2 * Block device elevator/IO-scheduler.
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
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"
47 static DEFINE_SPINLOCK(elv_list_lock);
48 static LIST_HEAD(elv_list);
53 #define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq))
56 * Query io scheduler to see if the current process issuing bio may be
59 static int elv_iosched_allow_bio_merge(struct request *rq, struct bio *bio)
61 struct request_queue *q = rq->q;
62 struct elevator_queue *e = q->elevator;
64 if (e->type->ops.allow_merge)
65 return e->type->ops.allow_merge(q, rq, bio);
71 * can we safely merge with this request?
73 bool elv_bio_merge_ok(struct request *rq, struct bio *bio)
75 if (!blk_rq_merge_ok(rq, bio))
78 if (!elv_iosched_allow_bio_merge(rq, bio))
83 EXPORT_SYMBOL(elv_bio_merge_ok);
85 static bool elevator_match(const struct elevator_type *e, const char *name)
87 if (!strcmp(e->elevator_name, name))
89 if (e->elevator_alias && !strcmp(e->elevator_alias, name))
96 * Return scheduler with name 'name'
98 static struct elevator_type *elevator_find(const char *name)
100 struct elevator_type *e;
102 list_for_each_entry(e, &elv_list, list) {
103 if (elevator_match(e, name))
110 static void elevator_put(struct elevator_type *e)
112 module_put(e->elevator_owner);
115 static struct elevator_type *elevator_get(struct request_queue *q,
116 const char *name, bool try_loading)
118 struct elevator_type *e;
120 spin_lock(&elv_list_lock);
122 e = elevator_find(name);
123 if (!e && try_loading) {
124 spin_unlock(&elv_list_lock);
125 request_module("%s-iosched", name);
126 spin_lock(&elv_list_lock);
127 e = elevator_find(name);
130 if (e && !try_module_get(e->elevator_owner))
133 spin_unlock(&elv_list_lock);
137 static char chosen_elevator[ELV_NAME_MAX];
139 static int __init elevator_setup(char *str)
142 * Be backwards-compatible with previous kernels, so users
143 * won't get the wrong elevator.
145 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
149 __setup("elevator=", elevator_setup);
151 static struct kobj_type elv_ktype;
153 struct elevator_queue *elevator_alloc(struct request_queue *q,
154 struct elevator_type *e)
156 struct elevator_queue *eq;
158 eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
163 kobject_init(&eq->kobj, &elv_ktype);
164 mutex_init(&eq->sysfs_lock);
169 EXPORT_SYMBOL(elevator_alloc);
171 static void elevator_release(struct kobject *kobj)
173 struct elevator_queue *e;
175 e = container_of(kobj, struct elevator_queue, kobj);
176 elevator_put(e->type);
180 void elevator_exit(struct request_queue *q, struct elevator_queue *e)
182 mutex_lock(&e->sysfs_lock);
183 if (e->type->ops.exit_sched)
184 blk_mq_exit_sched(q, e);
185 mutex_unlock(&e->sysfs_lock);
187 kobject_put(&e->kobj);
190 static inline void __elv_rqhash_del(struct request *rq)
193 rq->rq_flags &= ~RQF_HASHED;
196 void elv_rqhash_del(struct request_queue *q, struct request *rq)
199 __elv_rqhash_del(rq);
201 EXPORT_SYMBOL_GPL(elv_rqhash_del);
203 void elv_rqhash_add(struct request_queue *q, struct request *rq)
205 struct elevator_queue *e = q->elevator;
207 BUG_ON(ELV_ON_HASH(rq));
208 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
209 rq->rq_flags |= RQF_HASHED;
211 EXPORT_SYMBOL_GPL(elv_rqhash_add);
213 void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
215 __elv_rqhash_del(rq);
216 elv_rqhash_add(q, rq);
219 struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
221 struct elevator_queue *e = q->elevator;
222 struct hlist_node *next;
225 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
226 BUG_ON(!ELV_ON_HASH(rq));
228 if (unlikely(!rq_mergeable(rq))) {
229 __elv_rqhash_del(rq);
233 if (rq_hash_key(rq) == offset)
241 * RB-tree support functions for inserting/lookup/removal of requests
242 * in a sorted RB tree.
244 void elv_rb_add(struct rb_root *root, struct request *rq)
246 struct rb_node **p = &root->rb_node;
247 struct rb_node *parent = NULL;
248 struct request *__rq;
252 __rq = rb_entry(parent, struct request, rb_node);
254 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
256 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
260 rb_link_node(&rq->rb_node, parent, p);
261 rb_insert_color(&rq->rb_node, root);
263 EXPORT_SYMBOL(elv_rb_add);
265 void elv_rb_del(struct rb_root *root, struct request *rq)
267 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
268 rb_erase(&rq->rb_node, root);
269 RB_CLEAR_NODE(&rq->rb_node);
271 EXPORT_SYMBOL(elv_rb_del);
273 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
275 struct rb_node *n = root->rb_node;
279 rq = rb_entry(n, struct request, rb_node);
281 if (sector < blk_rq_pos(rq))
283 else if (sector > blk_rq_pos(rq))
291 EXPORT_SYMBOL(elv_rb_find);
293 enum elv_merge elv_merge(struct request_queue *q, struct request **req,
296 struct elevator_queue *e = q->elevator;
297 struct request *__rq;
301 * nomerges: No merges at all attempted
302 * noxmerges: Only simple one-hit cache try
303 * merges: All merge tries attempted
305 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
306 return ELEVATOR_NO_MERGE;
309 * First try one-hit cache.
311 if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
312 enum elv_merge ret = blk_try_merge(q->last_merge, bio);
314 if (ret != ELEVATOR_NO_MERGE) {
315 *req = q->last_merge;
320 if (blk_queue_noxmerges(q))
321 return ELEVATOR_NO_MERGE;
324 * See if our hash lookup can find a potential backmerge.
326 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
327 if (__rq && elv_bio_merge_ok(__rq, bio)) {
329 return ELEVATOR_BACK_MERGE;
332 if (e->type->ops.request_merge)
333 return e->type->ops.request_merge(q, req, bio);
335 return ELEVATOR_NO_MERGE;
339 * Attempt to do an insertion back merge. Only check for the case where
340 * we can append 'rq' to an existing request, so we can throw 'rq' away
343 * Returns true if we merged, false otherwise
345 bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq)
347 struct request *__rq;
350 if (blk_queue_nomerges(q))
354 * First try one-hit cache.
356 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
359 if (blk_queue_noxmerges(q))
364 * See if our hash lookup can find a potential backmerge.
367 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
368 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
371 /* The merged request could be merged with others, try again */
379 void elv_merged_request(struct request_queue *q, struct request *rq,
382 struct elevator_queue *e = q->elevator;
384 if (e->type->ops.request_merged)
385 e->type->ops.request_merged(q, rq, type);
387 if (type == ELEVATOR_BACK_MERGE)
388 elv_rqhash_reposition(q, rq);
393 void elv_merge_requests(struct request_queue *q, struct request *rq,
394 struct request *next)
396 struct elevator_queue *e = q->elevator;
398 if (e->type->ops.requests_merged)
399 e->type->ops.requests_merged(q, rq, next);
401 elv_rqhash_reposition(q, rq);
405 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
407 struct elevator_queue *e = q->elevator;
409 if (e->type->ops.next_request)
410 return e->type->ops.next_request(q, rq);
415 struct request *elv_former_request(struct request_queue *q, struct request *rq)
417 struct elevator_queue *e = q->elevator;
419 if (e->type->ops.former_request)
420 return e->type->ops.former_request(q, rq);
425 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
428 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
430 struct elv_fs_entry *entry = to_elv(attr);
431 struct elevator_queue *e;
437 e = container_of(kobj, struct elevator_queue, kobj);
438 mutex_lock(&e->sysfs_lock);
439 error = e->type ? entry->show(e, page) : -ENOENT;
440 mutex_unlock(&e->sysfs_lock);
445 elv_attr_store(struct kobject *kobj, struct attribute *attr,
446 const char *page, size_t length)
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->store(e, page, length) : -ENOENT;
458 mutex_unlock(&e->sysfs_lock);
462 static const struct sysfs_ops elv_sysfs_ops = {
463 .show = elv_attr_show,
464 .store = elv_attr_store,
467 static struct kobj_type elv_ktype = {
468 .sysfs_ops = &elv_sysfs_ops,
469 .release = elevator_release,
472 int elv_register_queue(struct request_queue *q)
474 struct elevator_queue *e = q->elevator;
477 lockdep_assert_held(&q->sysfs_lock);
479 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
481 struct elv_fs_entry *attr = e->type->elevator_attrs;
483 while (attr->attr.name) {
484 if (sysfs_create_file(&e->kobj, &attr->attr))
489 kobject_uevent(&e->kobj, KOBJ_ADD);
495 void elv_unregister_queue(struct request_queue *q)
497 lockdep_assert_held(&q->sysfs_lock);
500 struct elevator_queue *e = q->elevator;
502 kobject_uevent(&e->kobj, KOBJ_REMOVE);
503 kobject_del(&e->kobj);
505 /* Re-enable throttling in case elevator disabled it */
506 wbt_enable_default(q);
510 int elv_register(struct elevator_type *e)
514 /* create icq_cache if requested */
516 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
517 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
520 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
521 "%s_io_cq", e->elevator_name);
522 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
523 e->icq_align, 0, NULL);
528 /* register, don't allow duplicate names */
529 spin_lock(&elv_list_lock);
530 if (elevator_find(e->elevator_name)) {
531 spin_unlock(&elv_list_lock);
532 kmem_cache_destroy(e->icq_cache);
535 list_add_tail(&e->list, &elv_list);
536 spin_unlock(&elv_list_lock);
538 printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
542 EXPORT_SYMBOL_GPL(elv_register);
544 void elv_unregister(struct elevator_type *e)
547 spin_lock(&elv_list_lock);
548 list_del_init(&e->list);
549 spin_unlock(&elv_list_lock);
552 * Destroy icq_cache if it exists. icq's are RCU managed. Make
553 * sure all RCU operations are complete before proceeding.
557 kmem_cache_destroy(e->icq_cache);
561 EXPORT_SYMBOL_GPL(elv_unregister);
563 int elevator_switch_mq(struct request_queue *q,
564 struct elevator_type *new_e)
568 lockdep_assert_held(&q->sysfs_lock);
571 if (q->elevator->registered)
572 elv_unregister_queue(q);
574 elevator_exit(q, q->elevator);
577 ret = blk_mq_init_sched(q, new_e);
582 ret = elv_register_queue(q);
584 elevator_exit(q, q->elevator);
590 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
592 blk_add_trace_msg(q, "elv switch: none");
599 * For blk-mq devices, we default to using mq-deadline, if available, for single
600 * queue devices. If deadline isn't available OR we have multiple queues,
603 int elevator_init_mq(struct request_queue *q)
605 struct elevator_type *e;
608 if (q->nr_hw_queues != 1)
612 * q->sysfs_lock must be held to provide mutual exclusion between
613 * elevator_switch() and here.
615 mutex_lock(&q->sysfs_lock);
616 if (unlikely(q->elevator))
619 e = elevator_get(q, "mq-deadline", false);
623 err = blk_mq_init_sched(q, e);
627 mutex_unlock(&q->sysfs_lock);
633 * switch to new_e io scheduler. be careful not to introduce deadlocks -
634 * we don't free the old io scheduler, before we have allocated what we
635 * need for the new one. this way we have a chance of going back to the old
636 * one, if the new one fails init for some reason.
638 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
642 lockdep_assert_held(&q->sysfs_lock);
644 blk_mq_freeze_queue(q);
645 blk_mq_quiesce_queue(q);
647 err = elevator_switch_mq(q, new_e);
649 blk_mq_unquiesce_queue(q);
650 blk_mq_unfreeze_queue(q);
656 * Switch this queue to the given IO scheduler.
658 static int __elevator_change(struct request_queue *q, const char *name)
660 char elevator_name[ELV_NAME_MAX];
661 struct elevator_type *e;
663 /* Make sure queue is not in the middle of being removed */
664 if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
668 * Special case for mq, turn off scheduling
670 if (!strncmp(name, "none", 4)) {
673 return elevator_switch(q, NULL);
676 strlcpy(elevator_name, name, sizeof(elevator_name));
677 e = elevator_get(q, strstrip(elevator_name), true);
681 if (q->elevator && elevator_match(q->elevator->type, elevator_name)) {
686 return elevator_switch(q, e);
689 static inline bool elv_support_iosched(struct request_queue *q)
691 if (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED))
696 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
701 if (!queue_is_mq(q) || !elv_support_iosched(q))
704 ret = __elevator_change(q, name);
711 ssize_t elv_iosched_show(struct request_queue *q, char *name)
713 struct elevator_queue *e = q->elevator;
714 struct elevator_type *elv = NULL;
715 struct elevator_type *__e;
719 return sprintf(name, "none\n");
722 len += sprintf(name+len, "[none] ");
726 spin_lock(&elv_list_lock);
727 list_for_each_entry(__e, &elv_list, list) {
728 if (elv && elevator_match(elv, __e->elevator_name)) {
729 len += sprintf(name+len, "[%s] ", elv->elevator_name);
732 if (elv_support_iosched(q))
733 len += sprintf(name+len, "%s ", __e->elevator_name);
735 spin_unlock(&elv_list_lock);
738 len += sprintf(name+len, "none");
740 len += sprintf(len+name, "\n");
744 struct request *elv_rb_former_request(struct request_queue *q,
747 struct rb_node *rbprev = rb_prev(&rq->rb_node);
750 return rb_entry_rq(rbprev);
754 EXPORT_SYMBOL(elv_rb_former_request);
756 struct request *elv_rb_latter_request(struct request_queue *q,
759 struct rb_node *rbnext = rb_next(&rq->rb_node);
762 return rb_entry_rq(rbnext);
766 EXPORT_SYMBOL(elv_rb_latter_request);