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 struct elevator_type *elevator_find(const char *name)
88 struct elevator_type *e;
90 list_for_each_entry(e, &elv_list, list) {
91 if (!strcmp(e->elevator_name, name))
98 static void elevator_put(struct elevator_type *e)
100 module_put(e->elevator_owner);
103 static struct elevator_type *elevator_get(const char *name, bool try_loading)
105 struct elevator_type *e;
107 spin_lock(&elv_list_lock);
109 e = elevator_find(name);
110 if (!e && try_loading) {
111 spin_unlock(&elv_list_lock);
112 request_module("%s-iosched", name);
113 spin_lock(&elv_list_lock);
114 e = elevator_find(name);
117 if (e && !try_module_get(e->elevator_owner))
120 spin_unlock(&elv_list_lock);
125 static char chosen_elevator[ELV_NAME_MAX];
127 static int __init elevator_setup(char *str)
130 * Be backwards-compatible with previous kernels, so users
131 * won't get the wrong elevator.
133 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
137 __setup("elevator=", elevator_setup);
139 /* called during boot to load the elevator chosen by the elevator param */
140 void __init load_default_elevator_module(void)
142 struct elevator_type *e;
144 if (!chosen_elevator[0])
147 spin_lock(&elv_list_lock);
148 e = elevator_find(chosen_elevator);
149 spin_unlock(&elv_list_lock);
152 request_module("%s-iosched", chosen_elevator);
155 static struct kobj_type elv_ktype;
157 struct elevator_queue *elevator_alloc(struct request_queue *q,
158 struct elevator_type *e)
160 struct elevator_queue *eq;
162 eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
167 kobject_init(&eq->kobj, &elv_ktype);
168 mutex_init(&eq->sysfs_lock);
170 eq->uses_mq = e->uses_mq;
174 EXPORT_SYMBOL(elevator_alloc);
176 static void elevator_release(struct kobject *kobj)
178 struct elevator_queue *e;
180 e = container_of(kobj, struct elevator_queue, kobj);
181 elevator_put(e->type);
185 int elevator_init(struct request_queue *q, char *name)
187 struct elevator_type *e = NULL;
191 * q->sysfs_lock must be held to provide mutual exclusion between
192 * elevator_switch() and here.
194 lockdep_assert_held(&q->sysfs_lock);
196 if (unlikely(q->elevator))
199 INIT_LIST_HEAD(&q->queue_head);
200 q->last_merge = NULL;
202 q->boundary_rq = NULL;
205 e = elevator_get(name, true);
211 * Use the default elevator specified by config boot param for
212 * non-mq devices, or by config option. Don't try to load modules
213 * as we could be running off async and request_module() isn't
214 * allowed from async.
216 if (!e && !q->mq_ops && *chosen_elevator) {
217 e = elevator_get(chosen_elevator, false);
219 printk(KERN_ERR "I/O scheduler %s not found\n",
225 * For blk-mq devices, we default to using mq-deadline,
226 * if available, for single queue devices. If deadline
227 * isn't available OR we have multiple queues, default
231 if (q->nr_hw_queues == 1)
232 e = elevator_get("mq-deadline", false);
236 e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
240 "Default I/O scheduler not found. " \
242 e = elevator_get("noop", false);
247 err = blk_mq_init_sched(q, e);
249 err = e->ops.sq.elevator_init_fn(q, e);
254 EXPORT_SYMBOL(elevator_init);
256 void elevator_exit(struct request_queue *q, struct elevator_queue *e)
258 mutex_lock(&e->sysfs_lock);
259 if (e->uses_mq && e->type->ops.mq.exit_sched)
260 blk_mq_exit_sched(q, e);
261 else if (!e->uses_mq && e->type->ops.sq.elevator_exit_fn)
262 e->type->ops.sq.elevator_exit_fn(e);
263 mutex_unlock(&e->sysfs_lock);
265 kobject_put(&e->kobj);
267 EXPORT_SYMBOL(elevator_exit);
269 static inline void __elv_rqhash_del(struct request *rq)
272 rq->rq_flags &= ~RQF_HASHED;
275 void elv_rqhash_del(struct request_queue *q, struct request *rq)
278 __elv_rqhash_del(rq);
280 EXPORT_SYMBOL_GPL(elv_rqhash_del);
282 void elv_rqhash_add(struct request_queue *q, struct request *rq)
284 struct elevator_queue *e = q->elevator;
286 BUG_ON(ELV_ON_HASH(rq));
287 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
288 rq->rq_flags |= RQF_HASHED;
290 EXPORT_SYMBOL_GPL(elv_rqhash_add);
292 void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
294 __elv_rqhash_del(rq);
295 elv_rqhash_add(q, rq);
298 struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
300 struct elevator_queue *e = q->elevator;
301 struct hlist_node *next;
304 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
305 BUG_ON(!ELV_ON_HASH(rq));
307 if (unlikely(!rq_mergeable(rq))) {
308 __elv_rqhash_del(rq);
312 if (rq_hash_key(rq) == offset)
320 * RB-tree support functions for inserting/lookup/removal of requests
321 * in a sorted RB tree.
323 void elv_rb_add(struct rb_root *root, struct request *rq)
325 struct rb_node **p = &root->rb_node;
326 struct rb_node *parent = NULL;
327 struct request *__rq;
331 __rq = rb_entry(parent, struct request, rb_node);
333 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
335 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
339 rb_link_node(&rq->rb_node, parent, p);
340 rb_insert_color(&rq->rb_node, root);
342 EXPORT_SYMBOL(elv_rb_add);
344 void elv_rb_del(struct rb_root *root, struct request *rq)
346 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
347 rb_erase(&rq->rb_node, root);
348 RB_CLEAR_NODE(&rq->rb_node);
350 EXPORT_SYMBOL(elv_rb_del);
352 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
354 struct rb_node *n = root->rb_node;
358 rq = rb_entry(n, struct request, rb_node);
360 if (sector < blk_rq_pos(rq))
362 else if (sector > blk_rq_pos(rq))
370 EXPORT_SYMBOL(elv_rb_find);
373 * Insert rq into dispatch queue of q. Queue lock must be held on
374 * entry. rq is sort instead into the dispatch queue. To be used by
375 * specific elevators.
377 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
380 struct list_head *entry;
382 if (q->last_merge == rq)
383 q->last_merge = NULL;
385 elv_rqhash_del(q, rq);
389 boundary = q->end_sector;
390 list_for_each_prev(entry, &q->queue_head) {
391 struct request *pos = list_entry_rq(entry);
393 if (req_op(rq) != req_op(pos))
395 if (rq_data_dir(rq) != rq_data_dir(pos))
397 if (pos->rq_flags & (RQF_STARTED | RQF_SOFTBARRIER))
399 if (blk_rq_pos(rq) >= boundary) {
400 if (blk_rq_pos(pos) < boundary)
403 if (blk_rq_pos(pos) >= boundary)
406 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
410 list_add(&rq->queuelist, entry);
412 EXPORT_SYMBOL(elv_dispatch_sort);
415 * Insert rq into dispatch queue of q. Queue lock must be held on
416 * entry. rq is added to the back of the dispatch queue. To be used by
417 * specific elevators.
419 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
421 if (q->last_merge == rq)
422 q->last_merge = NULL;
424 elv_rqhash_del(q, rq);
428 q->end_sector = rq_end_sector(rq);
430 list_add_tail(&rq->queuelist, &q->queue_head);
432 EXPORT_SYMBOL(elv_dispatch_add_tail);
434 enum elv_merge elv_merge(struct request_queue *q, struct request **req,
437 struct elevator_queue *e = q->elevator;
438 struct request *__rq;
442 * nomerges: No merges at all attempted
443 * noxmerges: Only simple one-hit cache try
444 * merges: All merge tries attempted
446 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
447 return ELEVATOR_NO_MERGE;
450 * First try one-hit cache.
452 if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
453 enum elv_merge ret = blk_try_merge(q->last_merge, bio);
455 if (ret != ELEVATOR_NO_MERGE) {
456 *req = q->last_merge;
461 if (blk_queue_noxmerges(q))
462 return ELEVATOR_NO_MERGE;
465 * See if our hash lookup can find a potential backmerge.
467 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
468 if (__rq && elv_bio_merge_ok(__rq, bio)) {
470 return ELEVATOR_BACK_MERGE;
473 if (e->uses_mq && e->type->ops.mq.request_merge)
474 return e->type->ops.mq.request_merge(q, req, bio);
475 else if (!e->uses_mq && e->type->ops.sq.elevator_merge_fn)
476 return e->type->ops.sq.elevator_merge_fn(q, req, bio);
478 return ELEVATOR_NO_MERGE;
482 * Attempt to do an insertion back merge. Only check for the case where
483 * we can append 'rq' to an existing request, so we can throw 'rq' away
486 * Returns true if we merged, false otherwise
488 bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq)
490 struct request *__rq;
493 if (blk_queue_nomerges(q))
497 * First try one-hit cache.
499 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
502 if (blk_queue_noxmerges(q))
507 * See if our hash lookup can find a potential backmerge.
510 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
511 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
514 /* The merged request could be merged with others, try again */
522 void elv_merged_request(struct request_queue *q, struct request *rq,
525 struct elevator_queue *e = q->elevator;
527 if (e->uses_mq && e->type->ops.mq.request_merged)
528 e->type->ops.mq.request_merged(q, rq, type);
529 else if (!e->uses_mq && e->type->ops.sq.elevator_merged_fn)
530 e->type->ops.sq.elevator_merged_fn(q, rq, type);
532 if (type == ELEVATOR_BACK_MERGE)
533 elv_rqhash_reposition(q, rq);
538 void elv_merge_requests(struct request_queue *q, struct request *rq,
539 struct request *next)
541 struct elevator_queue *e = q->elevator;
542 bool next_sorted = false;
544 if (e->uses_mq && e->type->ops.mq.requests_merged)
545 e->type->ops.mq.requests_merged(q, rq, next);
546 else if (e->type->ops.sq.elevator_merge_req_fn) {
547 next_sorted = (__force bool)(next->rq_flags & RQF_SORTED);
549 e->type->ops.sq.elevator_merge_req_fn(q, rq, next);
552 elv_rqhash_reposition(q, rq);
555 elv_rqhash_del(q, next);
562 void elv_bio_merged(struct request_queue *q, struct request *rq,
565 struct elevator_queue *e = q->elevator;
567 if (WARN_ON_ONCE(e->uses_mq))
570 if (e->type->ops.sq.elevator_bio_merged_fn)
571 e->type->ops.sq.elevator_bio_merged_fn(q, rq, bio);
575 static void blk_pm_requeue_request(struct request *rq)
577 if (rq->q->dev && !(rq->rq_flags & RQF_PM))
581 static void blk_pm_add_request(struct request_queue *q, struct request *rq)
583 if (q->dev && !(rq->rq_flags & RQF_PM) && q->nr_pending++ == 0 &&
584 (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
585 pm_request_resume(q->dev);
588 static inline void blk_pm_requeue_request(struct request *rq) {}
589 static inline void blk_pm_add_request(struct request_queue *q,
595 void elv_requeue_request(struct request_queue *q, struct request *rq)
598 * it already went through dequeue, we need to decrement the
599 * in_flight count again
601 if (blk_account_rq(rq)) {
602 q->in_flight[rq_is_sync(rq)]--;
603 if (rq->rq_flags & RQF_SORTED)
604 elv_deactivate_rq(q, rq);
607 rq->rq_flags &= ~RQF_STARTED;
609 blk_pm_requeue_request(rq);
611 __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
614 void elv_drain_elevator(struct request_queue *q)
616 struct elevator_queue *e = q->elevator;
619 if (WARN_ON_ONCE(e->uses_mq))
622 lockdep_assert_held(q->queue_lock);
624 while (e->type->ops.sq.elevator_dispatch_fn(q, 1))
626 if (q->nr_sorted && printed++ < 10) {
627 printk(KERN_ERR "%s: forced dispatching is broken "
628 "(nr_sorted=%u), please report this\n",
629 q->elevator->type->elevator_name, q->nr_sorted);
633 void __elv_add_request(struct request_queue *q, struct request *rq, int where)
635 trace_block_rq_insert(q, rq);
637 blk_pm_add_request(q, rq);
641 if (rq->rq_flags & RQF_SOFTBARRIER) {
642 /* barriers are scheduling boundary, update end_sector */
643 if (!blk_rq_is_passthrough(rq)) {
644 q->end_sector = rq_end_sector(rq);
647 } else if (!(rq->rq_flags & RQF_ELVPRIV) &&
648 (where == ELEVATOR_INSERT_SORT ||
649 where == ELEVATOR_INSERT_SORT_MERGE))
650 where = ELEVATOR_INSERT_BACK;
653 case ELEVATOR_INSERT_REQUEUE:
654 case ELEVATOR_INSERT_FRONT:
655 rq->rq_flags |= RQF_SOFTBARRIER;
656 list_add(&rq->queuelist, &q->queue_head);
659 case ELEVATOR_INSERT_BACK:
660 rq->rq_flags |= RQF_SOFTBARRIER;
661 elv_drain_elevator(q);
662 list_add_tail(&rq->queuelist, &q->queue_head);
664 * We kick the queue here for the following reasons.
665 * - The elevator might have returned NULL previously
666 * to delay requests and returned them now. As the
667 * queue wasn't empty before this request, ll_rw_blk
668 * won't run the queue on return, resulting in hang.
669 * - Usually, back inserted requests won't be merged
670 * with anything. There's no point in delaying queue
676 case ELEVATOR_INSERT_SORT_MERGE:
678 * If we succeed in merging this request with one in the
679 * queue already, we are done - rq has now been freed,
680 * so no need to do anything further.
682 if (elv_attempt_insert_merge(q, rq))
685 case ELEVATOR_INSERT_SORT:
686 BUG_ON(blk_rq_is_passthrough(rq));
687 rq->rq_flags |= RQF_SORTED;
689 if (rq_mergeable(rq)) {
690 elv_rqhash_add(q, rq);
696 * Some ioscheds (cfq) run q->request_fn directly, so
697 * rq cannot be accessed after calling
698 * elevator_add_req_fn.
700 q->elevator->type->ops.sq.elevator_add_req_fn(q, rq);
703 case ELEVATOR_INSERT_FLUSH:
704 rq->rq_flags |= RQF_SOFTBARRIER;
705 blk_insert_flush(rq);
708 printk(KERN_ERR "%s: bad insertion point %d\n",
713 EXPORT_SYMBOL(__elv_add_request);
715 void elv_add_request(struct request_queue *q, struct request *rq, int where)
719 spin_lock_irqsave(q->queue_lock, flags);
720 __elv_add_request(q, rq, where);
721 spin_unlock_irqrestore(q->queue_lock, flags);
723 EXPORT_SYMBOL(elv_add_request);
725 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
727 struct elevator_queue *e = q->elevator;
729 if (e->uses_mq && e->type->ops.mq.next_request)
730 return e->type->ops.mq.next_request(q, rq);
731 else if (!e->uses_mq && e->type->ops.sq.elevator_latter_req_fn)
732 return e->type->ops.sq.elevator_latter_req_fn(q, rq);
737 struct request *elv_former_request(struct request_queue *q, struct request *rq)
739 struct elevator_queue *e = q->elevator;
741 if (e->uses_mq && e->type->ops.mq.former_request)
742 return e->type->ops.mq.former_request(q, rq);
743 if (!e->uses_mq && e->type->ops.sq.elevator_former_req_fn)
744 return e->type->ops.sq.elevator_former_req_fn(q, rq);
748 int elv_set_request(struct request_queue *q, struct request *rq,
749 struct bio *bio, gfp_t gfp_mask)
751 struct elevator_queue *e = q->elevator;
753 if (WARN_ON_ONCE(e->uses_mq))
756 if (e->type->ops.sq.elevator_set_req_fn)
757 return e->type->ops.sq.elevator_set_req_fn(q, rq, bio, gfp_mask);
761 void elv_put_request(struct request_queue *q, struct request *rq)
763 struct elevator_queue *e = q->elevator;
765 if (WARN_ON_ONCE(e->uses_mq))
768 if (e->type->ops.sq.elevator_put_req_fn)
769 e->type->ops.sq.elevator_put_req_fn(rq);
772 int elv_may_queue(struct request_queue *q, unsigned int op)
774 struct elevator_queue *e = q->elevator;
776 if (WARN_ON_ONCE(e->uses_mq))
779 if (e->type->ops.sq.elevator_may_queue_fn)
780 return e->type->ops.sq.elevator_may_queue_fn(q, op);
782 return ELV_MQUEUE_MAY;
785 void elv_completed_request(struct request_queue *q, struct request *rq)
787 struct elevator_queue *e = q->elevator;
789 if (WARN_ON_ONCE(e->uses_mq))
793 * request is released from the driver, io must be done
795 if (blk_account_rq(rq)) {
796 q->in_flight[rq_is_sync(rq)]--;
797 if ((rq->rq_flags & RQF_SORTED) &&
798 e->type->ops.sq.elevator_completed_req_fn)
799 e->type->ops.sq.elevator_completed_req_fn(q, rq);
803 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
806 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
808 struct elv_fs_entry *entry = to_elv(attr);
809 struct elevator_queue *e;
815 e = container_of(kobj, struct elevator_queue, kobj);
816 mutex_lock(&e->sysfs_lock);
817 error = e->type ? entry->show(e, page) : -ENOENT;
818 mutex_unlock(&e->sysfs_lock);
823 elv_attr_store(struct kobject *kobj, struct attribute *attr,
824 const char *page, size_t length)
826 struct elv_fs_entry *entry = to_elv(attr);
827 struct elevator_queue *e;
833 e = container_of(kobj, struct elevator_queue, kobj);
834 mutex_lock(&e->sysfs_lock);
835 error = e->type ? entry->store(e, page, length) : -ENOENT;
836 mutex_unlock(&e->sysfs_lock);
840 static const struct sysfs_ops elv_sysfs_ops = {
841 .show = elv_attr_show,
842 .store = elv_attr_store,
845 static struct kobj_type elv_ktype = {
846 .sysfs_ops = &elv_sysfs_ops,
847 .release = elevator_release,
850 int elv_register_queue(struct request_queue *q)
852 struct elevator_queue *e = q->elevator;
855 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
857 struct elv_fs_entry *attr = e->type->elevator_attrs;
859 while (attr->attr.name) {
860 if (sysfs_create_file(&e->kobj, &attr->attr))
865 kobject_uevent(&e->kobj, KOBJ_ADD);
867 if (!e->uses_mq && e->type->ops.sq.elevator_registered_fn)
868 e->type->ops.sq.elevator_registered_fn(q);
872 EXPORT_SYMBOL(elv_register_queue);
874 void elv_unregister_queue(struct request_queue *q)
877 struct elevator_queue *e = q->elevator;
879 kobject_uevent(&e->kobj, KOBJ_REMOVE);
880 kobject_del(&e->kobj);
882 /* Re-enable throttling in case elevator disabled it */
883 wbt_enable_default(q);
886 EXPORT_SYMBOL(elv_unregister_queue);
888 int elv_register(struct elevator_type *e)
892 /* create icq_cache if requested */
894 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
895 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
898 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
899 "%s_io_cq", e->elevator_name);
900 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
901 e->icq_align, 0, NULL);
906 /* register, don't allow duplicate names */
907 spin_lock(&elv_list_lock);
908 if (elevator_find(e->elevator_name)) {
909 spin_unlock(&elv_list_lock);
911 kmem_cache_destroy(e->icq_cache);
914 list_add_tail(&e->list, &elv_list);
915 spin_unlock(&elv_list_lock);
917 /* print pretty message */
918 if (!strcmp(e->elevator_name, chosen_elevator) ||
919 (!*chosen_elevator &&
920 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
923 printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
927 EXPORT_SYMBOL_GPL(elv_register);
929 void elv_unregister(struct elevator_type *e)
932 spin_lock(&elv_list_lock);
933 list_del_init(&e->list);
934 spin_unlock(&elv_list_lock);
937 * Destroy icq_cache if it exists. icq's are RCU managed. Make
938 * sure all RCU operations are complete before proceeding.
942 kmem_cache_destroy(e->icq_cache);
946 EXPORT_SYMBOL_GPL(elv_unregister);
948 static int elevator_switch_mq(struct request_queue *q,
949 struct elevator_type *new_e)
953 blk_mq_freeze_queue(q);
956 if (q->elevator->registered)
957 elv_unregister_queue(q);
959 elevator_exit(q, q->elevator);
962 ret = blk_mq_init_sched(q, new_e);
967 ret = elv_register_queue(q);
969 elevator_exit(q, q->elevator);
975 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
977 blk_add_trace_msg(q, "elv switch: none");
980 blk_mq_unfreeze_queue(q);
985 * switch to new_e io scheduler. be careful not to introduce deadlocks -
986 * we don't free the old io scheduler, before we have allocated what we
987 * need for the new one. this way we have a chance of going back to the old
988 * one, if the new one fails init for some reason.
990 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
992 struct elevator_queue *old = q->elevator;
993 bool old_registered = false;
997 return elevator_switch_mq(q, new_e);
1000 * Turn on BYPASS and drain all requests w/ elevator private data.
1001 * Block layer doesn't call into a quiesced elevator - all requests
1002 * are directly put on the dispatch list without elevator data
1003 * using INSERT_BACK. All requests have SOFTBARRIER set and no
1004 * merge happens either.
1007 old_registered = old->registered;
1009 blk_queue_bypass_start(q);
1011 /* unregister and clear all auxiliary data of the old elevator */
1013 elv_unregister_queue(q);
1018 /* allocate, init and register new elevator */
1019 err = new_e->ops.sq.elevator_init_fn(q, new_e);
1023 err = elv_register_queue(q);
1027 /* done, kill the old one and finish */
1029 elevator_exit(q, old);
1030 blk_queue_bypass_end(q);
1033 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
1038 elevator_exit(q, q->elevator);
1040 /* switch failed, restore and re-register old elevator */
1043 elv_register_queue(q);
1044 blk_queue_bypass_end(q);
1051 * Switch this queue to the given IO scheduler.
1053 static int __elevator_change(struct request_queue *q, const char *name)
1055 char elevator_name[ELV_NAME_MAX];
1056 struct elevator_type *e;
1059 * Special case for mq, turn off scheduling
1061 if (q->mq_ops && !strncmp(name, "none", 4))
1062 return elevator_switch(q, NULL);
1064 strlcpy(elevator_name, name, sizeof(elevator_name));
1065 e = elevator_get(strstrip(elevator_name), true);
1070 !strcmp(elevator_name, q->elevator->type->elevator_name)) {
1075 if (!e->uses_mq && q->mq_ops) {
1079 if (e->uses_mq && !q->mq_ops) {
1084 return elevator_switch(q, e);
1087 static inline bool elv_support_iosched(struct request_queue *q)
1089 if (q->mq_ops && q->tag_set && (q->tag_set->flags &
1095 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1100 if (!(q->mq_ops || q->request_fn) || !elv_support_iosched(q))
1103 ret = __elevator_change(q, name);
1110 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1112 struct elevator_queue *e = q->elevator;
1113 struct elevator_type *elv = NULL;
1114 struct elevator_type *__e;
1117 if (!blk_queue_stackable(q))
1118 return sprintf(name, "none\n");
1121 len += sprintf(name+len, "[none] ");
1125 spin_lock(&elv_list_lock);
1126 list_for_each_entry(__e, &elv_list, list) {
1127 if (elv && !strcmp(elv->elevator_name, __e->elevator_name)) {
1128 len += sprintf(name+len, "[%s] ", elv->elevator_name);
1131 if (__e->uses_mq && q->mq_ops && elv_support_iosched(q))
1132 len += sprintf(name+len, "%s ", __e->elevator_name);
1133 else if (!__e->uses_mq && !q->mq_ops)
1134 len += sprintf(name+len, "%s ", __e->elevator_name);
1136 spin_unlock(&elv_list_lock);
1138 if (q->mq_ops && q->elevator)
1139 len += sprintf(name+len, "none");
1141 len += sprintf(len+name, "\n");
1145 struct request *elv_rb_former_request(struct request_queue *q,
1148 struct rb_node *rbprev = rb_prev(&rq->rb_node);
1151 return rb_entry_rq(rbprev);
1155 EXPORT_SYMBOL(elv_rb_former_request);
1157 struct request *elv_rb_latter_request(struct request_queue *q,
1160 struct rb_node *rbnext = rb_next(&rq->rb_node);
1163 return rb_entry_rq(rbnext);
1167 EXPORT_SYMBOL(elv_rb_latter_request);