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, struct elevator_queue *e)
193 mutex_lock(&e->sysfs_lock);
194 blk_mq_exit_sched(q, e);
195 mutex_unlock(&e->sysfs_lock);
197 kobject_put(&e->kobj);
200 static inline void __elv_rqhash_del(struct request *rq)
203 rq->rq_flags &= ~RQF_HASHED;
206 void elv_rqhash_del(struct request_queue *q, struct request *rq)
209 __elv_rqhash_del(rq);
211 EXPORT_SYMBOL_GPL(elv_rqhash_del);
213 void elv_rqhash_add(struct request_queue *q, struct request *rq)
215 struct elevator_queue *e = q->elevator;
217 BUG_ON(ELV_ON_HASH(rq));
218 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
219 rq->rq_flags |= RQF_HASHED;
221 EXPORT_SYMBOL_GPL(elv_rqhash_add);
223 void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
225 __elv_rqhash_del(rq);
226 elv_rqhash_add(q, rq);
229 struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
231 struct elevator_queue *e = q->elevator;
232 struct hlist_node *next;
235 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
236 BUG_ON(!ELV_ON_HASH(rq));
238 if (unlikely(!rq_mergeable(rq))) {
239 __elv_rqhash_del(rq);
243 if (rq_hash_key(rq) == offset)
251 * RB-tree support functions for inserting/lookup/removal of requests
252 * in a sorted RB tree.
254 void elv_rb_add(struct rb_root *root, struct request *rq)
256 struct rb_node **p = &root->rb_node;
257 struct rb_node *parent = NULL;
258 struct request *__rq;
262 __rq = rb_entry(parent, struct request, rb_node);
264 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
266 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
270 rb_link_node(&rq->rb_node, parent, p);
271 rb_insert_color(&rq->rb_node, root);
273 EXPORT_SYMBOL(elv_rb_add);
275 void elv_rb_del(struct rb_root *root, struct request *rq)
277 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
278 rb_erase(&rq->rb_node, root);
279 RB_CLEAR_NODE(&rq->rb_node);
281 EXPORT_SYMBOL(elv_rb_del);
283 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
285 struct rb_node *n = root->rb_node;
289 rq = rb_entry(n, struct request, rb_node);
291 if (sector < blk_rq_pos(rq))
293 else if (sector > blk_rq_pos(rq))
301 EXPORT_SYMBOL(elv_rb_find);
303 enum elv_merge elv_merge(struct request_queue *q, struct request **req,
306 struct elevator_queue *e = q->elevator;
307 struct request *__rq;
311 * nomerges: No merges at all attempted
312 * noxmerges: Only simple one-hit cache try
313 * merges: All merge tries attempted
315 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
316 return ELEVATOR_NO_MERGE;
319 * First try one-hit cache.
321 if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
322 enum elv_merge ret = blk_try_merge(q->last_merge, bio);
324 if (ret != ELEVATOR_NO_MERGE) {
325 *req = q->last_merge;
330 if (blk_queue_noxmerges(q))
331 return ELEVATOR_NO_MERGE;
334 * See if our hash lookup can find a potential backmerge.
336 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
337 if (__rq && elv_bio_merge_ok(__rq, bio)) {
340 if (blk_discard_mergable(__rq))
341 return ELEVATOR_DISCARD_MERGE;
342 return ELEVATOR_BACK_MERGE;
345 if (e->type->ops.request_merge)
346 return e->type->ops.request_merge(q, req, bio);
348 return ELEVATOR_NO_MERGE;
352 * Attempt to do an insertion back merge. Only check for the case where
353 * we can append 'rq' to an existing request, so we can throw 'rq' away
356 * Returns true if we merged, false otherwise. 'free' will contain all
357 * requests that need to be freed.
359 bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq,
360 struct list_head *free)
362 struct request *__rq;
365 if (blk_queue_nomerges(q))
369 * First try one-hit cache.
371 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) {
372 list_add(&rq->queuelist, free);
376 if (blk_queue_noxmerges(q))
381 * See if our hash lookup can find a potential backmerge.
384 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
385 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
388 list_add(&rq->queuelist, free);
389 /* The merged request could be merged with others, try again */
397 void elv_merged_request(struct request_queue *q, struct request *rq,
400 struct elevator_queue *e = q->elevator;
402 if (e->type->ops.request_merged)
403 e->type->ops.request_merged(q, rq, type);
405 if (type == ELEVATOR_BACK_MERGE)
406 elv_rqhash_reposition(q, rq);
411 void elv_merge_requests(struct request_queue *q, struct request *rq,
412 struct request *next)
414 struct elevator_queue *e = q->elevator;
416 if (e->type->ops.requests_merged)
417 e->type->ops.requests_merged(q, rq, next);
419 elv_rqhash_reposition(q, rq);
423 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
425 struct elevator_queue *e = q->elevator;
427 if (e->type->ops.next_request)
428 return e->type->ops.next_request(q, rq);
433 struct request *elv_former_request(struct request_queue *q, struct request *rq)
435 struct elevator_queue *e = q->elevator;
437 if (e->type->ops.former_request)
438 return e->type->ops.former_request(q, rq);
443 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
446 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
448 struct elv_fs_entry *entry = to_elv(attr);
449 struct elevator_queue *e;
455 e = container_of(kobj, struct elevator_queue, kobj);
456 mutex_lock(&e->sysfs_lock);
457 error = e->type ? entry->show(e, page) : -ENOENT;
458 mutex_unlock(&e->sysfs_lock);
463 elv_attr_store(struct kobject *kobj, struct attribute *attr,
464 const char *page, size_t length)
466 struct elv_fs_entry *entry = to_elv(attr);
467 struct elevator_queue *e;
473 e = container_of(kobj, struct elevator_queue, kobj);
474 mutex_lock(&e->sysfs_lock);
475 error = e->type ? entry->store(e, page, length) : -ENOENT;
476 mutex_unlock(&e->sysfs_lock);
480 static const struct sysfs_ops elv_sysfs_ops = {
481 .show = elv_attr_show,
482 .store = elv_attr_store,
485 static struct kobj_type elv_ktype = {
486 .sysfs_ops = &elv_sysfs_ops,
487 .release = elevator_release,
490 int elv_register_queue(struct request_queue *q, bool uevent)
492 struct elevator_queue *e = q->elevator;
495 lockdep_assert_held(&q->sysfs_lock);
497 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
499 struct elv_fs_entry *attr = e->type->elevator_attrs;
501 while (attr->attr.name) {
502 if (sysfs_create_file(&e->kobj, &attr->attr))
508 kobject_uevent(&e->kobj, KOBJ_ADD);
515 void elv_unregister_queue(struct request_queue *q)
517 lockdep_assert_held(&q->sysfs_lock);
520 struct elevator_queue *e = q->elevator;
522 kobject_uevent(&e->kobj, KOBJ_REMOVE);
523 kobject_del(&e->kobj);
526 /* Re-enable throttling in case elevator disabled it */
527 wbt_enable_default(q);
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 elevator_exit(q, q->elevator);
601 ret = blk_mq_init_sched(q, new_e);
606 ret = elv_register_queue(q, true);
608 elevator_exit(q, q->elevator);
614 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
616 blk_add_trace_msg(q, "elv switch: none");
622 static inline bool elv_support_iosched(struct request_queue *q)
624 if (!queue_is_mq(q) ||
625 (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED)))
631 * For single queue devices, default to using mq-deadline. If we have multiple
632 * queues or mq-deadline is not available, default to "none".
634 static struct elevator_type *elevator_get_default(struct request_queue *q)
636 if (q->tag_set && q->tag_set->flags & BLK_MQ_F_NO_SCHED_BY_DEFAULT)
639 if (q->nr_hw_queues != 1 &&
640 !blk_mq_is_shared_tags(q->tag_set->flags))
643 return elevator_get(q, "mq-deadline", false);
647 * Get the first elevator providing the features required by the request queue.
648 * Default to "none" if no matching elevator is found.
650 static struct elevator_type *elevator_get_by_features(struct request_queue *q)
652 struct elevator_type *e, *found = NULL;
654 spin_lock(&elv_list_lock);
656 list_for_each_entry(e, &elv_list, list) {
657 if (elv_support_features(e->elevator_features,
658 q->required_elevator_features)) {
664 if (found && !try_module_get(found->elevator_owner))
667 spin_unlock(&elv_list_lock);
672 * For a device queue that has no required features, use the default elevator
673 * settings. Otherwise, use the first elevator available matching the required
674 * features. If no suitable elevator is find or if the chosen elevator
675 * initialization fails, fall back to the "none" elevator (no elevator).
677 void elevator_init_mq(struct request_queue *q)
679 struct elevator_type *e;
682 if (!elv_support_iosched(q))
685 WARN_ON_ONCE(blk_queue_registered(q));
687 if (unlikely(q->elevator))
690 if (!q->required_elevator_features)
691 e = elevator_get_default(q);
693 e = elevator_get_by_features(q);
697 blk_mq_freeze_queue(q);
698 blk_mq_quiesce_queue(q);
700 err = blk_mq_init_sched(q, e);
702 blk_mq_unquiesce_queue(q);
703 blk_mq_unfreeze_queue(q);
706 pr_warn("\"%s\" elevator initialization failed, "
707 "falling back to \"none\"\n", e->elevator_name);
713 * switch to new_e io scheduler. be careful not to introduce deadlocks -
714 * we don't free the old io scheduler, before we have allocated what we
715 * need for the new one. this way we have a chance of going back to the old
716 * one, if the new one fails init for some reason.
718 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
722 lockdep_assert_held(&q->sysfs_lock);
724 blk_mq_freeze_queue(q);
725 blk_mq_quiesce_queue(q);
727 err = elevator_switch_mq(q, new_e);
729 blk_mq_unquiesce_queue(q);
730 blk_mq_unfreeze_queue(q);
736 * Switch this queue to the given IO scheduler.
738 static int __elevator_change(struct request_queue *q, const char *name)
740 char elevator_name[ELV_NAME_MAX];
741 struct elevator_type *e;
743 /* Make sure queue is not in the middle of being removed */
744 if (!blk_queue_registered(q))
748 * Special case for mq, turn off scheduling
750 if (!strncmp(name, "none", 4)) {
753 return elevator_switch(q, NULL);
756 strlcpy(elevator_name, name, sizeof(elevator_name));
757 e = elevator_get(q, strstrip(elevator_name), true);
762 elevator_match(q->elevator->type, elevator_name, 0)) {
767 return elevator_switch(q, e);
770 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
775 if (!elv_support_iosched(q))
778 ret = __elevator_change(q, name);
785 ssize_t elv_iosched_show(struct request_queue *q, char *name)
787 struct elevator_queue *e = q->elevator;
788 struct elevator_type *elv = NULL;
789 struct elevator_type *__e;
793 return sprintf(name, "none\n");
796 len += sprintf(name+len, "[none] ");
800 spin_lock(&elv_list_lock);
801 list_for_each_entry(__e, &elv_list, list) {
802 if (elv && elevator_match(elv, __e->elevator_name, 0)) {
803 len += sprintf(name+len, "[%s] ", elv->elevator_name);
806 if (elv_support_iosched(q) &&
807 elevator_match(__e, __e->elevator_name,
808 q->required_elevator_features))
809 len += sprintf(name+len, "%s ", __e->elevator_name);
811 spin_unlock(&elv_list_lock);
814 len += sprintf(name+len, "none");
816 len += sprintf(len+name, "\n");
820 struct request *elv_rb_former_request(struct request_queue *q,
823 struct rb_node *rbprev = rb_prev(&rq->rb_node);
826 return rb_entry_rq(rbprev);
830 EXPORT_SYMBOL(elv_rb_former_request);
832 struct request *elv_rb_latter_request(struct request_queue *q,
835 struct rb_node *rbnext = rb_next(&rq->rb_node);
838 return rb_entry_rq(rbnext);
842 EXPORT_SYMBOL(elv_rb_latter_request);
844 static int __init elevator_setup(char *str)
846 pr_warn("Kernel parameter elevator= does not have any effect anymore.\n"
847 "Please use sysfs to set IO scheduler for individual devices.\n");
851 __setup("elevator=", elevator_setup);