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 inline bool elv_support_features(unsigned int elv_features
,
87 unsigned int required_features
)
89 return (required_features
& elv_features
) == required_features
;
93 * elevator_match - Test an elevator name and features
94 * @e: Scheduler to test
95 * @name: Elevator name to test
96 * @required_features: Features that the elevator must provide
98 * Return true is the elevator @e name matches @name and if @e provides all the
99 * the feratures spcified by @required_features.
101 static bool elevator_match(const struct elevator_type
*e
, const char *name
,
102 unsigned int required_features
)
104 if (!elv_support_features(e
->elevator_features
, required_features
))
106 if (!strcmp(e
->elevator_name
, name
))
108 if (e
->elevator_alias
&& !strcmp(e
->elevator_alias
, name
))
115 * elevator_find - Find an elevator
116 * @name: Name of the elevator to find
117 * @required_features: Features that the elevator must provide
119 * Return the first registered scheduler with name @name and supporting the
120 * features @required_features and NULL otherwise.
122 static struct elevator_type
*elevator_find(const char *name
,
123 unsigned int required_features
)
125 struct elevator_type
*e
;
127 list_for_each_entry(e
, &elv_list
, list
) {
128 if (elevator_match(e
, name
, required_features
))
135 static void elevator_put(struct elevator_type
*e
)
137 module_put(e
->elevator_owner
);
140 static struct elevator_type
*elevator_get(struct request_queue
*q
,
141 const char *name
, bool try_loading
)
143 struct elevator_type
*e
;
145 spin_lock(&elv_list_lock
);
147 e
= elevator_find(name
, q
->required_elevator_features
);
148 if (!e
&& try_loading
) {
149 spin_unlock(&elv_list_lock
);
150 request_module("%s-iosched", name
);
151 spin_lock(&elv_list_lock
);
152 e
= elevator_find(name
, q
->required_elevator_features
);
155 if (e
&& !try_module_get(e
->elevator_owner
))
158 spin_unlock(&elv_list_lock
);
162 static struct kobj_type elv_ktype
;
164 struct elevator_queue
*elevator_alloc(struct request_queue
*q
,
165 struct elevator_type
*e
)
167 struct elevator_queue
*eq
;
169 eq
= kzalloc_node(sizeof(*eq
), GFP_KERNEL
, q
->node
);
174 kobject_init(&eq
->kobj
, &elv_ktype
);
175 mutex_init(&eq
->sysfs_lock
);
180 EXPORT_SYMBOL(elevator_alloc
);
182 static void elevator_release(struct kobject
*kobj
)
184 struct elevator_queue
*e
;
186 e
= container_of(kobj
, struct elevator_queue
, kobj
);
187 elevator_put(e
->type
);
191 void __elevator_exit(struct request_queue
*q
, struct elevator_queue
*e
)
193 mutex_lock(&e
->sysfs_lock
);
194 if (e
->type
->ops
.exit_sched
)
195 blk_mq_exit_sched(q
, e
);
196 mutex_unlock(&e
->sysfs_lock
);
198 kobject_put(&e
->kobj
);
201 static inline void __elv_rqhash_del(struct request
*rq
)
204 rq
->rq_flags
&= ~RQF_HASHED
;
207 void elv_rqhash_del(struct request_queue
*q
, struct request
*rq
)
210 __elv_rqhash_del(rq
);
212 EXPORT_SYMBOL_GPL(elv_rqhash_del
);
214 void elv_rqhash_add(struct request_queue
*q
, struct request
*rq
)
216 struct elevator_queue
*e
= q
->elevator
;
218 BUG_ON(ELV_ON_HASH(rq
));
219 hash_add(e
->hash
, &rq
->hash
, rq_hash_key(rq
));
220 rq
->rq_flags
|= RQF_HASHED
;
222 EXPORT_SYMBOL_GPL(elv_rqhash_add
);
224 void elv_rqhash_reposition(struct request_queue
*q
, struct request
*rq
)
226 __elv_rqhash_del(rq
);
227 elv_rqhash_add(q
, rq
);
230 struct request
*elv_rqhash_find(struct request_queue
*q
, sector_t offset
)
232 struct elevator_queue
*e
= q
->elevator
;
233 struct hlist_node
*next
;
236 hash_for_each_possible_safe(e
->hash
, rq
, next
, hash
, offset
) {
237 BUG_ON(!ELV_ON_HASH(rq
));
239 if (unlikely(!rq_mergeable(rq
))) {
240 __elv_rqhash_del(rq
);
244 if (rq_hash_key(rq
) == offset
)
252 * RB-tree support functions for inserting/lookup/removal of requests
253 * in a sorted RB tree.
255 void elv_rb_add(struct rb_root
*root
, struct request
*rq
)
257 struct rb_node
**p
= &root
->rb_node
;
258 struct rb_node
*parent
= NULL
;
259 struct request
*__rq
;
263 __rq
= rb_entry(parent
, struct request
, rb_node
);
265 if (blk_rq_pos(rq
) < blk_rq_pos(__rq
))
267 else if (blk_rq_pos(rq
) >= blk_rq_pos(__rq
))
271 rb_link_node(&rq
->rb_node
, parent
, p
);
272 rb_insert_color(&rq
->rb_node
, root
);
274 EXPORT_SYMBOL(elv_rb_add
);
276 void elv_rb_del(struct rb_root
*root
, struct request
*rq
)
278 BUG_ON(RB_EMPTY_NODE(&rq
->rb_node
));
279 rb_erase(&rq
->rb_node
, root
);
280 RB_CLEAR_NODE(&rq
->rb_node
);
282 EXPORT_SYMBOL(elv_rb_del
);
284 struct request
*elv_rb_find(struct rb_root
*root
, sector_t sector
)
286 struct rb_node
*n
= root
->rb_node
;
290 rq
= rb_entry(n
, struct request
, rb_node
);
292 if (sector
< blk_rq_pos(rq
))
294 else if (sector
> blk_rq_pos(rq
))
302 EXPORT_SYMBOL(elv_rb_find
);
304 enum elv_merge
elv_merge(struct request_queue
*q
, struct request
**req
,
307 struct elevator_queue
*e
= q
->elevator
;
308 struct request
*__rq
;
312 * nomerges: No merges at all attempted
313 * noxmerges: Only simple one-hit cache try
314 * merges: All merge tries attempted
316 if (blk_queue_nomerges(q
) || !bio_mergeable(bio
))
317 return ELEVATOR_NO_MERGE
;
320 * First try one-hit cache.
322 if (q
->last_merge
&& elv_bio_merge_ok(q
->last_merge
, bio
)) {
323 enum elv_merge ret
= blk_try_merge(q
->last_merge
, bio
);
325 if (ret
!= ELEVATOR_NO_MERGE
) {
326 *req
= q
->last_merge
;
331 if (blk_queue_noxmerges(q
))
332 return ELEVATOR_NO_MERGE
;
335 * See if our hash lookup can find a potential backmerge.
337 __rq
= elv_rqhash_find(q
, bio
->bi_iter
.bi_sector
);
338 if (__rq
&& elv_bio_merge_ok(__rq
, bio
)) {
340 return ELEVATOR_BACK_MERGE
;
343 if (e
->type
->ops
.request_merge
)
344 return e
->type
->ops
.request_merge(q
, req
, bio
);
346 return ELEVATOR_NO_MERGE
;
350 * Attempt to do an insertion back merge. Only check for the case where
351 * we can append 'rq' to an existing request, so we can throw 'rq' away
354 * Returns true if we merged, false otherwise
356 bool elv_attempt_insert_merge(struct request_queue
*q
, struct request
*rq
)
358 struct request
*__rq
;
361 if (blk_queue_nomerges(q
))
365 * First try one-hit cache.
367 if (q
->last_merge
&& blk_attempt_req_merge(q
, q
->last_merge
, rq
))
370 if (blk_queue_noxmerges(q
))
375 * See if our hash lookup can find a potential backmerge.
378 __rq
= elv_rqhash_find(q
, blk_rq_pos(rq
));
379 if (!__rq
|| !blk_attempt_req_merge(q
, __rq
, rq
))
382 /* The merged request could be merged with others, try again */
390 void elv_merged_request(struct request_queue
*q
, struct request
*rq
,
393 struct elevator_queue
*e
= q
->elevator
;
395 if (e
->type
->ops
.request_merged
)
396 e
->type
->ops
.request_merged(q
, rq
, type
);
398 if (type
== ELEVATOR_BACK_MERGE
)
399 elv_rqhash_reposition(q
, rq
);
404 void elv_merge_requests(struct request_queue
*q
, struct request
*rq
,
405 struct request
*next
)
407 struct elevator_queue
*e
= q
->elevator
;
409 if (e
->type
->ops
.requests_merged
)
410 e
->type
->ops
.requests_merged(q
, rq
, next
);
412 elv_rqhash_reposition(q
, rq
);
416 struct request
*elv_latter_request(struct request_queue
*q
, struct request
*rq
)
418 struct elevator_queue
*e
= q
->elevator
;
420 if (e
->type
->ops
.next_request
)
421 return e
->type
->ops
.next_request(q
, rq
);
426 struct request
*elv_former_request(struct request_queue
*q
, struct request
*rq
)
428 struct elevator_queue
*e
= q
->elevator
;
430 if (e
->type
->ops
.former_request
)
431 return e
->type
->ops
.former_request(q
, rq
);
436 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
439 elv_attr_show(struct kobject
*kobj
, struct attribute
*attr
, char *page
)
441 struct elv_fs_entry
*entry
= to_elv(attr
);
442 struct elevator_queue
*e
;
448 e
= container_of(kobj
, struct elevator_queue
, kobj
);
449 mutex_lock(&e
->sysfs_lock
);
450 error
= e
->type
? entry
->show(e
, page
) : -ENOENT
;
451 mutex_unlock(&e
->sysfs_lock
);
456 elv_attr_store(struct kobject
*kobj
, struct attribute
*attr
,
457 const char *page
, size_t length
)
459 struct elv_fs_entry
*entry
= to_elv(attr
);
460 struct elevator_queue
*e
;
466 e
= container_of(kobj
, struct elevator_queue
, kobj
);
467 mutex_lock(&e
->sysfs_lock
);
468 error
= e
->type
? entry
->store(e
, page
, length
) : -ENOENT
;
469 mutex_unlock(&e
->sysfs_lock
);
473 static const struct sysfs_ops elv_sysfs_ops
= {
474 .show
= elv_attr_show
,
475 .store
= elv_attr_store
,
478 static struct kobj_type elv_ktype
= {
479 .sysfs_ops
= &elv_sysfs_ops
,
480 .release
= elevator_release
,
484 * elv_register_queue is called from either blk_register_queue or
485 * elevator_switch, elevator switch is prevented from being happen
486 * in the two paths, so it is safe to not hold q->sysfs_lock.
488 int elv_register_queue(struct request_queue
*q
, bool uevent
)
490 struct elevator_queue
*e
= q
->elevator
;
493 error
= kobject_add(&e
->kobj
, &q
->kobj
, "%s", "iosched");
495 struct elv_fs_entry
*attr
= e
->type
->elevator_attrs
;
497 while (attr
->attr
.name
) {
498 if (sysfs_create_file(&e
->kobj
, &attr
->attr
))
504 kobject_uevent(&e
->kobj
, KOBJ_ADD
);
512 * elv_unregister_queue is called from either blk_unregister_queue or
513 * elevator_switch, elevator switch is prevented from being happen
514 * in the two paths, so it is safe to not hold q->sysfs_lock.
516 void elv_unregister_queue(struct request_queue
*q
)
519 struct elevator_queue
*e
= q
->elevator
;
521 kobject_uevent(&e
->kobj
, KOBJ_REMOVE
);
522 kobject_del(&e
->kobj
);
525 /* Re-enable throttling in case elevator disabled it */
526 wbt_enable_default(q
);
530 int elv_register(struct elevator_type
*e
)
532 /* create icq_cache if requested */
534 if (WARN_ON(e
->icq_size
< sizeof(struct io_cq
)) ||
535 WARN_ON(e
->icq_align
< __alignof__(struct io_cq
)))
538 snprintf(e
->icq_cache_name
, sizeof(e
->icq_cache_name
),
539 "%s_io_cq", e
->elevator_name
);
540 e
->icq_cache
= kmem_cache_create(e
->icq_cache_name
, e
->icq_size
,
541 e
->icq_align
, 0, NULL
);
546 /* register, don't allow duplicate names */
547 spin_lock(&elv_list_lock
);
548 if (elevator_find(e
->elevator_name
, 0)) {
549 spin_unlock(&elv_list_lock
);
550 kmem_cache_destroy(e
->icq_cache
);
553 list_add_tail(&e
->list
, &elv_list
);
554 spin_unlock(&elv_list_lock
);
556 printk(KERN_INFO
"io scheduler %s registered\n", e
->elevator_name
);
560 EXPORT_SYMBOL_GPL(elv_register
);
562 void elv_unregister(struct elevator_type
*e
)
565 spin_lock(&elv_list_lock
);
566 list_del_init(&e
->list
);
567 spin_unlock(&elv_list_lock
);
570 * Destroy icq_cache if it exists. icq's are RCU managed. Make
571 * sure all RCU operations are complete before proceeding.
575 kmem_cache_destroy(e
->icq_cache
);
579 EXPORT_SYMBOL_GPL(elv_unregister
);
581 int elevator_switch_mq(struct request_queue
*q
,
582 struct elevator_type
*new_e
)
586 lockdep_assert_held(&q
->sysfs_lock
);
589 if (q
->elevator
->registered
)
590 elv_unregister_queue(q
);
593 elevator_exit(q
, q
->elevator
);
596 ret
= blk_mq_init_sched(q
, new_e
);
601 ret
= elv_register_queue(q
, true);
603 elevator_exit(q
, q
->elevator
);
609 blk_add_trace_msg(q
, "elv switch: %s", new_e
->elevator_name
);
611 blk_add_trace_msg(q
, "elv switch: none");
617 static inline bool elv_support_iosched(struct request_queue
*q
)
620 (q
->tag_set
&& (q
->tag_set
->flags
& BLK_MQ_F_NO_SCHED
)))
626 * For single queue devices, default to using mq-deadline. If we have multiple
627 * queues or mq-deadline is not available, default to "none".
629 static struct elevator_type
*elevator_get_default(struct request_queue
*q
)
631 if (q
->nr_hw_queues
!= 1)
634 return elevator_get(q
, "mq-deadline", false);
638 * Get the first elevator providing the features required by the request queue.
639 * Default to "none" if no matching elevator is found.
641 static struct elevator_type
*elevator_get_by_features(struct request_queue
*q
)
643 struct elevator_type
*e
, *found
= NULL
;
645 spin_lock(&elv_list_lock
);
647 list_for_each_entry(e
, &elv_list
, list
) {
648 if (elv_support_features(e
->elevator_features
,
649 q
->required_elevator_features
)) {
655 if (found
&& !try_module_get(found
->elevator_owner
))
658 spin_unlock(&elv_list_lock
);
663 * For a device queue that has no required features, use the default elevator
664 * settings. Otherwise, use the first elevator available matching the required
665 * features. If no suitable elevator is find or if the chosen elevator
666 * initialization fails, fall back to the "none" elevator (no elevator).
668 void elevator_init_mq(struct request_queue
*q
)
670 struct elevator_type
*e
;
673 if (!elv_support_iosched(q
))
676 WARN_ON_ONCE(test_bit(QUEUE_FLAG_REGISTERED
, &q
->queue_flags
));
678 if (unlikely(q
->elevator
))
681 if (!q
->required_elevator_features
)
682 e
= elevator_get_default(q
);
684 e
= elevator_get_by_features(q
);
688 blk_mq_freeze_queue(q
);
689 blk_mq_quiesce_queue(q
);
691 err
= blk_mq_init_sched(q
, e
);
693 blk_mq_unquiesce_queue(q
);
694 blk_mq_unfreeze_queue(q
);
697 pr_warn("\"%s\" elevator initialization failed, "
698 "falling back to \"none\"\n", e
->elevator_name
);
705 * switch to new_e io scheduler. be careful not to introduce deadlocks -
706 * we don't free the old io scheduler, before we have allocated what we
707 * need for the new one. this way we have a chance of going back to the old
708 * one, if the new one fails init for some reason.
710 static int elevator_switch(struct request_queue
*q
, struct elevator_type
*new_e
)
714 lockdep_assert_held(&q
->sysfs_lock
);
716 blk_mq_freeze_queue(q
);
717 blk_mq_quiesce_queue(q
);
719 err
= elevator_switch_mq(q
, new_e
);
721 blk_mq_unquiesce_queue(q
);
722 blk_mq_unfreeze_queue(q
);
728 * Switch this queue to the given IO scheduler.
730 static int __elevator_change(struct request_queue
*q
, const char *name
)
732 char elevator_name
[ELV_NAME_MAX
];
733 struct elevator_type
*e
;
735 /* Make sure queue is not in the middle of being removed */
736 if (!blk_queue_registered(q
))
740 * Special case for mq, turn off scheduling
742 if (!strncmp(name
, "none", 4)) {
745 return elevator_switch(q
, NULL
);
748 strlcpy(elevator_name
, name
, sizeof(elevator_name
));
749 e
= elevator_get(q
, strstrip(elevator_name
), true);
754 elevator_match(q
->elevator
->type
, elevator_name
, 0)) {
759 return elevator_switch(q
, e
);
762 ssize_t
elv_iosched_store(struct request_queue
*q
, const char *name
,
767 if (!queue_is_mq(q
) || !elv_support_iosched(q
))
770 ret
= __elevator_change(q
, name
);
777 ssize_t
elv_iosched_show(struct request_queue
*q
, char *name
)
779 struct elevator_queue
*e
= q
->elevator
;
780 struct elevator_type
*elv
= NULL
;
781 struct elevator_type
*__e
;
785 return sprintf(name
, "none\n");
788 len
+= sprintf(name
+len
, "[none] ");
792 spin_lock(&elv_list_lock
);
793 list_for_each_entry(__e
, &elv_list
, list
) {
794 if (elv
&& elevator_match(elv
, __e
->elevator_name
, 0)) {
795 len
+= sprintf(name
+len
, "[%s] ", elv
->elevator_name
);
798 if (elv_support_iosched(q
) &&
799 elevator_match(__e
, __e
->elevator_name
,
800 q
->required_elevator_features
))
801 len
+= sprintf(name
+len
, "%s ", __e
->elevator_name
);
803 spin_unlock(&elv_list_lock
);
806 len
+= sprintf(name
+len
, "none");
808 len
+= sprintf(len
+name
, "\n");
812 struct request
*elv_rb_former_request(struct request_queue
*q
,
815 struct rb_node
*rbprev
= rb_prev(&rq
->rb_node
);
818 return rb_entry_rq(rbprev
);
822 EXPORT_SYMBOL(elv_rb_former_request
);
824 struct request
*elv_rb_latter_request(struct request_queue
*q
,
827 struct rb_node
*rbnext
= rb_next(&rq
->rb_node
);
830 return rb_entry_rq(rbnext
);
834 EXPORT_SYMBOL(elv_rb_latter_request
);
836 static int __init
elevator_setup(char *str
)
838 pr_warn("Kernel parameter elevator= does not have any effect anymore.\n"
839 "Please use sysfs to set IO scheduler for individual devices.\n");
843 __setup("elevator=", elevator_setup
);