2 * Block device elevator/IO-scheduler.
4 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
6 * 30042000 Jens Axboe <axboe@kernel.dk> :
8 * Split the elevator a bit so that it is possible to choose a different
9 * one or even write a new "plug in". There are three pieces:
10 * - elevator_fn, inserts a new request in the queue list
11 * - elevator_merge_fn, decides whether a new buffer can be merged with
13 * - elevator_dequeue_fn, called when a request is taken off the active list
15 * 20082000 Dave Jones <davej@suse.de> :
16 * Removed tests for max-bomb-segments, which was breaking elvtune
17 * when run without -bN
20 * - Rework again to work with bio instead of buffer_heads
21 * - loose bi_dev comparisons, partition handling is right now
22 * - completely modularize elevator setup and teardown
25 #include <linux/kernel.h>
27 #include <linux/blkdev.h>
28 #include <linux/elevator.h>
29 #include <linux/bio.h>
30 #include <linux/module.h>
31 #include <linux/slab.h>
32 #include <linux/init.h>
33 #include <linux/compiler.h>
34 #include <linux/blktrace_api.h>
35 #include <linux/hash.h>
36 #include <linux/uaccess.h>
37 #include <linux/pm_runtime.h>
38 #include <linux/blk-cgroup.h>
40 #include <trace/events/block.h>
43 #include "blk-mq-sched.h"
46 static DEFINE_SPINLOCK(elv_list_lock);
47 static LIST_HEAD(elv_list);
52 #define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq))
55 * Query io scheduler to see if the current process issuing bio may be
58 static int elv_iosched_allow_bio_merge(struct request *rq, struct bio *bio)
60 struct request_queue *q = rq->q;
61 struct elevator_queue *e = q->elevator;
63 if (e->uses_mq && e->type->ops.mq.allow_merge)
64 return e->type->ops.mq.allow_merge(q, rq, bio);
65 else if (!e->uses_mq && e->type->ops.sq.elevator_allow_bio_merge_fn)
66 return e->type->ops.sq.elevator_allow_bio_merge_fn(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' and with matching 'mq capability
99 static struct elevator_type *elevator_find(const char *name, bool mq)
101 struct elevator_type *e;
103 list_for_each_entry(e, &elv_list, list) {
104 if (elevator_match(e, name) && (mq == e->uses_mq))
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, q->mq_ops != NULL);
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, q->mq_ops != NULL);
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 /* called during boot to load the elevator chosen by the elevator param */
153 void __init load_default_elevator_module(void)
155 struct elevator_type *e;
157 if (!chosen_elevator[0])
161 * Boot parameter is deprecated, we haven't supported that for MQ.
162 * Only look for non-mq schedulers from here.
164 spin_lock(&elv_list_lock);
165 e = elevator_find(chosen_elevator, false);
166 spin_unlock(&elv_list_lock);
169 request_module("%s-iosched", chosen_elevator);
172 static struct kobj_type elv_ktype;
174 struct elevator_queue *elevator_alloc(struct request_queue *q,
175 struct elevator_type *e)
177 struct elevator_queue *eq;
179 eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
184 kobject_init(&eq->kobj, &elv_ktype);
185 mutex_init(&eq->sysfs_lock);
187 eq->uses_mq = e->uses_mq;
191 EXPORT_SYMBOL(elevator_alloc);
193 static void elevator_release(struct kobject *kobj)
195 struct elevator_queue *e;
197 e = container_of(kobj, struct elevator_queue, kobj);
198 elevator_put(e->type);
203 * Use the default elevator specified by config boot param for non-mq devices,
204 * or by config option. Don't try to load modules as we could be running off
205 * async and request_module() isn't allowed from async.
207 int elevator_init(struct request_queue *q)
209 struct elevator_type *e = NULL;
213 * q->sysfs_lock must be held to provide mutual exclusion between
214 * elevator_switch() and here.
216 mutex_lock(&q->sysfs_lock);
217 if (unlikely(q->elevator))
220 if (*chosen_elevator) {
221 e = elevator_get(q, chosen_elevator, false);
223 printk(KERN_ERR "I/O scheduler %s not found\n",
228 e = elevator_get(q, CONFIG_DEFAULT_IOSCHED, false);
231 "Default I/O scheduler not found. Using noop.\n");
232 e = elevator_get(q, "noop", false);
235 err = e->ops.sq.elevator_init_fn(q, e);
239 mutex_unlock(&q->sysfs_lock);
243 void elevator_exit(struct request_queue *q, struct elevator_queue *e)
245 mutex_lock(&e->sysfs_lock);
246 if (e->uses_mq && e->type->ops.mq.exit_sched)
247 blk_mq_exit_sched(q, e);
248 else if (!e->uses_mq && e->type->ops.sq.elevator_exit_fn)
249 e->type->ops.sq.elevator_exit_fn(e);
250 mutex_unlock(&e->sysfs_lock);
252 kobject_put(&e->kobj);
255 static inline void __elv_rqhash_del(struct request *rq)
258 rq->rq_flags &= ~RQF_HASHED;
261 void elv_rqhash_del(struct request_queue *q, struct request *rq)
264 __elv_rqhash_del(rq);
266 EXPORT_SYMBOL_GPL(elv_rqhash_del);
268 void elv_rqhash_add(struct request_queue *q, struct request *rq)
270 struct elevator_queue *e = q->elevator;
272 BUG_ON(ELV_ON_HASH(rq));
273 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
274 rq->rq_flags |= RQF_HASHED;
276 EXPORT_SYMBOL_GPL(elv_rqhash_add);
278 void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
280 __elv_rqhash_del(rq);
281 elv_rqhash_add(q, rq);
284 struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
286 struct elevator_queue *e = q->elevator;
287 struct hlist_node *next;
290 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
291 BUG_ON(!ELV_ON_HASH(rq));
293 if (unlikely(!rq_mergeable(rq))) {
294 __elv_rqhash_del(rq);
298 if (rq_hash_key(rq) == offset)
306 * RB-tree support functions for inserting/lookup/removal of requests
307 * in a sorted RB tree.
309 void elv_rb_add(struct rb_root *root, struct request *rq)
311 struct rb_node **p = &root->rb_node;
312 struct rb_node *parent = NULL;
313 struct request *__rq;
317 __rq = rb_entry(parent, struct request, rb_node);
319 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
321 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
325 rb_link_node(&rq->rb_node, parent, p);
326 rb_insert_color(&rq->rb_node, root);
328 EXPORT_SYMBOL(elv_rb_add);
330 void elv_rb_del(struct rb_root *root, struct request *rq)
332 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
333 rb_erase(&rq->rb_node, root);
334 RB_CLEAR_NODE(&rq->rb_node);
336 EXPORT_SYMBOL(elv_rb_del);
338 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
340 struct rb_node *n = root->rb_node;
344 rq = rb_entry(n, struct request, rb_node);
346 if (sector < blk_rq_pos(rq))
348 else if (sector > blk_rq_pos(rq))
356 EXPORT_SYMBOL(elv_rb_find);
359 * Insert rq into dispatch queue of q. Queue lock must be held on
360 * entry. rq is sort instead into the dispatch queue. To be used by
361 * specific elevators.
363 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
366 struct list_head *entry;
368 if (q->last_merge == rq)
369 q->last_merge = NULL;
371 elv_rqhash_del(q, rq);
375 boundary = q->end_sector;
376 list_for_each_prev(entry, &q->queue_head) {
377 struct request *pos = list_entry_rq(entry);
379 if (req_op(rq) != req_op(pos))
381 if (rq_data_dir(rq) != rq_data_dir(pos))
383 if (pos->rq_flags & (RQF_STARTED | RQF_SOFTBARRIER))
385 if (blk_rq_pos(rq) >= boundary) {
386 if (blk_rq_pos(pos) < boundary)
389 if (blk_rq_pos(pos) >= boundary)
392 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
396 list_add(&rq->queuelist, entry);
398 EXPORT_SYMBOL(elv_dispatch_sort);
401 * Insert rq into dispatch queue of q. Queue lock must be held on
402 * entry. rq is added to the back of the dispatch queue. To be used by
403 * specific elevators.
405 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
407 if (q->last_merge == rq)
408 q->last_merge = NULL;
410 elv_rqhash_del(q, rq);
414 q->end_sector = rq_end_sector(rq);
416 list_add_tail(&rq->queuelist, &q->queue_head);
418 EXPORT_SYMBOL(elv_dispatch_add_tail);
420 enum elv_merge elv_merge(struct request_queue *q, struct request **req,
423 struct elevator_queue *e = q->elevator;
424 struct request *__rq;
428 * nomerges: No merges at all attempted
429 * noxmerges: Only simple one-hit cache try
430 * merges: All merge tries attempted
432 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
433 return ELEVATOR_NO_MERGE;
436 * First try one-hit cache.
438 if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
439 enum elv_merge ret = blk_try_merge(q->last_merge, bio);
441 if (ret != ELEVATOR_NO_MERGE) {
442 *req = q->last_merge;
447 if (blk_queue_noxmerges(q))
448 return ELEVATOR_NO_MERGE;
451 * See if our hash lookup can find a potential backmerge.
453 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
454 if (__rq && elv_bio_merge_ok(__rq, bio)) {
456 return ELEVATOR_BACK_MERGE;
459 if (e->uses_mq && e->type->ops.mq.request_merge)
460 return e->type->ops.mq.request_merge(q, req, bio);
461 else if (!e->uses_mq && e->type->ops.sq.elevator_merge_fn)
462 return e->type->ops.sq.elevator_merge_fn(q, req, bio);
464 return ELEVATOR_NO_MERGE;
468 * Attempt to do an insertion back merge. Only check for the case where
469 * we can append 'rq' to an existing request, so we can throw 'rq' away
472 * Returns true if we merged, false otherwise
474 bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq)
476 struct request *__rq;
479 if (blk_queue_nomerges(q))
483 * First try one-hit cache.
485 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
488 if (blk_queue_noxmerges(q))
493 * See if our hash lookup can find a potential backmerge.
496 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
497 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
500 /* The merged request could be merged with others, try again */
508 void elv_merged_request(struct request_queue *q, struct request *rq,
511 struct elevator_queue *e = q->elevator;
513 if (e->uses_mq && e->type->ops.mq.request_merged)
514 e->type->ops.mq.request_merged(q, rq, type);
515 else if (!e->uses_mq && e->type->ops.sq.elevator_merged_fn)
516 e->type->ops.sq.elevator_merged_fn(q, rq, type);
518 if (type == ELEVATOR_BACK_MERGE)
519 elv_rqhash_reposition(q, rq);
524 void elv_merge_requests(struct request_queue *q, struct request *rq,
525 struct request *next)
527 struct elevator_queue *e = q->elevator;
528 bool next_sorted = false;
530 if (e->uses_mq && e->type->ops.mq.requests_merged)
531 e->type->ops.mq.requests_merged(q, rq, next);
532 else if (e->type->ops.sq.elevator_merge_req_fn) {
533 next_sorted = (__force bool)(next->rq_flags & RQF_SORTED);
535 e->type->ops.sq.elevator_merge_req_fn(q, rq, next);
538 elv_rqhash_reposition(q, rq);
541 elv_rqhash_del(q, next);
548 void elv_bio_merged(struct request_queue *q, struct request *rq,
551 struct elevator_queue *e = q->elevator;
553 if (WARN_ON_ONCE(e->uses_mq))
556 if (e->type->ops.sq.elevator_bio_merged_fn)
557 e->type->ops.sq.elevator_bio_merged_fn(q, rq, bio);
561 static void blk_pm_requeue_request(struct request *rq)
563 if (rq->q->dev && !(rq->rq_flags & RQF_PM))
567 static void blk_pm_add_request(struct request_queue *q, struct request *rq)
569 if (q->dev && !(rq->rq_flags & RQF_PM) && q->nr_pending++ == 0 &&
570 (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
571 pm_request_resume(q->dev);
574 static inline void blk_pm_requeue_request(struct request *rq) {}
575 static inline void blk_pm_add_request(struct request_queue *q,
581 void elv_requeue_request(struct request_queue *q, struct request *rq)
584 * it already went through dequeue, we need to decrement the
585 * in_flight count again
587 if (blk_account_rq(rq)) {
588 q->in_flight[rq_is_sync(rq)]--;
589 if (rq->rq_flags & RQF_SORTED)
590 elv_deactivate_rq(q, rq);
593 rq->rq_flags &= ~RQF_STARTED;
595 blk_pm_requeue_request(rq);
597 __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
600 void elv_drain_elevator(struct request_queue *q)
602 struct elevator_queue *e = q->elevator;
605 if (WARN_ON_ONCE(e->uses_mq))
608 lockdep_assert_held(q->queue_lock);
610 while (e->type->ops.sq.elevator_dispatch_fn(q, 1))
612 if (q->nr_sorted && !blk_queue_is_zoned(q) && printed++ < 10 ) {
613 printk(KERN_ERR "%s: forced dispatching is broken "
614 "(nr_sorted=%u), please report this\n",
615 q->elevator->type->elevator_name, q->nr_sorted);
619 void __elv_add_request(struct request_queue *q, struct request *rq, int where)
621 trace_block_rq_insert(q, rq);
623 blk_pm_add_request(q, rq);
627 if (rq->rq_flags & RQF_SOFTBARRIER) {
628 /* barriers are scheduling boundary, update end_sector */
629 if (!blk_rq_is_passthrough(rq)) {
630 q->end_sector = rq_end_sector(rq);
633 } else if (!(rq->rq_flags & RQF_ELVPRIV) &&
634 (where == ELEVATOR_INSERT_SORT ||
635 where == ELEVATOR_INSERT_SORT_MERGE))
636 where = ELEVATOR_INSERT_BACK;
639 case ELEVATOR_INSERT_REQUEUE:
640 case ELEVATOR_INSERT_FRONT:
641 rq->rq_flags |= RQF_SOFTBARRIER;
642 list_add(&rq->queuelist, &q->queue_head);
645 case ELEVATOR_INSERT_BACK:
646 rq->rq_flags |= RQF_SOFTBARRIER;
647 elv_drain_elevator(q);
648 list_add_tail(&rq->queuelist, &q->queue_head);
650 * We kick the queue here for the following reasons.
651 * - The elevator might have returned NULL previously
652 * to delay requests and returned them now. As the
653 * queue wasn't empty before this request, ll_rw_blk
654 * won't run the queue on return, resulting in hang.
655 * - Usually, back inserted requests won't be merged
656 * with anything. There's no point in delaying queue
662 case ELEVATOR_INSERT_SORT_MERGE:
664 * If we succeed in merging this request with one in the
665 * queue already, we are done - rq has now been freed,
666 * so no need to do anything further.
668 if (elv_attempt_insert_merge(q, rq))
671 case ELEVATOR_INSERT_SORT:
672 BUG_ON(blk_rq_is_passthrough(rq));
673 rq->rq_flags |= RQF_SORTED;
675 if (rq_mergeable(rq)) {
676 elv_rqhash_add(q, rq);
682 * Some ioscheds (cfq) run q->request_fn directly, so
683 * rq cannot be accessed after calling
684 * elevator_add_req_fn.
686 q->elevator->type->ops.sq.elevator_add_req_fn(q, rq);
689 case ELEVATOR_INSERT_FLUSH:
690 rq->rq_flags |= RQF_SOFTBARRIER;
691 blk_insert_flush(rq);
694 printk(KERN_ERR "%s: bad insertion point %d\n",
699 EXPORT_SYMBOL(__elv_add_request);
701 void elv_add_request(struct request_queue *q, struct request *rq, int where)
705 spin_lock_irqsave(q->queue_lock, flags);
706 __elv_add_request(q, rq, where);
707 spin_unlock_irqrestore(q->queue_lock, flags);
709 EXPORT_SYMBOL(elv_add_request);
711 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
713 struct elevator_queue *e = q->elevator;
715 if (e->uses_mq && e->type->ops.mq.next_request)
716 return e->type->ops.mq.next_request(q, rq);
717 else if (!e->uses_mq && e->type->ops.sq.elevator_latter_req_fn)
718 return e->type->ops.sq.elevator_latter_req_fn(q, rq);
723 struct request *elv_former_request(struct request_queue *q, struct request *rq)
725 struct elevator_queue *e = q->elevator;
727 if (e->uses_mq && e->type->ops.mq.former_request)
728 return e->type->ops.mq.former_request(q, rq);
729 if (!e->uses_mq && e->type->ops.sq.elevator_former_req_fn)
730 return e->type->ops.sq.elevator_former_req_fn(q, rq);
734 int elv_set_request(struct request_queue *q, struct request *rq,
735 struct bio *bio, gfp_t gfp_mask)
737 struct elevator_queue *e = q->elevator;
739 if (WARN_ON_ONCE(e->uses_mq))
742 if (e->type->ops.sq.elevator_set_req_fn)
743 return e->type->ops.sq.elevator_set_req_fn(q, rq, bio, gfp_mask);
747 void elv_put_request(struct request_queue *q, struct request *rq)
749 struct elevator_queue *e = q->elevator;
751 if (WARN_ON_ONCE(e->uses_mq))
754 if (e->type->ops.sq.elevator_put_req_fn)
755 e->type->ops.sq.elevator_put_req_fn(rq);
758 int elv_may_queue(struct request_queue *q, unsigned int op)
760 struct elevator_queue *e = q->elevator;
762 if (WARN_ON_ONCE(e->uses_mq))
765 if (e->type->ops.sq.elevator_may_queue_fn)
766 return e->type->ops.sq.elevator_may_queue_fn(q, op);
768 return ELV_MQUEUE_MAY;
771 void elv_completed_request(struct request_queue *q, struct request *rq)
773 struct elevator_queue *e = q->elevator;
775 if (WARN_ON_ONCE(e->uses_mq))
779 * request is released from the driver, io must be done
781 if (blk_account_rq(rq)) {
782 q->in_flight[rq_is_sync(rq)]--;
783 if ((rq->rq_flags & RQF_SORTED) &&
784 e->type->ops.sq.elevator_completed_req_fn)
785 e->type->ops.sq.elevator_completed_req_fn(q, rq);
789 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
792 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
794 struct elv_fs_entry *entry = to_elv(attr);
795 struct elevator_queue *e;
801 e = container_of(kobj, struct elevator_queue, kobj);
802 mutex_lock(&e->sysfs_lock);
803 error = e->type ? entry->show(e, page) : -ENOENT;
804 mutex_unlock(&e->sysfs_lock);
809 elv_attr_store(struct kobject *kobj, struct attribute *attr,
810 const char *page, size_t length)
812 struct elv_fs_entry *entry = to_elv(attr);
813 struct elevator_queue *e;
819 e = container_of(kobj, struct elevator_queue, kobj);
820 mutex_lock(&e->sysfs_lock);
821 error = e->type ? entry->store(e, page, length) : -ENOENT;
822 mutex_unlock(&e->sysfs_lock);
826 static const struct sysfs_ops elv_sysfs_ops = {
827 .show = elv_attr_show,
828 .store = elv_attr_store,
831 static struct kobj_type elv_ktype = {
832 .sysfs_ops = &elv_sysfs_ops,
833 .release = elevator_release,
837 * elv_register_queue is called from either blk_register_queue or
838 * elevator_switch, elevator switch is prevented from being happen
839 * in the two paths, so it is safe to not hold q->sysfs_lock.
841 int elv_register_queue(struct request_queue *q, bool uevent)
843 struct elevator_queue *e = q->elevator;
846 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
848 struct elv_fs_entry *attr = e->type->elevator_attrs;
850 while (attr->attr.name) {
851 if (sysfs_create_file(&e->kobj, &attr->attr))
857 kobject_uevent(&e->kobj, KOBJ_ADD);
860 if (!e->uses_mq && e->type->ops.sq.elevator_registered_fn)
861 e->type->ops.sq.elevator_registered_fn(q);
867 * elv_unregister_queue is called from either blk_unregister_queue or
868 * elevator_switch, elevator switch is prevented from being happen
869 * in the two paths, so it is safe to not hold q->sysfs_lock.
871 void elv_unregister_queue(struct request_queue *q)
874 struct elevator_queue *e = q->elevator;
876 kobject_uevent(&e->kobj, KOBJ_REMOVE);
877 kobject_del(&e->kobj);
880 /* Re-enable throttling in case elevator disabled it */
881 wbt_enable_default(q);
885 int elv_register(struct elevator_type *e)
889 /* create icq_cache if requested */
891 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
892 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
895 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
896 "%s_io_cq", e->elevator_name);
897 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
898 e->icq_align, 0, NULL);
903 /* register, don't allow duplicate names */
904 spin_lock(&elv_list_lock);
905 if (elevator_find(e->elevator_name, e->uses_mq)) {
906 spin_unlock(&elv_list_lock);
907 kmem_cache_destroy(e->icq_cache);
910 list_add_tail(&e->list, &elv_list);
911 spin_unlock(&elv_list_lock);
913 /* print pretty message */
914 if (elevator_match(e, chosen_elevator) ||
915 (!*chosen_elevator &&
916 elevator_match(e, CONFIG_DEFAULT_IOSCHED)))
919 printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
923 EXPORT_SYMBOL_GPL(elv_register);
925 void elv_unregister(struct elevator_type *e)
928 spin_lock(&elv_list_lock);
929 list_del_init(&e->list);
930 spin_unlock(&elv_list_lock);
933 * Destroy icq_cache if it exists. icq's are RCU managed. Make
934 * sure all RCU operations are complete before proceeding.
938 kmem_cache_destroy(e->icq_cache);
942 EXPORT_SYMBOL_GPL(elv_unregister);
944 int elevator_switch_mq(struct request_queue *q,
945 struct elevator_type *new_e)
949 lockdep_assert_held(&q->sysfs_lock);
952 if (q->elevator->registered)
953 elv_unregister_queue(q);
956 elevator_exit(q, q->elevator);
959 ret = blk_mq_init_sched(q, new_e);
964 ret = elv_register_queue(q, true);
966 elevator_exit(q, q->elevator);
972 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
974 blk_add_trace_msg(q, "elv switch: none");
981 * For blk-mq devices, we default to using mq-deadline, if available, for single
982 * queue devices. If deadline isn't available OR we have multiple queues,
985 int elevator_init_mq(struct request_queue *q)
987 struct elevator_type *e;
990 if (q->nr_hw_queues != 1)
993 WARN_ON_ONCE(test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags));
995 if (unlikely(q->elevator))
998 e = elevator_get(q, "mq-deadline", false);
1002 err = blk_mq_init_sched(q, e);
1011 * switch to new_e io scheduler. be careful not to introduce deadlocks -
1012 * we don't free the old io scheduler, before we have allocated what we
1013 * need for the new one. this way we have a chance of going back to the old
1014 * one, if the new one fails init for some reason.
1016 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1018 struct elevator_queue *old = q->elevator;
1019 bool old_registered = false;
1022 lockdep_assert_held(&q->sysfs_lock);
1025 blk_mq_freeze_queue(q);
1026 blk_mq_quiesce_queue(q);
1028 err = elevator_switch_mq(q, new_e);
1030 blk_mq_unquiesce_queue(q);
1031 blk_mq_unfreeze_queue(q);
1037 * Turn on BYPASS and drain all requests w/ elevator private data.
1038 * Block layer doesn't call into a quiesced elevator - all requests
1039 * are directly put on the dispatch list without elevator data
1040 * using INSERT_BACK. All requests have SOFTBARRIER set and no
1041 * merge happens either.
1044 old_registered = old->registered;
1046 blk_queue_bypass_start(q);
1048 /* unregister and clear all auxiliary data of the old elevator */
1050 elv_unregister_queue(q);
1055 /* allocate, init and register new elevator */
1056 err = new_e->ops.sq.elevator_init_fn(q, new_e);
1060 err = elv_register_queue(q, true);
1064 /* done, kill the old one and finish */
1066 elevator_exit(q, old);
1067 blk_queue_bypass_end(q);
1070 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
1075 elevator_exit(q, q->elevator);
1077 /* switch failed, restore and re-register old elevator */
1080 elv_register_queue(q, true);
1081 blk_queue_bypass_end(q);
1088 * Switch this queue to the given IO scheduler.
1090 static int __elevator_change(struct request_queue *q, const char *name)
1092 char elevator_name[ELV_NAME_MAX];
1093 struct elevator_type *e;
1095 /* Make sure queue is not in the middle of being removed */
1096 if (!blk_queue_registered(q))
1100 * Special case for mq, turn off scheduling
1102 if (q->mq_ops && !strncmp(name, "none", 4))
1103 return elevator_switch(q, NULL);
1105 strlcpy(elevator_name, name, sizeof(elevator_name));
1106 e = elevator_get(q, strstrip(elevator_name), true);
1110 if (q->elevator && elevator_match(q->elevator->type, elevator_name)) {
1115 return elevator_switch(q, e);
1118 static inline bool elv_support_iosched(struct request_queue *q)
1120 if (q->mq_ops && q->tag_set && (q->tag_set->flags &
1126 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1131 if (!(q->mq_ops || q->request_fn) || !elv_support_iosched(q))
1134 ret = __elevator_change(q, name);
1141 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1143 struct elevator_queue *e = q->elevator;
1144 struct elevator_type *elv = NULL;
1145 struct elevator_type *__e;
1146 bool uses_mq = q->mq_ops != NULL;
1149 if (!queue_is_rq_based(q))
1150 return sprintf(name, "none\n");
1153 len += sprintf(name+len, "[none] ");
1157 spin_lock(&elv_list_lock);
1158 list_for_each_entry(__e, &elv_list, list) {
1159 if (elv && elevator_match(elv, __e->elevator_name) &&
1160 (__e->uses_mq == uses_mq)) {
1161 len += sprintf(name+len, "[%s] ", elv->elevator_name);
1164 if (__e->uses_mq && q->mq_ops && elv_support_iosched(q))
1165 len += sprintf(name+len, "%s ", __e->elevator_name);
1166 else if (!__e->uses_mq && !q->mq_ops)
1167 len += sprintf(name+len, "%s ", __e->elevator_name);
1169 spin_unlock(&elv_list_lock);
1171 if (q->mq_ops && q->elevator)
1172 len += sprintf(name+len, "none");
1174 len += sprintf(len+name, "\n");
1178 struct request *elv_rb_former_request(struct request_queue *q,
1181 struct rb_node *rbprev = rb_prev(&rq->rb_node);
1184 return rb_entry_rq(rbprev);
1188 EXPORT_SYMBOL(elv_rb_former_request);
1190 struct request *elv_rb_latter_request(struct request_queue *q,
1193 struct rb_node *rbnext = rb_next(&rq->rb_node);
1196 return rb_entry_rq(rbnext);
1200 EXPORT_SYMBOL(elv_rb_latter_request);