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/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>
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 if the elevator @e name matches @name and if @e provides all
99 * the features specified 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)
193 struct elevator_queue *e = q->elevator;
195 mutex_lock(&e->sysfs_lock);
196 blk_mq_exit_sched(q, e);
197 mutex_unlock(&e->sysfs_lock);
199 kobject_put(&e->kobj);
202 static inline void __elv_rqhash_del(struct request *rq)
205 rq->rq_flags &= ~RQF_HASHED;
208 void elv_rqhash_del(struct request_queue *q, struct request *rq)
211 __elv_rqhash_del(rq);
213 EXPORT_SYMBOL_GPL(elv_rqhash_del);
215 void elv_rqhash_add(struct request_queue *q, struct request *rq)
217 struct elevator_queue *e = q->elevator;
219 BUG_ON(ELV_ON_HASH(rq));
220 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
221 rq->rq_flags |= RQF_HASHED;
223 EXPORT_SYMBOL_GPL(elv_rqhash_add);
225 void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
227 __elv_rqhash_del(rq);
228 elv_rqhash_add(q, rq);
231 struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
233 struct elevator_queue *e = q->elevator;
234 struct hlist_node *next;
237 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
238 BUG_ON(!ELV_ON_HASH(rq));
240 if (unlikely(!rq_mergeable(rq))) {
241 __elv_rqhash_del(rq);
245 if (rq_hash_key(rq) == offset)
253 * RB-tree support functions for inserting/lookup/removal of requests
254 * in a sorted RB tree.
256 void elv_rb_add(struct rb_root *root, struct request *rq)
258 struct rb_node **p = &root->rb_node;
259 struct rb_node *parent = NULL;
260 struct request *__rq;
264 __rq = rb_entry(parent, struct request, rb_node);
266 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
268 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
272 rb_link_node(&rq->rb_node, parent, p);
273 rb_insert_color(&rq->rb_node, root);
275 EXPORT_SYMBOL(elv_rb_add);
277 void elv_rb_del(struct rb_root *root, struct request *rq)
279 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
280 rb_erase(&rq->rb_node, root);
281 RB_CLEAR_NODE(&rq->rb_node);
283 EXPORT_SYMBOL(elv_rb_del);
285 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
287 struct rb_node *n = root->rb_node;
291 rq = rb_entry(n, struct request, rb_node);
293 if (sector < blk_rq_pos(rq))
295 else if (sector > blk_rq_pos(rq))
303 EXPORT_SYMBOL(elv_rb_find);
305 enum elv_merge elv_merge(struct request_queue *q, struct request **req,
308 struct elevator_queue *e = q->elevator;
309 struct request *__rq;
313 * nomerges: No merges at all attempted
314 * noxmerges: Only simple one-hit cache try
315 * merges: All merge tries attempted
317 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
318 return ELEVATOR_NO_MERGE;
321 * First try one-hit cache.
323 if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
324 enum elv_merge ret = blk_try_merge(q->last_merge, bio);
326 if (ret != ELEVATOR_NO_MERGE) {
327 *req = q->last_merge;
332 if (blk_queue_noxmerges(q))
333 return ELEVATOR_NO_MERGE;
336 * See if our hash lookup can find a potential backmerge.
338 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
339 if (__rq && elv_bio_merge_ok(__rq, bio)) {
342 if (blk_discard_mergable(__rq))
343 return ELEVATOR_DISCARD_MERGE;
344 return ELEVATOR_BACK_MERGE;
347 if (e->type->ops.request_merge)
348 return e->type->ops.request_merge(q, req, bio);
350 return ELEVATOR_NO_MERGE;
354 * Attempt to do an insertion back merge. Only check for the case where
355 * we can append 'rq' to an existing request, so we can throw 'rq' away
358 * Returns true if we merged, false otherwise. 'free' will contain all
359 * requests that need to be freed.
361 bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq,
362 struct list_head *free)
364 struct request *__rq;
367 if (blk_queue_nomerges(q))
371 * First try one-hit cache.
373 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) {
374 list_add(&rq->queuelist, free);
378 if (blk_queue_noxmerges(q))
383 * See if our hash lookup can find a potential backmerge.
386 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
387 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
390 list_add(&rq->queuelist, free);
391 /* The merged request could be merged with others, try again */
399 void elv_merged_request(struct request_queue *q, struct request *rq,
402 struct elevator_queue *e = q->elevator;
404 if (e->type->ops.request_merged)
405 e->type->ops.request_merged(q, rq, type);
407 if (type == ELEVATOR_BACK_MERGE)
408 elv_rqhash_reposition(q, rq);
413 void elv_merge_requests(struct request_queue *q, struct request *rq,
414 struct request *next)
416 struct elevator_queue *e = q->elevator;
418 if (e->type->ops.requests_merged)
419 e->type->ops.requests_merged(q, rq, next);
421 elv_rqhash_reposition(q, rq);
425 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
427 struct elevator_queue *e = q->elevator;
429 if (e->type->ops.next_request)
430 return e->type->ops.next_request(q, rq);
435 struct request *elv_former_request(struct request_queue *q, struct request *rq)
437 struct elevator_queue *e = q->elevator;
439 if (e->type->ops.former_request)
440 return e->type->ops.former_request(q, rq);
445 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
448 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
450 struct elv_fs_entry *entry = to_elv(attr);
451 struct elevator_queue *e;
457 e = container_of(kobj, struct elevator_queue, kobj);
458 mutex_lock(&e->sysfs_lock);
459 error = e->type ? entry->show(e, page) : -ENOENT;
460 mutex_unlock(&e->sysfs_lock);
465 elv_attr_store(struct kobject *kobj, struct attribute *attr,
466 const char *page, size_t length)
468 struct elv_fs_entry *entry = to_elv(attr);
469 struct elevator_queue *e;
475 e = container_of(kobj, struct elevator_queue, kobj);
476 mutex_lock(&e->sysfs_lock);
477 error = e->type ? entry->store(e, page, length) : -ENOENT;
478 mutex_unlock(&e->sysfs_lock);
482 static const struct sysfs_ops elv_sysfs_ops = {
483 .show = elv_attr_show,
484 .store = elv_attr_store,
487 static struct kobj_type elv_ktype = {
488 .sysfs_ops = &elv_sysfs_ops,
489 .release = elevator_release,
492 int elv_register_queue(struct request_queue *q, bool uevent)
494 struct elevator_queue *e = q->elevator;
497 lockdep_assert_held(&q->sysfs_lock);
499 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
501 struct elv_fs_entry *attr = e->type->elevator_attrs;
503 while (attr->attr.name) {
504 if (sysfs_create_file(&e->kobj, &attr->attr))
510 kobject_uevent(&e->kobj, KOBJ_ADD);
517 void elv_unregister_queue(struct request_queue *q)
519 lockdep_assert_held(&q->sysfs_lock);
522 struct elevator_queue *e = q->elevator;
524 kobject_uevent(&e->kobj, KOBJ_REMOVE);
525 kobject_del(&e->kobj);
531 int elv_register(struct elevator_type *e)
533 /* insert_requests and dispatch_request are mandatory */
534 if (WARN_ON_ONCE(!e->ops.insert_requests || !e->ops.dispatch_request))
537 /* create icq_cache if requested */
539 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
540 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
543 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
544 "%s_io_cq", e->elevator_name);
545 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
546 e->icq_align, 0, NULL);
551 /* register, don't allow duplicate names */
552 spin_lock(&elv_list_lock);
553 if (elevator_find(e->elevator_name, 0)) {
554 spin_unlock(&elv_list_lock);
555 kmem_cache_destroy(e->icq_cache);
558 list_add_tail(&e->list, &elv_list);
559 spin_unlock(&elv_list_lock);
561 printk(KERN_INFO "io scheduler %s registered\n", e->elevator_name);
565 EXPORT_SYMBOL_GPL(elv_register);
567 void elv_unregister(struct elevator_type *e)
570 spin_lock(&elv_list_lock);
571 list_del_init(&e->list);
572 spin_unlock(&elv_list_lock);
575 * Destroy icq_cache if it exists. icq's are RCU managed. Make
576 * sure all RCU operations are complete before proceeding.
580 kmem_cache_destroy(e->icq_cache);
584 EXPORT_SYMBOL_GPL(elv_unregister);
586 int elevator_switch_mq(struct request_queue *q,
587 struct elevator_type *new_e)
591 lockdep_assert_held(&q->sysfs_lock);
594 if (q->elevator->registered)
595 elv_unregister_queue(q);
598 blk_mq_sched_free_rqs(q);
602 ret = blk_mq_init_sched(q, new_e);
607 ret = elv_register_queue(q, true);
609 blk_mq_sched_free_rqs(q);
616 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
618 blk_add_trace_msg(q, "elv switch: none");
624 static inline bool elv_support_iosched(struct request_queue *q)
626 if (!queue_is_mq(q) ||
627 (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED)))
633 * For single queue devices, default to using mq-deadline. If we have multiple
634 * queues or mq-deadline is not available, default to "none".
636 static struct elevator_type *elevator_get_default(struct request_queue *q)
638 if (q->tag_set && q->tag_set->flags & BLK_MQ_F_NO_SCHED_BY_DEFAULT)
641 if (q->nr_hw_queues != 1 &&
642 !blk_mq_is_shared_tags(q->tag_set->flags))
645 return elevator_get(q, "mq-deadline", false);
649 * Get the first elevator providing the features required by the request queue.
650 * Default to "none" if no matching elevator is found.
652 static struct elevator_type *elevator_get_by_features(struct request_queue *q)
654 struct elevator_type *e, *found = NULL;
656 spin_lock(&elv_list_lock);
658 list_for_each_entry(e, &elv_list, list) {
659 if (elv_support_features(e->elevator_features,
660 q->required_elevator_features)) {
666 if (found && !try_module_get(found->elevator_owner))
669 spin_unlock(&elv_list_lock);
674 * For a device queue that has no required features, use the default elevator
675 * settings. Otherwise, use the first elevator available matching the required
676 * features. If no suitable elevator is find or if the chosen elevator
677 * initialization fails, fall back to the "none" elevator (no elevator).
679 void elevator_init_mq(struct request_queue *q)
681 struct elevator_type *e;
684 if (!elv_support_iosched(q))
687 WARN_ON_ONCE(blk_queue_registered(q));
689 if (unlikely(q->elevator))
692 if (!q->required_elevator_features)
693 e = elevator_get_default(q);
695 e = elevator_get_by_features(q);
700 * We are called before adding disk, when there isn't any FS I/O,
701 * so freezing queue plus canceling dispatch work is enough to
702 * drain any dispatch activities originated from passthrough
703 * requests, then no need to quiesce queue which may add long boot
704 * latency, especially when lots of disks are involved.
706 blk_mq_freeze_queue(q);
707 blk_mq_cancel_work_sync(q);
709 err = blk_mq_init_sched(q, e);
711 blk_mq_unfreeze_queue(q);
714 pr_warn("\"%s\" elevator initialization failed, "
715 "falling back to \"none\"\n", e->elevator_name);
721 * switch to new_e io scheduler. be careful not to introduce deadlocks -
722 * we don't free the old io scheduler, before we have allocated what we
723 * need for the new one. this way we have a chance of going back to the old
724 * one, if the new one fails init for some reason.
726 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
730 lockdep_assert_held(&q->sysfs_lock);
732 blk_mq_freeze_queue(q);
733 blk_mq_quiesce_queue(q);
735 err = elevator_switch_mq(q, new_e);
737 blk_mq_unquiesce_queue(q);
738 blk_mq_unfreeze_queue(q);
744 * Switch this queue to the given IO scheduler.
746 static int __elevator_change(struct request_queue *q, const char *name)
748 char elevator_name[ELV_NAME_MAX];
749 struct elevator_type *e;
751 /* Make sure queue is not in the middle of being removed */
752 if (!blk_queue_registered(q))
756 * Special case for mq, turn off scheduling
758 if (!strncmp(name, "none", 4)) {
761 return elevator_switch(q, NULL);
764 strlcpy(elevator_name, name, sizeof(elevator_name));
765 e = elevator_get(q, strstrip(elevator_name), true);
770 elevator_match(q->elevator->type, elevator_name, 0)) {
775 return elevator_switch(q, e);
778 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
783 if (!elv_support_iosched(q))
786 ret = __elevator_change(q, name);
793 ssize_t elv_iosched_show(struct request_queue *q, char *name)
795 struct elevator_queue *e = q->elevator;
796 struct elevator_type *elv = NULL;
797 struct elevator_type *__e;
801 return sprintf(name, "none\n");
804 len += sprintf(name+len, "[none] ");
808 spin_lock(&elv_list_lock);
809 list_for_each_entry(__e, &elv_list, list) {
810 if (elv && elevator_match(elv, __e->elevator_name, 0)) {
811 len += sprintf(name+len, "[%s] ", elv->elevator_name);
814 if (elv_support_iosched(q) &&
815 elevator_match(__e, __e->elevator_name,
816 q->required_elevator_features))
817 len += sprintf(name+len, "%s ", __e->elevator_name);
819 spin_unlock(&elv_list_lock);
822 len += sprintf(name+len, "none");
824 len += sprintf(len+name, "\n");
828 struct request *elv_rb_former_request(struct request_queue *q,
831 struct rb_node *rbprev = rb_prev(&rq->rb_node);
834 return rb_entry_rq(rbprev);
838 EXPORT_SYMBOL(elv_rb_former_request);
840 struct request *elv_rb_latter_request(struct request_queue *q,
843 struct rb_node *rbnext = rb_next(&rq->rb_node);
846 return rb_entry_rq(rbnext);
850 EXPORT_SYMBOL(elv_rb_latter_request);
852 static int __init elevator_setup(char *str)
854 pr_warn("Kernel parameter elevator= does not have any effect anymore.\n"
855 "Please use sysfs to set IO scheduler for individual devices.\n");
859 __setup("elevator=", elevator_setup);