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>
39 #include <trace/events/block.h>
43 #include "blk-mq-sched.h"
46 #include "blk-cgroup.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(struct request_queue *q,
87 const struct elevator_type *e)
89 return (q->required_elevator_features & e->elevator_features) ==
90 q->required_elevator_features;
94 * elevator_match - Test an elevator name and features
95 * @e: Scheduler to test
96 * @name: Elevator name to test
97 * @required_features: Features that the elevator must provide
99 * Return true if the elevator @e name matches @name and if @e provides all
100 * the features specified by @required_features.
102 static bool elevator_match(const struct elevator_type *e, const char *name)
104 return !strcmp(e->elevator_name, name) ||
105 (e->elevator_alias && !strcmp(e->elevator_alias, name));
108 static struct elevator_type *__elevator_find(const char *name)
110 struct elevator_type *e;
112 list_for_each_entry(e, &elv_list, list)
113 if (elevator_match(e, name))
118 static struct elevator_type *elevator_find_get(struct request_queue *q,
121 struct elevator_type *e;
123 spin_lock(&elv_list_lock);
124 e = __elevator_find(name);
125 if (e && (!elv_support_features(q, e) || !elevator_tryget(e)))
127 spin_unlock(&elv_list_lock);
131 static struct kobj_type elv_ktype;
133 struct elevator_queue *elevator_alloc(struct request_queue *q,
134 struct elevator_type *e)
136 struct elevator_queue *eq;
138 eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
144 kobject_init(&eq->kobj, &elv_ktype);
145 mutex_init(&eq->sysfs_lock);
150 EXPORT_SYMBOL(elevator_alloc);
152 static void elevator_release(struct kobject *kobj)
154 struct elevator_queue *e;
156 e = container_of(kobj, struct elevator_queue, kobj);
157 elevator_put(e->type);
161 void elevator_exit(struct request_queue *q)
163 struct elevator_queue *e = q->elevator;
166 blk_mq_sched_free_rqs(q);
168 mutex_lock(&e->sysfs_lock);
169 blk_mq_exit_sched(q, e);
170 mutex_unlock(&e->sysfs_lock);
172 kobject_put(&e->kobj);
175 static inline void __elv_rqhash_del(struct request *rq)
178 rq->rq_flags &= ~RQF_HASHED;
181 void elv_rqhash_del(struct request_queue *q, struct request *rq)
184 __elv_rqhash_del(rq);
186 EXPORT_SYMBOL_GPL(elv_rqhash_del);
188 void elv_rqhash_add(struct request_queue *q, struct request *rq)
190 struct elevator_queue *e = q->elevator;
192 BUG_ON(ELV_ON_HASH(rq));
193 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
194 rq->rq_flags |= RQF_HASHED;
196 EXPORT_SYMBOL_GPL(elv_rqhash_add);
198 void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
200 __elv_rqhash_del(rq);
201 elv_rqhash_add(q, rq);
204 struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
206 struct elevator_queue *e = q->elevator;
207 struct hlist_node *next;
210 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
211 BUG_ON(!ELV_ON_HASH(rq));
213 if (unlikely(!rq_mergeable(rq))) {
214 __elv_rqhash_del(rq);
218 if (rq_hash_key(rq) == offset)
226 * RB-tree support functions for inserting/lookup/removal of requests
227 * in a sorted RB tree.
229 void elv_rb_add(struct rb_root *root, struct request *rq)
231 struct rb_node **p = &root->rb_node;
232 struct rb_node *parent = NULL;
233 struct request *__rq;
237 __rq = rb_entry(parent, struct request, rb_node);
239 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
241 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
245 rb_link_node(&rq->rb_node, parent, p);
246 rb_insert_color(&rq->rb_node, root);
248 EXPORT_SYMBOL(elv_rb_add);
250 void elv_rb_del(struct rb_root *root, struct request *rq)
252 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
253 rb_erase(&rq->rb_node, root);
254 RB_CLEAR_NODE(&rq->rb_node);
256 EXPORT_SYMBOL(elv_rb_del);
258 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
260 struct rb_node *n = root->rb_node;
264 rq = rb_entry(n, struct request, rb_node);
266 if (sector < blk_rq_pos(rq))
268 else if (sector > blk_rq_pos(rq))
276 EXPORT_SYMBOL(elv_rb_find);
278 enum elv_merge elv_merge(struct request_queue *q, struct request **req,
281 struct elevator_queue *e = q->elevator;
282 struct request *__rq;
286 * nomerges: No merges at all attempted
287 * noxmerges: Only simple one-hit cache try
288 * merges: All merge tries attempted
290 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
291 return ELEVATOR_NO_MERGE;
294 * First try one-hit cache.
296 if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
297 enum elv_merge ret = blk_try_merge(q->last_merge, bio);
299 if (ret != ELEVATOR_NO_MERGE) {
300 *req = q->last_merge;
305 if (blk_queue_noxmerges(q))
306 return ELEVATOR_NO_MERGE;
309 * See if our hash lookup can find a potential backmerge.
311 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
312 if (__rq && elv_bio_merge_ok(__rq, bio)) {
315 if (blk_discard_mergable(__rq))
316 return ELEVATOR_DISCARD_MERGE;
317 return ELEVATOR_BACK_MERGE;
320 if (e->type->ops.request_merge)
321 return e->type->ops.request_merge(q, req, bio);
323 return ELEVATOR_NO_MERGE;
327 * Attempt to do an insertion back merge. Only check for the case where
328 * we can append 'rq' to an existing request, so we can throw 'rq' away
331 * Returns true if we merged, false otherwise. 'free' will contain all
332 * requests that need to be freed.
334 bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq,
335 struct list_head *free)
337 struct request *__rq;
340 if (blk_queue_nomerges(q))
344 * First try one-hit cache.
346 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) {
347 list_add(&rq->queuelist, free);
351 if (blk_queue_noxmerges(q))
356 * See if our hash lookup can find a potential backmerge.
359 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
360 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
363 list_add(&rq->queuelist, free);
364 /* The merged request could be merged with others, try again */
372 void elv_merged_request(struct request_queue *q, struct request *rq,
375 struct elevator_queue *e = q->elevator;
377 if (e->type->ops.request_merged)
378 e->type->ops.request_merged(q, rq, type);
380 if (type == ELEVATOR_BACK_MERGE)
381 elv_rqhash_reposition(q, rq);
386 void elv_merge_requests(struct request_queue *q, struct request *rq,
387 struct request *next)
389 struct elevator_queue *e = q->elevator;
391 if (e->type->ops.requests_merged)
392 e->type->ops.requests_merged(q, rq, next);
394 elv_rqhash_reposition(q, rq);
398 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
400 struct elevator_queue *e = q->elevator;
402 if (e->type->ops.next_request)
403 return e->type->ops.next_request(q, rq);
408 struct request *elv_former_request(struct request_queue *q, struct request *rq)
410 struct elevator_queue *e = q->elevator;
412 if (e->type->ops.former_request)
413 return e->type->ops.former_request(q, rq);
418 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
421 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
423 struct elv_fs_entry *entry = to_elv(attr);
424 struct elevator_queue *e;
430 e = container_of(kobj, struct elevator_queue, kobj);
431 mutex_lock(&e->sysfs_lock);
432 error = e->type ? entry->show(e, page) : -ENOENT;
433 mutex_unlock(&e->sysfs_lock);
438 elv_attr_store(struct kobject *kobj, struct attribute *attr,
439 const char *page, size_t length)
441 struct elv_fs_entry *entry = to_elv(attr);
442 struct elevator_queue *e;
448 e = container_of(kobj, struct elevator_queue, kobj);
449 mutex_lock(&e->sysfs_lock);
450 error = e->type ? entry->store(e, page, length) : -ENOENT;
451 mutex_unlock(&e->sysfs_lock);
455 static const struct sysfs_ops elv_sysfs_ops = {
456 .show = elv_attr_show,
457 .store = elv_attr_store,
460 static struct kobj_type elv_ktype = {
461 .sysfs_ops = &elv_sysfs_ops,
462 .release = elevator_release,
465 int elv_register_queue(struct request_queue *q, bool uevent)
467 struct elevator_queue *e = q->elevator;
470 lockdep_assert_held(&q->sysfs_lock);
472 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
474 struct elv_fs_entry *attr = e->type->elevator_attrs;
476 while (attr->attr.name) {
477 if (sysfs_create_file(&e->kobj, &attr->attr))
483 kobject_uevent(&e->kobj, KOBJ_ADD);
485 set_bit(ELEVATOR_FLAG_REGISTERED, &e->flags);
490 void elv_unregister_queue(struct request_queue *q)
492 struct elevator_queue *e = q->elevator;
494 lockdep_assert_held(&q->sysfs_lock);
496 if (e && test_and_clear_bit(ELEVATOR_FLAG_REGISTERED, &e->flags)) {
497 kobject_uevent(&e->kobj, KOBJ_REMOVE);
498 kobject_del(&e->kobj);
502 int elv_register(struct elevator_type *e)
504 /* insert_requests and dispatch_request are mandatory */
505 if (WARN_ON_ONCE(!e->ops.insert_requests || !e->ops.dispatch_request))
508 /* create icq_cache if requested */
510 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
511 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
514 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
515 "%s_io_cq", e->elevator_name);
516 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
517 e->icq_align, 0, NULL);
522 /* register, don't allow duplicate names */
523 spin_lock(&elv_list_lock);
524 if (__elevator_find(e->elevator_name)) {
525 spin_unlock(&elv_list_lock);
526 kmem_cache_destroy(e->icq_cache);
529 list_add_tail(&e->list, &elv_list);
530 spin_unlock(&elv_list_lock);
532 printk(KERN_INFO "io scheduler %s registered\n", e->elevator_name);
536 EXPORT_SYMBOL_GPL(elv_register);
538 void elv_unregister(struct elevator_type *e)
541 spin_lock(&elv_list_lock);
542 list_del_init(&e->list);
543 spin_unlock(&elv_list_lock);
546 * Destroy icq_cache if it exists. icq's are RCU managed. Make
547 * sure all RCU operations are complete before proceeding.
551 kmem_cache_destroy(e->icq_cache);
555 EXPORT_SYMBOL_GPL(elv_unregister);
557 static int elevator_switch_mq(struct request_queue *q,
558 struct elevator_type *new_e)
562 lockdep_assert_held(&q->sysfs_lock);
565 elv_unregister_queue(q);
569 ret = blk_mq_init_sched(q, new_e);
574 ret = elv_register_queue(q, true);
582 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
584 blk_add_trace_msg(q, "elv switch: none");
590 static inline bool elv_support_iosched(struct request_queue *q)
592 if (!queue_is_mq(q) ||
593 (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED)))
599 * For single queue devices, default to using mq-deadline. If we have multiple
600 * queues or mq-deadline is not available, default to "none".
602 static struct elevator_type *elevator_get_default(struct request_queue *q)
604 if (q->tag_set && q->tag_set->flags & BLK_MQ_F_NO_SCHED_BY_DEFAULT)
607 if (q->nr_hw_queues != 1 &&
608 !blk_mq_is_shared_tags(q->tag_set->flags))
611 return elevator_find_get(q, "mq-deadline");
615 * Get the first elevator providing the features required by the request queue.
616 * Default to "none" if no matching elevator is found.
618 static struct elevator_type *elevator_get_by_features(struct request_queue *q)
620 struct elevator_type *e, *found = NULL;
622 spin_lock(&elv_list_lock);
624 list_for_each_entry(e, &elv_list, list) {
625 if (elv_support_features(q, e)) {
631 if (found && !elevator_tryget(found))
634 spin_unlock(&elv_list_lock);
639 * For a device queue that has no required features, use the default elevator
640 * settings. Otherwise, use the first elevator available matching the required
641 * features. If no suitable elevator is find or if the chosen elevator
642 * initialization fails, fall back to the "none" elevator (no elevator).
644 void elevator_init_mq(struct request_queue *q)
646 struct elevator_type *e;
649 if (!elv_support_iosched(q))
652 WARN_ON_ONCE(blk_queue_registered(q));
654 if (unlikely(q->elevator))
657 if (!q->required_elevator_features)
658 e = elevator_get_default(q);
660 e = elevator_get_by_features(q);
665 * We are called before adding disk, when there isn't any FS I/O,
666 * so freezing queue plus canceling dispatch work is enough to
667 * drain any dispatch activities originated from passthrough
668 * requests, then no need to quiesce queue which may add long boot
669 * latency, especially when lots of disks are involved.
671 blk_mq_freeze_queue(q);
672 blk_mq_cancel_work_sync(q);
674 err = blk_mq_init_sched(q, e);
676 blk_mq_unfreeze_queue(q);
679 pr_warn("\"%s\" elevator initialization failed, "
680 "falling back to \"none\"\n", e->elevator_name);
687 * switch to new_e io scheduler. be careful not to introduce deadlocks -
688 * we don't free the old io scheduler, before we have allocated what we
689 * need for the new one. this way we have a chance of going back to the old
690 * one, if the new one fails init for some reason.
692 int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
696 lockdep_assert_held(&q->sysfs_lock);
698 blk_mq_freeze_queue(q);
699 blk_mq_quiesce_queue(q);
701 err = elevator_switch_mq(q, new_e);
703 blk_mq_unquiesce_queue(q);
704 blk_mq_unfreeze_queue(q);
710 * Switch this queue to the given IO scheduler.
712 static int elevator_change(struct request_queue *q, const char *elevator_name)
714 struct elevator_type *e;
717 /* Make sure queue is not in the middle of being removed */
718 if (!blk_queue_registered(q))
722 * Special case for mq, turn off scheduling
724 if (!strncmp(elevator_name, "none", 4)) {
727 return elevator_switch(q, NULL);
730 if (q->elevator && elevator_match(q->elevator->type, elevator_name))
733 e = elevator_find_get(q, elevator_name);
735 request_module("%s-iosched", elevator_name);
736 e = elevator_find_get(q, elevator_name);
740 ret = elevator_switch(q, e);
745 ssize_t elv_iosched_store(struct request_queue *q, const char *buf,
748 char elevator_name[ELV_NAME_MAX];
751 if (!elv_support_iosched(q))
754 strlcpy(elevator_name, buf, sizeof(elevator_name));
755 ret = elevator_change(q, strstrip(elevator_name));
761 ssize_t elv_iosched_show(struct request_queue *q, char *name)
763 struct elevator_queue *eq = q->elevator;
764 struct elevator_type *cur = NULL, *e;
767 if (!elv_support_iosched(q))
768 return sprintf(name, "none\n");
771 len += sprintf(name+len, "[none] ");
775 spin_lock(&elv_list_lock);
776 list_for_each_entry(e, &elv_list, list) {
778 len += sprintf(name+len, "[%s] ", cur->elevator_name);
781 if (elv_support_features(q, e))
782 len += sprintf(name+len, "%s ", e->elevator_name);
784 spin_unlock(&elv_list_lock);
787 len += sprintf(name+len, "none");
789 len += sprintf(len+name, "\n");
793 struct request *elv_rb_former_request(struct request_queue *q,
796 struct rb_node *rbprev = rb_prev(&rq->rb_node);
799 return rb_entry_rq(rbprev);
803 EXPORT_SYMBOL(elv_rb_former_request);
805 struct request *elv_rb_latter_request(struct request_queue *q,
808 struct rb_node *rbnext = rb_next(&rq->rb_node);
811 return rb_entry_rq(rbnext);
815 EXPORT_SYMBOL(elv_rb_latter_request);
817 static int __init elevator_setup(char *str)
819 pr_warn("Kernel parameter elevator= does not have any effect anymore.\n"
820 "Please use sysfs to set IO scheduler for individual devices.\n");
824 __setup("elevator=", elevator_setup);