6 #include <sys/sysmacros.h>
11 #ifdef CONFIG_VALGRIND_DEV
12 #include <valgrind/drd.h>
14 #define DRD_IGNORE_VAR(x) do { } while (0)
20 #include "helper_thread.h"
22 static int last_majdev, last_mindev;
23 static struct disk_util *last_du;
25 static struct fio_sem *disk_util_sem;
27 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
28 int majdev, int mindev, char *path);
30 static void disk_util_free(struct disk_util *du)
35 while (!flist_empty(&du->slaves)) {
36 struct disk_util *slave;
38 slave = flist_first_entry(&du->slaves, struct disk_util, slavelist);
39 flist_del(&slave->slavelist);
43 fio_sem_remove(du->lock);
48 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
51 unsigned long long sectors[2];
57 dprint(FD_DISKUTIL, "open stat file: %s\n", du->path);
59 f = fopen(du->path, "r");
63 p = fgets(line, sizeof(line), f);
69 dprint(FD_DISKUTIL, "%s: %s", du->path, p);
71 ret = sscanf(p, "%llu %llu %llu %llu %llu %llu %llu %llu %u %llu %llu\n",
72 (unsigned long long *) &dus->s.ios[0],
73 (unsigned long long *) &dus->s.merges[0],
75 (unsigned long long *) &dus->s.ticks[0],
76 (unsigned long long *) &dus->s.ios[1],
77 (unsigned long long *) &dus->s.merges[1],
79 (unsigned long long *) &dus->s.ticks[1],
81 (unsigned long long *) &dus->s.io_ticks,
82 (unsigned long long *) &dus->s.time_in_queue);
84 dprint(FD_DISKUTIL, "%s: stat read ok? %d\n", du->path, ret == 1);
85 dus->s.sectors[0] = sectors[0];
86 dus->s.sectors[1] = sectors[1];
90 static void update_io_tick_disk(struct disk_util *du)
92 struct disk_util_stat __dus, *dus, *ldus;
97 if (get_io_ticks(du, &__dus))
101 ldus = &du->last_dus;
103 dus->s.sectors[0] += (__dus.s.sectors[0] - ldus->s.sectors[0]);
104 dus->s.sectors[1] += (__dus.s.sectors[1] - ldus->s.sectors[1]);
105 dus->s.ios[0] += (__dus.s.ios[0] - ldus->s.ios[0]);
106 dus->s.ios[1] += (__dus.s.ios[1] - ldus->s.ios[1]);
107 dus->s.merges[0] += (__dus.s.merges[0] - ldus->s.merges[0]);
108 dus->s.merges[1] += (__dus.s.merges[1] - ldus->s.merges[1]);
109 dus->s.ticks[0] += (__dus.s.ticks[0] - ldus->s.ticks[0]);
110 dus->s.ticks[1] += (__dus.s.ticks[1] - ldus->s.ticks[1]);
111 dus->s.io_ticks += (__dus.s.io_ticks - ldus->s.io_ticks);
112 dus->s.time_in_queue += (__dus.s.time_in_queue - ldus->s.time_in_queue);
114 fio_gettime(&t, NULL);
115 dus->s.msec += mtime_since(&du->time, &t);
116 memcpy(&du->time, &t, sizeof(t));
117 memcpy(&ldus->s, &__dus.s, sizeof(__dus.s));
120 int update_io_ticks(void)
122 struct flist_head *entry;
123 struct disk_util *du;
126 dprint(FD_DISKUTIL, "update io ticks\n");
128 fio_sem_down(disk_util_sem);
130 if (!helper_should_exit()) {
131 flist_for_each(entry, &disk_list) {
132 du = flist_entry(entry, struct disk_util, list);
133 update_io_tick_disk(du);
138 fio_sem_up(disk_util_sem);
142 static struct disk_util *disk_util_exists(int major, int minor)
144 struct flist_head *entry;
145 struct disk_util *du;
147 fio_sem_down(disk_util_sem);
149 flist_for_each(entry, &disk_list) {
150 du = flist_entry(entry, struct disk_util, list);
152 if (major == du->major && minor == du->minor) {
153 fio_sem_up(disk_util_sem);
158 fio_sem_up(disk_util_sem);
162 static int get_device_numbers(char *file_name, int *maj, int *min)
166 char tempname[PATH_MAX], *p;
168 if (!lstat(file_name, &st)) {
169 if (S_ISBLK(st.st_mode)) {
170 majdev = major(st.st_rdev);
171 mindev = minor(st.st_rdev);
172 } else if (S_ISCHR(st.st_mode)) {
173 majdev = major(st.st_rdev);
174 mindev = minor(st.st_rdev);
175 if (fio_lookup_raw(st.st_rdev, &majdev, &mindev))
177 } else if (S_ISFIFO(st.st_mode))
180 majdev = major(st.st_dev);
181 mindev = minor(st.st_dev);
185 * must be a file, open "." in that path
187 tempname[PATH_MAX - 1] = '\0';
188 strncpy(tempname, file_name, PATH_MAX - 1);
189 p = dirname(tempname);
191 perror("disk util stat");
195 majdev = major(st.st_dev);
196 mindev = minor(st.st_dev);
205 static int read_block_dev_entry(char *path, int *maj, int *min)
210 f = fopen(path, "r");
216 p = fgets(line, sizeof(line), f);
222 if (sscanf(p, "%u:%u", maj, min) != 2)
228 static void find_add_disk_slaves(struct thread_data *td, char *path,
229 struct disk_util *masterdu)
231 DIR *dirhandle = NULL;
232 struct dirent *dirent = NULL;
233 char slavesdir[PATH_MAX], temppath[PATH_MAX], slavepath[PATH_MAX];
234 struct disk_util *slavedu = NULL;
238 sprintf(slavesdir, "%s/%s", path, "slaves");
239 dirhandle = opendir(slavesdir);
243 while ((dirent = readdir(dirhandle)) != NULL) {
244 if (!strcmp(dirent->d_name, ".") ||
245 !strcmp(dirent->d_name, ".."))
248 sprintf(temppath, "%s/%s", slavesdir, dirent->d_name);
249 /* Can we always assume that the slaves device entries
250 * are links to the real directories for the slave
253 linklen = readlink(temppath, slavepath, PATH_MAX - 1);
255 perror("readlink() for slave device.");
259 slavepath[linklen] = '\0';
261 sprintf(temppath, "%s/%s/dev", slavesdir, slavepath);
262 if (access(temppath, F_OK) != 0)
263 sprintf(temppath, "%s/%s/device/dev", slavesdir, slavepath);
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 sprintf(temppath, "%s/%s", slavesdir, slavepath);
278 __init_per_file_disk_util(td, majdev, mindev, temppath);
279 slavedu = disk_util_exists(majdev, mindev);
281 /* Should probably use an assert here. slavedu should
282 * always be present at this point. */
285 flist_add_tail(&slavedu->slavelist, &masterdu->slaves);
292 static struct disk_util *disk_util_add(struct thread_data *td, int majdev,
293 int mindev, char *path)
295 struct disk_util *du, *__du;
296 struct flist_head *entry;
299 dprint(FD_DISKUTIL, "add maj/min %d/%d: %s\n", majdev, mindev, path);
301 du = smalloc(sizeof(*du));
305 DRD_IGNORE_VAR(du->users);
306 memset(du, 0, sizeof(*du));
307 INIT_FLIST_HEAD(&du->list);
308 l = snprintf(du->path, sizeof(du->path), "%s/stat", path);
309 if (l < 0 || l >= sizeof(du->path)) {
310 log_err("constructed path \"%.100s[...]/stat\" larger than buffer (%zu bytes)\n",
311 path, sizeof(du->path) - 1);
315 strncpy((char *) du->dus.name, basename(path), FIO_DU_NAME_SZ - 1);
316 du->sysfs_root = strdup(path);
319 INIT_FLIST_HEAD(&du->slavelist);
320 INIT_FLIST_HEAD(&du->slaves);
321 du->lock = fio_sem_init(FIO_SEM_UNLOCKED);
324 fio_sem_down(disk_util_sem);
326 flist_for_each(entry, &disk_list) {
327 __du = flist_entry(entry, struct disk_util, list);
329 dprint(FD_DISKUTIL, "found %s in list\n", __du->dus.name);
331 if (!strcmp((char *) du->dus.name, (char *) __du->dus.name)) {
333 fio_sem_up(disk_util_sem);
338 dprint(FD_DISKUTIL, "add %s to list\n", du->dus.name);
340 fio_gettime(&du->time, NULL);
341 get_io_ticks(du, &du->last_dus);
343 flist_add_tail(&du->list, &disk_list);
344 fio_sem_up(disk_util_sem);
346 find_add_disk_slaves(td, path, du);
350 static int check_dev_match(int majdev, int mindev, char *path)
354 if (read_block_dev_entry(path, &major, &minor))
357 if (majdev == major && mindev == minor)
363 static int find_block_dir(int majdev, int mindev, char *path, int link_ok)
374 while ((dir = readdir(D)) != NULL) {
377 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
380 sprintf(full_path, "%s/%s", path, dir->d_name);
382 if (!strcmp(dir->d_name, "dev")) {
383 if (!check_dev_match(majdev, mindev, full_path)) {
390 if (stat(full_path, &st) == -1) {
395 if (lstat(full_path, &st) == -1) {
401 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
404 found = find_block_dir(majdev, mindev, full_path, 0);
406 strcpy(path, full_path);
415 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
416 int majdev, int mindev,
424 * If there's a ../queue/ directory there, we are inside a partition.
425 * Check if that is the case and jump back. For loop/md/dm etc we
426 * are already in the right spot.
428 sprintf(tmp, "%s/../queue", path);
429 if (!stat(tmp, &st)) {
431 sprintf(tmp, "%s/queue", p);
432 if (stat(tmp, &st)) {
433 log_err("unknown sysfs layout\n");
436 tmp[PATH_MAX - 1] = '\0';
437 strncpy(tmp, p, PATH_MAX - 1);
438 sprintf(path, "%s", tmp);
441 return disk_util_add(td, majdev, mindev, path);
444 static struct disk_util *init_per_file_disk_util(struct thread_data *td,
449 struct disk_util *du;
452 if (get_device_numbers(filename, &majdev, &mindev))
455 dprint(FD_DISKUTIL, "%s belongs to maj/min %d/%d\n", filename, majdev,
458 du = disk_util_exists(majdev, mindev);
463 * for an fs without a device, we will repeatedly stat through
464 * sysfs which can take oodles of time for thousands of files. so
465 * cache the last lookup and compare with that before going through
468 if (mindev == last_mindev && majdev == last_majdev)
471 last_mindev = mindev;
472 last_majdev = majdev;
474 sprintf(foo, "/sys/block");
475 if (!find_block_dir(majdev, mindev, foo, 1))
478 return __init_per_file_disk_util(td, majdev, mindev, foo);
481 static struct disk_util *__init_disk_util(struct thread_data *td,
484 return init_per_file_disk_util(td, f->file_name);
487 void init_disk_util(struct thread_data *td)
492 if (!td->o.do_disk_util ||
493 td_ioengine_flagged(td, FIO_DISKLESSIO | FIO_NODISKUTIL))
496 for_each_file(td, f, i)
497 f->du = __init_disk_util(td, f);
500 static void show_agg_stats(struct disk_util_agg *agg, int terse,
501 struct buf_output *out)
503 if (!agg->slavecount)
507 log_buf(out, ", aggrios=%llu/%llu, aggrmerge=%llu/%llu, "
508 "aggrticks=%llu/%llu, aggrin_queue=%llu, "
510 (unsigned long long) agg->ios[0] / agg->slavecount,
511 (unsigned long long) agg->ios[1] / agg->slavecount,
512 (unsigned long long) agg->merges[0] / agg->slavecount,
513 (unsigned long long) agg->merges[1] / agg->slavecount,
514 (unsigned long long) agg->ticks[0] / agg->slavecount,
515 (unsigned long long) agg->ticks[1] / agg->slavecount,
516 (unsigned long long) agg->time_in_queue / agg->slavecount,
519 log_buf(out, ";slaves;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
520 (unsigned long long) agg->ios[0] / agg->slavecount,
521 (unsigned long long) agg->ios[1] / agg->slavecount,
522 (unsigned long long) agg->merges[0] / agg->slavecount,
523 (unsigned long long) agg->merges[1] / agg->slavecount,
524 (unsigned long long) agg->ticks[0] / agg->slavecount,
525 (unsigned long long) agg->ticks[1] / agg->slavecount,
526 (unsigned long long) agg->time_in_queue / agg->slavecount,
531 static void aggregate_slaves_stats(struct disk_util *masterdu)
533 struct disk_util_agg *agg = &masterdu->agg;
534 struct disk_util_stat *dus;
535 struct flist_head *entry;
536 struct disk_util *slavedu;
539 flist_for_each(entry, &masterdu->slaves) {
540 slavedu = flist_entry(entry, struct disk_util, slavelist);
542 agg->ios[0] += dus->s.ios[0];
543 agg->ios[1] += dus->s.ios[1];
544 agg->merges[0] += dus->s.merges[0];
545 agg->merges[1] += dus->s.merges[1];
546 agg->sectors[0] += dus->s.sectors[0];
547 agg->sectors[1] += dus->s.sectors[1];
548 agg->ticks[0] += dus->s.ticks[0];
549 agg->ticks[1] += dus->s.ticks[1];
550 agg->time_in_queue += dus->s.time_in_queue;
553 util = (double) (100 * dus->s.io_ticks / (double) slavedu->dus.s.msec);
554 /* System utilization is the utilization of the
555 * component with the highest utilization.
557 if (util > agg->max_util.u.f)
558 agg->max_util.u.f = util;
562 if (agg->max_util.u.f > 100.0)
563 agg->max_util.u.f = 100.0;
566 void disk_util_prune_entries(void)
568 fio_sem_down(disk_util_sem);
570 while (!flist_empty(&disk_list)) {
571 struct disk_util *du;
573 du = flist_first_entry(&disk_list, struct disk_util, list);
574 flist_del(&du->list);
578 last_majdev = last_mindev = -1;
579 fio_sem_up(disk_util_sem);
580 fio_sem_remove(disk_util_sem);
583 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
584 int terse, struct buf_output *out)
589 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
597 log_buf(out, " %s: ios=%llu/%llu, merge=%llu/%llu, "
598 "ticks=%llu/%llu, in_queue=%llu, util=%3.2f%%",
600 (unsigned long long) dus->s.ios[0],
601 (unsigned long long) dus->s.ios[1],
602 (unsigned long long) dus->s.merges[0],
603 (unsigned long long) dus->s.merges[1],
604 (unsigned long long) dus->s.ticks[0],
605 (unsigned long long) dus->s.ticks[1],
606 (unsigned long long) dus->s.time_in_queue,
609 log_buf(out, ";%s;%llu;%llu;%llu;%llu;%llu;%llu;%llu;%3.2f%%",
611 (unsigned long long) dus->s.ios[0],
612 (unsigned long long) dus->s.ios[1],
613 (unsigned long long) dus->s.merges[0],
614 (unsigned long long) dus->s.merges[1],
615 (unsigned long long) dus->s.ticks[0],
616 (unsigned long long) dus->s.ticks[1],
617 (unsigned long long) dus->s.time_in_queue,
622 * If the device has slaves, aggregate the stats for
623 * those slave devices also.
625 show_agg_stats(agg, terse, out);
631 void json_array_add_disk_util(struct disk_util_stat *dus,
632 struct disk_util_agg *agg, struct json_array *array)
634 struct json_object *obj;
638 util = (double) 100 * dus->s.io_ticks / (double) dus->s.msec;
642 obj = json_create_object();
643 json_array_add_value_object(array, obj);
645 json_object_add_value_string(obj, "name", dus->name);
646 json_object_add_value_int(obj, "read_ios", dus->s.ios[0]);
647 json_object_add_value_int(obj, "write_ios", dus->s.ios[1]);
648 json_object_add_value_int(obj, "read_merges", dus->s.merges[0]);
649 json_object_add_value_int(obj, "write_merges", dus->s.merges[1]);
650 json_object_add_value_int(obj, "read_ticks", dus->s.ticks[0]);
651 json_object_add_value_int(obj, "write_ticks", dus->s.ticks[1]);
652 json_object_add_value_int(obj, "in_queue", dus->s.time_in_queue);
653 json_object_add_value_float(obj, "util", util);
656 * If the device has slaves, aggregate the stats for
657 * those slave devices also.
659 if (!agg->slavecount)
661 json_object_add_value_int(obj, "aggr_read_ios",
662 agg->ios[0] / agg->slavecount);
663 json_object_add_value_int(obj, "aggr_write_ios",
664 agg->ios[1] / agg->slavecount);
665 json_object_add_value_int(obj, "aggr_read_merges",
666 agg->merges[0] / agg->slavecount);
667 json_object_add_value_int(obj, "aggr_write_merge",
668 agg->merges[1] / agg->slavecount);
669 json_object_add_value_int(obj, "aggr_read_ticks",
670 agg->ticks[0] / agg->slavecount);
671 json_object_add_value_int(obj, "aggr_write_ticks",
672 agg->ticks[1] / agg->slavecount);
673 json_object_add_value_int(obj, "aggr_in_queue",
674 agg->time_in_queue / agg->slavecount);
675 json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
678 static void json_object_add_disk_utils(struct json_object *obj,
679 struct flist_head *head)
681 struct json_array *array = json_create_array();
682 struct flist_head *entry;
683 struct disk_util *du;
685 json_object_add_value_array(obj, "disk_util", array);
687 flist_for_each(entry, head) {
688 du = flist_entry(entry, struct disk_util, list);
690 aggregate_slaves_stats(du);
691 json_array_add_disk_util(&du->dus, &du->agg, array);
695 void show_disk_util(int terse, struct json_object *parent,
696 struct buf_output *out)
698 struct flist_head *entry;
699 struct disk_util *du;
705 fio_sem_down(disk_util_sem);
707 if (flist_empty(&disk_list)) {
708 fio_sem_up(disk_util_sem);
712 if ((output_format & FIO_OUTPUT_JSON) && parent)
717 if (!terse && !do_json)
718 log_buf(out, "\nDisk stats (read/write):\n");
721 json_object_add_disk_utils(parent, &disk_list);
722 else if (output_format & ~(FIO_OUTPUT_JSON | FIO_OUTPUT_JSON_PLUS)) {
723 flist_for_each(entry, &disk_list) {
724 du = flist_entry(entry, struct disk_util, list);
726 aggregate_slaves_stats(du);
727 print_disk_util(&du->dus, &du->agg, terse, out);
731 fio_sem_up(disk_util_sem);
734 void setup_disk_util(void)
736 disk_util_sem = fio_sem_init(FIO_SEM_UNLOCKED);