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);
87 * Return scheduler with name 'name' and with matching 'mq capability
89 static struct elevator_type *elevator_find(const char *name, bool mq)
91 struct elevator_type *e;
93 list_for_each_entry(e, &elv_list, list) {
94 if (!strcmp(e->elevator_name, name) && (mq == e->uses_mq))
101 static void elevator_put(struct elevator_type *e)
103 module_put(e->elevator_owner);
106 static struct elevator_type *elevator_get(struct request_queue *q,
107 const char *name, bool try_loading)
109 struct elevator_type *e;
111 spin_lock(&elv_list_lock);
113 e = elevator_find(name, q->mq_ops != NULL);
114 if (!e && try_loading) {
115 spin_unlock(&elv_list_lock);
116 request_module("%s-iosched", name);
117 spin_lock(&elv_list_lock);
118 e = elevator_find(name, q->mq_ops != NULL);
121 if (e && !try_module_get(e->elevator_owner))
124 spin_unlock(&elv_list_lock);
128 static char chosen_elevator[ELV_NAME_MAX];
130 static int __init elevator_setup(char *str)
133 * Be backwards-compatible with previous kernels, so users
134 * won't get the wrong elevator.
136 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
140 __setup("elevator=", elevator_setup);
142 /* called during boot to load the elevator chosen by the elevator param */
143 void __init load_default_elevator_module(void)
145 struct elevator_type *e;
147 if (!chosen_elevator[0])
151 * Boot parameter is deprecated, we haven't supported that for MQ.
152 * Only look for non-mq schedulers from here.
154 spin_lock(&elv_list_lock);
155 e = elevator_find(chosen_elevator, false);
156 spin_unlock(&elv_list_lock);
159 request_module("%s-iosched", chosen_elevator);
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);
177 eq->uses_mq = e->uses_mq;
181 EXPORT_SYMBOL(elevator_alloc);
183 static void elevator_release(struct kobject *kobj)
185 struct elevator_queue *e;
187 e = container_of(kobj, struct elevator_queue, kobj);
188 elevator_put(e->type);
192 int elevator_init(struct request_queue *q, char *name)
194 struct elevator_type *e = NULL;
198 * q->sysfs_lock must be held to provide mutual exclusion between
199 * elevator_switch() and here.
201 lockdep_assert_held(&q->sysfs_lock);
203 if (unlikely(q->elevator))
206 INIT_LIST_HEAD(&q->queue_head);
207 q->last_merge = NULL;
209 q->boundary_rq = NULL;
212 e = elevator_get(q, name, true);
218 * Use the default elevator specified by config boot param for
219 * non-mq devices, or by config option. Don't try to load modules
220 * as we could be running off async and request_module() isn't
221 * allowed from async.
223 if (!e && !q->mq_ops && *chosen_elevator) {
224 e = elevator_get(q, chosen_elevator, false);
226 printk(KERN_ERR "I/O scheduler %s not found\n",
232 * For blk-mq devices, we default to using mq-deadline,
233 * if available, for single queue devices. If deadline
234 * isn't available OR we have multiple queues, default
238 if (q->nr_hw_queues == 1)
239 e = elevator_get(q, "mq-deadline", false);
243 e = elevator_get(q, CONFIG_DEFAULT_IOSCHED, false);
247 "Default I/O scheduler not found. " \
249 e = elevator_get(q, "noop", false);
254 err = blk_mq_init_sched(q, e);
256 err = e->ops.sq.elevator_init_fn(q, e);
261 EXPORT_SYMBOL(elevator_init);
263 void elevator_exit(struct request_queue *q, struct elevator_queue *e)
265 mutex_lock(&e->sysfs_lock);
266 if (e->uses_mq && e->type->ops.mq.exit_sched)
267 blk_mq_exit_sched(q, e);
268 else if (!e->uses_mq && e->type->ops.sq.elevator_exit_fn)
269 e->type->ops.sq.elevator_exit_fn(e);
270 mutex_unlock(&e->sysfs_lock);
272 kobject_put(&e->kobj);
274 EXPORT_SYMBOL(elevator_exit);
276 static inline void __elv_rqhash_del(struct request *rq)
279 rq->rq_flags &= ~RQF_HASHED;
282 void elv_rqhash_del(struct request_queue *q, struct request *rq)
285 __elv_rqhash_del(rq);
287 EXPORT_SYMBOL_GPL(elv_rqhash_del);
289 void elv_rqhash_add(struct request_queue *q, struct request *rq)
291 struct elevator_queue *e = q->elevator;
293 BUG_ON(ELV_ON_HASH(rq));
294 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
295 rq->rq_flags |= RQF_HASHED;
297 EXPORT_SYMBOL_GPL(elv_rqhash_add);
299 void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
301 __elv_rqhash_del(rq);
302 elv_rqhash_add(q, rq);
305 struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
307 struct elevator_queue *e = q->elevator;
308 struct hlist_node *next;
311 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
312 BUG_ON(!ELV_ON_HASH(rq));
314 if (unlikely(!rq_mergeable(rq))) {
315 __elv_rqhash_del(rq);
319 if (rq_hash_key(rq) == offset)
327 * RB-tree support functions for inserting/lookup/removal of requests
328 * in a sorted RB tree.
330 void elv_rb_add(struct rb_root *root, struct request *rq)
332 struct rb_node **p = &root->rb_node;
333 struct rb_node *parent = NULL;
334 struct request *__rq;
338 __rq = rb_entry(parent, struct request, rb_node);
340 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
342 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
346 rb_link_node(&rq->rb_node, parent, p);
347 rb_insert_color(&rq->rb_node, root);
349 EXPORT_SYMBOL(elv_rb_add);
351 void elv_rb_del(struct rb_root *root, struct request *rq)
353 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
354 rb_erase(&rq->rb_node, root);
355 RB_CLEAR_NODE(&rq->rb_node);
357 EXPORT_SYMBOL(elv_rb_del);
359 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
361 struct rb_node *n = root->rb_node;
365 rq = rb_entry(n, struct request, rb_node);
367 if (sector < blk_rq_pos(rq))
369 else if (sector > blk_rq_pos(rq))
377 EXPORT_SYMBOL(elv_rb_find);
380 * Insert rq into dispatch queue of q. Queue lock must be held on
381 * entry. rq is sort instead into the dispatch queue. To be used by
382 * specific elevators.
384 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
387 struct list_head *entry;
389 if (q->last_merge == rq)
390 q->last_merge = NULL;
392 elv_rqhash_del(q, rq);
396 boundary = q->end_sector;
397 list_for_each_prev(entry, &q->queue_head) {
398 struct request *pos = list_entry_rq(entry);
400 if (req_op(rq) != req_op(pos))
402 if (rq_data_dir(rq) != rq_data_dir(pos))
404 if (pos->rq_flags & (RQF_STARTED | RQF_SOFTBARRIER))
406 if (blk_rq_pos(rq) >= boundary) {
407 if (blk_rq_pos(pos) < boundary)
410 if (blk_rq_pos(pos) >= boundary)
413 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
417 list_add(&rq->queuelist, entry);
419 EXPORT_SYMBOL(elv_dispatch_sort);
422 * Insert rq into dispatch queue of q. Queue lock must be held on
423 * entry. rq is added to the back of the dispatch queue. To be used by
424 * specific elevators.
426 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
428 if (q->last_merge == rq)
429 q->last_merge = NULL;
431 elv_rqhash_del(q, rq);
435 q->end_sector = rq_end_sector(rq);
437 list_add_tail(&rq->queuelist, &q->queue_head);
439 EXPORT_SYMBOL(elv_dispatch_add_tail);
441 enum elv_merge elv_merge(struct request_queue *q, struct request **req,
444 struct elevator_queue *e = q->elevator;
445 struct request *__rq;
449 * nomerges: No merges at all attempted
450 * noxmerges: Only simple one-hit cache try
451 * merges: All merge tries attempted
453 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
454 return ELEVATOR_NO_MERGE;
457 * First try one-hit cache.
459 if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
460 enum elv_merge ret = blk_try_merge(q->last_merge, bio);
462 if (ret != ELEVATOR_NO_MERGE) {
463 *req = q->last_merge;
468 if (blk_queue_noxmerges(q))
469 return ELEVATOR_NO_MERGE;
472 * See if our hash lookup can find a potential backmerge.
474 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
475 if (__rq && elv_bio_merge_ok(__rq, bio)) {
477 return ELEVATOR_BACK_MERGE;
480 if (e->uses_mq && e->type->ops.mq.request_merge)
481 return e->type->ops.mq.request_merge(q, req, bio);
482 else if (!e->uses_mq && e->type->ops.sq.elevator_merge_fn)
483 return e->type->ops.sq.elevator_merge_fn(q, req, bio);
485 return ELEVATOR_NO_MERGE;
489 * Attempt to do an insertion back merge. Only check for the case where
490 * we can append 'rq' to an existing request, so we can throw 'rq' away
493 * Returns true if we merged, false otherwise
495 bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq)
497 struct request *__rq;
500 if (blk_queue_nomerges(q))
504 * First try one-hit cache.
506 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
509 if (blk_queue_noxmerges(q))
514 * See if our hash lookup can find a potential backmerge.
517 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
518 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
521 /* The merged request could be merged with others, try again */
529 void elv_merged_request(struct request_queue *q, struct request *rq,
532 struct elevator_queue *e = q->elevator;
534 if (e->uses_mq && e->type->ops.mq.request_merged)
535 e->type->ops.mq.request_merged(q, rq, type);
536 else if (!e->uses_mq && e->type->ops.sq.elevator_merged_fn)
537 e->type->ops.sq.elevator_merged_fn(q, rq, type);
539 if (type == ELEVATOR_BACK_MERGE)
540 elv_rqhash_reposition(q, rq);
545 void elv_merge_requests(struct request_queue *q, struct request *rq,
546 struct request *next)
548 struct elevator_queue *e = q->elevator;
549 bool next_sorted = false;
551 if (e->uses_mq && e->type->ops.mq.requests_merged)
552 e->type->ops.mq.requests_merged(q, rq, next);
553 else if (e->type->ops.sq.elevator_merge_req_fn) {
554 next_sorted = (__force bool)(next->rq_flags & RQF_SORTED);
556 e->type->ops.sq.elevator_merge_req_fn(q, rq, next);
559 elv_rqhash_reposition(q, rq);
562 elv_rqhash_del(q, next);
569 void elv_bio_merged(struct request_queue *q, struct request *rq,
572 struct elevator_queue *e = q->elevator;
574 if (WARN_ON_ONCE(e->uses_mq))
577 if (e->type->ops.sq.elevator_bio_merged_fn)
578 e->type->ops.sq.elevator_bio_merged_fn(q, rq, bio);
582 static void blk_pm_requeue_request(struct request *rq)
584 if (rq->q->dev && !(rq->rq_flags & RQF_PM))
588 static void blk_pm_add_request(struct request_queue *q, struct request *rq)
590 if (q->dev && !(rq->rq_flags & RQF_PM) && q->nr_pending++ == 0 &&
591 (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
592 pm_request_resume(q->dev);
595 static inline void blk_pm_requeue_request(struct request *rq) {}
596 static inline void blk_pm_add_request(struct request_queue *q,
602 void elv_requeue_request(struct request_queue *q, struct request *rq)
605 * it already went through dequeue, we need to decrement the
606 * in_flight count again
608 if (blk_account_rq(rq)) {
609 q->in_flight[rq_is_sync(rq)]--;
610 if (rq->rq_flags & RQF_SORTED)
611 elv_deactivate_rq(q, rq);
614 rq->rq_flags &= ~RQF_STARTED;
616 blk_pm_requeue_request(rq);
618 __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
621 void elv_drain_elevator(struct request_queue *q)
623 struct elevator_queue *e = q->elevator;
626 if (WARN_ON_ONCE(e->uses_mq))
629 lockdep_assert_held(q->queue_lock);
631 while (e->type->ops.sq.elevator_dispatch_fn(q, 1))
633 if (q->nr_sorted && printed++ < 10) {
634 printk(KERN_ERR "%s: forced dispatching is broken "
635 "(nr_sorted=%u), please report this\n",
636 q->elevator->type->elevator_name, q->nr_sorted);
640 void __elv_add_request(struct request_queue *q, struct request *rq, int where)
642 trace_block_rq_insert(q, rq);
644 blk_pm_add_request(q, rq);
648 if (rq->rq_flags & RQF_SOFTBARRIER) {
649 /* barriers are scheduling boundary, update end_sector */
650 if (!blk_rq_is_passthrough(rq)) {
651 q->end_sector = rq_end_sector(rq);
654 } else if (!(rq->rq_flags & RQF_ELVPRIV) &&
655 (where == ELEVATOR_INSERT_SORT ||
656 where == ELEVATOR_INSERT_SORT_MERGE))
657 where = ELEVATOR_INSERT_BACK;
660 case ELEVATOR_INSERT_REQUEUE:
661 case ELEVATOR_INSERT_FRONT:
662 rq->rq_flags |= RQF_SOFTBARRIER;
663 list_add(&rq->queuelist, &q->queue_head);
666 case ELEVATOR_INSERT_BACK:
667 rq->rq_flags |= RQF_SOFTBARRIER;
668 elv_drain_elevator(q);
669 list_add_tail(&rq->queuelist, &q->queue_head);
671 * We kick the queue here for the following reasons.
672 * - The elevator might have returned NULL previously
673 * to delay requests and returned them now. As the
674 * queue wasn't empty before this request, ll_rw_blk
675 * won't run the queue on return, resulting in hang.
676 * - Usually, back inserted requests won't be merged
677 * with anything. There's no point in delaying queue
683 case ELEVATOR_INSERT_SORT_MERGE:
685 * If we succeed in merging this request with one in the
686 * queue already, we are done - rq has now been freed,
687 * so no need to do anything further.
689 if (elv_attempt_insert_merge(q, rq))
692 case ELEVATOR_INSERT_SORT:
693 BUG_ON(blk_rq_is_passthrough(rq));
694 rq->rq_flags |= RQF_SORTED;
696 if (rq_mergeable(rq)) {
697 elv_rqhash_add(q, rq);
703 * Some ioscheds (cfq) run q->request_fn directly, so
704 * rq cannot be accessed after calling
705 * elevator_add_req_fn.
707 q->elevator->type->ops.sq.elevator_add_req_fn(q, rq);
710 case ELEVATOR_INSERT_FLUSH:
711 rq->rq_flags |= RQF_SOFTBARRIER;
712 blk_insert_flush(rq);
715 printk(KERN_ERR "%s: bad insertion point %d\n",
720 EXPORT_SYMBOL(__elv_add_request);
722 void elv_add_request(struct request_queue *q, struct request *rq, int where)
726 spin_lock_irqsave(q->queue_lock, flags);
727 __elv_add_request(q, rq, where);
728 spin_unlock_irqrestore(q->queue_lock, flags);
730 EXPORT_SYMBOL(elv_add_request);
732 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
734 struct elevator_queue *e = q->elevator;
736 if (e->uses_mq && e->type->ops.mq.next_request)
737 return e->type->ops.mq.next_request(q, rq);
738 else if (!e->uses_mq && e->type->ops.sq.elevator_latter_req_fn)
739 return e->type->ops.sq.elevator_latter_req_fn(q, rq);
744 struct request *elv_former_request(struct request_queue *q, struct request *rq)
746 struct elevator_queue *e = q->elevator;
748 if (e->uses_mq && e->type->ops.mq.former_request)
749 return e->type->ops.mq.former_request(q, rq);
750 if (!e->uses_mq && e->type->ops.sq.elevator_former_req_fn)
751 return e->type->ops.sq.elevator_former_req_fn(q, rq);
755 int elv_set_request(struct request_queue *q, struct request *rq,
756 struct bio *bio, gfp_t gfp_mask)
758 struct elevator_queue *e = q->elevator;
760 if (WARN_ON_ONCE(e->uses_mq))
763 if (e->type->ops.sq.elevator_set_req_fn)
764 return e->type->ops.sq.elevator_set_req_fn(q, rq, bio, gfp_mask);
768 void elv_put_request(struct request_queue *q, struct request *rq)
770 struct elevator_queue *e = q->elevator;
772 if (WARN_ON_ONCE(e->uses_mq))
775 if (e->type->ops.sq.elevator_put_req_fn)
776 e->type->ops.sq.elevator_put_req_fn(rq);
779 int elv_may_queue(struct request_queue *q, unsigned int op)
781 struct elevator_queue *e = q->elevator;
783 if (WARN_ON_ONCE(e->uses_mq))
786 if (e->type->ops.sq.elevator_may_queue_fn)
787 return e->type->ops.sq.elevator_may_queue_fn(q, op);
789 return ELV_MQUEUE_MAY;
792 void elv_completed_request(struct request_queue *q, struct request *rq)
794 struct elevator_queue *e = q->elevator;
796 if (WARN_ON_ONCE(e->uses_mq))
800 * request is released from the driver, io must be done
802 if (blk_account_rq(rq)) {
803 q->in_flight[rq_is_sync(rq)]--;
804 if ((rq->rq_flags & RQF_SORTED) &&
805 e->type->ops.sq.elevator_completed_req_fn)
806 e->type->ops.sq.elevator_completed_req_fn(q, rq);
810 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
813 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
815 struct elv_fs_entry *entry = to_elv(attr);
816 struct elevator_queue *e;
822 e = container_of(kobj, struct elevator_queue, kobj);
823 mutex_lock(&e->sysfs_lock);
824 error = e->type ? entry->show(e, page) : -ENOENT;
825 mutex_unlock(&e->sysfs_lock);
830 elv_attr_store(struct kobject *kobj, struct attribute *attr,
831 const char *page, size_t length)
833 struct elv_fs_entry *entry = to_elv(attr);
834 struct elevator_queue *e;
840 e = container_of(kobj, struct elevator_queue, kobj);
841 mutex_lock(&e->sysfs_lock);
842 error = e->type ? entry->store(e, page, length) : -ENOENT;
843 mutex_unlock(&e->sysfs_lock);
847 static const struct sysfs_ops elv_sysfs_ops = {
848 .show = elv_attr_show,
849 .store = elv_attr_store,
852 static struct kobj_type elv_ktype = {
853 .sysfs_ops = &elv_sysfs_ops,
854 .release = elevator_release,
857 int elv_register_queue(struct request_queue *q)
859 struct elevator_queue *e = q->elevator;
862 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
864 struct elv_fs_entry *attr = e->type->elevator_attrs;
866 while (attr->attr.name) {
867 if (sysfs_create_file(&e->kobj, &attr->attr))
872 kobject_uevent(&e->kobj, KOBJ_ADD);
874 if (!e->uses_mq && e->type->ops.sq.elevator_registered_fn)
875 e->type->ops.sq.elevator_registered_fn(q);
879 EXPORT_SYMBOL(elv_register_queue);
881 void elv_unregister_queue(struct request_queue *q)
884 struct elevator_queue *e = q->elevator;
886 kobject_uevent(&e->kobj, KOBJ_REMOVE);
887 kobject_del(&e->kobj);
889 /* Re-enable throttling in case elevator disabled it */
890 wbt_enable_default(q);
893 EXPORT_SYMBOL(elv_unregister_queue);
895 int elv_register(struct elevator_type *e)
899 /* create icq_cache if requested */
901 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
902 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
905 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
906 "%s_io_cq", e->elevator_name);
907 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
908 e->icq_align, 0, NULL);
913 /* register, don't allow duplicate names */
914 spin_lock(&elv_list_lock);
915 if (elevator_find(e->elevator_name, e->uses_mq)) {
916 spin_unlock(&elv_list_lock);
918 kmem_cache_destroy(e->icq_cache);
921 list_add_tail(&e->list, &elv_list);
922 spin_unlock(&elv_list_lock);
924 /* print pretty message */
925 if (!strcmp(e->elevator_name, chosen_elevator) ||
926 (!*chosen_elevator &&
927 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
930 printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
934 EXPORT_SYMBOL_GPL(elv_register);
936 void elv_unregister(struct elevator_type *e)
939 spin_lock(&elv_list_lock);
940 list_del_init(&e->list);
941 spin_unlock(&elv_list_lock);
944 * Destroy icq_cache if it exists. icq's are RCU managed. Make
945 * sure all RCU operations are complete before proceeding.
949 kmem_cache_destroy(e->icq_cache);
953 EXPORT_SYMBOL_GPL(elv_unregister);
955 static int elevator_switch_mq(struct request_queue *q,
956 struct elevator_type *new_e)
960 blk_mq_freeze_queue(q);
963 if (q->elevator->registered)
964 elv_unregister_queue(q);
966 elevator_exit(q, q->elevator);
969 ret = blk_mq_init_sched(q, new_e);
974 ret = elv_register_queue(q);
976 elevator_exit(q, q->elevator);
982 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
984 blk_add_trace_msg(q, "elv switch: none");
987 blk_mq_unfreeze_queue(q);
992 * switch to new_e io scheduler. be careful not to introduce deadlocks -
993 * we don't free the old io scheduler, before we have allocated what we
994 * need for the new one. this way we have a chance of going back to the old
995 * one, if the new one fails init for some reason.
997 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
999 struct elevator_queue *old = q->elevator;
1000 bool old_registered = false;
1004 return elevator_switch_mq(q, new_e);
1007 * Turn on BYPASS and drain all requests w/ elevator private data.
1008 * Block layer doesn't call into a quiesced elevator - all requests
1009 * are directly put on the dispatch list without elevator data
1010 * using INSERT_BACK. All requests have SOFTBARRIER set and no
1011 * merge happens either.
1014 old_registered = old->registered;
1016 blk_queue_bypass_start(q);
1018 /* unregister and clear all auxiliary data of the old elevator */
1020 elv_unregister_queue(q);
1025 /* allocate, init and register new elevator */
1026 err = new_e->ops.sq.elevator_init_fn(q, new_e);
1030 err = elv_register_queue(q);
1034 /* done, kill the old one and finish */
1036 elevator_exit(q, old);
1037 blk_queue_bypass_end(q);
1040 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
1045 elevator_exit(q, q->elevator);
1047 /* switch failed, restore and re-register old elevator */
1050 elv_register_queue(q);
1051 blk_queue_bypass_end(q);
1058 * Switch this queue to the given IO scheduler.
1060 static int __elevator_change(struct request_queue *q, const char *name)
1062 char elevator_name[ELV_NAME_MAX];
1063 struct elevator_type *e;
1065 /* Make sure queue is not in the middle of being removed */
1066 if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
1070 * Special case for mq, turn off scheduling
1072 if (q->mq_ops && !strncmp(name, "none", 4))
1073 return elevator_switch(q, NULL);
1075 strlcpy(elevator_name, name, sizeof(elevator_name));
1076 e = elevator_get(q, strstrip(elevator_name), true);
1081 !strcmp(elevator_name, q->elevator->type->elevator_name)) {
1086 return elevator_switch(q, e);
1089 static inline bool elv_support_iosched(struct request_queue *q)
1091 if (q->mq_ops && q->tag_set && (q->tag_set->flags &
1097 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1102 if (!(q->mq_ops || q->request_fn) || !elv_support_iosched(q))
1105 ret = __elevator_change(q, name);
1112 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1114 struct elevator_queue *e = q->elevator;
1115 struct elevator_type *elv = NULL;
1116 struct elevator_type *__e;
1119 if (!blk_queue_stackable(q))
1120 return sprintf(name, "none\n");
1123 len += sprintf(name+len, "[none] ");
1127 spin_lock(&elv_list_lock);
1128 list_for_each_entry(__e, &elv_list, list) {
1129 if (elv && !strcmp(elv->elevator_name, __e->elevator_name)) {
1130 len += sprintf(name+len, "[%s] ", elv->elevator_name);
1133 if (__e->uses_mq && q->mq_ops && elv_support_iosched(q))
1134 len += sprintf(name+len, "%s ", __e->elevator_name);
1135 else if (!__e->uses_mq && !q->mq_ops)
1136 len += sprintf(name+len, "%s ", __e->elevator_name);
1138 spin_unlock(&elv_list_lock);
1140 if (q->mq_ops && q->elevator)
1141 len += sprintf(name+len, "none");
1143 len += sprintf(len+name, "\n");
1147 struct request *elv_rb_former_request(struct request_queue *q,
1150 struct rb_node *rbprev = rb_prev(&rq->rb_node);
1153 return rb_entry_rq(rbprev);
1157 EXPORT_SYMBOL(elv_rb_former_request);
1159 struct request *elv_rb_latter_request(struct request_queue *q,
1162 struct rb_node *rbnext = rb_next(&rq->rb_node);
1165 return rb_entry_rq(rbnext);
1169 EXPORT_SYMBOL(elv_rb_latter_request);