Loading...
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
38#include <trace/events/block.h>
39
40#include "blk.h"
41#include "blk-cgroup.h"
42
43static DEFINE_SPINLOCK(elv_list_lock);
44static LIST_HEAD(elv_list);
45
46/*
47 * Merge hash stuff.
48 */
49static const int elv_hash_shift = 6;
50#define ELV_HASH_BLOCK(sec) ((sec) >> 3)
51#define ELV_HASH_FN(sec) \
52 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
53#define ELV_HASH_ENTRIES (1 << elv_hash_shift)
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 int elv_iosched_allow_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.elevator_allow_merge_fn)
66 return e->type->ops.elevator_allow_merge_fn(q, rq, bio);
67
68 return 1;
69}
70
71/*
72 * can we safely merge with this request?
73 */
74bool elv_rq_merge_ok(struct request *rq, struct bio *bio)
75{
76 if (!blk_rq_merge_ok(rq, bio))
77 return 0;
78
79 if (!elv_iosched_allow_merge(rq, bio))
80 return 0;
81
82 return 1;
83}
84EXPORT_SYMBOL(elv_rq_merge_ok);
85
86static struct elevator_type *elevator_find(const char *name)
87{
88 struct elevator_type *e;
89
90 list_for_each_entry(e, &elv_list, list) {
91 if (!strcmp(e->elevator_name, name))
92 return e;
93 }
94
95 return NULL;
96}
97
98static void elevator_put(struct elevator_type *e)
99{
100 module_put(e->elevator_owner);
101}
102
103static struct elevator_type *elevator_get(const char *name)
104{
105 struct elevator_type *e;
106
107 spin_lock(&elv_list_lock);
108
109 e = elevator_find(name);
110 if (!e) {
111 spin_unlock(&elv_list_lock);
112 request_module("%s-iosched", name);
113 spin_lock(&elv_list_lock);
114 e = elevator_find(name);
115 }
116
117 if (e && !try_module_get(e->elevator_owner))
118 e = NULL;
119
120 spin_unlock(&elv_list_lock);
121
122 return e;
123}
124
125static char chosen_elevator[ELV_NAME_MAX];
126
127static int __init elevator_setup(char *str)
128{
129 /*
130 * Be backwards-compatible with previous kernels, so users
131 * won't get the wrong elevator.
132 */
133 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
134 return 1;
135}
136
137__setup("elevator=", elevator_setup);
138
139static struct kobj_type elv_ktype;
140
141static struct elevator_queue *elevator_alloc(struct request_queue *q,
142 struct elevator_type *e)
143{
144 struct elevator_queue *eq;
145 int i;
146
147 eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
148 if (unlikely(!eq))
149 goto err;
150
151 eq->type = e;
152 kobject_init(&eq->kobj, &elv_ktype);
153 mutex_init(&eq->sysfs_lock);
154
155 eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
156 GFP_KERNEL, q->node);
157 if (!eq->hash)
158 goto err;
159
160 for (i = 0; i < ELV_HASH_ENTRIES; i++)
161 INIT_HLIST_HEAD(&eq->hash[i]);
162
163 return eq;
164err:
165 kfree(eq);
166 elevator_put(e);
167 return NULL;
168}
169
170static void elevator_release(struct kobject *kobj)
171{
172 struct elevator_queue *e;
173
174 e = container_of(kobj, struct elevator_queue, kobj);
175 elevator_put(e->type);
176 kfree(e->hash);
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 if (unlikely(q->elevator))
186 return 0;
187
188 INIT_LIST_HEAD(&q->queue_head);
189 q->last_merge = NULL;
190 q->end_sector = 0;
191 q->boundary_rq = NULL;
192
193 if (name) {
194 e = elevator_get(name);
195 if (!e)
196 return -EINVAL;
197 }
198
199 if (!e && *chosen_elevator) {
200 e = elevator_get(chosen_elevator);
201 if (!e)
202 printk(KERN_ERR "I/O scheduler %s not found\n",
203 chosen_elevator);
204 }
205
206 if (!e) {
207 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
208 if (!e) {
209 printk(KERN_ERR
210 "Default I/O scheduler not found. " \
211 "Using noop.\n");
212 e = elevator_get("noop");
213 }
214 }
215
216 q->elevator = elevator_alloc(q, e);
217 if (!q->elevator)
218 return -ENOMEM;
219
220 err = e->ops.elevator_init_fn(q);
221 if (err) {
222 kobject_put(&q->elevator->kobj);
223 return err;
224 }
225
226 return 0;
227}
228EXPORT_SYMBOL(elevator_init);
229
230void elevator_exit(struct elevator_queue *e)
231{
232 mutex_lock(&e->sysfs_lock);
233 if (e->type->ops.elevator_exit_fn)
234 e->type->ops.elevator_exit_fn(e);
235 mutex_unlock(&e->sysfs_lock);
236
237 kobject_put(&e->kobj);
238}
239EXPORT_SYMBOL(elevator_exit);
240
241static inline void __elv_rqhash_del(struct request *rq)
242{
243 hlist_del_init(&rq->hash);
244}
245
246static void elv_rqhash_del(struct request_queue *q, struct request *rq)
247{
248 if (ELV_ON_HASH(rq))
249 __elv_rqhash_del(rq);
250}
251
252static void elv_rqhash_add(struct request_queue *q, struct request *rq)
253{
254 struct elevator_queue *e = q->elevator;
255
256 BUG_ON(ELV_ON_HASH(rq));
257 hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
258}
259
260static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
261{
262 __elv_rqhash_del(rq);
263 elv_rqhash_add(q, rq);
264}
265
266static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
267{
268 struct elevator_queue *e = q->elevator;
269 struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
270 struct hlist_node *entry, *next;
271 struct request *rq;
272
273 hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
274 BUG_ON(!ELV_ON_HASH(rq));
275
276 if (unlikely(!rq_mergeable(rq))) {
277 __elv_rqhash_del(rq);
278 continue;
279 }
280
281 if (rq_hash_key(rq) == offset)
282 return rq;
283 }
284
285 return NULL;
286}
287
288/*
289 * RB-tree support functions for inserting/lookup/removal of requests
290 * in a sorted RB tree.
291 */
292void elv_rb_add(struct rb_root *root, struct request *rq)
293{
294 struct rb_node **p = &root->rb_node;
295 struct rb_node *parent = NULL;
296 struct request *__rq;
297
298 while (*p) {
299 parent = *p;
300 __rq = rb_entry(parent, struct request, rb_node);
301
302 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
303 p = &(*p)->rb_left;
304 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
305 p = &(*p)->rb_right;
306 }
307
308 rb_link_node(&rq->rb_node, parent, p);
309 rb_insert_color(&rq->rb_node, root);
310}
311EXPORT_SYMBOL(elv_rb_add);
312
313void elv_rb_del(struct rb_root *root, struct request *rq)
314{
315 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
316 rb_erase(&rq->rb_node, root);
317 RB_CLEAR_NODE(&rq->rb_node);
318}
319EXPORT_SYMBOL(elv_rb_del);
320
321struct request *elv_rb_find(struct rb_root *root, sector_t sector)
322{
323 struct rb_node *n = root->rb_node;
324 struct request *rq;
325
326 while (n) {
327 rq = rb_entry(n, struct request, rb_node);
328
329 if (sector < blk_rq_pos(rq))
330 n = n->rb_left;
331 else if (sector > blk_rq_pos(rq))
332 n = n->rb_right;
333 else
334 return rq;
335 }
336
337 return NULL;
338}
339EXPORT_SYMBOL(elv_rb_find);
340
341/*
342 * Insert rq into dispatch queue of q. Queue lock must be held on
343 * entry. rq is sort instead into the dispatch queue. To be used by
344 * specific elevators.
345 */
346void elv_dispatch_sort(struct request_queue *q, struct request *rq)
347{
348 sector_t boundary;
349 struct list_head *entry;
350 int stop_flags;
351
352 if (q->last_merge == rq)
353 q->last_merge = NULL;
354
355 elv_rqhash_del(q, rq);
356
357 q->nr_sorted--;
358
359 boundary = q->end_sector;
360 stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
361 list_for_each_prev(entry, &q->queue_head) {
362 struct request *pos = list_entry_rq(entry);
363
364 if ((rq->cmd_flags & REQ_DISCARD) !=
365 (pos->cmd_flags & REQ_DISCARD))
366 break;
367 if (rq_data_dir(rq) != rq_data_dir(pos))
368 break;
369 if (pos->cmd_flags & stop_flags)
370 break;
371 if (blk_rq_pos(rq) >= boundary) {
372 if (blk_rq_pos(pos) < boundary)
373 continue;
374 } else {
375 if (blk_rq_pos(pos) >= boundary)
376 break;
377 }
378 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
379 break;
380 }
381
382 list_add(&rq->queuelist, entry);
383}
384EXPORT_SYMBOL(elv_dispatch_sort);
385
386/*
387 * Insert rq into dispatch queue of q. Queue lock must be held on
388 * entry. rq is added to the back of the dispatch queue. To be used by
389 * specific elevators.
390 */
391void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
392{
393 if (q->last_merge == rq)
394 q->last_merge = NULL;
395
396 elv_rqhash_del(q, rq);
397
398 q->nr_sorted--;
399
400 q->end_sector = rq_end_sector(rq);
401 q->boundary_rq = rq;
402 list_add_tail(&rq->queuelist, &q->queue_head);
403}
404EXPORT_SYMBOL(elv_dispatch_add_tail);
405
406int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
407{
408 struct elevator_queue *e = q->elevator;
409 struct request *__rq;
410 int ret;
411
412 /*
413 * Levels of merges:
414 * nomerges: No merges at all attempted
415 * noxmerges: Only simple one-hit cache try
416 * merges: All merge tries attempted
417 */
418 if (blk_queue_nomerges(q))
419 return ELEVATOR_NO_MERGE;
420
421 /*
422 * First try one-hit cache.
423 */
424 if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) {
425 ret = blk_try_merge(q->last_merge, bio);
426 if (ret != ELEVATOR_NO_MERGE) {
427 *req = q->last_merge;
428 return ret;
429 }
430 }
431
432 if (blk_queue_noxmerges(q))
433 return ELEVATOR_NO_MERGE;
434
435 /*
436 * See if our hash lookup can find a potential backmerge.
437 */
438 __rq = elv_rqhash_find(q, bio->bi_sector);
439 if (__rq && elv_rq_merge_ok(__rq, bio)) {
440 *req = __rq;
441 return ELEVATOR_BACK_MERGE;
442 }
443
444 if (e->type->ops.elevator_merge_fn)
445 return e->type->ops.elevator_merge_fn(q, req, bio);
446
447 return ELEVATOR_NO_MERGE;
448}
449
450/*
451 * Attempt to do an insertion back merge. Only check for the case where
452 * we can append 'rq' to an existing request, so we can throw 'rq' away
453 * afterwards.
454 *
455 * Returns true if we merged, false otherwise
456 */
457static bool elv_attempt_insert_merge(struct request_queue *q,
458 struct request *rq)
459{
460 struct request *__rq;
461
462 if (blk_queue_nomerges(q))
463 return false;
464
465 /*
466 * First try one-hit cache.
467 */
468 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
469 return true;
470
471 if (blk_queue_noxmerges(q))
472 return false;
473
474 /*
475 * See if our hash lookup can find a potential backmerge.
476 */
477 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
478 if (__rq && blk_attempt_req_merge(q, __rq, rq))
479 return true;
480
481 return false;
482}
483
484void elv_merged_request(struct request_queue *q, struct request *rq, int type)
485{
486 struct elevator_queue *e = q->elevator;
487
488 if (e->type->ops.elevator_merged_fn)
489 e->type->ops.elevator_merged_fn(q, rq, type);
490
491 if (type == ELEVATOR_BACK_MERGE)
492 elv_rqhash_reposition(q, rq);
493
494 q->last_merge = rq;
495}
496
497void elv_merge_requests(struct request_queue *q, struct request *rq,
498 struct request *next)
499{
500 struct elevator_queue *e = q->elevator;
501 const int next_sorted = next->cmd_flags & REQ_SORTED;
502
503 if (next_sorted && e->type->ops.elevator_merge_req_fn)
504 e->type->ops.elevator_merge_req_fn(q, rq, next);
505
506 elv_rqhash_reposition(q, rq);
507
508 if (next_sorted) {
509 elv_rqhash_del(q, next);
510 q->nr_sorted--;
511 }
512
513 q->last_merge = rq;
514}
515
516void elv_bio_merged(struct request_queue *q, struct request *rq,
517 struct bio *bio)
518{
519 struct elevator_queue *e = q->elevator;
520
521 if (e->type->ops.elevator_bio_merged_fn)
522 e->type->ops.elevator_bio_merged_fn(q, rq, bio);
523}
524
525void elv_requeue_request(struct request_queue *q, struct request *rq)
526{
527 /*
528 * it already went through dequeue, we need to decrement the
529 * in_flight count again
530 */
531 if (blk_account_rq(rq)) {
532 q->in_flight[rq_is_sync(rq)]--;
533 if (rq->cmd_flags & REQ_SORTED)
534 elv_deactivate_rq(q, rq);
535 }
536
537 rq->cmd_flags &= ~REQ_STARTED;
538
539 __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
540}
541
542void elv_drain_elevator(struct request_queue *q)
543{
544 static int printed;
545
546 lockdep_assert_held(q->queue_lock);
547
548 while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
549 ;
550 if (q->nr_sorted && printed++ < 10) {
551 printk(KERN_ERR "%s: forced dispatching is broken "
552 "(nr_sorted=%u), please report this\n",
553 q->elevator->type->elevator_name, q->nr_sorted);
554 }
555}
556
557void __elv_add_request(struct request_queue *q, struct request *rq, int where)
558{
559 trace_block_rq_insert(q, rq);
560
561 rq->q = q;
562
563 if (rq->cmd_flags & REQ_SOFTBARRIER) {
564 /* barriers are scheduling boundary, update end_sector */
565 if (rq->cmd_type == REQ_TYPE_FS ||
566 (rq->cmd_flags & REQ_DISCARD)) {
567 q->end_sector = rq_end_sector(rq);
568 q->boundary_rq = rq;
569 }
570 } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
571 (where == ELEVATOR_INSERT_SORT ||
572 where == ELEVATOR_INSERT_SORT_MERGE))
573 where = ELEVATOR_INSERT_BACK;
574
575 switch (where) {
576 case ELEVATOR_INSERT_REQUEUE:
577 case ELEVATOR_INSERT_FRONT:
578 rq->cmd_flags |= REQ_SOFTBARRIER;
579 list_add(&rq->queuelist, &q->queue_head);
580 break;
581
582 case ELEVATOR_INSERT_BACK:
583 rq->cmd_flags |= REQ_SOFTBARRIER;
584 elv_drain_elevator(q);
585 list_add_tail(&rq->queuelist, &q->queue_head);
586 /*
587 * We kick the queue here for the following reasons.
588 * - The elevator might have returned NULL previously
589 * to delay requests and returned them now. As the
590 * queue wasn't empty before this request, ll_rw_blk
591 * won't run the queue on return, resulting in hang.
592 * - Usually, back inserted requests won't be merged
593 * with anything. There's no point in delaying queue
594 * processing.
595 */
596 __blk_run_queue(q);
597 break;
598
599 case ELEVATOR_INSERT_SORT_MERGE:
600 /*
601 * If we succeed in merging this request with one in the
602 * queue already, we are done - rq has now been freed,
603 * so no need to do anything further.
604 */
605 if (elv_attempt_insert_merge(q, rq))
606 break;
607 case ELEVATOR_INSERT_SORT:
608 BUG_ON(rq->cmd_type != REQ_TYPE_FS &&
609 !(rq->cmd_flags & REQ_DISCARD));
610 rq->cmd_flags |= REQ_SORTED;
611 q->nr_sorted++;
612 if (rq_mergeable(rq)) {
613 elv_rqhash_add(q, rq);
614 if (!q->last_merge)
615 q->last_merge = rq;
616 }
617
618 /*
619 * Some ioscheds (cfq) run q->request_fn directly, so
620 * rq cannot be accessed after calling
621 * elevator_add_req_fn.
622 */
623 q->elevator->type->ops.elevator_add_req_fn(q, rq);
624 break;
625
626 case ELEVATOR_INSERT_FLUSH:
627 rq->cmd_flags |= REQ_SOFTBARRIER;
628 blk_insert_flush(rq);
629 break;
630 default:
631 printk(KERN_ERR "%s: bad insertion point %d\n",
632 __func__, where);
633 BUG();
634 }
635}
636EXPORT_SYMBOL(__elv_add_request);
637
638void elv_add_request(struct request_queue *q, struct request *rq, int where)
639{
640 unsigned long flags;
641
642 spin_lock_irqsave(q->queue_lock, flags);
643 __elv_add_request(q, rq, where);
644 spin_unlock_irqrestore(q->queue_lock, flags);
645}
646EXPORT_SYMBOL(elv_add_request);
647
648struct request *elv_latter_request(struct request_queue *q, struct request *rq)
649{
650 struct elevator_queue *e = q->elevator;
651
652 if (e->type->ops.elevator_latter_req_fn)
653 return e->type->ops.elevator_latter_req_fn(q, rq);
654 return NULL;
655}
656
657struct request *elv_former_request(struct request_queue *q, struct request *rq)
658{
659 struct elevator_queue *e = q->elevator;
660
661 if (e->type->ops.elevator_former_req_fn)
662 return e->type->ops.elevator_former_req_fn(q, rq);
663 return NULL;
664}
665
666int elv_set_request(struct request_queue *q, struct request *rq,
667 struct bio *bio, gfp_t gfp_mask)
668{
669 struct elevator_queue *e = q->elevator;
670
671 if (e->type->ops.elevator_set_req_fn)
672 return e->type->ops.elevator_set_req_fn(q, rq, bio, gfp_mask);
673 return 0;
674}
675
676void elv_put_request(struct request_queue *q, struct request *rq)
677{
678 struct elevator_queue *e = q->elevator;
679
680 if (e->type->ops.elevator_put_req_fn)
681 e->type->ops.elevator_put_req_fn(rq);
682}
683
684int elv_may_queue(struct request_queue *q, int rw)
685{
686 struct elevator_queue *e = q->elevator;
687
688 if (e->type->ops.elevator_may_queue_fn)
689 return e->type->ops.elevator_may_queue_fn(q, rw);
690
691 return ELV_MQUEUE_MAY;
692}
693
694void elv_abort_queue(struct request_queue *q)
695{
696 struct request *rq;
697
698 blk_abort_flushes(q);
699
700 while (!list_empty(&q->queue_head)) {
701 rq = list_entry_rq(q->queue_head.next);
702 rq->cmd_flags |= REQ_QUIET;
703 trace_block_rq_abort(q, rq);
704 /*
705 * Mark this request as started so we don't trigger
706 * any debug logic in the end I/O path.
707 */
708 blk_start_request(rq);
709 __blk_end_request_all(rq, -EIO);
710 }
711}
712EXPORT_SYMBOL(elv_abort_queue);
713
714void elv_completed_request(struct request_queue *q, struct request *rq)
715{
716 struct elevator_queue *e = q->elevator;
717
718 /*
719 * request is released from the driver, io must be done
720 */
721 if (blk_account_rq(rq)) {
722 q->in_flight[rq_is_sync(rq)]--;
723 if ((rq->cmd_flags & REQ_SORTED) &&
724 e->type->ops.elevator_completed_req_fn)
725 e->type->ops.elevator_completed_req_fn(q, rq);
726 }
727}
728
729#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
730
731static ssize_t
732elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
733{
734 struct elv_fs_entry *entry = to_elv(attr);
735 struct elevator_queue *e;
736 ssize_t error;
737
738 if (!entry->show)
739 return -EIO;
740
741 e = container_of(kobj, struct elevator_queue, kobj);
742 mutex_lock(&e->sysfs_lock);
743 error = e->type ? entry->show(e, page) : -ENOENT;
744 mutex_unlock(&e->sysfs_lock);
745 return error;
746}
747
748static ssize_t
749elv_attr_store(struct kobject *kobj, struct attribute *attr,
750 const char *page, size_t length)
751{
752 struct elv_fs_entry *entry = to_elv(attr);
753 struct elevator_queue *e;
754 ssize_t error;
755
756 if (!entry->store)
757 return -EIO;
758
759 e = container_of(kobj, struct elevator_queue, kobj);
760 mutex_lock(&e->sysfs_lock);
761 error = e->type ? entry->store(e, page, length) : -ENOENT;
762 mutex_unlock(&e->sysfs_lock);
763 return error;
764}
765
766static const struct sysfs_ops elv_sysfs_ops = {
767 .show = elv_attr_show,
768 .store = elv_attr_store,
769};
770
771static struct kobj_type elv_ktype = {
772 .sysfs_ops = &elv_sysfs_ops,
773 .release = elevator_release,
774};
775
776int elv_register_queue(struct request_queue *q)
777{
778 struct elevator_queue *e = q->elevator;
779 int error;
780
781 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
782 if (!error) {
783 struct elv_fs_entry *attr = e->type->elevator_attrs;
784 if (attr) {
785 while (attr->attr.name) {
786 if (sysfs_create_file(&e->kobj, &attr->attr))
787 break;
788 attr++;
789 }
790 }
791 kobject_uevent(&e->kobj, KOBJ_ADD);
792 e->registered = 1;
793 }
794 return error;
795}
796EXPORT_SYMBOL(elv_register_queue);
797
798void elv_unregister_queue(struct request_queue *q)
799{
800 if (q) {
801 struct elevator_queue *e = q->elevator;
802
803 kobject_uevent(&e->kobj, KOBJ_REMOVE);
804 kobject_del(&e->kobj);
805 e->registered = 0;
806 }
807}
808EXPORT_SYMBOL(elv_unregister_queue);
809
810int elv_register(struct elevator_type *e)
811{
812 char *def = "";
813
814 /* create icq_cache if requested */
815 if (e->icq_size) {
816 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
817 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
818 return -EINVAL;
819
820 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
821 "%s_io_cq", e->elevator_name);
822 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
823 e->icq_align, 0, NULL);
824 if (!e->icq_cache)
825 return -ENOMEM;
826 }
827
828 /* register, don't allow duplicate names */
829 spin_lock(&elv_list_lock);
830 if (elevator_find(e->elevator_name)) {
831 spin_unlock(&elv_list_lock);
832 if (e->icq_cache)
833 kmem_cache_destroy(e->icq_cache);
834 return -EBUSY;
835 }
836 list_add_tail(&e->list, &elv_list);
837 spin_unlock(&elv_list_lock);
838
839 /* print pretty message */
840 if (!strcmp(e->elevator_name, chosen_elevator) ||
841 (!*chosen_elevator &&
842 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
843 def = " (default)";
844
845 printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
846 def);
847 return 0;
848}
849EXPORT_SYMBOL_GPL(elv_register);
850
851void elv_unregister(struct elevator_type *e)
852{
853 /* unregister */
854 spin_lock(&elv_list_lock);
855 list_del_init(&e->list);
856 spin_unlock(&elv_list_lock);
857
858 /*
859 * Destroy icq_cache if it exists. icq's are RCU managed. Make
860 * sure all RCU operations are complete before proceeding.
861 */
862 if (e->icq_cache) {
863 rcu_barrier();
864 kmem_cache_destroy(e->icq_cache);
865 e->icq_cache = NULL;
866 }
867}
868EXPORT_SYMBOL_GPL(elv_unregister);
869
870/*
871 * switch to new_e io scheduler. be careful not to introduce deadlocks -
872 * we don't free the old io scheduler, before we have allocated what we
873 * need for the new one. this way we have a chance of going back to the old
874 * one, if the new one fails init for some reason.
875 */
876static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
877{
878 struct elevator_queue *old = q->elevator;
879 bool registered = old->registered;
880 int err;
881
882 /*
883 * Turn on BYPASS and drain all requests w/ elevator private data.
884 * Block layer doesn't call into a quiesced elevator - all requests
885 * are directly put on the dispatch list without elevator data
886 * using INSERT_BACK. All requests have SOFTBARRIER set and no
887 * merge happens either.
888 */
889 blk_queue_bypass_start(q);
890
891 /* unregister and clear all auxiliary data of the old elevator */
892 if (registered)
893 elv_unregister_queue(q);
894
895 spin_lock_irq(q->queue_lock);
896 ioc_clear_queue(q);
897 spin_unlock_irq(q->queue_lock);
898
899 /* allocate, init and register new elevator */
900 err = -ENOMEM;
901 q->elevator = elevator_alloc(q, new_e);
902 if (!q->elevator)
903 goto fail_init;
904
905 err = new_e->ops.elevator_init_fn(q);
906 if (err) {
907 kobject_put(&q->elevator->kobj);
908 goto fail_init;
909 }
910
911 if (registered) {
912 err = elv_register_queue(q);
913 if (err)
914 goto fail_register;
915 }
916
917 /* done, kill the old one and finish */
918 elevator_exit(old);
919 blk_queue_bypass_end(q);
920
921 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
922
923 return 0;
924
925fail_register:
926 elevator_exit(q->elevator);
927fail_init:
928 /* switch failed, restore and re-register old elevator */
929 q->elevator = old;
930 elv_register_queue(q);
931 blk_queue_bypass_end(q);
932
933 return err;
934}
935
936/*
937 * Switch this queue to the given IO scheduler.
938 */
939int elevator_change(struct request_queue *q, const char *name)
940{
941 char elevator_name[ELV_NAME_MAX];
942 struct elevator_type *e;
943
944 if (!q->elevator)
945 return -ENXIO;
946
947 strlcpy(elevator_name, name, sizeof(elevator_name));
948 e = elevator_get(strstrip(elevator_name));
949 if (!e) {
950 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
951 return -EINVAL;
952 }
953
954 if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
955 elevator_put(e);
956 return 0;
957 }
958
959 return elevator_switch(q, e);
960}
961EXPORT_SYMBOL(elevator_change);
962
963ssize_t elv_iosched_store(struct request_queue *q, const char *name,
964 size_t count)
965{
966 int ret;
967
968 if (!q->elevator)
969 return count;
970
971 ret = elevator_change(q, name);
972 if (!ret)
973 return count;
974
975 printk(KERN_ERR "elevator: switch to %s failed\n", name);
976 return ret;
977}
978
979ssize_t elv_iosched_show(struct request_queue *q, char *name)
980{
981 struct elevator_queue *e = q->elevator;
982 struct elevator_type *elv;
983 struct elevator_type *__e;
984 int len = 0;
985
986 if (!q->elevator || !blk_queue_stackable(q))
987 return sprintf(name, "none\n");
988
989 elv = e->type;
990
991 spin_lock(&elv_list_lock);
992 list_for_each_entry(__e, &elv_list, list) {
993 if (!strcmp(elv->elevator_name, __e->elevator_name))
994 len += sprintf(name+len, "[%s] ", elv->elevator_name);
995 else
996 len += sprintf(name+len, "%s ", __e->elevator_name);
997 }
998 spin_unlock(&elv_list_lock);
999
1000 len += sprintf(len+name, "\n");
1001 return len;
1002}
1003
1004struct request *elv_rb_former_request(struct request_queue *q,
1005 struct request *rq)
1006{
1007 struct rb_node *rbprev = rb_prev(&rq->rb_node);
1008
1009 if (rbprev)
1010 return rb_entry_rq(rbprev);
1011
1012 return NULL;
1013}
1014EXPORT_SYMBOL(elv_rb_former_request);
1015
1016struct request *elv_rb_latter_request(struct request_queue *q,
1017 struct request *rq)
1018{
1019 struct rb_node *rbnext = rb_next(&rq->rb_node);
1020
1021 if (rbnext)
1022 return rb_entry_rq(rbnext);
1023
1024 return NULL;
1025}
1026EXPORT_SYMBOL(elv_rb_latter_request);
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#include "blk-mq-sched.h"
44#include "blk-wbt.h"
45
46static DEFINE_SPINLOCK(elv_list_lock);
47static LIST_HEAD(elv_list);
48
49/*
50 * Merge hash stuff.
51 */
52#define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq))
53
54/*
55 * Query io scheduler to see if the current process issuing bio may be
56 * merged with rq.
57 */
58static int elv_iosched_allow_bio_merge(struct request *rq, struct bio *bio)
59{
60 struct request_queue *q = rq->q;
61 struct elevator_queue *e = q->elevator;
62
63 if (e->uses_mq && e->type->ops.mq.allow_merge)
64 return e->type->ops.mq.allow_merge(q, rq, bio);
65 else if (!e->uses_mq && e->type->ops.sq.elevator_allow_bio_merge_fn)
66 return e->type->ops.sq.elevator_allow_bio_merge_fn(q, rq, bio);
67
68 return 1;
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 bool elevator_match(const struct elevator_type *e, const char *name)
87{
88 if (!strcmp(e->elevator_name, name))
89 return true;
90 if (e->elevator_alias && !strcmp(e->elevator_alias, name))
91 return true;
92
93 return false;
94}
95
96/*
97 * Return scheduler with name 'name' and with matching 'mq capability
98 */
99static struct elevator_type *elevator_find(const char *name, bool mq)
100{
101 struct elevator_type *e;
102
103 list_for_each_entry(e, &elv_list, list) {
104 if (elevator_match(e, name) && (mq == e->uses_mq))
105 return e;
106 }
107
108 return NULL;
109}
110
111static void elevator_put(struct elevator_type *e)
112{
113 module_put(e->elevator_owner);
114}
115
116static struct elevator_type *elevator_get(struct request_queue *q,
117 const char *name, bool try_loading)
118{
119 struct elevator_type *e;
120
121 spin_lock(&elv_list_lock);
122
123 e = elevator_find(name, q->mq_ops != NULL);
124 if (!e && try_loading) {
125 spin_unlock(&elv_list_lock);
126 request_module("%s-iosched", name);
127 spin_lock(&elv_list_lock);
128 e = elevator_find(name, q->mq_ops != NULL);
129 }
130
131 if (e && !try_module_get(e->elevator_owner))
132 e = NULL;
133
134 spin_unlock(&elv_list_lock);
135 return e;
136}
137
138static char chosen_elevator[ELV_NAME_MAX];
139
140static int __init elevator_setup(char *str)
141{
142 /*
143 * Be backwards-compatible with previous kernels, so users
144 * won't get the wrong elevator.
145 */
146 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
147 return 1;
148}
149
150__setup("elevator=", elevator_setup);
151
152/* called during boot to load the elevator chosen by the elevator param */
153void __init load_default_elevator_module(void)
154{
155 struct elevator_type *e;
156
157 if (!chosen_elevator[0])
158 return;
159
160 /*
161 * Boot parameter is deprecated, we haven't supported that for MQ.
162 * Only look for non-mq schedulers from here.
163 */
164 spin_lock(&elv_list_lock);
165 e = elevator_find(chosen_elevator, false);
166 spin_unlock(&elv_list_lock);
167
168 if (!e)
169 request_module("%s-iosched", chosen_elevator);
170}
171
172static struct kobj_type elv_ktype;
173
174struct elevator_queue *elevator_alloc(struct request_queue *q,
175 struct elevator_type *e)
176{
177 struct elevator_queue *eq;
178
179 eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
180 if (unlikely(!eq))
181 return NULL;
182
183 eq->type = e;
184 kobject_init(&eq->kobj, &elv_ktype);
185 mutex_init(&eq->sysfs_lock);
186 hash_init(eq->hash);
187 eq->uses_mq = e->uses_mq;
188
189 return eq;
190}
191EXPORT_SYMBOL(elevator_alloc);
192
193static void elevator_release(struct kobject *kobj)
194{
195 struct elevator_queue *e;
196
197 e = container_of(kobj, struct elevator_queue, kobj);
198 elevator_put(e->type);
199 kfree(e);
200}
201
202int elevator_init(struct request_queue *q, char *name)
203{
204 struct elevator_type *e = NULL;
205 int err;
206
207 /*
208 * q->sysfs_lock must be held to provide mutual exclusion between
209 * elevator_switch() and here.
210 */
211 lockdep_assert_held(&q->sysfs_lock);
212
213 if (unlikely(q->elevator))
214 return 0;
215
216 INIT_LIST_HEAD(&q->queue_head);
217 q->last_merge = NULL;
218 q->end_sector = 0;
219 q->boundary_rq = NULL;
220
221 if (name) {
222 e = elevator_get(q, name, true);
223 if (!e)
224 return -EINVAL;
225 }
226
227 /*
228 * Use the default elevator specified by config boot param for
229 * non-mq devices, or by config option. Don't try to load modules
230 * as we could be running off async and request_module() isn't
231 * allowed from async.
232 */
233 if (!e && !q->mq_ops && *chosen_elevator) {
234 e = elevator_get(q, chosen_elevator, false);
235 if (!e)
236 printk(KERN_ERR "I/O scheduler %s not found\n",
237 chosen_elevator);
238 }
239
240 if (!e) {
241 /*
242 * For blk-mq devices, we default to using mq-deadline,
243 * if available, for single queue devices. If deadline
244 * isn't available OR we have multiple queues, default
245 * to "none".
246 */
247 if (q->mq_ops) {
248 if (q->nr_hw_queues == 1)
249 e = elevator_get(q, "mq-deadline", false);
250 if (!e)
251 return 0;
252 } else
253 e = elevator_get(q, CONFIG_DEFAULT_IOSCHED, false);
254
255 if (!e) {
256 printk(KERN_ERR
257 "Default I/O scheduler not found. " \
258 "Using noop.\n");
259 e = elevator_get(q, "noop", false);
260 }
261 }
262
263 if (e->uses_mq)
264 err = blk_mq_init_sched(q, e);
265 else
266 err = e->ops.sq.elevator_init_fn(q, e);
267 if (err)
268 elevator_put(e);
269 return err;
270}
271EXPORT_SYMBOL(elevator_init);
272
273void elevator_exit(struct request_queue *q, struct elevator_queue *e)
274{
275 mutex_lock(&e->sysfs_lock);
276 if (e->uses_mq && e->type->ops.mq.exit_sched)
277 blk_mq_exit_sched(q, e);
278 else if (!e->uses_mq && e->type->ops.sq.elevator_exit_fn)
279 e->type->ops.sq.elevator_exit_fn(e);
280 mutex_unlock(&e->sysfs_lock);
281
282 kobject_put(&e->kobj);
283}
284EXPORT_SYMBOL(elevator_exit);
285
286static inline void __elv_rqhash_del(struct request *rq)
287{
288 hash_del(&rq->hash);
289 rq->rq_flags &= ~RQF_HASHED;
290}
291
292void elv_rqhash_del(struct request_queue *q, struct request *rq)
293{
294 if (ELV_ON_HASH(rq))
295 __elv_rqhash_del(rq);
296}
297EXPORT_SYMBOL_GPL(elv_rqhash_del);
298
299void elv_rqhash_add(struct request_queue *q, struct request *rq)
300{
301 struct elevator_queue *e = q->elevator;
302
303 BUG_ON(ELV_ON_HASH(rq));
304 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
305 rq->rq_flags |= RQF_HASHED;
306}
307EXPORT_SYMBOL_GPL(elv_rqhash_add);
308
309void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
310{
311 __elv_rqhash_del(rq);
312 elv_rqhash_add(q, rq);
313}
314
315struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
316{
317 struct elevator_queue *e = q->elevator;
318 struct hlist_node *next;
319 struct request *rq;
320
321 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
322 BUG_ON(!ELV_ON_HASH(rq));
323
324 if (unlikely(!rq_mergeable(rq))) {
325 __elv_rqhash_del(rq);
326 continue;
327 }
328
329 if (rq_hash_key(rq) == offset)
330 return rq;
331 }
332
333 return NULL;
334}
335
336/*
337 * RB-tree support functions for inserting/lookup/removal of requests
338 * in a sorted RB tree.
339 */
340void elv_rb_add(struct rb_root *root, struct request *rq)
341{
342 struct rb_node **p = &root->rb_node;
343 struct rb_node *parent = NULL;
344 struct request *__rq;
345
346 while (*p) {
347 parent = *p;
348 __rq = rb_entry(parent, struct request, rb_node);
349
350 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
351 p = &(*p)->rb_left;
352 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
353 p = &(*p)->rb_right;
354 }
355
356 rb_link_node(&rq->rb_node, parent, p);
357 rb_insert_color(&rq->rb_node, root);
358}
359EXPORT_SYMBOL(elv_rb_add);
360
361void elv_rb_del(struct rb_root *root, struct request *rq)
362{
363 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
364 rb_erase(&rq->rb_node, root);
365 RB_CLEAR_NODE(&rq->rb_node);
366}
367EXPORT_SYMBOL(elv_rb_del);
368
369struct request *elv_rb_find(struct rb_root *root, sector_t sector)
370{
371 struct rb_node *n = root->rb_node;
372 struct request *rq;
373
374 while (n) {
375 rq = rb_entry(n, struct request, rb_node);
376
377 if (sector < blk_rq_pos(rq))
378 n = n->rb_left;
379 else if (sector > blk_rq_pos(rq))
380 n = n->rb_right;
381 else
382 return rq;
383 }
384
385 return NULL;
386}
387EXPORT_SYMBOL(elv_rb_find);
388
389/*
390 * Insert rq into dispatch queue of q. Queue lock must be held on
391 * entry. rq is sort instead into the dispatch queue. To be used by
392 * specific elevators.
393 */
394void elv_dispatch_sort(struct request_queue *q, struct request *rq)
395{
396 sector_t boundary;
397 struct list_head *entry;
398
399 if (q->last_merge == rq)
400 q->last_merge = NULL;
401
402 elv_rqhash_del(q, rq);
403
404 q->nr_sorted--;
405
406 boundary = q->end_sector;
407 list_for_each_prev(entry, &q->queue_head) {
408 struct request *pos = list_entry_rq(entry);
409
410 if (req_op(rq) != req_op(pos))
411 break;
412 if (rq_data_dir(rq) != rq_data_dir(pos))
413 break;
414 if (pos->rq_flags & (RQF_STARTED | RQF_SOFTBARRIER))
415 break;
416 if (blk_rq_pos(rq) >= boundary) {
417 if (blk_rq_pos(pos) < boundary)
418 continue;
419 } else {
420 if (blk_rq_pos(pos) >= boundary)
421 break;
422 }
423 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
424 break;
425 }
426
427 list_add(&rq->queuelist, entry);
428}
429EXPORT_SYMBOL(elv_dispatch_sort);
430
431/*
432 * Insert rq into dispatch queue of q. Queue lock must be held on
433 * entry. rq is added to the back of the dispatch queue. To be used by
434 * specific elevators.
435 */
436void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
437{
438 if (q->last_merge == rq)
439 q->last_merge = NULL;
440
441 elv_rqhash_del(q, rq);
442
443 q->nr_sorted--;
444
445 q->end_sector = rq_end_sector(rq);
446 q->boundary_rq = rq;
447 list_add_tail(&rq->queuelist, &q->queue_head);
448}
449EXPORT_SYMBOL(elv_dispatch_add_tail);
450
451enum elv_merge elv_merge(struct request_queue *q, struct request **req,
452 struct bio *bio)
453{
454 struct elevator_queue *e = q->elevator;
455 struct request *__rq;
456
457 /*
458 * Levels of merges:
459 * nomerges: No merges at all attempted
460 * noxmerges: Only simple one-hit cache try
461 * merges: All merge tries attempted
462 */
463 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
464 return ELEVATOR_NO_MERGE;
465
466 /*
467 * First try one-hit cache.
468 */
469 if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
470 enum elv_merge ret = blk_try_merge(q->last_merge, bio);
471
472 if (ret != ELEVATOR_NO_MERGE) {
473 *req = q->last_merge;
474 return ret;
475 }
476 }
477
478 if (blk_queue_noxmerges(q))
479 return ELEVATOR_NO_MERGE;
480
481 /*
482 * See if our hash lookup can find a potential backmerge.
483 */
484 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
485 if (__rq && elv_bio_merge_ok(__rq, bio)) {
486 *req = __rq;
487 return ELEVATOR_BACK_MERGE;
488 }
489
490 if (e->uses_mq && e->type->ops.mq.request_merge)
491 return e->type->ops.mq.request_merge(q, req, bio);
492 else if (!e->uses_mq && e->type->ops.sq.elevator_merge_fn)
493 return e->type->ops.sq.elevator_merge_fn(q, req, bio);
494
495 return ELEVATOR_NO_MERGE;
496}
497
498/*
499 * Attempt to do an insertion back merge. Only check for the case where
500 * we can append 'rq' to an existing request, so we can throw 'rq' away
501 * afterwards.
502 *
503 * Returns true if we merged, false otherwise
504 */
505bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq)
506{
507 struct request *__rq;
508 bool ret;
509
510 if (blk_queue_nomerges(q))
511 return false;
512
513 /*
514 * First try one-hit cache.
515 */
516 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
517 return true;
518
519 if (blk_queue_noxmerges(q))
520 return false;
521
522 ret = false;
523 /*
524 * See if our hash lookup can find a potential backmerge.
525 */
526 while (1) {
527 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
528 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
529 break;
530
531 /* The merged request could be merged with others, try again */
532 ret = true;
533 rq = __rq;
534 }
535
536 return ret;
537}
538
539void elv_merged_request(struct request_queue *q, struct request *rq,
540 enum elv_merge type)
541{
542 struct elevator_queue *e = q->elevator;
543
544 if (e->uses_mq && e->type->ops.mq.request_merged)
545 e->type->ops.mq.request_merged(q, rq, type);
546 else if (!e->uses_mq && e->type->ops.sq.elevator_merged_fn)
547 e->type->ops.sq.elevator_merged_fn(q, rq, type);
548
549 if (type == ELEVATOR_BACK_MERGE)
550 elv_rqhash_reposition(q, rq);
551
552 q->last_merge = rq;
553}
554
555void elv_merge_requests(struct request_queue *q, struct request *rq,
556 struct request *next)
557{
558 struct elevator_queue *e = q->elevator;
559 bool next_sorted = false;
560
561 if (e->uses_mq && e->type->ops.mq.requests_merged)
562 e->type->ops.mq.requests_merged(q, rq, next);
563 else if (e->type->ops.sq.elevator_merge_req_fn) {
564 next_sorted = (__force bool)(next->rq_flags & RQF_SORTED);
565 if (next_sorted)
566 e->type->ops.sq.elevator_merge_req_fn(q, rq, next);
567 }
568
569 elv_rqhash_reposition(q, rq);
570
571 if (next_sorted) {
572 elv_rqhash_del(q, next);
573 q->nr_sorted--;
574 }
575
576 q->last_merge = rq;
577}
578
579void elv_bio_merged(struct request_queue *q, struct request *rq,
580 struct bio *bio)
581{
582 struct elevator_queue *e = q->elevator;
583
584 if (WARN_ON_ONCE(e->uses_mq))
585 return;
586
587 if (e->type->ops.sq.elevator_bio_merged_fn)
588 e->type->ops.sq.elevator_bio_merged_fn(q, rq, bio);
589}
590
591#ifdef CONFIG_PM
592static void blk_pm_requeue_request(struct request *rq)
593{
594 if (rq->q->dev && !(rq->rq_flags & RQF_PM))
595 rq->q->nr_pending--;
596}
597
598static void blk_pm_add_request(struct request_queue *q, struct request *rq)
599{
600 if (q->dev && !(rq->rq_flags & RQF_PM) && q->nr_pending++ == 0 &&
601 (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
602 pm_request_resume(q->dev);
603}
604#else
605static inline void blk_pm_requeue_request(struct request *rq) {}
606static inline void blk_pm_add_request(struct request_queue *q,
607 struct request *rq)
608{
609}
610#endif
611
612void elv_requeue_request(struct request_queue *q, struct request *rq)
613{
614 /*
615 * it already went through dequeue, we need to decrement the
616 * in_flight count again
617 */
618 if (blk_account_rq(rq)) {
619 q->in_flight[rq_is_sync(rq)]--;
620 if (rq->rq_flags & RQF_SORTED)
621 elv_deactivate_rq(q, rq);
622 }
623
624 rq->rq_flags &= ~RQF_STARTED;
625
626 blk_pm_requeue_request(rq);
627
628 __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
629}
630
631void elv_drain_elevator(struct request_queue *q)
632{
633 struct elevator_queue *e = q->elevator;
634 static int printed;
635
636 if (WARN_ON_ONCE(e->uses_mq))
637 return;
638
639 lockdep_assert_held(q->queue_lock);
640
641 while (e->type->ops.sq.elevator_dispatch_fn(q, 1))
642 ;
643 if (q->nr_sorted && printed++ < 10) {
644 printk(KERN_ERR "%s: forced dispatching is broken "
645 "(nr_sorted=%u), please report this\n",
646 q->elevator->type->elevator_name, q->nr_sorted);
647 }
648}
649
650void __elv_add_request(struct request_queue *q, struct request *rq, int where)
651{
652 trace_block_rq_insert(q, rq);
653
654 blk_pm_add_request(q, rq);
655
656 rq->q = q;
657
658 if (rq->rq_flags & RQF_SOFTBARRIER) {
659 /* barriers are scheduling boundary, update end_sector */
660 if (!blk_rq_is_passthrough(rq)) {
661 q->end_sector = rq_end_sector(rq);
662 q->boundary_rq = rq;
663 }
664 } else if (!(rq->rq_flags & RQF_ELVPRIV) &&
665 (where == ELEVATOR_INSERT_SORT ||
666 where == ELEVATOR_INSERT_SORT_MERGE))
667 where = ELEVATOR_INSERT_BACK;
668
669 switch (where) {
670 case ELEVATOR_INSERT_REQUEUE:
671 case ELEVATOR_INSERT_FRONT:
672 rq->rq_flags |= RQF_SOFTBARRIER;
673 list_add(&rq->queuelist, &q->queue_head);
674 break;
675
676 case ELEVATOR_INSERT_BACK:
677 rq->rq_flags |= RQF_SOFTBARRIER;
678 elv_drain_elevator(q);
679 list_add_tail(&rq->queuelist, &q->queue_head);
680 /*
681 * We kick the queue here for the following reasons.
682 * - The elevator might have returned NULL previously
683 * to delay requests and returned them now. As the
684 * queue wasn't empty before this request, ll_rw_blk
685 * won't run the queue on return, resulting in hang.
686 * - Usually, back inserted requests won't be merged
687 * with anything. There's no point in delaying queue
688 * processing.
689 */
690 __blk_run_queue(q);
691 break;
692
693 case ELEVATOR_INSERT_SORT_MERGE:
694 /*
695 * If we succeed in merging this request with one in the
696 * queue already, we are done - rq has now been freed,
697 * so no need to do anything further.
698 */
699 if (elv_attempt_insert_merge(q, rq))
700 break;
701 /* fall through */
702 case ELEVATOR_INSERT_SORT:
703 BUG_ON(blk_rq_is_passthrough(rq));
704 rq->rq_flags |= RQF_SORTED;
705 q->nr_sorted++;
706 if (rq_mergeable(rq)) {
707 elv_rqhash_add(q, rq);
708 if (!q->last_merge)
709 q->last_merge = rq;
710 }
711
712 /*
713 * Some ioscheds (cfq) run q->request_fn directly, so
714 * rq cannot be accessed after calling
715 * elevator_add_req_fn.
716 */
717 q->elevator->type->ops.sq.elevator_add_req_fn(q, rq);
718 break;
719
720 case ELEVATOR_INSERT_FLUSH:
721 rq->rq_flags |= RQF_SOFTBARRIER;
722 blk_insert_flush(rq);
723 break;
724 default:
725 printk(KERN_ERR "%s: bad insertion point %d\n",
726 __func__, where);
727 BUG();
728 }
729}
730EXPORT_SYMBOL(__elv_add_request);
731
732void elv_add_request(struct request_queue *q, struct request *rq, int where)
733{
734 unsigned long flags;
735
736 spin_lock_irqsave(q->queue_lock, flags);
737 __elv_add_request(q, rq, where);
738 spin_unlock_irqrestore(q->queue_lock, flags);
739}
740EXPORT_SYMBOL(elv_add_request);
741
742struct request *elv_latter_request(struct request_queue *q, struct request *rq)
743{
744 struct elevator_queue *e = q->elevator;
745
746 if (e->uses_mq && e->type->ops.mq.next_request)
747 return e->type->ops.mq.next_request(q, rq);
748 else if (!e->uses_mq && e->type->ops.sq.elevator_latter_req_fn)
749 return e->type->ops.sq.elevator_latter_req_fn(q, rq);
750
751 return NULL;
752}
753
754struct request *elv_former_request(struct request_queue *q, struct request *rq)
755{
756 struct elevator_queue *e = q->elevator;
757
758 if (e->uses_mq && e->type->ops.mq.former_request)
759 return e->type->ops.mq.former_request(q, rq);
760 if (!e->uses_mq && e->type->ops.sq.elevator_former_req_fn)
761 return e->type->ops.sq.elevator_former_req_fn(q, rq);
762 return NULL;
763}
764
765int elv_set_request(struct request_queue *q, struct request *rq,
766 struct bio *bio, gfp_t gfp_mask)
767{
768 struct elevator_queue *e = q->elevator;
769
770 if (WARN_ON_ONCE(e->uses_mq))
771 return 0;
772
773 if (e->type->ops.sq.elevator_set_req_fn)
774 return e->type->ops.sq.elevator_set_req_fn(q, rq, bio, gfp_mask);
775 return 0;
776}
777
778void elv_put_request(struct request_queue *q, struct request *rq)
779{
780 struct elevator_queue *e = q->elevator;
781
782 if (WARN_ON_ONCE(e->uses_mq))
783 return;
784
785 if (e->type->ops.sq.elevator_put_req_fn)
786 e->type->ops.sq.elevator_put_req_fn(rq);
787}
788
789int elv_may_queue(struct request_queue *q, unsigned int op)
790{
791 struct elevator_queue *e = q->elevator;
792
793 if (WARN_ON_ONCE(e->uses_mq))
794 return 0;
795
796 if (e->type->ops.sq.elevator_may_queue_fn)
797 return e->type->ops.sq.elevator_may_queue_fn(q, op);
798
799 return ELV_MQUEUE_MAY;
800}
801
802void elv_completed_request(struct request_queue *q, struct request *rq)
803{
804 struct elevator_queue *e = q->elevator;
805
806 if (WARN_ON_ONCE(e->uses_mq))
807 return;
808
809 /*
810 * request is released from the driver, io must be done
811 */
812 if (blk_account_rq(rq)) {
813 q->in_flight[rq_is_sync(rq)]--;
814 if ((rq->rq_flags & RQF_SORTED) &&
815 e->type->ops.sq.elevator_completed_req_fn)
816 e->type->ops.sq.elevator_completed_req_fn(q, rq);
817 }
818}
819
820#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
821
822static ssize_t
823elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
824{
825 struct elv_fs_entry *entry = to_elv(attr);
826 struct elevator_queue *e;
827 ssize_t error;
828
829 if (!entry->show)
830 return -EIO;
831
832 e = container_of(kobj, struct elevator_queue, kobj);
833 mutex_lock(&e->sysfs_lock);
834 error = e->type ? entry->show(e, page) : -ENOENT;
835 mutex_unlock(&e->sysfs_lock);
836 return error;
837}
838
839static ssize_t
840elv_attr_store(struct kobject *kobj, struct attribute *attr,
841 const char *page, size_t length)
842{
843 struct elv_fs_entry *entry = to_elv(attr);
844 struct elevator_queue *e;
845 ssize_t error;
846
847 if (!entry->store)
848 return -EIO;
849
850 e = container_of(kobj, struct elevator_queue, kobj);
851 mutex_lock(&e->sysfs_lock);
852 error = e->type ? entry->store(e, page, length) : -ENOENT;
853 mutex_unlock(&e->sysfs_lock);
854 return error;
855}
856
857static const struct sysfs_ops elv_sysfs_ops = {
858 .show = elv_attr_show,
859 .store = elv_attr_store,
860};
861
862static struct kobj_type elv_ktype = {
863 .sysfs_ops = &elv_sysfs_ops,
864 .release = elevator_release,
865};
866
867int elv_register_queue(struct request_queue *q)
868{
869 struct elevator_queue *e = q->elevator;
870 int error;
871
872 lockdep_assert_held(&q->sysfs_lock);
873
874 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
875 if (!error) {
876 struct elv_fs_entry *attr = e->type->elevator_attrs;
877 if (attr) {
878 while (attr->attr.name) {
879 if (sysfs_create_file(&e->kobj, &attr->attr))
880 break;
881 attr++;
882 }
883 }
884 kobject_uevent(&e->kobj, KOBJ_ADD);
885 e->registered = 1;
886 if (!e->uses_mq && e->type->ops.sq.elevator_registered_fn)
887 e->type->ops.sq.elevator_registered_fn(q);
888 }
889 return error;
890}
891
892void elv_unregister_queue(struct request_queue *q)
893{
894 lockdep_assert_held(&q->sysfs_lock);
895
896 if (q) {
897 struct elevator_queue *e = q->elevator;
898
899 kobject_uevent(&e->kobj, KOBJ_REMOVE);
900 kobject_del(&e->kobj);
901 e->registered = 0;
902 /* Re-enable throttling in case elevator disabled it */
903 wbt_enable_default(q);
904 }
905}
906
907int elv_register(struct elevator_type *e)
908{
909 char *def = "";
910
911 /* create icq_cache if requested */
912 if (e->icq_size) {
913 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
914 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
915 return -EINVAL;
916
917 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
918 "%s_io_cq", e->elevator_name);
919 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
920 e->icq_align, 0, NULL);
921 if (!e->icq_cache)
922 return -ENOMEM;
923 }
924
925 /* register, don't allow duplicate names */
926 spin_lock(&elv_list_lock);
927 if (elevator_find(e->elevator_name, e->uses_mq)) {
928 spin_unlock(&elv_list_lock);
929 if (e->icq_cache)
930 kmem_cache_destroy(e->icq_cache);
931 return -EBUSY;
932 }
933 list_add_tail(&e->list, &elv_list);
934 spin_unlock(&elv_list_lock);
935
936 /* print pretty message */
937 if (elevator_match(e, chosen_elevator) ||
938 (!*chosen_elevator &&
939 elevator_match(e, CONFIG_DEFAULT_IOSCHED)))
940 def = " (default)";
941
942 printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
943 def);
944 return 0;
945}
946EXPORT_SYMBOL_GPL(elv_register);
947
948void elv_unregister(struct elevator_type *e)
949{
950 /* unregister */
951 spin_lock(&elv_list_lock);
952 list_del_init(&e->list);
953 spin_unlock(&elv_list_lock);
954
955 /*
956 * Destroy icq_cache if it exists. icq's are RCU managed. Make
957 * sure all RCU operations are complete before proceeding.
958 */
959 if (e->icq_cache) {
960 rcu_barrier();
961 kmem_cache_destroy(e->icq_cache);
962 e->icq_cache = NULL;
963 }
964}
965EXPORT_SYMBOL_GPL(elv_unregister);
966
967static int elevator_switch_mq(struct request_queue *q,
968 struct elevator_type *new_e)
969{
970 int ret;
971
972 lockdep_assert_held(&q->sysfs_lock);
973
974 blk_mq_freeze_queue(q);
975 blk_mq_quiesce_queue(q);
976
977 if (q->elevator) {
978 if (q->elevator->registered)
979 elv_unregister_queue(q);
980 ioc_clear_queue(q);
981 elevator_exit(q, q->elevator);
982 }
983
984 ret = blk_mq_init_sched(q, new_e);
985 if (ret)
986 goto out;
987
988 if (new_e) {
989 ret = elv_register_queue(q);
990 if (ret) {
991 elevator_exit(q, q->elevator);
992 goto out;
993 }
994 }
995
996 if (new_e)
997 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
998 else
999 blk_add_trace_msg(q, "elv switch: none");
1000
1001out:
1002 blk_mq_unquiesce_queue(q);
1003 blk_mq_unfreeze_queue(q);
1004 return ret;
1005}
1006
1007/*
1008 * switch to new_e io scheduler. be careful not to introduce deadlocks -
1009 * we don't free the old io scheduler, before we have allocated what we
1010 * need for the new one. this way we have a chance of going back to the old
1011 * one, if the new one fails init for some reason.
1012 */
1013static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1014{
1015 struct elevator_queue *old = q->elevator;
1016 bool old_registered = false;
1017 int err;
1018
1019 lockdep_assert_held(&q->sysfs_lock);
1020
1021 if (q->mq_ops)
1022 return elevator_switch_mq(q, new_e);
1023
1024 /*
1025 * Turn on BYPASS and drain all requests w/ elevator private data.
1026 * Block layer doesn't call into a quiesced elevator - all requests
1027 * are directly put on the dispatch list without elevator data
1028 * using INSERT_BACK. All requests have SOFTBARRIER set and no
1029 * merge happens either.
1030 */
1031 if (old) {
1032 old_registered = old->registered;
1033
1034 blk_queue_bypass_start(q);
1035
1036 /* unregister and clear all auxiliary data of the old elevator */
1037 if (old_registered)
1038 elv_unregister_queue(q);
1039
1040 ioc_clear_queue(q);
1041 }
1042
1043 /* allocate, init and register new elevator */
1044 err = new_e->ops.sq.elevator_init_fn(q, new_e);
1045 if (err)
1046 goto fail_init;
1047
1048 err = elv_register_queue(q);
1049 if (err)
1050 goto fail_register;
1051
1052 /* done, kill the old one and finish */
1053 if (old) {
1054 elevator_exit(q, old);
1055 blk_queue_bypass_end(q);
1056 }
1057
1058 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
1059
1060 return 0;
1061
1062fail_register:
1063 elevator_exit(q, q->elevator);
1064fail_init:
1065 /* switch failed, restore and re-register old elevator */
1066 if (old) {
1067 q->elevator = old;
1068 elv_register_queue(q);
1069 blk_queue_bypass_end(q);
1070 }
1071
1072 return err;
1073}
1074
1075/*
1076 * Switch this queue to the given IO scheduler.
1077 */
1078static int __elevator_change(struct request_queue *q, const char *name)
1079{
1080 char elevator_name[ELV_NAME_MAX];
1081 struct elevator_type *e;
1082
1083 /* Make sure queue is not in the middle of being removed */
1084 if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
1085 return -ENOENT;
1086
1087 /*
1088 * Special case for mq, turn off scheduling
1089 */
1090 if (q->mq_ops && !strncmp(name, "none", 4))
1091 return elevator_switch(q, NULL);
1092
1093 strlcpy(elevator_name, name, sizeof(elevator_name));
1094 e = elevator_get(q, strstrip(elevator_name), true);
1095 if (!e)
1096 return -EINVAL;
1097
1098 if (q->elevator && elevator_match(q->elevator->type, elevator_name)) {
1099 elevator_put(e);
1100 return 0;
1101 }
1102
1103 return elevator_switch(q, e);
1104}
1105
1106static inline bool elv_support_iosched(struct request_queue *q)
1107{
1108 if (q->mq_ops && q->tag_set && (q->tag_set->flags &
1109 BLK_MQ_F_NO_SCHED))
1110 return false;
1111 return true;
1112}
1113
1114ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1115 size_t count)
1116{
1117 int ret;
1118
1119 if (!(q->mq_ops || q->request_fn) || !elv_support_iosched(q))
1120 return count;
1121
1122 ret = __elevator_change(q, name);
1123 if (!ret)
1124 return count;
1125
1126 return ret;
1127}
1128
1129ssize_t elv_iosched_show(struct request_queue *q, char *name)
1130{
1131 struct elevator_queue *e = q->elevator;
1132 struct elevator_type *elv = NULL;
1133 struct elevator_type *__e;
1134 bool uses_mq = q->mq_ops != NULL;
1135 int len = 0;
1136
1137 if (!queue_is_rq_based(q))
1138 return sprintf(name, "none\n");
1139
1140 if (!q->elevator)
1141 len += sprintf(name+len, "[none] ");
1142 else
1143 elv = e->type;
1144
1145 spin_lock(&elv_list_lock);
1146 list_for_each_entry(__e, &elv_list, list) {
1147 if (elv && elevator_match(elv, __e->elevator_name) &&
1148 (__e->uses_mq == uses_mq)) {
1149 len += sprintf(name+len, "[%s] ", elv->elevator_name);
1150 continue;
1151 }
1152 if (__e->uses_mq && q->mq_ops && elv_support_iosched(q))
1153 len += sprintf(name+len, "%s ", __e->elevator_name);
1154 else if (!__e->uses_mq && !q->mq_ops)
1155 len += sprintf(name+len, "%s ", __e->elevator_name);
1156 }
1157 spin_unlock(&elv_list_lock);
1158
1159 if (q->mq_ops && q->elevator)
1160 len += sprintf(name+len, "none");
1161
1162 len += sprintf(len+name, "\n");
1163 return len;
1164}
1165
1166struct request *elv_rb_former_request(struct request_queue *q,
1167 struct request *rq)
1168{
1169 struct rb_node *rbprev = rb_prev(&rq->rb_node);
1170
1171 if (rbprev)
1172 return rb_entry_rq(rbprev);
1173
1174 return NULL;
1175}
1176EXPORT_SYMBOL(elv_rb_former_request);
1177
1178struct request *elv_rb_latter_request(struct request_queue *q,
1179 struct request *rq)
1180{
1181 struct rb_node *rbnext = rb_next(&rq->rb_node);
1182
1183 if (rbnext)
1184 return rb_entry_rq(rbnext);
1185
1186 return NULL;
1187}
1188EXPORT_SYMBOL(elv_rb_latter_request);