1 // SPDX-License-Identifier: GPL-2.0
3 * Block device elevator/IO-scheduler.
5 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
7 * 30042000 Jens Axboe <axboe@kernel.dk> :
9 * Split the elevator a bit so that it is possible to choose a different
10 * one or even write a new "plug in". There are three pieces:
11 * - elevator_fn, inserts a new request in the queue list
12 * - elevator_merge_fn, decides whether a new buffer can be merged with
14 * - elevator_dequeue_fn, called when a request is taken off the active list
16 * 20082000 Dave Jones <davej@suse.de> :
17 * Removed tests for max-bomb-segments, which was breaking elvtune
18 * when run without -bN
21 * - Rework again to work with bio instead of buffer_heads
22 * - loose bi_dev comparisons, partition handling is right now
23 * - completely modularize elevator setup and teardown
26 #include <linux/kernel.h>
28 #include <linux/blkdev.h>
29 #include <linux/elevator.h>
30 #include <linux/bio.h>
31 #include <linux/module.h>
32 #include <linux/slab.h>
33 #include <linux/init.h>
34 #include <linux/compiler.h>
35 #include <linux/blktrace_api.h>
36 #include <linux/hash.h>
37 #include <linux/uaccess.h>
38 #include <linux/pm_runtime.h>
39 #include <linux/blk-cgroup.h>
41 #include <trace/events/block.h>
44 #include "blk-mq-sched.h"
48 static DEFINE_SPINLOCK(elv_list_lock
);
49 static LIST_HEAD(elv_list
);
54 #define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq))
57 * Query io scheduler to see if the current process issuing bio may be
60 static int elv_iosched_allow_bio_merge(struct request
*rq
, struct bio
*bio
)
62 struct request_queue
*q
= rq
->q
;
63 struct elevator_queue
*e
= q
->elevator
;
65 if (e
->type
->ops
.allow_merge
)
66 return e
->type
->ops
.allow_merge(q
, rq
, bio
);
72 * can we safely merge with this request?
74 bool elv_bio_merge_ok(struct request
*rq
, struct bio
*bio
)
76 if (!blk_rq_merge_ok(rq
, bio
))
79 if (!elv_iosched_allow_bio_merge(rq
, bio
))
84 EXPORT_SYMBOL(elv_bio_merge_ok
);
86 static bool elevator_match(const struct elevator_type
*e
, const char *name
)
88 if (!strcmp(e
->elevator_name
, name
))
90 if (e
->elevator_alias
&& !strcmp(e
->elevator_alias
, name
))
97 * Return scheduler with name 'name'
99 static struct elevator_type
*elevator_find(const char *name
)
101 struct elevator_type
*e
;
103 list_for_each_entry(e
, &elv_list
, list
) {
104 if (elevator_match(e
, name
))
111 static void elevator_put(struct elevator_type
*e
)
113 module_put(e
->elevator_owner
);
116 static struct elevator_type
*elevator_get(struct request_queue
*q
,
117 const char *name
, bool try_loading
)
119 struct elevator_type
*e
;
121 spin_lock(&elv_list_lock
);
123 e
= elevator_find(name
);
124 if (!e
&& try_loading
) {
125 spin_unlock(&elv_list_lock
);
126 request_module("%s-iosched", name
);
127 spin_lock(&elv_list_lock
);
128 e
= elevator_find(name
);
131 if (e
&& !try_module_get(e
->elevator_owner
))
134 spin_unlock(&elv_list_lock
);
138 static char chosen_elevator
[ELV_NAME_MAX
];
140 static int __init
elevator_setup(char *str
)
143 * Be backwards-compatible with previous kernels, so users
144 * won't get the wrong elevator.
146 strncpy(chosen_elevator
, str
, sizeof(chosen_elevator
) - 1);
150 __setup("elevator=", elevator_setup
);
152 static struct kobj_type elv_ktype
;
154 struct elevator_queue
*elevator_alloc(struct request_queue
*q
,
155 struct elevator_type
*e
)
157 struct elevator_queue
*eq
;
159 eq
= kzalloc_node(sizeof(*eq
), GFP_KERNEL
, q
->node
);
164 kobject_init(&eq
->kobj
, &elv_ktype
);
165 mutex_init(&eq
->sysfs_lock
);
170 EXPORT_SYMBOL(elevator_alloc
);
172 static void elevator_release(struct kobject
*kobj
)
174 struct elevator_queue
*e
;
176 e
= container_of(kobj
, struct elevator_queue
, kobj
);
177 elevator_put(e
->type
);
181 void __elevator_exit(struct request_queue
*q
, struct elevator_queue
*e
)
183 mutex_lock(&e
->sysfs_lock
);
184 if (e
->type
->ops
.exit_sched
)
185 blk_mq_exit_sched(q
, e
);
186 mutex_unlock(&e
->sysfs_lock
);
188 kobject_put(&e
->kobj
);
191 static inline void __elv_rqhash_del(struct request
*rq
)
194 rq
->rq_flags
&= ~RQF_HASHED
;
197 void elv_rqhash_del(struct request_queue
*q
, struct request
*rq
)
200 __elv_rqhash_del(rq
);
202 EXPORT_SYMBOL_GPL(elv_rqhash_del
);
204 void elv_rqhash_add(struct request_queue
*q
, struct request
*rq
)
206 struct elevator_queue
*e
= q
->elevator
;
208 BUG_ON(ELV_ON_HASH(rq
));
209 hash_add(e
->hash
, &rq
->hash
, rq_hash_key(rq
));
210 rq
->rq_flags
|= RQF_HASHED
;
212 EXPORT_SYMBOL_GPL(elv_rqhash_add
);
214 void elv_rqhash_reposition(struct request_queue
*q
, struct request
*rq
)
216 __elv_rqhash_del(rq
);
217 elv_rqhash_add(q
, rq
);
220 struct request
*elv_rqhash_find(struct request_queue
*q
, sector_t offset
)
222 struct elevator_queue
*e
= q
->elevator
;
223 struct hlist_node
*next
;
226 hash_for_each_possible_safe(e
->hash
, rq
, next
, hash
, offset
) {
227 BUG_ON(!ELV_ON_HASH(rq
));
229 if (unlikely(!rq_mergeable(rq
))) {
230 __elv_rqhash_del(rq
);
234 if (rq_hash_key(rq
) == offset
)
242 * RB-tree support functions for inserting/lookup/removal of requests
243 * in a sorted RB tree.
245 void elv_rb_add(struct rb_root
*root
, struct request
*rq
)
247 struct rb_node
**p
= &root
->rb_node
;
248 struct rb_node
*parent
= NULL
;
249 struct request
*__rq
;
253 __rq
= rb_entry(parent
, struct request
, rb_node
);
255 if (blk_rq_pos(rq
) < blk_rq_pos(__rq
))
257 else if (blk_rq_pos(rq
) >= blk_rq_pos(__rq
))
261 rb_link_node(&rq
->rb_node
, parent
, p
);
262 rb_insert_color(&rq
->rb_node
, root
);
264 EXPORT_SYMBOL(elv_rb_add
);
266 void elv_rb_del(struct rb_root
*root
, struct request
*rq
)
268 BUG_ON(RB_EMPTY_NODE(&rq
->rb_node
));
269 rb_erase(&rq
->rb_node
, root
);
270 RB_CLEAR_NODE(&rq
->rb_node
);
272 EXPORT_SYMBOL(elv_rb_del
);
274 struct request
*elv_rb_find(struct rb_root
*root
, sector_t sector
)
276 struct rb_node
*n
= root
->rb_node
;
280 rq
= rb_entry(n
, struct request
, rb_node
);
282 if (sector
< blk_rq_pos(rq
))
284 else if (sector
> blk_rq_pos(rq
))
292 EXPORT_SYMBOL(elv_rb_find
);
294 enum elv_merge
elv_merge(struct request_queue
*q
, struct request
**req
,
297 struct elevator_queue
*e
= q
->elevator
;
298 struct request
*__rq
;
302 * nomerges: No merges at all attempted
303 * noxmerges: Only simple one-hit cache try
304 * merges: All merge tries attempted
306 if (blk_queue_nomerges(q
) || !bio_mergeable(bio
))
307 return ELEVATOR_NO_MERGE
;
310 * First try one-hit cache.
312 if (q
->last_merge
&& elv_bio_merge_ok(q
->last_merge
, bio
)) {
313 enum elv_merge ret
= blk_try_merge(q
->last_merge
, bio
);
315 if (ret
!= ELEVATOR_NO_MERGE
) {
316 *req
= q
->last_merge
;
321 if (blk_queue_noxmerges(q
))
322 return ELEVATOR_NO_MERGE
;
325 * See if our hash lookup can find a potential backmerge.
327 __rq
= elv_rqhash_find(q
, bio
->bi_iter
.bi_sector
);
328 if (__rq
&& elv_bio_merge_ok(__rq
, bio
)) {
330 return ELEVATOR_BACK_MERGE
;
333 if (e
->type
->ops
.request_merge
)
334 return e
->type
->ops
.request_merge(q
, req
, bio
);
336 return ELEVATOR_NO_MERGE
;
340 * Attempt to do an insertion back merge. Only check for the case where
341 * we can append 'rq' to an existing request, so we can throw 'rq' away
344 * Returns true if we merged, false otherwise
346 bool elv_attempt_insert_merge(struct request_queue
*q
, struct request
*rq
)
348 struct request
*__rq
;
351 if (blk_queue_nomerges(q
))
355 * First try one-hit cache.
357 if (q
->last_merge
&& blk_attempt_req_merge(q
, q
->last_merge
, rq
))
360 if (blk_queue_noxmerges(q
))
365 * See if our hash lookup can find a potential backmerge.
368 __rq
= elv_rqhash_find(q
, blk_rq_pos(rq
));
369 if (!__rq
|| !blk_attempt_req_merge(q
, __rq
, rq
))
372 /* The merged request could be merged with others, try again */
380 void elv_merged_request(struct request_queue
*q
, struct request
*rq
,
383 struct elevator_queue
*e
= q
->elevator
;
385 if (e
->type
->ops
.request_merged
)
386 e
->type
->ops
.request_merged(q
, rq
, type
);
388 if (type
== ELEVATOR_BACK_MERGE
)
389 elv_rqhash_reposition(q
, rq
);
394 void elv_merge_requests(struct request_queue
*q
, struct request
*rq
,
395 struct request
*next
)
397 struct elevator_queue
*e
= q
->elevator
;
399 if (e
->type
->ops
.requests_merged
)
400 e
->type
->ops
.requests_merged(q
, rq
, next
);
402 elv_rqhash_reposition(q
, rq
);
406 struct request
*elv_latter_request(struct request_queue
*q
, struct request
*rq
)
408 struct elevator_queue
*e
= q
->elevator
;
410 if (e
->type
->ops
.next_request
)
411 return e
->type
->ops
.next_request(q
, rq
);
416 struct request
*elv_former_request(struct request_queue
*q
, struct request
*rq
)
418 struct elevator_queue
*e
= q
->elevator
;
420 if (e
->type
->ops
.former_request
)
421 return e
->type
->ops
.former_request(q
, rq
);
426 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
429 elv_attr_show(struct kobject
*kobj
, struct attribute
*attr
, char *page
)
431 struct elv_fs_entry
*entry
= to_elv(attr
);
432 struct elevator_queue
*e
;
438 e
= container_of(kobj
, struct elevator_queue
, kobj
);
439 mutex_lock(&e
->sysfs_lock
);
440 error
= e
->type
? entry
->show(e
, page
) : -ENOENT
;
441 mutex_unlock(&e
->sysfs_lock
);
446 elv_attr_store(struct kobject
*kobj
, struct attribute
*attr
,
447 const char *page
, size_t length
)
449 struct elv_fs_entry
*entry
= to_elv(attr
);
450 struct elevator_queue
*e
;
456 e
= container_of(kobj
, struct elevator_queue
, kobj
);
457 mutex_lock(&e
->sysfs_lock
);
458 error
= e
->type
? entry
->store(e
, page
, length
) : -ENOENT
;
459 mutex_unlock(&e
->sysfs_lock
);
463 static const struct sysfs_ops elv_sysfs_ops
= {
464 .show
= elv_attr_show
,
465 .store
= elv_attr_store
,
468 static struct kobj_type elv_ktype
= {
469 .sysfs_ops
= &elv_sysfs_ops
,
470 .release
= elevator_release
,
473 int elv_register_queue(struct request_queue
*q
)
475 struct elevator_queue
*e
= q
->elevator
;
478 lockdep_assert_held(&q
->sysfs_lock
);
480 error
= kobject_add(&e
->kobj
, &q
->kobj
, "%s", "iosched");
482 struct elv_fs_entry
*attr
= e
->type
->elevator_attrs
;
484 while (attr
->attr
.name
) {
485 if (sysfs_create_file(&e
->kobj
, &attr
->attr
))
490 kobject_uevent(&e
->kobj
, KOBJ_ADD
);
496 void elv_unregister_queue(struct request_queue
*q
)
498 lockdep_assert_held(&q
->sysfs_lock
);
501 struct elevator_queue
*e
= q
->elevator
;
503 kobject_uevent(&e
->kobj
, KOBJ_REMOVE
);
504 kobject_del(&e
->kobj
);
506 /* Re-enable throttling in case elevator disabled it */
507 wbt_enable_default(q
);
511 int elv_register(struct elevator_type
*e
)
513 /* create icq_cache if requested */
515 if (WARN_ON(e
->icq_size
< sizeof(struct io_cq
)) ||
516 WARN_ON(e
->icq_align
< __alignof__(struct io_cq
)))
519 snprintf(e
->icq_cache_name
, sizeof(e
->icq_cache_name
),
520 "%s_io_cq", e
->elevator_name
);
521 e
->icq_cache
= kmem_cache_create(e
->icq_cache_name
, e
->icq_size
,
522 e
->icq_align
, 0, NULL
);
527 /* register, don't allow duplicate names */
528 spin_lock(&elv_list_lock
);
529 if (elevator_find(e
->elevator_name
)) {
530 spin_unlock(&elv_list_lock
);
531 kmem_cache_destroy(e
->icq_cache
);
534 list_add_tail(&e
->list
, &elv_list
);
535 spin_unlock(&elv_list_lock
);
537 printk(KERN_INFO
"io scheduler %s registered\n", e
->elevator_name
);
541 EXPORT_SYMBOL_GPL(elv_register
);
543 void elv_unregister(struct elevator_type
*e
)
546 spin_lock(&elv_list_lock
);
547 list_del_init(&e
->list
);
548 spin_unlock(&elv_list_lock
);
551 * Destroy icq_cache if it exists. icq's are RCU managed. Make
552 * sure all RCU operations are complete before proceeding.
556 kmem_cache_destroy(e
->icq_cache
);
560 EXPORT_SYMBOL_GPL(elv_unregister
);
562 int elevator_switch_mq(struct request_queue
*q
,
563 struct elevator_type
*new_e
)
567 lockdep_assert_held(&q
->sysfs_lock
);
570 if (q
->elevator
->registered
)
571 elv_unregister_queue(q
);
573 elevator_exit(q
, q
->elevator
);
576 ret
= blk_mq_init_sched(q
, new_e
);
581 ret
= elv_register_queue(q
);
583 elevator_exit(q
, q
->elevator
);
589 blk_add_trace_msg(q
, "elv switch: %s", new_e
->elevator_name
);
591 blk_add_trace_msg(q
, "elv switch: none");
598 * For blk-mq devices, we default to using mq-deadline, if available, for single
599 * queue devices. If deadline isn't available OR we have multiple queues,
602 int elevator_init_mq(struct request_queue
*q
)
604 struct elevator_type
*e
;
607 if (q
->nr_hw_queues
!= 1)
611 * q->sysfs_lock must be held to provide mutual exclusion between
612 * elevator_switch() and here.
614 mutex_lock(&q
->sysfs_lock
);
615 if (unlikely(q
->elevator
))
618 e
= elevator_get(q
, "mq-deadline", false);
622 err
= blk_mq_init_sched(q
, e
);
626 mutex_unlock(&q
->sysfs_lock
);
632 * switch to new_e io scheduler. be careful not to introduce deadlocks -
633 * we don't free the old io scheduler, before we have allocated what we
634 * need for the new one. this way we have a chance of going back to the old
635 * one, if the new one fails init for some reason.
637 static int elevator_switch(struct request_queue
*q
, struct elevator_type
*new_e
)
641 lockdep_assert_held(&q
->sysfs_lock
);
643 blk_mq_freeze_queue(q
);
644 blk_mq_quiesce_queue(q
);
646 err
= elevator_switch_mq(q
, new_e
);
648 blk_mq_unquiesce_queue(q
);
649 blk_mq_unfreeze_queue(q
);
655 * Switch this queue to the given IO scheduler.
657 static int __elevator_change(struct request_queue
*q
, const char *name
)
659 char elevator_name
[ELV_NAME_MAX
];
660 struct elevator_type
*e
;
662 /* Make sure queue is not in the middle of being removed */
663 if (!test_bit(QUEUE_FLAG_REGISTERED
, &q
->queue_flags
))
667 * Special case for mq, turn off scheduling
669 if (!strncmp(name
, "none", 4)) {
672 return elevator_switch(q
, NULL
);
675 strlcpy(elevator_name
, name
, sizeof(elevator_name
));
676 e
= elevator_get(q
, strstrip(elevator_name
), true);
680 if (q
->elevator
&& elevator_match(q
->elevator
->type
, elevator_name
)) {
685 return elevator_switch(q
, e
);
688 static inline bool elv_support_iosched(struct request_queue
*q
)
690 if (q
->tag_set
&& (q
->tag_set
->flags
& BLK_MQ_F_NO_SCHED
))
695 ssize_t
elv_iosched_store(struct request_queue
*q
, const char *name
,
700 if (!queue_is_mq(q
) || !elv_support_iosched(q
))
703 ret
= __elevator_change(q
, name
);
710 ssize_t
elv_iosched_show(struct request_queue
*q
, char *name
)
712 struct elevator_queue
*e
= q
->elevator
;
713 struct elevator_type
*elv
= NULL
;
714 struct elevator_type
*__e
;
718 return sprintf(name
, "none\n");
721 len
+= sprintf(name
+len
, "[none] ");
725 spin_lock(&elv_list_lock
);
726 list_for_each_entry(__e
, &elv_list
, list
) {
727 if (elv
&& elevator_match(elv
, __e
->elevator_name
)) {
728 len
+= sprintf(name
+len
, "[%s] ", elv
->elevator_name
);
731 if (elv_support_iosched(q
))
732 len
+= sprintf(name
+len
, "%s ", __e
->elevator_name
);
734 spin_unlock(&elv_list_lock
);
737 len
+= sprintf(name
+len
, "none");
739 len
+= sprintf(len
+name
, "\n");
743 struct request
*elv_rb_former_request(struct request_queue
*q
,
746 struct rb_node
*rbprev
= rb_prev(&rq
->rb_node
);
749 return rb_entry_rq(rbprev
);
753 EXPORT_SYMBOL(elv_rb_former_request
);
755 struct request
*elv_rb_latter_request(struct request_queue
*q
,
758 struct rb_node
*rbnext
= rb_next(&rq
->rb_node
);
761 return rb_entry_rq(rbnext
);
765 EXPORT_SYMBOL(elv_rb_latter_request
);