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 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
22 int majdev, int mindev, char *path);
24 static void disk_util_free(struct disk_util *du)
29 while (!flist_empty(&du->slaves)) {
30 struct disk_util *slave;
32 slave = flist_first_entry(&du->slaves, struct disk_util, slavelist);
33 flist_del(&slave->slavelist);
37 fio_mutex_remove(du->lock);
42 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
45 unsigned long long sectors[2];
51 dprint(FD_DISKUTIL, "open stat file: %s\n", du->path);
53 f = fopen(du->path, "r");
57 p = fgets(line, sizeof(line), f);
63 dprint(FD_DISKUTIL, "%s: %s", du->path, p);
65 ret = sscanf(p, "%llu %llu %llu %llu %llu %llu %llu %llu %u %llu %llu\n",
66 (unsigned long long *) &dus->s.ios[0],
67 (unsigned long long *) &dus->s.merges[0],
69 (unsigned long long *) &dus->s.ticks[0],
70 (unsigned long long *) &dus->s.ios[1],
71 (unsigned long long *) &dus->s.merges[1],
73 (unsigned long long *) &dus->s.ticks[1],
75 (unsigned long long *) &dus->s.io_ticks,
76 (unsigned long long *) &dus->s.time_in_queue);
78 dprint(FD_DISKUTIL, "%s: stat read ok? %d\n", du->path, ret == 1);
79 dus->s.sectors[0] = sectors[0];
80 dus->s.sectors[1] = sectors[1];
84 static void update_io_tick_disk(struct disk_util *du)
86 struct disk_util_stat __dus, *dus, *ldus;
91 if (get_io_ticks(du, &__dus))
97 dus->s.sectors[0] += (__dus.s.sectors[0] - ldus->s.sectors[0]);
98 dus->s.sectors[1] += (__dus.s.sectors[1] - ldus->s.sectors[1]);
99 dus->s.ios[0] += (__dus.s.ios[0] - ldus->s.ios[0]);
100 dus->s.ios[1] += (__dus.s.ios[1] - ldus->s.ios[1]);
101 dus->s.merges[0] += (__dus.s.merges[0] - ldus->s.merges[0]);
102 dus->s.merges[1] += (__dus.s.merges[1] - ldus->s.merges[1]);
103 dus->s.ticks[0] += (__dus.s.ticks[0] - ldus->s.ticks[0]);
104 dus->s.ticks[1] += (__dus.s.ticks[1] - ldus->s.ticks[1]);
105 dus->s.io_ticks += (__dus.s.io_ticks - ldus->s.io_ticks);
106 dus->s.time_in_queue += (__dus.s.time_in_queue - ldus->s.time_in_queue);
108 fio_gettime(&t, NULL);
109 dus->s.msec += mtime_since(&du->time, &t);
110 memcpy(&du->time, &t, sizeof(t));
111 memcpy(&ldus->s, &__dus.s, sizeof(__dus.s));
114 int update_io_ticks(void)
116 struct flist_head *entry;
117 struct disk_util *du;
120 dprint(FD_DISKUTIL, "update io ticks\n");
122 fio_mutex_down(disk_util_mutex);
124 if (!helper_should_exit()) {
125 flist_for_each(entry, &disk_list) {
126 du = flist_entry(entry, struct disk_util, list);
127 update_io_tick_disk(du);
132 fio_mutex_up(disk_util_mutex);
136 static struct disk_util *disk_util_exists(int major, int minor)
138 struct flist_head *entry;
139 struct disk_util *du;
141 fio_mutex_down(disk_util_mutex);
143 flist_for_each(entry, &disk_list) {
144 du = flist_entry(entry, struct disk_util, list);
146 if (major == du->major && minor == du->minor) {
147 fio_mutex_up(disk_util_mutex);
152 fio_mutex_up(disk_util_mutex);
156 static int get_device_numbers(char *file_name, int *maj, int *min)
160 char tempname[PATH_MAX], *p;
162 if (!lstat(file_name, &st)) {
163 if (S_ISBLK(st.st_mode)) {
164 majdev = major(st.st_rdev);
165 mindev = minor(st.st_rdev);
166 } else if (S_ISCHR(st.st_mode)) {
167 majdev = major(st.st_rdev);
168 mindev = minor(st.st_rdev);
169 if (fio_lookup_raw(st.st_rdev, &majdev, &mindev))
171 } else if (S_ISFIFO(st.st_mode))
174 majdev = major(st.st_dev);
175 mindev = minor(st.st_dev);
179 * must be a file, open "." in that path
181 tempname[PATH_MAX - 1] = '\0';
182 strncpy(tempname, file_name, PATH_MAX - 1);
183 p = dirname(tempname);
185 perror("disk util stat");
189 majdev = major(st.st_dev);
190 mindev = minor(st.st_dev);
199 static int read_block_dev_entry(char *path, int *maj, int *min)
204 f = fopen(path, "r");
210 p = fgets(line, sizeof(line), f);
216 if (sscanf(p, "%u:%u", maj, min) != 2)
222 static void find_add_disk_slaves(struct thread_data *td, char *path,
223 struct disk_util *masterdu)
225 DIR *dirhandle = NULL;
226 struct dirent *dirent = NULL;
227 char slavesdir[PATH_MAX], temppath[PATH_MAX], slavepath[PATH_MAX];
228 struct disk_util *slavedu = NULL;
232 sprintf(slavesdir, "%s/%s", path, "slaves");
233 dirhandle = opendir(slavesdir);
237 while ((dirent = readdir(dirhandle)) != NULL) {
238 if (!strcmp(dirent->d_name, ".") ||
239 !strcmp(dirent->d_name, ".."))
242 sprintf(temppath, "%s/%s", slavesdir, dirent->d_name);
243 /* Can we always assume that the slaves device entries
244 * are links to the real directories for the slave
247 linklen = readlink(temppath, slavepath, PATH_MAX - 1);
249 perror("readlink() for slave device.");
253 slavepath[linklen] = '\0';
255 sprintf(temppath, "%s/%s/dev", slavesdir, slavepath);
256 if (read_block_dev_entry(temppath, &majdev, &mindev)) {
257 perror("Error getting slave device numbers.");
263 * See if this maj,min already exists
265 slavedu = disk_util_exists(majdev, mindev);
269 sprintf(temppath, "%s/%s", slavesdir, slavepath);
270 __init_per_file_disk_util(td, majdev, mindev, temppath);
271 slavedu = disk_util_exists(majdev, mindev);
273 /* Should probably use an assert here. slavedu should
274 * always be present at this point. */
277 flist_add_tail(&slavedu->slavelist, &masterdu->slaves);
284 static struct disk_util *disk_util_add(struct thread_data *td, int majdev,
285 int mindev, char *path)
287 struct disk_util *du, *__du;
288 struct flist_head *entry;
291 dprint(FD_DISKUTIL, "add maj/min %d/%d: %s\n", majdev, mindev, path);
293 du = smalloc(sizeof(*du));
297 memset(du, 0, sizeof(*du));
298 INIT_FLIST_HEAD(&du->list);
299 l = snprintf(du->path, sizeof(du->path), "%s/stat", path);
300 if (l < 0 || l >= sizeof(du->path)) {
301 log_err("constructed path \"%.100s[...]/stat\" larger than buffer (%zu bytes)\n",
302 path, sizeof(du->path) - 1);
306 strncpy((char *) du->dus.name, basename(path), FIO_DU_NAME_SZ - 1);
307 du->sysfs_root = strdup(path);
310 INIT_FLIST_HEAD(&du->slavelist);
311 INIT_FLIST_HEAD(&du->slaves);
312 du->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
315 fio_mutex_down(disk_util_mutex);
317 flist_for_each(entry, &disk_list) {
318 __du = flist_entry(entry, struct disk_util, list);
320 dprint(FD_DISKUTIL, "found %s in list\n", __du->dus.name);
322 if (!strcmp((char *) du->dus.name, (char *) __du->dus.name)) {
324 fio_mutex_up(disk_util_mutex);
329 dprint(FD_DISKUTIL, "add %s to list\n", du->dus.name);
331 fio_gettime(&du->time, NULL);
332 get_io_ticks(du, &du->last_dus);
334 flist_add_tail(&du->list, &disk_list);
335 fio_mutex_up(disk_util_mutex);
337 find_add_disk_slaves(td, path, du);
341 static int check_dev_match(int majdev, int mindev, char *path)
345 if (read_block_dev_entry(path, &major, &minor))
348 if (majdev == major && mindev == minor)
354 static int find_block_dir(int majdev, int mindev, char *path, int link_ok)
365 while ((dir = readdir(D)) != NULL) {
368 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
371 sprintf(full_path, "%s/%s", path, dir->d_name);
373 if (!strcmp(dir->d_name, "dev")) {
374 if (!check_dev_match(majdev, mindev, full_path)) {
381 if (stat(full_path, &st) == -1) {
386 if (lstat(full_path, &st) == -1) {
392 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
395 found = find_block_dir(majdev, mindev, full_path, 0);
397 strcpy(path, full_path);
406 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
407 int majdev, int mindev,
415 * If there's a ../queue/ directory there, we are inside a partition.
416 * Check if that is the case and jump back. For loop/md/dm etc we
417 * are already in the right spot.
419 sprintf(tmp, "%s/../queue", path);
420 if (!stat(tmp, &st)) {
422 sprintf(tmp, "%s/queue", p);
423 if (stat(tmp, &st)) {
424 log_err("unknown sysfs layout\n");
427 tmp[PATH_MAX - 1] = '\0';
428 strncpy(tmp, p, PATH_MAX - 1);
429 sprintf(path, "%s", tmp);
432 return disk_util_add(td, majdev, mindev, path);
435 static struct disk_util *init_per_file_disk_util(struct thread_data *td,
440 struct disk_util *du;
443 if (get_device_numbers(filename, &majdev, &mindev))
446 dprint(FD_DISKUTIL, "%s belongs to maj/min %d/%d\n", filename, majdev,
449 du = disk_util_exists(majdev, mindev);
454 * for an fs without a device, we will repeatedly stat through
455 * sysfs which can take oodles of time for thousands of files. so
456 * cache the last lookup and compare with that before going through
459 if (mindev == last_mindev && majdev == last_majdev)
462 last_mindev = mindev;
463 last_majdev = majdev;
465 sprintf(foo, "/sys/block");
466 if (!find_block_dir(majdev, mindev, foo, 1))
469 return __init_per_file_disk_util(td, majdev, mindev, foo);
472 static struct disk_util *__init_disk_util(struct thread_data *td,
475 return init_per_file_disk_util(td, f->file_name);
478 void init_disk_util(struct thread_data *td)
483 if (!td->o.do_disk_util ||
484 td_ioengine_flagged(td, FIO_DISKLESSIO | FIO_NODISKUTIL))
487 for_each_file(td, f, i)
488 f->du = __init_disk_util(td, f);
491 static void show_agg_stats(struct disk_util_agg *agg, int terse,
492 struct buf_output *out)
494 if (!agg->slavecount)
498 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
499 "aggrticks=%llu/%llu, aggrin_queue=%llu, "
501 (unsigned long long) agg->ios[0] / agg->slavecount,
502 (unsigned long long) agg->ios[1] / agg->slavecount,
503 (unsigned long long) agg->merges[0] / agg->slavecount,
504 (unsigned long long) agg->merges[1] / agg->slavecount,
505 (unsigned long long) agg->ticks[0] / agg->slavecount,
506 (unsigned long long) agg->ticks[1] / agg->slavecount,
507 (unsigned long long) agg->time_in_queue / agg->slavecount,
510 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
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,
522 static void aggregate_slaves_stats(struct disk_util *masterdu)
524 struct disk_util_agg *agg = &masterdu->agg;
525 struct disk_util_stat *dus;
526 struct flist_head *entry;
527 struct disk_util *slavedu;
530 flist_for_each(entry, &masterdu->slaves) {
531 slavedu = flist_entry(entry, struct disk_util, slavelist);
533 agg->ios[0] += dus->s.ios[0];
534 agg->ios[1] += dus->s.ios[1];
535 agg->merges[0] += dus->s.merges[0];
536 agg->merges[1] += dus->s.merges[1];
537 agg->sectors[0] += dus->s.sectors[0];
538 agg->sectors[1] += dus->s.sectors[1];
539 agg->ticks[0] += dus->s.ticks[0];
540 agg->ticks[1] += dus->s.ticks[1];
541 agg->time_in_queue += dus->s.time_in_queue;
544 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
545 /* System utilization is the utilization of the
546 * component with the highest utilization.
548 if (util > agg->max_util.u.f)
549 agg->max_util.u.f = util;
553 if (agg->max_util.u.f > 100.0)
554 agg->max_util.u.f = 100.0;
557 void disk_util_prune_entries(void)
559 fio_mutex_down(disk_util_mutex);
561 while (!flist_empty(&disk_list)) {
562 struct disk_util *du;
564 du = flist_first_entry(&disk_list, struct disk_util, list);
565 flist_del(&du->list);
569 last_majdev = last_mindev = -1;
570 fio_mutex_up(disk_util_mutex);
571 fio_mutex_remove(disk_util_mutex);
574 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
575 int terse, struct buf_output *out)
580 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
588 log_buf(out, " %s: ios=%llu/%llu, merge=%llu/%llu, "
589 "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
591 (unsigned long long) dus->s.ios[0],
592 (unsigned long long) dus->s.ios[1],
593 (unsigned long long) dus->s.merges[0],
594 (unsigned long long) dus->s.merges[1],
595 (unsigned long long) dus->s.ticks[0],
596 (unsigned long long) dus->s.ticks[1],
597 (unsigned long long) dus->s.time_in_queue,
600 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
602 (unsigned long long) dus->s.ios[0],
603 (unsigned long long) dus->s.ios[1],
604 (unsigned long long) dus->s.merges[0],
605 (unsigned long long) dus->s.merges[1],
606 (unsigned long long) dus->s.ticks[0],
607 (unsigned long long) dus->s.ticks[1],
608 (unsigned long long) dus->s.time_in_queue,
613 * If the device has slaves, aggregate the stats for
614 * those slave devices also.
616 show_agg_stats(agg, terse, out);
622 void json_array_add_disk_util(struct disk_util_stat *dus,
623 struct disk_util_agg *agg, struct json_array *array)
625 struct json_object *obj;
629 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
633 obj = json_create_object();
634 json_array_add_value_object(array, obj);
636 json_object_add_value_string(obj, "name", dus->name);
637 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
638 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
639 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
640 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
641 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
642 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
643 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
644 json_object_add_value_float(obj, "util", util);
647 * If the device has slaves, aggregate the stats for
648 * those slave devices also.
650 if (!agg->slavecount)
652 json_object_add_value_int(obj, "aggr_read_ios",
653 agg->ios[0] / agg->slavecount);
654 json_object_add_value_int(obj, "aggr_write_ios",
655 agg->ios[1] / agg->slavecount);
656 json_object_add_value_int(obj, "aggr_read_merges",
657 agg->merges[0] / agg->slavecount);
658 json_object_add_value_int(obj, "aggr_write_merge",
659 agg->merges[1] / agg->slavecount);
660 json_object_add_value_int(obj, "aggr_read_ticks",
661 agg->ticks[0] / agg->slavecount);
662 json_object_add_value_int(obj, "aggr_write_ticks",
663 agg->ticks[1] / agg->slavecount);
664 json_object_add_value_int(obj, "aggr_in_queue",
665 agg->time_in_queue / agg->slavecount);
666 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
669 static void json_object_add_disk_utils(struct json_object *obj,
670 struct flist_head *head)
672 struct json_array *array = json_create_array();
673 struct flist_head *entry;
674 struct disk_util *du;
676 json_object_add_value_array(obj, "disk_util", array);
678 flist_for_each(entry, head) {
679 du = flist_entry(entry, struct disk_util, list);
681 aggregate_slaves_stats(du);
682 json_array_add_disk_util(&du->dus, &du->agg, array);
686 void show_disk_util(int terse, struct json_object *parent,
687 struct buf_output *out)
689 struct flist_head *entry;
690 struct disk_util *du;
693 if (!disk_util_mutex)
696 fio_mutex_down(disk_util_mutex);
698 if (flist_empty(&disk_list)) {
699 fio_mutex_up(disk_util_mutex);
703 if ((output_format & FIO_OUTPUT_JSON) && parent)
708 if (!terse && !do_json)
709 log_buf(out, "\nDisk stats (read/write):\n");
712 json_object_add_disk_utils(parent, &disk_list);
713 else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
714 flist_for_each(entry, &disk_list) {
715 du = flist_entry(entry, struct disk_util, list);
717 aggregate_slaves_stats(du);
718 print_disk_util(&du->dus, &du->agg, terse, out);
722 fio_mutex_up(disk_util_mutex);
725 void setup_disk_util(void)
727 disk_util_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);