Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Block device elevator/IO-scheduler.
4 *
5 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
6 *
7 * 30042000 Jens Axboe <axboe@kernel.dk> :
8 *
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
13 * an existing request
14 * - elevator_dequeue_fn, called when a request is taken off the active list
15 *
16 * 20082000 Dave Jones <davej@suse.de> :
17 * Removed tests for max-bomb-segments, which was breaking elvtune
18 * when run without -bN
19 *
20 * Jens:
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
24 *
25 */
26#include <linux/kernel.h>
27#include <linux/fs.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
39#include <trace/events/block.h>
40
41#include "elevator.h"
42#include "blk.h"
43#include "blk-mq-sched.h"
44#include "blk-pm.h"
45#include "blk-wbt.h"
46#include "blk-cgroup.h"
47
48static DEFINE_SPINLOCK(elv_list_lock);
49static LIST_HEAD(elv_list);
50
51/*
52 * Merge hash stuff.
53 */
54#define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq))
55
56/*
57 * Query io scheduler to see if the current process issuing bio may be
58 * merged with rq.
59 */
60static bool elv_iosched_allow_bio_merge(struct request *rq, struct bio *bio)
61{
62 struct request_queue *q = rq->q;
63 struct elevator_queue *e = q->elevator;
64
65 if (e->type->ops.allow_merge)
66 return e->type->ops.allow_merge(q, rq, bio);
67
68 return true;
69}
70
71/*
72 * can we safely merge with this request?
73 */
74bool elv_bio_merge_ok(struct request *rq, struct bio *bio)
75{
76 if (!blk_rq_merge_ok(rq, bio))
77 return false;
78
79 if (!elv_iosched_allow_bio_merge(rq, bio))
80 return false;
81
82 return true;
83}
84EXPORT_SYMBOL(elv_bio_merge_ok);
85
86static inline bool elv_support_features(struct request_queue *q,
87 const struct elevator_type *e)
88{
89 return (q->required_elevator_features & e->elevator_features) ==
90 q->required_elevator_features;
91}
92
93/**
94 * elevator_match - Check whether @e's name or alias matches @name
95 * @e: Scheduler to test
96 * @name: Elevator name to test
97 *
98 * Return true if the elevator @e's name or alias matches @name.
99 */
100static bool elevator_match(const struct elevator_type *e, const char *name)
101{
102 return !strcmp(e->elevator_name, name) ||
103 (e->elevator_alias && !strcmp(e->elevator_alias, name));
104}
105
106static struct elevator_type *__elevator_find(const char *name)
107{
108 struct elevator_type *e;
109
110 list_for_each_entry(e, &elv_list, list)
111 if (elevator_match(e, name))
112 return e;
113 return NULL;
114}
115
116static struct elevator_type *elevator_find_get(struct request_queue *q,
117 const char *name)
118{
119 struct elevator_type *e;
120
121 spin_lock(&elv_list_lock);
122 e = __elevator_find(name);
123 if (e && (!elv_support_features(q, e) || !elevator_tryget(e)))
124 e = NULL;
125 spin_unlock(&elv_list_lock);
126 return e;
127}
128
129static struct kobj_type elv_ktype;
130
131struct elevator_queue *elevator_alloc(struct request_queue *q,
132 struct elevator_type *e)
133{
134 struct elevator_queue *eq;
135
136 eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
137 if (unlikely(!eq))
138 return NULL;
139
140 __elevator_get(e);
141 eq->type = e;
142 kobject_init(&eq->kobj, &elv_ktype);
143 mutex_init(&eq->sysfs_lock);
144 hash_init(eq->hash);
145
146 return eq;
147}
148EXPORT_SYMBOL(elevator_alloc);
149
150static void elevator_release(struct kobject *kobj)
151{
152 struct elevator_queue *e;
153
154 e = container_of(kobj, struct elevator_queue, kobj);
155 elevator_put(e->type);
156 kfree(e);
157}
158
159void elevator_exit(struct request_queue *q)
160{
161 struct elevator_queue *e = q->elevator;
162
163 ioc_clear_queue(q);
164 blk_mq_sched_free_rqs(q);
165
166 mutex_lock(&e->sysfs_lock);
167 blk_mq_exit_sched(q, e);
168 mutex_unlock(&e->sysfs_lock);
169
170 kobject_put(&e->kobj);
171}
172
173static inline void __elv_rqhash_del(struct request *rq)
174{
175 hash_del(&rq->hash);
176 rq->rq_flags &= ~RQF_HASHED;
177}
178
179void elv_rqhash_del(struct request_queue *q, struct request *rq)
180{
181 if (ELV_ON_HASH(rq))
182 __elv_rqhash_del(rq);
183}
184EXPORT_SYMBOL_GPL(elv_rqhash_del);
185
186void elv_rqhash_add(struct request_queue *q, struct request *rq)
187{
188 struct elevator_queue *e = q->elevator;
189
190 BUG_ON(ELV_ON_HASH(rq));
191 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
192 rq->rq_flags |= RQF_HASHED;
193}
194EXPORT_SYMBOL_GPL(elv_rqhash_add);
195
196void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
197{
198 __elv_rqhash_del(rq);
199 elv_rqhash_add(q, rq);
200}
201
202struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
203{
204 struct elevator_queue *e = q->elevator;
205 struct hlist_node *next;
206 struct request *rq;
207
208 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
209 BUG_ON(!ELV_ON_HASH(rq));
210
211 if (unlikely(!rq_mergeable(rq))) {
212 __elv_rqhash_del(rq);
213 continue;
214 }
215
216 if (rq_hash_key(rq) == offset)
217 return rq;
218 }
219
220 return NULL;
221}
222
223/*
224 * RB-tree support functions for inserting/lookup/removal of requests
225 * in a sorted RB tree.
226 */
227void elv_rb_add(struct rb_root *root, struct request *rq)
228{
229 struct rb_node **p = &root->rb_node;
230 struct rb_node *parent = NULL;
231 struct request *__rq;
232
233 while (*p) {
234 parent = *p;
235 __rq = rb_entry(parent, struct request, rb_node);
236
237 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
238 p = &(*p)->rb_left;
239 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
240 p = &(*p)->rb_right;
241 }
242
243 rb_link_node(&rq->rb_node, parent, p);
244 rb_insert_color(&rq->rb_node, root);
245}
246EXPORT_SYMBOL(elv_rb_add);
247
248void elv_rb_del(struct rb_root *root, struct request *rq)
249{
250 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
251 rb_erase(&rq->rb_node, root);
252 RB_CLEAR_NODE(&rq->rb_node);
253}
254EXPORT_SYMBOL(elv_rb_del);
255
256struct request *elv_rb_find(struct rb_root *root, sector_t sector)
257{
258 struct rb_node *n = root->rb_node;
259 struct request *rq;
260
261 while (n) {
262 rq = rb_entry(n, struct request, rb_node);
263
264 if (sector < blk_rq_pos(rq))
265 n = n->rb_left;
266 else if (sector > blk_rq_pos(rq))
267 n = n->rb_right;
268 else
269 return rq;
270 }
271
272 return NULL;
273}
274EXPORT_SYMBOL(elv_rb_find);
275
276enum elv_merge elv_merge(struct request_queue *q, struct request **req,
277 struct bio *bio)
278{
279 struct elevator_queue *e = q->elevator;
280 struct request *__rq;
281
282 /*
283 * Levels of merges:
284 * nomerges: No merges at all attempted
285 * noxmerges: Only simple one-hit cache try
286 * merges: All merge tries attempted
287 */
288 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
289 return ELEVATOR_NO_MERGE;
290
291 /*
292 * First try one-hit cache.
293 */
294 if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
295 enum elv_merge ret = blk_try_merge(q->last_merge, bio);
296
297 if (ret != ELEVATOR_NO_MERGE) {
298 *req = q->last_merge;
299 return ret;
300 }
301 }
302
303 if (blk_queue_noxmerges(q))
304 return ELEVATOR_NO_MERGE;
305
306 /*
307 * See if our hash lookup can find a potential backmerge.
308 */
309 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
310 if (__rq && elv_bio_merge_ok(__rq, bio)) {
311 *req = __rq;
312
313 if (blk_discard_mergable(__rq))
314 return ELEVATOR_DISCARD_MERGE;
315 return ELEVATOR_BACK_MERGE;
316 }
317
318 if (e->type->ops.request_merge)
319 return e->type->ops.request_merge(q, req, bio);
320
321 return ELEVATOR_NO_MERGE;
322}
323
324/*
325 * Attempt to do an insertion back merge. Only check for the case where
326 * we can append 'rq' to an existing request, so we can throw 'rq' away
327 * afterwards.
328 *
329 * Returns true if we merged, false otherwise. 'free' will contain all
330 * requests that need to be freed.
331 */
332bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq,
333 struct list_head *free)
334{
335 struct request *__rq;
336 bool ret;
337
338 if (blk_queue_nomerges(q))
339 return false;
340
341 /*
342 * First try one-hit cache.
343 */
344 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) {
345 list_add(&rq->queuelist, free);
346 return true;
347 }
348
349 if (blk_queue_noxmerges(q))
350 return false;
351
352 ret = false;
353 /*
354 * See if our hash lookup can find a potential backmerge.
355 */
356 while (1) {
357 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
358 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
359 break;
360
361 list_add(&rq->queuelist, free);
362 /* The merged request could be merged with others, try again */
363 ret = true;
364 rq = __rq;
365 }
366
367 return ret;
368}
369
370void elv_merged_request(struct request_queue *q, struct request *rq,
371 enum elv_merge type)
372{
373 struct elevator_queue *e = q->elevator;
374
375 if (e->type->ops.request_merged)
376 e->type->ops.request_merged(q, rq, type);
377
378 if (type == ELEVATOR_BACK_MERGE)
379 elv_rqhash_reposition(q, rq);
380
381 q->last_merge = rq;
382}
383
384void elv_merge_requests(struct request_queue *q, struct request *rq,
385 struct request *next)
386{
387 struct elevator_queue *e = q->elevator;
388
389 if (e->type->ops.requests_merged)
390 e->type->ops.requests_merged(q, rq, next);
391
392 elv_rqhash_reposition(q, rq);
393 q->last_merge = rq;
394}
395
396struct request *elv_latter_request(struct request_queue *q, struct request *rq)
397{
398 struct elevator_queue *e = q->elevator;
399
400 if (e->type->ops.next_request)
401 return e->type->ops.next_request(q, rq);
402
403 return NULL;
404}
405
406struct request *elv_former_request(struct request_queue *q, struct request *rq)
407{
408 struct elevator_queue *e = q->elevator;
409
410 if (e->type->ops.former_request)
411 return e->type->ops.former_request(q, rq);
412
413 return NULL;
414}
415
416#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
417
418static ssize_t
419elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
420{
421 struct elv_fs_entry *entry = to_elv(attr);
422 struct elevator_queue *e;
423 ssize_t error;
424
425 if (!entry->show)
426 return -EIO;
427
428 e = container_of(kobj, struct elevator_queue, kobj);
429 mutex_lock(&e->sysfs_lock);
430 error = e->type ? entry->show(e, page) : -ENOENT;
431 mutex_unlock(&e->sysfs_lock);
432 return error;
433}
434
435static ssize_t
436elv_attr_store(struct kobject *kobj, struct attribute *attr,
437 const char *page, size_t length)
438{
439 struct elv_fs_entry *entry = to_elv(attr);
440 struct elevator_queue *e;
441 ssize_t error;
442
443 if (!entry->store)
444 return -EIO;
445
446 e = container_of(kobj, struct elevator_queue, kobj);
447 mutex_lock(&e->sysfs_lock);
448 error = e->type ? entry->store(e, page, length) : -ENOENT;
449 mutex_unlock(&e->sysfs_lock);
450 return error;
451}
452
453static const struct sysfs_ops elv_sysfs_ops = {
454 .show = elv_attr_show,
455 .store = elv_attr_store,
456};
457
458static struct kobj_type elv_ktype = {
459 .sysfs_ops = &elv_sysfs_ops,
460 .release = elevator_release,
461};
462
463int elv_register_queue(struct request_queue *q, bool uevent)
464{
465 struct elevator_queue *e = q->elevator;
466 int error;
467
468 lockdep_assert_held(&q->sysfs_lock);
469
470 error = kobject_add(&e->kobj, &q->disk->queue_kobj, "iosched");
471 if (!error) {
472 struct elv_fs_entry *attr = e->type->elevator_attrs;
473 if (attr) {
474 while (attr->attr.name) {
475 if (sysfs_create_file(&e->kobj, &attr->attr))
476 break;
477 attr++;
478 }
479 }
480 if (uevent)
481 kobject_uevent(&e->kobj, KOBJ_ADD);
482
483 set_bit(ELEVATOR_FLAG_REGISTERED, &e->flags);
484 }
485 return error;
486}
487
488void elv_unregister_queue(struct request_queue *q)
489{
490 struct elevator_queue *e = q->elevator;
491
492 lockdep_assert_held(&q->sysfs_lock);
493
494 if (e && test_and_clear_bit(ELEVATOR_FLAG_REGISTERED, &e->flags)) {
495 kobject_uevent(&e->kobj, KOBJ_REMOVE);
496 kobject_del(&e->kobj);
497 }
498}
499
500int elv_register(struct elevator_type *e)
501{
502 /* insert_requests and dispatch_request are mandatory */
503 if (WARN_ON_ONCE(!e->ops.insert_requests || !e->ops.dispatch_request))
504 return -EINVAL;
505
506 /* create icq_cache if requested */
507 if (e->icq_size) {
508 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
509 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
510 return -EINVAL;
511
512 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
513 "%s_io_cq", e->elevator_name);
514 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
515 e->icq_align, 0, NULL);
516 if (!e->icq_cache)
517 return -ENOMEM;
518 }
519
520 /* register, don't allow duplicate names */
521 spin_lock(&elv_list_lock);
522 if (__elevator_find(e->elevator_name)) {
523 spin_unlock(&elv_list_lock);
524 kmem_cache_destroy(e->icq_cache);
525 return -EBUSY;
526 }
527 list_add_tail(&e->list, &elv_list);
528 spin_unlock(&elv_list_lock);
529
530 printk(KERN_INFO "io scheduler %s registered\n", e->elevator_name);
531
532 return 0;
533}
534EXPORT_SYMBOL_GPL(elv_register);
535
536void elv_unregister(struct elevator_type *e)
537{
538 /* unregister */
539 spin_lock(&elv_list_lock);
540 list_del_init(&e->list);
541 spin_unlock(&elv_list_lock);
542
543 /*
544 * Destroy icq_cache if it exists. icq's are RCU managed. Make
545 * sure all RCU operations are complete before proceeding.
546 */
547 if (e->icq_cache) {
548 rcu_barrier();
549 kmem_cache_destroy(e->icq_cache);
550 e->icq_cache = NULL;
551 }
552}
553EXPORT_SYMBOL_GPL(elv_unregister);
554
555static inline bool elv_support_iosched(struct request_queue *q)
556{
557 if (!queue_is_mq(q) ||
558 (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED)))
559 return false;
560 return true;
561}
562
563/*
564 * For single queue devices, default to using mq-deadline. If we have multiple
565 * queues or mq-deadline is not available, default to "none".
566 */
567static struct elevator_type *elevator_get_default(struct request_queue *q)
568{
569 if (q->tag_set && q->tag_set->flags & BLK_MQ_F_NO_SCHED_BY_DEFAULT)
570 return NULL;
571
572 if (q->nr_hw_queues != 1 &&
573 !blk_mq_is_shared_tags(q->tag_set->flags))
574 return NULL;
575
576 return elevator_find_get(q, "mq-deadline");
577}
578
579/*
580 * Get the first elevator providing the features required by the request queue.
581 * Default to "none" if no matching elevator is found.
582 */
583static struct elevator_type *elevator_get_by_features(struct request_queue *q)
584{
585 struct elevator_type *e, *found = NULL;
586
587 spin_lock(&elv_list_lock);
588
589 list_for_each_entry(e, &elv_list, list) {
590 if (elv_support_features(q, e)) {
591 found = e;
592 break;
593 }
594 }
595
596 if (found && !elevator_tryget(found))
597 found = NULL;
598
599 spin_unlock(&elv_list_lock);
600 return found;
601}
602
603/*
604 * For a device queue that has no required features, use the default elevator
605 * settings. Otherwise, use the first elevator available matching the required
606 * features. If no suitable elevator is find or if the chosen elevator
607 * initialization fails, fall back to the "none" elevator (no elevator).
608 */
609void elevator_init_mq(struct request_queue *q)
610{
611 struct elevator_type *e;
612 int err;
613
614 if (!elv_support_iosched(q))
615 return;
616
617 WARN_ON_ONCE(blk_queue_registered(q));
618
619 if (unlikely(q->elevator))
620 return;
621
622 if (!q->required_elevator_features)
623 e = elevator_get_default(q);
624 else
625 e = elevator_get_by_features(q);
626 if (!e)
627 return;
628
629 /*
630 * We are called before adding disk, when there isn't any FS I/O,
631 * so freezing queue plus canceling dispatch work is enough to
632 * drain any dispatch activities originated from passthrough
633 * requests, then no need to quiesce queue which may add long boot
634 * latency, especially when lots of disks are involved.
635 */
636 blk_mq_freeze_queue(q);
637 blk_mq_cancel_work_sync(q);
638
639 err = blk_mq_init_sched(q, e);
640
641 blk_mq_unfreeze_queue(q);
642
643 if (err) {
644 pr_warn("\"%s\" elevator initialization failed, "
645 "falling back to \"none\"\n", e->elevator_name);
646 }
647
648 elevator_put(e);
649}
650
651/*
652 * Switch to new_e io scheduler.
653 *
654 * If switching fails, we are most likely running out of memory and not able
655 * to restore the old io scheduler, so leaving the io scheduler being none.
656 */
657int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
658{
659 int ret;
660
661 lockdep_assert_held(&q->sysfs_lock);
662
663 blk_mq_freeze_queue(q);
664 blk_mq_quiesce_queue(q);
665
666 if (q->elevator) {
667 elv_unregister_queue(q);
668 elevator_exit(q);
669 }
670
671 ret = blk_mq_init_sched(q, new_e);
672 if (ret)
673 goto out_unfreeze;
674
675 ret = elv_register_queue(q, true);
676 if (ret) {
677 elevator_exit(q);
678 goto out_unfreeze;
679 }
680 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
681
682out_unfreeze:
683 blk_mq_unquiesce_queue(q);
684 blk_mq_unfreeze_queue(q);
685
686 if (ret) {
687 pr_warn("elv: switch to \"%s\" failed, falling back to \"none\"\n",
688 new_e->elevator_name);
689 }
690
691 return ret;
692}
693
694void elevator_disable(struct request_queue *q)
695{
696 lockdep_assert_held(&q->sysfs_lock);
697
698 blk_mq_freeze_queue(q);
699 blk_mq_quiesce_queue(q);
700
701 elv_unregister_queue(q);
702 elevator_exit(q);
703 blk_queue_flag_clear(QUEUE_FLAG_SQ_SCHED, q);
704 q->elevator = NULL;
705 q->nr_requests = q->tag_set->queue_depth;
706 blk_add_trace_msg(q, "elv switch: none");
707
708 blk_mq_unquiesce_queue(q);
709 blk_mq_unfreeze_queue(q);
710}
711
712/*
713 * Switch this queue to the given IO scheduler.
714 */
715static int elevator_change(struct request_queue *q, const char *elevator_name)
716{
717 struct elevator_type *e;
718 int ret;
719
720 /* Make sure queue is not in the middle of being removed */
721 if (!blk_queue_registered(q))
722 return -ENOENT;
723
724 if (!strncmp(elevator_name, "none", 4)) {
725 if (q->elevator)
726 elevator_disable(q);
727 return 0;
728 }
729
730 if (q->elevator && elevator_match(q->elevator->type, elevator_name))
731 return 0;
732
733 e = elevator_find_get(q, elevator_name);
734 if (!e) {
735 request_module("%s-iosched", elevator_name);
736 e = elevator_find_get(q, elevator_name);
737 if (!e)
738 return -EINVAL;
739 }
740 ret = elevator_switch(q, e);
741 elevator_put(e);
742 return ret;
743}
744
745ssize_t elv_iosched_store(struct request_queue *q, const char *buf,
746 size_t count)
747{
748 char elevator_name[ELV_NAME_MAX];
749 int ret;
750
751 if (!elv_support_iosched(q))
752 return count;
753
754 strlcpy(elevator_name, buf, sizeof(elevator_name));
755 ret = elevator_change(q, strstrip(elevator_name));
756 if (!ret)
757 return count;
758 return ret;
759}
760
761ssize_t elv_iosched_show(struct request_queue *q, char *name)
762{
763 struct elevator_queue *eq = q->elevator;
764 struct elevator_type *cur = NULL, *e;
765 int len = 0;
766
767 if (!elv_support_iosched(q))
768 return sprintf(name, "none\n");
769
770 if (!q->elevator) {
771 len += sprintf(name+len, "[none] ");
772 } else {
773 len += sprintf(name+len, "none ");
774 cur = eq->type;
775 }
776
777 spin_lock(&elv_list_lock);
778 list_for_each_entry(e, &elv_list, list) {
779 if (e == cur)
780 len += sprintf(name+len, "[%s] ", e->elevator_name);
781 else if (elv_support_features(q, e))
782 len += sprintf(name+len, "%s ", e->elevator_name);
783 }
784 spin_unlock(&elv_list_lock);
785
786 len += sprintf(name+len, "\n");
787 return len;
788}
789
790struct request *elv_rb_former_request(struct request_queue *q,
791 struct request *rq)
792{
793 struct rb_node *rbprev = rb_prev(&rq->rb_node);
794
795 if (rbprev)
796 return rb_entry_rq(rbprev);
797
798 return NULL;
799}
800EXPORT_SYMBOL(elv_rb_former_request);
801
802struct request *elv_rb_latter_request(struct request_queue *q,
803 struct request *rq)
804{
805 struct rb_node *rbnext = rb_next(&rq->rb_node);
806
807 if (rbnext)
808 return rb_entry_rq(rbnext);
809
810 return NULL;
811}
812EXPORT_SYMBOL(elv_rb_latter_request);
813
814static int __init elevator_setup(char *str)
815{
816 pr_warn("Kernel parameter elevator= does not have any effect anymore.\n"
817 "Please use sysfs to set IO scheduler for individual devices.\n");
818 return 1;
819}
820
821__setup("elevator=", elevator_setup);
1/*
2 * Block device elevator/IO-scheduler.
3 *
4 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
5 *
6 * 30042000 Jens Axboe <axboe@kernel.dk> :
7 *
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
12 * an existing request
13 * - elevator_dequeue_fn, called when a request is taken off the active list
14 *
15 * 20082000 Dave Jones <davej@suse.de> :
16 * Removed tests for max-bomb-segments, which was breaking elvtune
17 * when run without -bN
18 *
19 * Jens:
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
23 *
24 */
25#include <linux/kernel.h>
26#include <linux/fs.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>
39
40#include <trace/events/block.h>
41
42#include "blk.h"
43
44static DEFINE_SPINLOCK(elv_list_lock);
45static LIST_HEAD(elv_list);
46
47/*
48 * Merge hash stuff.
49 */
50#define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq))
51
52/*
53 * Query io scheduler to see if the current process issuing bio may be
54 * merged with rq.
55 */
56static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
57{
58 struct request_queue *q = rq->q;
59 struct elevator_queue *e = q->elevator;
60
61 if (e->type->ops.elevator_allow_merge_fn)
62 return e->type->ops.elevator_allow_merge_fn(q, rq, bio);
63
64 return 1;
65}
66
67/*
68 * can we safely merge with this request?
69 */
70bool elv_rq_merge_ok(struct request *rq, struct bio *bio)
71{
72 if (!blk_rq_merge_ok(rq, bio))
73 return 0;
74
75 if (!elv_iosched_allow_merge(rq, bio))
76 return 0;
77
78 return 1;
79}
80EXPORT_SYMBOL(elv_rq_merge_ok);
81
82static struct elevator_type *elevator_find(const char *name)
83{
84 struct elevator_type *e;
85
86 list_for_each_entry(e, &elv_list, list) {
87 if (!strcmp(e->elevator_name, name))
88 return e;
89 }
90
91 return NULL;
92}
93
94static void elevator_put(struct elevator_type *e)
95{
96 module_put(e->elevator_owner);
97}
98
99static struct elevator_type *elevator_get(const char *name, bool try_loading)
100{
101 struct elevator_type *e;
102
103 spin_lock(&elv_list_lock);
104
105 e = elevator_find(name);
106 if (!e && try_loading) {
107 spin_unlock(&elv_list_lock);
108 request_module("%s-iosched", name);
109 spin_lock(&elv_list_lock);
110 e = elevator_find(name);
111 }
112
113 if (e && !try_module_get(e->elevator_owner))
114 e = NULL;
115
116 spin_unlock(&elv_list_lock);
117
118 return e;
119}
120
121static char chosen_elevator[ELV_NAME_MAX];
122
123static int __init elevator_setup(char *str)
124{
125 /*
126 * Be backwards-compatible with previous kernels, so users
127 * won't get the wrong elevator.
128 */
129 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
130 return 1;
131}
132
133__setup("elevator=", elevator_setup);
134
135/* called during boot to load the elevator chosen by the elevator param */
136void __init load_default_elevator_module(void)
137{
138 struct elevator_type *e;
139
140 if (!chosen_elevator[0])
141 return;
142
143 spin_lock(&elv_list_lock);
144 e = elevator_find(chosen_elevator);
145 spin_unlock(&elv_list_lock);
146
147 if (!e)
148 request_module("%s-iosched", chosen_elevator);
149}
150
151static struct kobj_type elv_ktype;
152
153struct elevator_queue *elevator_alloc(struct request_queue *q,
154 struct elevator_type *e)
155{
156 struct elevator_queue *eq;
157
158 eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
159 if (unlikely(!eq))
160 return NULL;
161
162 eq->type = e;
163 kobject_init(&eq->kobj, &elv_ktype);
164 mutex_init(&eq->sysfs_lock);
165 hash_init(eq->hash);
166
167 return eq;
168}
169EXPORT_SYMBOL(elevator_alloc);
170
171static void elevator_release(struct kobject *kobj)
172{
173 struct elevator_queue *e;
174
175 e = container_of(kobj, struct elevator_queue, kobj);
176 elevator_put(e->type);
177 kfree(e);
178}
179
180int elevator_init(struct request_queue *q, char *name)
181{
182 struct elevator_type *e = NULL;
183 int err;
184
185 /*
186 * q->sysfs_lock must be held to provide mutual exclusion between
187 * elevator_switch() and here.
188 */
189 lockdep_assert_held(&q->sysfs_lock);
190
191 if (unlikely(q->elevator))
192 return 0;
193
194 INIT_LIST_HEAD(&q->queue_head);
195 q->last_merge = NULL;
196 q->end_sector = 0;
197 q->boundary_rq = NULL;
198
199 if (name) {
200 e = elevator_get(name, true);
201 if (!e)
202 return -EINVAL;
203 }
204
205 /*
206 * Use the default elevator specified by config boot param or
207 * config option. Don't try to load modules as we could be running
208 * off async and request_module() isn't allowed from async.
209 */
210 if (!e && *chosen_elevator) {
211 e = elevator_get(chosen_elevator, false);
212 if (!e)
213 printk(KERN_ERR "I/O scheduler %s not found\n",
214 chosen_elevator);
215 }
216
217 if (!e) {
218 e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
219 if (!e) {
220 printk(KERN_ERR
221 "Default I/O scheduler not found. " \
222 "Using noop.\n");
223 e = elevator_get("noop", false);
224 }
225 }
226
227 err = e->ops.elevator_init_fn(q, e);
228 if (err)
229 elevator_put(e);
230 return err;
231}
232EXPORT_SYMBOL(elevator_init);
233
234void elevator_exit(struct elevator_queue *e)
235{
236 mutex_lock(&e->sysfs_lock);
237 if (e->type->ops.elevator_exit_fn)
238 e->type->ops.elevator_exit_fn(e);
239 mutex_unlock(&e->sysfs_lock);
240
241 kobject_put(&e->kobj);
242}
243EXPORT_SYMBOL(elevator_exit);
244
245static inline void __elv_rqhash_del(struct request *rq)
246{
247 hash_del(&rq->hash);
248 rq->cmd_flags &= ~REQ_HASHED;
249}
250
251static void elv_rqhash_del(struct request_queue *q, struct request *rq)
252{
253 if (ELV_ON_HASH(rq))
254 __elv_rqhash_del(rq);
255}
256
257static void elv_rqhash_add(struct request_queue *q, struct request *rq)
258{
259 struct elevator_queue *e = q->elevator;
260
261 BUG_ON(ELV_ON_HASH(rq));
262 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
263 rq->cmd_flags |= REQ_HASHED;
264}
265
266static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
267{
268 __elv_rqhash_del(rq);
269 elv_rqhash_add(q, rq);
270}
271
272static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
273{
274 struct elevator_queue *e = q->elevator;
275 struct hlist_node *next;
276 struct request *rq;
277
278 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
279 BUG_ON(!ELV_ON_HASH(rq));
280
281 if (unlikely(!rq_mergeable(rq))) {
282 __elv_rqhash_del(rq);
283 continue;
284 }
285
286 if (rq_hash_key(rq) == offset)
287 return rq;
288 }
289
290 return NULL;
291}
292
293/*
294 * RB-tree support functions for inserting/lookup/removal of requests
295 * in a sorted RB tree.
296 */
297void elv_rb_add(struct rb_root *root, struct request *rq)
298{
299 struct rb_node **p = &root->rb_node;
300 struct rb_node *parent = NULL;
301 struct request *__rq;
302
303 while (*p) {
304 parent = *p;
305 __rq = rb_entry(parent, struct request, rb_node);
306
307 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
308 p = &(*p)->rb_left;
309 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
310 p = &(*p)->rb_right;
311 }
312
313 rb_link_node(&rq->rb_node, parent, p);
314 rb_insert_color(&rq->rb_node, root);
315}
316EXPORT_SYMBOL(elv_rb_add);
317
318void elv_rb_del(struct rb_root *root, struct request *rq)
319{
320 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
321 rb_erase(&rq->rb_node, root);
322 RB_CLEAR_NODE(&rq->rb_node);
323}
324EXPORT_SYMBOL(elv_rb_del);
325
326struct request *elv_rb_find(struct rb_root *root, sector_t sector)
327{
328 struct rb_node *n = root->rb_node;
329 struct request *rq;
330
331 while (n) {
332 rq = rb_entry(n, struct request, rb_node);
333
334 if (sector < blk_rq_pos(rq))
335 n = n->rb_left;
336 else if (sector > blk_rq_pos(rq))
337 n = n->rb_right;
338 else
339 return rq;
340 }
341
342 return NULL;
343}
344EXPORT_SYMBOL(elv_rb_find);
345
346/*
347 * Insert rq into dispatch queue of q. Queue lock must be held on
348 * entry. rq is sort instead into the dispatch queue. To be used by
349 * specific elevators.
350 */
351void elv_dispatch_sort(struct request_queue *q, struct request *rq)
352{
353 sector_t boundary;
354 struct list_head *entry;
355 int stop_flags;
356
357 if (q->last_merge == rq)
358 q->last_merge = NULL;
359
360 elv_rqhash_del(q, rq);
361
362 q->nr_sorted--;
363
364 boundary = q->end_sector;
365 stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
366 list_for_each_prev(entry, &q->queue_head) {
367 struct request *pos = list_entry_rq(entry);
368
369 if ((rq->cmd_flags & REQ_DISCARD) !=
370 (pos->cmd_flags & REQ_DISCARD))
371 break;
372 if (rq_data_dir(rq) != rq_data_dir(pos))
373 break;
374 if (pos->cmd_flags & stop_flags)
375 break;
376 if (blk_rq_pos(rq) >= boundary) {
377 if (blk_rq_pos(pos) < boundary)
378 continue;
379 } else {
380 if (blk_rq_pos(pos) >= boundary)
381 break;
382 }
383 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
384 break;
385 }
386
387 list_add(&rq->queuelist, entry);
388}
389EXPORT_SYMBOL(elv_dispatch_sort);
390
391/*
392 * Insert rq into dispatch queue of q. Queue lock must be held on
393 * entry. rq is added to the back of the dispatch queue. To be used by
394 * specific elevators.
395 */
396void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
397{
398 if (q->last_merge == rq)
399 q->last_merge = NULL;
400
401 elv_rqhash_del(q, rq);
402
403 q->nr_sorted--;
404
405 q->end_sector = rq_end_sector(rq);
406 q->boundary_rq = rq;
407 list_add_tail(&rq->queuelist, &q->queue_head);
408}
409EXPORT_SYMBOL(elv_dispatch_add_tail);
410
411int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
412{
413 struct elevator_queue *e = q->elevator;
414 struct request *__rq;
415 int ret;
416
417 /*
418 * Levels of merges:
419 * nomerges: No merges at all attempted
420 * noxmerges: Only simple one-hit cache try
421 * merges: All merge tries attempted
422 */
423 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
424 return ELEVATOR_NO_MERGE;
425
426 /*
427 * First try one-hit cache.
428 */
429 if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) {
430 ret = blk_try_merge(q->last_merge, bio);
431 if (ret != ELEVATOR_NO_MERGE) {
432 *req = q->last_merge;
433 return ret;
434 }
435 }
436
437 if (blk_queue_noxmerges(q))
438 return ELEVATOR_NO_MERGE;
439
440 /*
441 * See if our hash lookup can find a potential backmerge.
442 */
443 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
444 if (__rq && elv_rq_merge_ok(__rq, bio)) {
445 *req = __rq;
446 return ELEVATOR_BACK_MERGE;
447 }
448
449 if (e->type->ops.elevator_merge_fn)
450 return e->type->ops.elevator_merge_fn(q, req, bio);
451
452 return ELEVATOR_NO_MERGE;
453}
454
455/*
456 * Attempt to do an insertion back merge. Only check for the case where
457 * we can append 'rq' to an existing request, so we can throw 'rq' away
458 * afterwards.
459 *
460 * Returns true if we merged, false otherwise
461 */
462static bool elv_attempt_insert_merge(struct request_queue *q,
463 struct request *rq)
464{
465 struct request *__rq;
466 bool ret;
467
468 if (blk_queue_nomerges(q))
469 return false;
470
471 /*
472 * First try one-hit cache.
473 */
474 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
475 return true;
476
477 if (blk_queue_noxmerges(q))
478 return false;
479
480 ret = false;
481 /*
482 * See if our hash lookup can find a potential backmerge.
483 */
484 while (1) {
485 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
486 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
487 break;
488
489 /* The merged request could be merged with others, try again */
490 ret = true;
491 rq = __rq;
492 }
493
494 return ret;
495}
496
497void elv_merged_request(struct request_queue *q, struct request *rq, int type)
498{
499 struct elevator_queue *e = q->elevator;
500
501 if (e->type->ops.elevator_merged_fn)
502 e->type->ops.elevator_merged_fn(q, rq, type);
503
504 if (type == ELEVATOR_BACK_MERGE)
505 elv_rqhash_reposition(q, rq);
506
507 q->last_merge = rq;
508}
509
510void elv_merge_requests(struct request_queue *q, struct request *rq,
511 struct request *next)
512{
513 struct elevator_queue *e = q->elevator;
514 const int next_sorted = next->cmd_flags & REQ_SORTED;
515
516 if (next_sorted && e->type->ops.elevator_merge_req_fn)
517 e->type->ops.elevator_merge_req_fn(q, rq, next);
518
519 elv_rqhash_reposition(q, rq);
520
521 if (next_sorted) {
522 elv_rqhash_del(q, next);
523 q->nr_sorted--;
524 }
525
526 q->last_merge = rq;
527}
528
529void elv_bio_merged(struct request_queue *q, struct request *rq,
530 struct bio *bio)
531{
532 struct elevator_queue *e = q->elevator;
533
534 if (e->type->ops.elevator_bio_merged_fn)
535 e->type->ops.elevator_bio_merged_fn(q, rq, bio);
536}
537
538#ifdef CONFIG_PM
539static void blk_pm_requeue_request(struct request *rq)
540{
541 if (rq->q->dev && !(rq->cmd_flags & REQ_PM))
542 rq->q->nr_pending--;
543}
544
545static void blk_pm_add_request(struct request_queue *q, struct request *rq)
546{
547 if (q->dev && !(rq->cmd_flags & REQ_PM) && q->nr_pending++ == 0 &&
548 (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
549 pm_request_resume(q->dev);
550}
551#else
552static inline void blk_pm_requeue_request(struct request *rq) {}
553static inline void blk_pm_add_request(struct request_queue *q,
554 struct request *rq)
555{
556}
557#endif
558
559void elv_requeue_request(struct request_queue *q, struct request *rq)
560{
561 /*
562 * it already went through dequeue, we need to decrement the
563 * in_flight count again
564 */
565 if (blk_account_rq(rq)) {
566 q->in_flight[rq_is_sync(rq)]--;
567 if (rq->cmd_flags & REQ_SORTED)
568 elv_deactivate_rq(q, rq);
569 }
570
571 rq->cmd_flags &= ~REQ_STARTED;
572
573 blk_pm_requeue_request(rq);
574
575 __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
576}
577
578void elv_drain_elevator(struct request_queue *q)
579{
580 static int printed;
581
582 lockdep_assert_held(q->queue_lock);
583
584 while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
585 ;
586 if (q->nr_sorted && printed++ < 10) {
587 printk(KERN_ERR "%s: forced dispatching is broken "
588 "(nr_sorted=%u), please report this\n",
589 q->elevator->type->elevator_name, q->nr_sorted);
590 }
591}
592
593void __elv_add_request(struct request_queue *q, struct request *rq, int where)
594{
595 trace_block_rq_insert(q, rq);
596
597 blk_pm_add_request(q, rq);
598
599 rq->q = q;
600
601 if (rq->cmd_flags & REQ_SOFTBARRIER) {
602 /* barriers are scheduling boundary, update end_sector */
603 if (rq->cmd_type == REQ_TYPE_FS) {
604 q->end_sector = rq_end_sector(rq);
605 q->boundary_rq = rq;
606 }
607 } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
608 (where == ELEVATOR_INSERT_SORT ||
609 where == ELEVATOR_INSERT_SORT_MERGE))
610 where = ELEVATOR_INSERT_BACK;
611
612 switch (where) {
613 case ELEVATOR_INSERT_REQUEUE:
614 case ELEVATOR_INSERT_FRONT:
615 rq->cmd_flags |= REQ_SOFTBARRIER;
616 list_add(&rq->queuelist, &q->queue_head);
617 break;
618
619 case ELEVATOR_INSERT_BACK:
620 rq->cmd_flags |= REQ_SOFTBARRIER;
621 elv_drain_elevator(q);
622 list_add_tail(&rq->queuelist, &q->queue_head);
623 /*
624 * We kick the queue here for the following reasons.
625 * - The elevator might have returned NULL previously
626 * to delay requests and returned them now. As the
627 * queue wasn't empty before this request, ll_rw_blk
628 * won't run the queue on return, resulting in hang.
629 * - Usually, back inserted requests won't be merged
630 * with anything. There's no point in delaying queue
631 * processing.
632 */
633 __blk_run_queue(q);
634 break;
635
636 case ELEVATOR_INSERT_SORT_MERGE:
637 /*
638 * If we succeed in merging this request with one in the
639 * queue already, we are done - rq has now been freed,
640 * so no need to do anything further.
641 */
642 if (elv_attempt_insert_merge(q, rq))
643 break;
644 case ELEVATOR_INSERT_SORT:
645 BUG_ON(rq->cmd_type != REQ_TYPE_FS);
646 rq->cmd_flags |= REQ_SORTED;
647 q->nr_sorted++;
648 if (rq_mergeable(rq)) {
649 elv_rqhash_add(q, rq);
650 if (!q->last_merge)
651 q->last_merge = rq;
652 }
653
654 /*
655 * Some ioscheds (cfq) run q->request_fn directly, so
656 * rq cannot be accessed after calling
657 * elevator_add_req_fn.
658 */
659 q->elevator->type->ops.elevator_add_req_fn(q, rq);
660 break;
661
662 case ELEVATOR_INSERT_FLUSH:
663 rq->cmd_flags |= REQ_SOFTBARRIER;
664 blk_insert_flush(rq);
665 break;
666 default:
667 printk(KERN_ERR "%s: bad insertion point %d\n",
668 __func__, where);
669 BUG();
670 }
671}
672EXPORT_SYMBOL(__elv_add_request);
673
674void elv_add_request(struct request_queue *q, struct request *rq, int where)
675{
676 unsigned long flags;
677
678 spin_lock_irqsave(q->queue_lock, flags);
679 __elv_add_request(q, rq, where);
680 spin_unlock_irqrestore(q->queue_lock, flags);
681}
682EXPORT_SYMBOL(elv_add_request);
683
684struct request *elv_latter_request(struct request_queue *q, struct request *rq)
685{
686 struct elevator_queue *e = q->elevator;
687
688 if (e->type->ops.elevator_latter_req_fn)
689 return e->type->ops.elevator_latter_req_fn(q, rq);
690 return NULL;
691}
692
693struct request *elv_former_request(struct request_queue *q, struct request *rq)
694{
695 struct elevator_queue *e = q->elevator;
696
697 if (e->type->ops.elevator_former_req_fn)
698 return e->type->ops.elevator_former_req_fn(q, rq);
699 return NULL;
700}
701
702int elv_set_request(struct request_queue *q, struct request *rq,
703 struct bio *bio, gfp_t gfp_mask)
704{
705 struct elevator_queue *e = q->elevator;
706
707 if (e->type->ops.elevator_set_req_fn)
708 return e->type->ops.elevator_set_req_fn(q, rq, bio, gfp_mask);
709 return 0;
710}
711
712void elv_put_request(struct request_queue *q, struct request *rq)
713{
714 struct elevator_queue *e = q->elevator;
715
716 if (e->type->ops.elevator_put_req_fn)
717 e->type->ops.elevator_put_req_fn(rq);
718}
719
720int elv_may_queue(struct request_queue *q, int rw)
721{
722 struct elevator_queue *e = q->elevator;
723
724 if (e->type->ops.elevator_may_queue_fn)
725 return e->type->ops.elevator_may_queue_fn(q, rw);
726
727 return ELV_MQUEUE_MAY;
728}
729
730void elv_completed_request(struct request_queue *q, struct request *rq)
731{
732 struct elevator_queue *e = q->elevator;
733
734 /*
735 * request is released from the driver, io must be done
736 */
737 if (blk_account_rq(rq)) {
738 q->in_flight[rq_is_sync(rq)]--;
739 if ((rq->cmd_flags & REQ_SORTED) &&
740 e->type->ops.elevator_completed_req_fn)
741 e->type->ops.elevator_completed_req_fn(q, rq);
742 }
743}
744
745#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
746
747static ssize_t
748elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
749{
750 struct elv_fs_entry *entry = to_elv(attr);
751 struct elevator_queue *e;
752 ssize_t error;
753
754 if (!entry->show)
755 return -EIO;
756
757 e = container_of(kobj, struct elevator_queue, kobj);
758 mutex_lock(&e->sysfs_lock);
759 error = e->type ? entry->show(e, page) : -ENOENT;
760 mutex_unlock(&e->sysfs_lock);
761 return error;
762}
763
764static ssize_t
765elv_attr_store(struct kobject *kobj, struct attribute *attr,
766 const char *page, size_t length)
767{
768 struct elv_fs_entry *entry = to_elv(attr);
769 struct elevator_queue *e;
770 ssize_t error;
771
772 if (!entry->store)
773 return -EIO;
774
775 e = container_of(kobj, struct elevator_queue, kobj);
776 mutex_lock(&e->sysfs_lock);
777 error = e->type ? entry->store(e, page, length) : -ENOENT;
778 mutex_unlock(&e->sysfs_lock);
779 return error;
780}
781
782static const struct sysfs_ops elv_sysfs_ops = {
783 .show = elv_attr_show,
784 .store = elv_attr_store,
785};
786
787static struct kobj_type elv_ktype = {
788 .sysfs_ops = &elv_sysfs_ops,
789 .release = elevator_release,
790};
791
792int elv_register_queue(struct request_queue *q)
793{
794 struct elevator_queue *e = q->elevator;
795 int error;
796
797 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
798 if (!error) {
799 struct elv_fs_entry *attr = e->type->elevator_attrs;
800 if (attr) {
801 while (attr->attr.name) {
802 if (sysfs_create_file(&e->kobj, &attr->attr))
803 break;
804 attr++;
805 }
806 }
807 kobject_uevent(&e->kobj, KOBJ_ADD);
808 e->registered = 1;
809 if (e->type->ops.elevator_registered_fn)
810 e->type->ops.elevator_registered_fn(q);
811 }
812 return error;
813}
814EXPORT_SYMBOL(elv_register_queue);
815
816void elv_unregister_queue(struct request_queue *q)
817{
818 if (q) {
819 struct elevator_queue *e = q->elevator;
820
821 kobject_uevent(&e->kobj, KOBJ_REMOVE);
822 kobject_del(&e->kobj);
823 e->registered = 0;
824 }
825}
826EXPORT_SYMBOL(elv_unregister_queue);
827
828int elv_register(struct elevator_type *e)
829{
830 char *def = "";
831
832 /* create icq_cache if requested */
833 if (e->icq_size) {
834 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
835 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
836 return -EINVAL;
837
838 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
839 "%s_io_cq", e->elevator_name);
840 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
841 e->icq_align, 0, NULL);
842 if (!e->icq_cache)
843 return -ENOMEM;
844 }
845
846 /* register, don't allow duplicate names */
847 spin_lock(&elv_list_lock);
848 if (elevator_find(e->elevator_name)) {
849 spin_unlock(&elv_list_lock);
850 if (e->icq_cache)
851 kmem_cache_destroy(e->icq_cache);
852 return -EBUSY;
853 }
854 list_add_tail(&e->list, &elv_list);
855 spin_unlock(&elv_list_lock);
856
857 /* print pretty message */
858 if (!strcmp(e->elevator_name, chosen_elevator) ||
859 (!*chosen_elevator &&
860 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
861 def = " (default)";
862
863 printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
864 def);
865 return 0;
866}
867EXPORT_SYMBOL_GPL(elv_register);
868
869void elv_unregister(struct elevator_type *e)
870{
871 /* unregister */
872 spin_lock(&elv_list_lock);
873 list_del_init(&e->list);
874 spin_unlock(&elv_list_lock);
875
876 /*
877 * Destroy icq_cache if it exists. icq's are RCU managed. Make
878 * sure all RCU operations are complete before proceeding.
879 */
880 if (e->icq_cache) {
881 rcu_barrier();
882 kmem_cache_destroy(e->icq_cache);
883 e->icq_cache = NULL;
884 }
885}
886EXPORT_SYMBOL_GPL(elv_unregister);
887
888/*
889 * switch to new_e io scheduler. be careful not to introduce deadlocks -
890 * we don't free the old io scheduler, before we have allocated what we
891 * need for the new one. this way we have a chance of going back to the old
892 * one, if the new one fails init for some reason.
893 */
894static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
895{
896 struct elevator_queue *old = q->elevator;
897 bool registered = old->registered;
898 int err;
899
900 /*
901 * Turn on BYPASS and drain all requests w/ elevator private data.
902 * Block layer doesn't call into a quiesced elevator - all requests
903 * are directly put on the dispatch list without elevator data
904 * using INSERT_BACK. All requests have SOFTBARRIER set and no
905 * merge happens either.
906 */
907 blk_queue_bypass_start(q);
908
909 /* unregister and clear all auxiliary data of the old elevator */
910 if (registered)
911 elv_unregister_queue(q);
912
913 spin_lock_irq(q->queue_lock);
914 ioc_clear_queue(q);
915 spin_unlock_irq(q->queue_lock);
916
917 /* allocate, init and register new elevator */
918 err = new_e->ops.elevator_init_fn(q, new_e);
919 if (err)
920 goto fail_init;
921
922 if (registered) {
923 err = elv_register_queue(q);
924 if (err)
925 goto fail_register;
926 }
927
928 /* done, kill the old one and finish */
929 elevator_exit(old);
930 blk_queue_bypass_end(q);
931
932 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
933
934 return 0;
935
936fail_register:
937 elevator_exit(q->elevator);
938fail_init:
939 /* switch failed, restore and re-register old elevator */
940 q->elevator = old;
941 elv_register_queue(q);
942 blk_queue_bypass_end(q);
943
944 return err;
945}
946
947/*
948 * Switch this queue to the given IO scheduler.
949 */
950static int __elevator_change(struct request_queue *q, const char *name)
951{
952 char elevator_name[ELV_NAME_MAX];
953 struct elevator_type *e;
954
955 if (!q->elevator)
956 return -ENXIO;
957
958 strlcpy(elevator_name, name, sizeof(elevator_name));
959 e = elevator_get(strstrip(elevator_name), true);
960 if (!e) {
961 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
962 return -EINVAL;
963 }
964
965 if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
966 elevator_put(e);
967 return 0;
968 }
969
970 return elevator_switch(q, e);
971}
972
973int elevator_change(struct request_queue *q, const char *name)
974{
975 int ret;
976
977 /* Protect q->elevator from elevator_init() */
978 mutex_lock(&q->sysfs_lock);
979 ret = __elevator_change(q, name);
980 mutex_unlock(&q->sysfs_lock);
981
982 return ret;
983}
984EXPORT_SYMBOL(elevator_change);
985
986ssize_t elv_iosched_store(struct request_queue *q, const char *name,
987 size_t count)
988{
989 int ret;
990
991 if (!q->elevator)
992 return count;
993
994 ret = __elevator_change(q, name);
995 if (!ret)
996 return count;
997
998 printk(KERN_ERR "elevator: switch to %s failed\n", name);
999 return ret;
1000}
1001
1002ssize_t elv_iosched_show(struct request_queue *q, char *name)
1003{
1004 struct elevator_queue *e = q->elevator;
1005 struct elevator_type *elv;
1006 struct elevator_type *__e;
1007 int len = 0;
1008
1009 if (!q->elevator || !blk_queue_stackable(q))
1010 return sprintf(name, "none\n");
1011
1012 elv = e->type;
1013
1014 spin_lock(&elv_list_lock);
1015 list_for_each_entry(__e, &elv_list, list) {
1016 if (!strcmp(elv->elevator_name, __e->elevator_name))
1017 len += sprintf(name+len, "[%s] ", elv->elevator_name);
1018 else
1019 len += sprintf(name+len, "%s ", __e->elevator_name);
1020 }
1021 spin_unlock(&elv_list_lock);
1022
1023 len += sprintf(len+name, "\n");
1024 return len;
1025}
1026
1027struct request *elv_rb_former_request(struct request_queue *q,
1028 struct request *rq)
1029{
1030 struct rb_node *rbprev = rb_prev(&rq->rb_node);
1031
1032 if (rbprev)
1033 return rb_entry_rq(rbprev);
1034
1035 return NULL;
1036}
1037EXPORT_SYMBOL(elv_rb_former_request);
1038
1039struct request *elv_rb_latter_request(struct request_queue *q,
1040 struct request *rq)
1041{
1042 struct rb_node *rbnext = rb_next(&rq->rb_node);
1043
1044 if (rbnext)
1045 return rb_entry_rq(rbnext);
1046
1047 return NULL;
1048}
1049EXPORT_SYMBOL(elv_rb_latter_request);