155 fio_sem_up(disk_util_sem);
159 static int get_device_numbers(char *file_name, int *maj, int *min)
163 char tempname[PATH_MAX], *p;
165 if (!lstat(file_name, &st)) {
166 if (S_ISBLK(st.st_mode)) {
167 majdev = major(st.st_rdev);
168 mindev = minor(st.st_rdev);
169 } else if (S_ISCHR(st.st_mode)) {
170 majdev = major(st.st_rdev);
171 mindev = minor(st.st_rdev);
172 if (fio_lookup_raw(st.st_rdev, &majdev, &mindev))
174 } else if (S_ISFIFO(st.st_mode))
177 majdev = major(st.st_dev);
178 mindev = minor(st.st_dev);
182 * must be a file, open "." in that path
184 snprintf(tempname, ARRAY_SIZE(tempname), "%s", file_name);
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 nowarn_snprintf(temppath, sizeof(temppath), "%s/%s", slavesdir,
246 /* Can we always assume that the slaves device entries
247 * are links to the real directories for the slave
250 linklen = readlink(temppath, slavepath, PATH_MAX - 1);
252 perror("readlink() for slave device.");
256 slavepath[linklen] = '\0';
258 nowarn_snprintf(temppath, sizeof(temppath), "%s/%s/dev",
259 slavesdir, slavepath);
260 if (access(temppath, F_OK) != 0)
261 nowarn_snprintf(temppath, sizeof(temppath),
262 "%s/%s/device/dev", slavesdir,
264 if (read_block_dev_entry(temppath, &majdev, &mindev)) {
265 perror("Error getting slave device numbers");
271 * See if this maj,min already exists
273 slavedu = disk_util_exists(majdev, mindev);
277 nowarn_snprintf(temppath, sizeof(temppath), "%s/%s", slavesdir,
279 __init_per_file_disk_util(td, majdev, mindev, temppath);
280 slavedu = disk_util_exists(majdev, mindev);
282 /* Should probably use an assert here. slavedu should
283 * always be present at this point. */
286 flist_add_tail(&slavedu->slavelist, &masterdu->slaves);
293 static struct disk_util *disk_util_add(struct thread_data *td, int majdev,
294 int mindev, char *path)
296 struct disk_util *du, *__du;
297 struct flist_head *entry;
300 dprint(FD_DISKUTIL, "add maj/min %d/%d: %s\n", majdev, mindev, path);
302 du = smalloc(sizeof(*du));
306 DRD_IGNORE_VAR(du->users);
307 memset(du, 0, sizeof(*du));
308 INIT_FLIST_HEAD(&du->list);
309 l = snprintf(du->path, sizeof(du->path), "%s/stat", path);
310 if (l < 0 || l >= sizeof(du->path)) {
311 log_err("constructed path \"%.100s[...]/stat\" larger than buffer (%zu bytes)\n",
312 path, sizeof(du->path) - 1);
316 snprintf((char *) du->dus.name, ARRAY_SIZE(du->dus.name), "%s",
318 du->sysfs_root = strdup(path);
321 INIT_FLIST_HEAD(&du->slavelist);
322 INIT_FLIST_HEAD(&du->slaves);
323 du->lock = fio_sem_init(FIO_SEM_UNLOCKED);
326 fio_sem_down(disk_util_sem);
328 flist_for_each(entry, &disk_list) {
329 __du = flist_entry(entry, struct disk_util, list);
331 dprint(FD_DISKUTIL, "found %s in list\n", __du->dus.name);
333 if (!strcmp((char *) du->dus.name, (char *) __du->dus.name)) {
335 fio_sem_up(disk_util_sem);
340 dprint(FD_DISKUTIL, "add %s to list\n", du->dus.name);
342 fio_gettime(&du->time, NULL);
343 get_io_ticks(du, &du->last_dus);
345 flist_add_tail(&du->list, &disk_list);
346 fio_sem_up(disk_util_sem);
348 find_add_disk_slaves(td, path, du);
352 static int check_dev_match(int majdev, int mindev, char *path)
356 if (read_block_dev_entry(path, &major, &minor))
359 if (majdev == major && mindev == minor)
365 static int find_block_dir(int majdev, int mindev, char *path, int link_ok)
376 while ((dir = readdir(D)) != NULL) {
379 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
382 sprintf(full_path, "%s/%s", path, dir->d_name);
384 if (!strcmp(dir->d_name, "dev")) {
385 if (!check_dev_match(majdev, mindev, full_path)) {
392 if (stat(full_path, &st) == -1) {
397 if (lstat(full_path, &st) == -1) {
403 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
406 found = find_block_dir(majdev, mindev, full_path, 0);
408 strcpy(path, full_path);
417 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
418 int majdev, int mindev,
426 * If there's a ../queue/ directory there, we are inside a partition.
427 * Check if that is the case and jump back. For loop/md/dm etc we
428 * are already in the right spot.
430 sprintf(tmp, "%s/../queue", path);
431 if (!stat(tmp, &st)) {
433 sprintf(tmp, "%s/queue", p);
434 if (stat(tmp, &st)) {
435 log_err("unknown sysfs layout\n");
438 snprintf(tmp, ARRAY_SIZE(tmp), "%s", p);
439 sprintf(path, "%s", tmp);
442 return disk_util_add(td, majdev, mindev, path);
445 static struct disk_util *init_per_file_disk_util(struct thread_data *td,
450 struct disk_util *du;
453 if (get_device_numbers(filename, &majdev, &mindev))
456 dprint(FD_DISKUTIL, "%s belongs to maj/min %d/%d\n", filename, majdev,
459 du = disk_util_exists(majdev, mindev);
464 * for an fs without a device, we will repeatedly stat through
465 * sysfs which can take oodles of time for thousands of files. so
466 * cache the last lookup and compare with that before going through
469 if (mindev == last_mindev && majdev == last_majdev)
472 last_mindev = mindev;
473 last_majdev = majdev;
475 sprintf(foo, "/sys/block");
476 if (!find_block_dir(majdev, mindev, foo, 1))
479 return __init_per_file_disk_util(td, majdev, mindev, foo);
482 static struct disk_util *__init_disk_util(struct thread_data *td,
485 return init_per_file_disk_util(td, f->file_name);
488 void init_disk_util(struct thread_data *td)
493 if (!td->o.do_disk_util ||
494 td_ioengine_flagged(td, FIO_DISKLESSIO | FIO_NODISKUTIL))
497 for_each_file(td, f, i)
498 f->du = __init_disk_util(td, f);
501 static void show_agg_stats(struct disk_util_agg *agg, int terse,
502 struct buf_output *out)
504 if (!agg->slavecount)
508 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
509 "aggrticks=%llu/%llu, aggrin_queue=%llu, "
511 (unsigned long long) agg->ios[0] / agg->slavecount,
512 (unsigned long long) agg->ios[1] / agg->slavecount,
513 (unsigned long long) agg->merges[0] / agg->slavecount,
514 (unsigned long long) agg->merges[1] / agg->slavecount,
515 (unsigned long long) agg->ticks[0] / agg->slavecount,
516 (unsigned long long) agg->ticks[1] / agg->slavecount,
517 (unsigned long long) agg->time_in_queue / agg->slavecount,
520 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
521 (unsigned long long) agg->ios[0] / agg->slavecount,
522 (unsigned long long) agg->ios[1] / agg->slavecount,
523 (unsigned long long) agg->merges[0] / agg->slavecount,
524 (unsigned long long) agg->merges[1] / agg->slavecount,
525 (unsigned long long) agg->ticks[0] / agg->slavecount,
526 (unsigned long long) agg->ticks[1] / agg->slavecount,
527 (unsigned long long) agg->time_in_queue / agg->slavecount,
532 static void aggregate_slaves_stats(struct disk_util *masterdu)
534 struct disk_util_agg *agg = &masterdu->agg;
535 struct disk_util_stat *dus;
536 struct flist_head *entry;
537 struct disk_util *slavedu;
540 flist_for_each(entry, &masterdu->slaves) {
541 slavedu = flist_entry(entry, struct disk_util, slavelist);
543 agg->ios[0] += dus->s.ios[0];
544 agg->ios[1] += dus->s.ios[1];
545 agg->merges[0] += dus->s.merges[0];
546 agg->merges[1] += dus->s.merges[1];
547 agg->sectors[0] += dus->s.sectors[0];
548 agg->sectors[1] += dus->s.sectors[1];
549 agg->ticks[0] += dus->s.ticks[0];
550 agg->ticks[1] += dus->s.ticks[1];
551 agg->time_in_queue += dus->s.time_in_queue;
554 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
555 /* System utilization is the utilization of the
556 * component with the highest utilization.
558 if (util > agg->max_util.u.f)
559 agg->max_util.u.f = util;
563 if (agg->max_util.u.f > 100.0)
564 agg->max_util.u.f = 100.0;
567 void disk_util_prune_entries(void)
569 fio_sem_down(disk_util_sem);
571 while (!flist_empty(&disk_list)) {
572 struct disk_util *du;
574 du = flist_first_entry(&disk_list, struct disk_util, list);
575 flist_del(&du->list);
579 last_majdev = last_mindev = -1;
580 fio_sem_up(disk_util_sem);
581 fio_sem_remove(disk_util_sem);
584 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
585 int terse, struct buf_output *out)
590 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
598 log_buf(out, " %s: ios=%llu/%llu, merge=%llu/%llu, "
599 "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
601 (unsigned long long) dus->s.ios[0],
602 (unsigned long long) dus->s.ios[1],
603 (unsigned long long) dus->s.merges[0],
604 (unsigned long long) dus->s.merges[1],
605 (unsigned long long) dus->s.ticks[0],
606 (unsigned long long) dus->s.ticks[1],
607 (unsigned long long) dus->s.time_in_queue,
610 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
612 (unsigned long long) dus->s.ios[0],
613 (unsigned long long) dus->s.ios[1],
614 (unsigned long long) dus->s.merges[0],
615 (unsigned long long) dus->s.merges[1],
616 (unsigned long long) dus->s.ticks[0],
617 (unsigned long long) dus->s.ticks[1],
618 (unsigned long long) dus->s.time_in_queue,
623 * If the device has slaves, aggregate the stats for
624 * those slave devices also.
626 show_agg_stats(agg, terse, out);
632 void json_array_add_disk_util(struct disk_util_stat *dus,
633 struct disk_util_agg *agg, struct json_array *array)
635 struct json_object *obj;
639 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
643 obj = json_create_object();
644 json_array_add_value_object(array, obj);
646 json_object_add_value_string(obj, "name", dus->name);
647 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
648 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
649 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
650 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
651 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
652 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
653 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
654 json_object_add_value_float(obj, "util", util);
657 * If the device has slaves, aggregate the stats for
658 * those slave devices also.
660 if (!agg->slavecount)
662 json_object_add_value_int(obj, "aggr_read_ios",
663 agg->ios[0] / agg->slavecount);
664 json_object_add_value_int(obj, "aggr_write_ios",
665 agg->ios[1] / agg->slavecount);
666 json_object_add_value_int(obj, "aggr_read_merges",
667 agg->merges[0] / agg->slavecount);
668 json_object_add_value_int(obj, "aggr_write_merge",
669 agg->merges[1] / agg->slavecount);
670 json_object_add_value_int(obj, "aggr_read_ticks",
671 agg->ticks[0] / agg->slavecount);
672 json_object_add_value_int(obj, "aggr_write_ticks",
673 agg->ticks[1] / agg->slavecount);
674 json_object_add_value_int(obj, "aggr_in_queue",
675 agg->time_in_queue / agg->slavecount);
676 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
679 static void json_object_add_disk_utils(struct json_object *obj,
680 struct flist_head *head)
682 struct json_array *array = json_create_array();
683 struct flist_head *entry;
684 struct disk_util *du;
686 json_object_add_value_array(obj, "disk_util", array);
688 flist_for_each(entry, head) {
689 du = flist_entry(entry, struct disk_util, list);
691 aggregate_slaves_stats(du);
692 json_array_add_disk_util(&du->dus, &du->agg, array);
696 void show_disk_util(int terse, struct json_object *parent,
697 struct buf_output *out)
699 struct flist_head *entry;
700 struct disk_util *du;
703 if (!is_running_backend())
706 fio_sem_down(disk_util_sem);
708 if (flist_empty(&disk_list)) {
709 fio_sem_up(disk_util_sem);
713 if ((output_format & FIO_OUTPUT_JSON) && parent)
718 if (!terse && !do_json)
719 log_buf(out, "\nDisk stats (read/write):\n");
722 json_object_add_disk_utils(parent, &disk_list);
723 else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
724 flist_for_each(entry, &disk_list) {
725 du = flist_entry(entry, struct disk_util, list);
727 aggregate_slaves_stats(du);
728 print_disk_util(&du->dus, &du->agg, terse, out);
732 fio_sem_up(disk_util_sem);
735 void setup_disk_util(void)
737 disk_util_sem = fio_sem_init(FIO_SEM_UNLOCKED);