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
39#include <trace/events/block.h>
40
41#include "blk.h"
42#include "blk-cgroup.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 goto err;
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;
168err:
169 kfree(eq);
170 elevator_put(e);
171 return NULL;
172}
173EXPORT_SYMBOL(elevator_alloc);
174
175static void elevator_release(struct kobject *kobj)
176{
177 struct elevator_queue *e;
178
179 e = container_of(kobj, struct elevator_queue, kobj);
180 elevator_put(e->type);
181 kfree(e);
182}
183
184int elevator_init(struct request_queue *q, char *name)
185{
186 struct elevator_type *e = NULL;
187 int err;
188
189 /*
190 * q->sysfs_lock must be held to provide mutual exclusion between
191 * elevator_switch() and here.
192 */
193 lockdep_assert_held(&q->sysfs_lock);
194
195 if (unlikely(q->elevator))
196 return 0;
197
198 INIT_LIST_HEAD(&q->queue_head);
199 q->last_merge = NULL;
200 q->end_sector = 0;
201 q->boundary_rq = NULL;
202
203 if (name) {
204 e = elevator_get(name, true);
205 if (!e)
206 return -EINVAL;
207 }
208
209 /*
210 * Use the default elevator specified by config boot param or
211 * config option. Don't try to load modules as we could be running
212 * off async and request_module() isn't allowed from async.
213 */
214 if (!e && *chosen_elevator) {
215 e = elevator_get(chosen_elevator, false);
216 if (!e)
217 printk(KERN_ERR "I/O scheduler %s not found\n",
218 chosen_elevator);
219 }
220
221 if (!e) {
222 e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
223 if (!e) {
224 printk(KERN_ERR
225 "Default I/O scheduler not found. " \
226 "Using noop.\n");
227 e = elevator_get("noop", false);
228 }
229 }
230
231 err = e->ops.elevator_init_fn(q, e);
232 return 0;
233}
234EXPORT_SYMBOL(elevator_init);
235
236void elevator_exit(struct elevator_queue *e)
237{
238 mutex_lock(&e->sysfs_lock);
239 if (e->type->ops.elevator_exit_fn)
240 e->type->ops.elevator_exit_fn(e);
241 mutex_unlock(&e->sysfs_lock);
242
243 kobject_put(&e->kobj);
244}
245EXPORT_SYMBOL(elevator_exit);
246
247static inline void __elv_rqhash_del(struct request *rq)
248{
249 hash_del(&rq->hash);
250 rq->cmd_flags &= ~REQ_HASHED;
251}
252
253static void elv_rqhash_del(struct request_queue *q, struct request *rq)
254{
255 if (ELV_ON_HASH(rq))
256 __elv_rqhash_del(rq);
257}
258
259static void elv_rqhash_add(struct request_queue *q, struct request *rq)
260{
261 struct elevator_queue *e = q->elevator;
262
263 BUG_ON(ELV_ON_HASH(rq));
264 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
265 rq->cmd_flags |= REQ_HASHED;
266}
267
268static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
269{
270 __elv_rqhash_del(rq);
271 elv_rqhash_add(q, rq);
272}
273
274static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
275{
276 struct elevator_queue *e = q->elevator;
277 struct hlist_node *next;
278 struct request *rq;
279
280 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
281 BUG_ON(!ELV_ON_HASH(rq));
282
283 if (unlikely(!rq_mergeable(rq))) {
284 __elv_rqhash_del(rq);
285 continue;
286 }
287
288 if (rq_hash_key(rq) == offset)
289 return rq;
290 }
291
292 return NULL;
293}
294
295/*
296 * RB-tree support functions for inserting/lookup/removal of requests
297 * in a sorted RB tree.
298 */
299void elv_rb_add(struct rb_root *root, struct request *rq)
300{
301 struct rb_node **p = &root->rb_node;
302 struct rb_node *parent = NULL;
303 struct request *__rq;
304
305 while (*p) {
306 parent = *p;
307 __rq = rb_entry(parent, struct request, rb_node);
308
309 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
310 p = &(*p)->rb_left;
311 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
312 p = &(*p)->rb_right;
313 }
314
315 rb_link_node(&rq->rb_node, parent, p);
316 rb_insert_color(&rq->rb_node, root);
317}
318EXPORT_SYMBOL(elv_rb_add);
319
320void elv_rb_del(struct rb_root *root, struct request *rq)
321{
322 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
323 rb_erase(&rq->rb_node, root);
324 RB_CLEAR_NODE(&rq->rb_node);
325}
326EXPORT_SYMBOL(elv_rb_del);
327
328struct request *elv_rb_find(struct rb_root *root, sector_t sector)
329{
330 struct rb_node *n = root->rb_node;
331 struct request *rq;
332
333 while (n) {
334 rq = rb_entry(n, struct request, rb_node);
335
336 if (sector < blk_rq_pos(rq))
337 n = n->rb_left;
338 else if (sector > blk_rq_pos(rq))
339 n = n->rb_right;
340 else
341 return rq;
342 }
343
344 return NULL;
345}
346EXPORT_SYMBOL(elv_rb_find);
347
348/*
349 * Insert rq into dispatch queue of q. Queue lock must be held on
350 * entry. rq is sort instead into the dispatch queue. To be used by
351 * specific elevators.
352 */
353void elv_dispatch_sort(struct request_queue *q, struct request *rq)
354{
355 sector_t boundary;
356 struct list_head *entry;
357 int stop_flags;
358
359 if (q->last_merge == rq)
360 q->last_merge = NULL;
361
362 elv_rqhash_del(q, rq);
363
364 q->nr_sorted--;
365
366 boundary = q->end_sector;
367 stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
368 list_for_each_prev(entry, &q->queue_head) {
369 struct request *pos = list_entry_rq(entry);
370
371 if ((rq->cmd_flags & REQ_DISCARD) !=
372 (pos->cmd_flags & REQ_DISCARD))
373 break;
374 if (rq_data_dir(rq) != rq_data_dir(pos))
375 break;
376 if (pos->cmd_flags & stop_flags)
377 break;
378 if (blk_rq_pos(rq) >= boundary) {
379 if (blk_rq_pos(pos) < boundary)
380 continue;
381 } else {
382 if (blk_rq_pos(pos) >= boundary)
383 break;
384 }
385 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
386 break;
387 }
388
389 list_add(&rq->queuelist, entry);
390}
391EXPORT_SYMBOL(elv_dispatch_sort);
392
393/*
394 * Insert rq into dispatch queue of q. Queue lock must be held on
395 * entry. rq is added to the back of the dispatch queue. To be used by
396 * specific elevators.
397 */
398void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
399{
400 if (q->last_merge == rq)
401 q->last_merge = NULL;
402
403 elv_rqhash_del(q, rq);
404
405 q->nr_sorted--;
406
407 q->end_sector = rq_end_sector(rq);
408 q->boundary_rq = rq;
409 list_add_tail(&rq->queuelist, &q->queue_head);
410}
411EXPORT_SYMBOL(elv_dispatch_add_tail);
412
413int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
414{
415 struct elevator_queue *e = q->elevator;
416 struct request *__rq;
417 int ret;
418
419 /*
420 * Levels of merges:
421 * nomerges: No merges at all attempted
422 * noxmerges: Only simple one-hit cache try
423 * merges: All merge tries attempted
424 */
425 if (blk_queue_nomerges(q))
426 return ELEVATOR_NO_MERGE;
427
428 /*
429 * First try one-hit cache.
430 */
431 if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) {
432 ret = blk_try_merge(q->last_merge, bio);
433 if (ret != ELEVATOR_NO_MERGE) {
434 *req = q->last_merge;
435 return ret;
436 }
437 }
438
439 if (blk_queue_noxmerges(q))
440 return ELEVATOR_NO_MERGE;
441
442 /*
443 * See if our hash lookup can find a potential backmerge.
444 */
445 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
446 if (__rq && elv_rq_merge_ok(__rq, bio)) {
447 *req = __rq;
448 return ELEVATOR_BACK_MERGE;
449 }
450
451 if (e->type->ops.elevator_merge_fn)
452 return e->type->ops.elevator_merge_fn(q, req, bio);
453
454 return ELEVATOR_NO_MERGE;
455}
456
457/*
458 * Attempt to do an insertion back merge. Only check for the case where
459 * we can append 'rq' to an existing request, so we can throw 'rq' away
460 * afterwards.
461 *
462 * Returns true if we merged, false otherwise
463 */
464static bool elv_attempt_insert_merge(struct request_queue *q,
465 struct request *rq)
466{
467 struct request *__rq;
468 bool ret;
469
470 if (blk_queue_nomerges(q))
471 return false;
472
473 /*
474 * First try one-hit cache.
475 */
476 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
477 return true;
478
479 if (blk_queue_noxmerges(q))
480 return false;
481
482 ret = false;
483 /*
484 * See if our hash lookup can find a potential backmerge.
485 */
486 while (1) {
487 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
488 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
489 break;
490
491 /* The merged request could be merged with others, try again */
492 ret = true;
493 rq = __rq;
494 }
495
496 return ret;
497}
498
499void elv_merged_request(struct request_queue *q, struct request *rq, int type)
500{
501 struct elevator_queue *e = q->elevator;
502
503 if (e->type->ops.elevator_merged_fn)
504 e->type->ops.elevator_merged_fn(q, rq, type);
505
506 if (type == ELEVATOR_BACK_MERGE)
507 elv_rqhash_reposition(q, rq);
508
509 q->last_merge = rq;
510}
511
512void elv_merge_requests(struct request_queue *q, struct request *rq,
513 struct request *next)
514{
515 struct elevator_queue *e = q->elevator;
516 const int next_sorted = next->cmd_flags & REQ_SORTED;
517
518 if (next_sorted && e->type->ops.elevator_merge_req_fn)
519 e->type->ops.elevator_merge_req_fn(q, rq, next);
520
521 elv_rqhash_reposition(q, rq);
522
523 if (next_sorted) {
524 elv_rqhash_del(q, next);
525 q->nr_sorted--;
526 }
527
528 q->last_merge = rq;
529}
530
531void elv_bio_merged(struct request_queue *q, struct request *rq,
532 struct bio *bio)
533{
534 struct elevator_queue *e = q->elevator;
535
536 if (e->type->ops.elevator_bio_merged_fn)
537 e->type->ops.elevator_bio_merged_fn(q, rq, bio);
538}
539
540#ifdef CONFIG_PM_RUNTIME
541static void blk_pm_requeue_request(struct request *rq)
542{
543 if (rq->q->dev && !(rq->cmd_flags & REQ_PM))
544 rq->q->nr_pending--;
545}
546
547static void blk_pm_add_request(struct request_queue *q, struct request *rq)
548{
549 if (q->dev && !(rq->cmd_flags & REQ_PM) && q->nr_pending++ == 0 &&
550 (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
551 pm_request_resume(q->dev);
552}
553#else
554static inline void blk_pm_requeue_request(struct request *rq) {}
555static inline void blk_pm_add_request(struct request_queue *q,
556 struct request *rq)
557{
558}
559#endif
560
561void elv_requeue_request(struct request_queue *q, struct request *rq)
562{
563 /*
564 * it already went through dequeue, we need to decrement the
565 * in_flight count again
566 */
567 if (blk_account_rq(rq)) {
568 q->in_flight[rq_is_sync(rq)]--;
569 if (rq->cmd_flags & REQ_SORTED)
570 elv_deactivate_rq(q, rq);
571 }
572
573 rq->cmd_flags &= ~REQ_STARTED;
574
575 blk_pm_requeue_request(rq);
576
577 __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
578}
579
580void elv_drain_elevator(struct request_queue *q)
581{
582 static int printed;
583
584 lockdep_assert_held(q->queue_lock);
585
586 while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
587 ;
588 if (q->nr_sorted && printed++ < 10) {
589 printk(KERN_ERR "%s: forced dispatching is broken "
590 "(nr_sorted=%u), please report this\n",
591 q->elevator->type->elevator_name, q->nr_sorted);
592 }
593}
594
595void __elv_add_request(struct request_queue *q, struct request *rq, int where)
596{
597 trace_block_rq_insert(q, rq);
598
599 blk_pm_add_request(q, rq);
600
601 rq->q = q;
602
603 if (rq->cmd_flags & REQ_SOFTBARRIER) {
604 /* barriers are scheduling boundary, update end_sector */
605 if (rq->cmd_type == REQ_TYPE_FS) {
606 q->end_sector = rq_end_sector(rq);
607 q->boundary_rq = rq;
608 }
609 } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
610 (where == ELEVATOR_INSERT_SORT ||
611 where == ELEVATOR_INSERT_SORT_MERGE))
612 where = ELEVATOR_INSERT_BACK;
613
614 switch (where) {
615 case ELEVATOR_INSERT_REQUEUE:
616 case ELEVATOR_INSERT_FRONT:
617 rq->cmd_flags |= REQ_SOFTBARRIER;
618 list_add(&rq->queuelist, &q->queue_head);
619 break;
620
621 case ELEVATOR_INSERT_BACK:
622 rq->cmd_flags |= REQ_SOFTBARRIER;
623 elv_drain_elevator(q);
624 list_add_tail(&rq->queuelist, &q->queue_head);
625 /*
626 * We kick the queue here for the following reasons.
627 * - The elevator might have returned NULL previously
628 * to delay requests and returned them now. As the
629 * queue wasn't empty before this request, ll_rw_blk
630 * won't run the queue on return, resulting in hang.
631 * - Usually, back inserted requests won't be merged
632 * with anything. There's no point in delaying queue
633 * processing.
634 */
635 __blk_run_queue(q);
636 break;
637
638 case ELEVATOR_INSERT_SORT_MERGE:
639 /*
640 * If we succeed in merging this request with one in the
641 * queue already, we are done - rq has now been freed,
642 * so no need to do anything further.
643 */
644 if (elv_attempt_insert_merge(q, rq))
645 break;
646 case ELEVATOR_INSERT_SORT:
647 BUG_ON(rq->cmd_type != REQ_TYPE_FS);
648 rq->cmd_flags |= REQ_SORTED;
649 q->nr_sorted++;
650 if (rq_mergeable(rq)) {
651 elv_rqhash_add(q, rq);
652 if (!q->last_merge)
653 q->last_merge = rq;
654 }
655
656 /*
657 * Some ioscheds (cfq) run q->request_fn directly, so
658 * rq cannot be accessed after calling
659 * elevator_add_req_fn.
660 */
661 q->elevator->type->ops.elevator_add_req_fn(q, rq);
662 break;
663
664 case ELEVATOR_INSERT_FLUSH:
665 rq->cmd_flags |= REQ_SOFTBARRIER;
666 blk_insert_flush(rq);
667 break;
668 default:
669 printk(KERN_ERR "%s: bad insertion point %d\n",
670 __func__, where);
671 BUG();
672 }
673}
674EXPORT_SYMBOL(__elv_add_request);
675
676void elv_add_request(struct request_queue *q, struct request *rq, int where)
677{
678 unsigned long flags;
679
680 spin_lock_irqsave(q->queue_lock, flags);
681 __elv_add_request(q, rq, where);
682 spin_unlock_irqrestore(q->queue_lock, flags);
683}
684EXPORT_SYMBOL(elv_add_request);
685
686struct request *elv_latter_request(struct request_queue *q, struct request *rq)
687{
688 struct elevator_queue *e = q->elevator;
689
690 if (e->type->ops.elevator_latter_req_fn)
691 return e->type->ops.elevator_latter_req_fn(q, rq);
692 return NULL;
693}
694
695struct request *elv_former_request(struct request_queue *q, struct request *rq)
696{
697 struct elevator_queue *e = q->elevator;
698
699 if (e->type->ops.elevator_former_req_fn)
700 return e->type->ops.elevator_former_req_fn(q, rq);
701 return NULL;
702}
703
704int elv_set_request(struct request_queue *q, struct request *rq,
705 struct bio *bio, gfp_t gfp_mask)
706{
707 struct elevator_queue *e = q->elevator;
708
709 if (e->type->ops.elevator_set_req_fn)
710 return e->type->ops.elevator_set_req_fn(q, rq, bio, gfp_mask);
711 return 0;
712}
713
714void elv_put_request(struct request_queue *q, struct request *rq)
715{
716 struct elevator_queue *e = q->elevator;
717
718 if (e->type->ops.elevator_put_req_fn)
719 e->type->ops.elevator_put_req_fn(rq);
720}
721
722int elv_may_queue(struct request_queue *q, int rw)
723{
724 struct elevator_queue *e = q->elevator;
725
726 if (e->type->ops.elevator_may_queue_fn)
727 return e->type->ops.elevator_may_queue_fn(q, rw);
728
729 return ELV_MQUEUE_MAY;
730}
731
732void elv_abort_queue(struct request_queue *q)
733{
734 struct request *rq;
735
736 blk_abort_flushes(q);
737
738 while (!list_empty(&q->queue_head)) {
739 rq = list_entry_rq(q->queue_head.next);
740 rq->cmd_flags |= REQ_QUIET;
741 trace_block_rq_abort(q, rq);
742 /*
743 * Mark this request as started so we don't trigger
744 * any debug logic in the end I/O path.
745 */
746 blk_start_request(rq);
747 __blk_end_request_all(rq, -EIO);
748 }
749}
750EXPORT_SYMBOL(elv_abort_queue);
751
752void elv_completed_request(struct request_queue *q, struct request *rq)
753{
754 struct elevator_queue *e = q->elevator;
755
756 /*
757 * request is released from the driver, io must be done
758 */
759 if (blk_account_rq(rq)) {
760 q->in_flight[rq_is_sync(rq)]--;
761 if ((rq->cmd_flags & REQ_SORTED) &&
762 e->type->ops.elevator_completed_req_fn)
763 e->type->ops.elevator_completed_req_fn(q, rq);
764 }
765}
766
767#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
768
769static ssize_t
770elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
771{
772 struct elv_fs_entry *entry = to_elv(attr);
773 struct elevator_queue *e;
774 ssize_t error;
775
776 if (!entry->show)
777 return -EIO;
778
779 e = container_of(kobj, struct elevator_queue, kobj);
780 mutex_lock(&e->sysfs_lock);
781 error = e->type ? entry->show(e, page) : -ENOENT;
782 mutex_unlock(&e->sysfs_lock);
783 return error;
784}
785
786static ssize_t
787elv_attr_store(struct kobject *kobj, struct attribute *attr,
788 const char *page, size_t length)
789{
790 struct elv_fs_entry *entry = to_elv(attr);
791 struct elevator_queue *e;
792 ssize_t error;
793
794 if (!entry->store)
795 return -EIO;
796
797 e = container_of(kobj, struct elevator_queue, kobj);
798 mutex_lock(&e->sysfs_lock);
799 error = e->type ? entry->store(e, page, length) : -ENOENT;
800 mutex_unlock(&e->sysfs_lock);
801 return error;
802}
803
804static const struct sysfs_ops elv_sysfs_ops = {
805 .show = elv_attr_show,
806 .store = elv_attr_store,
807};
808
809static struct kobj_type elv_ktype = {
810 .sysfs_ops = &elv_sysfs_ops,
811 .release = elevator_release,
812};
813
814int elv_register_queue(struct request_queue *q)
815{
816 struct elevator_queue *e = q->elevator;
817 int error;
818
819 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
820 if (!error) {
821 struct elv_fs_entry *attr = e->type->elevator_attrs;
822 if (attr) {
823 while (attr->attr.name) {
824 if (sysfs_create_file(&e->kobj, &attr->attr))
825 break;
826 attr++;
827 }
828 }
829 kobject_uevent(&e->kobj, KOBJ_ADD);
830 e->registered = 1;
831 }
832 return error;
833}
834EXPORT_SYMBOL(elv_register_queue);
835
836void elv_unregister_queue(struct request_queue *q)
837{
838 if (q) {
839 struct elevator_queue *e = q->elevator;
840
841 kobject_uevent(&e->kobj, KOBJ_REMOVE);
842 kobject_del(&e->kobj);
843 e->registered = 0;
844 }
845}
846EXPORT_SYMBOL(elv_unregister_queue);
847
848int elv_register(struct elevator_type *e)
849{
850 char *def = "";
851
852 /* create icq_cache if requested */
853 if (e->icq_size) {
854 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
855 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
856 return -EINVAL;
857
858 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
859 "%s_io_cq", e->elevator_name);
860 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
861 e->icq_align, 0, NULL);
862 if (!e->icq_cache)
863 return -ENOMEM;
864 }
865
866 /* register, don't allow duplicate names */
867 spin_lock(&elv_list_lock);
868 if (elevator_find(e->elevator_name)) {
869 spin_unlock(&elv_list_lock);
870 if (e->icq_cache)
871 kmem_cache_destroy(e->icq_cache);
872 return -EBUSY;
873 }
874 list_add_tail(&e->list, &elv_list);
875 spin_unlock(&elv_list_lock);
876
877 /* print pretty message */
878 if (!strcmp(e->elevator_name, chosen_elevator) ||
879 (!*chosen_elevator &&
880 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
881 def = " (default)";
882
883 printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
884 def);
885 return 0;
886}
887EXPORT_SYMBOL_GPL(elv_register);
888
889void elv_unregister(struct elevator_type *e)
890{
891 /* unregister */
892 spin_lock(&elv_list_lock);
893 list_del_init(&e->list);
894 spin_unlock(&elv_list_lock);
895
896 /*
897 * Destroy icq_cache if it exists. icq's are RCU managed. Make
898 * sure all RCU operations are complete before proceeding.
899 */
900 if (e->icq_cache) {
901 rcu_barrier();
902 kmem_cache_destroy(e->icq_cache);
903 e->icq_cache = NULL;
904 }
905}
906EXPORT_SYMBOL_GPL(elv_unregister);
907
908/*
909 * switch to new_e io scheduler. be careful not to introduce deadlocks -
910 * we don't free the old io scheduler, before we have allocated what we
911 * need for the new one. this way we have a chance of going back to the old
912 * one, if the new one fails init for some reason.
913 */
914static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
915{
916 struct elevator_queue *old = q->elevator;
917 bool registered = old->registered;
918 int err;
919
920 /*
921 * Turn on BYPASS and drain all requests w/ elevator private data.
922 * Block layer doesn't call into a quiesced elevator - all requests
923 * are directly put on the dispatch list without elevator data
924 * using INSERT_BACK. All requests have SOFTBARRIER set and no
925 * merge happens either.
926 */
927 blk_queue_bypass_start(q);
928
929 /* unregister and clear all auxiliary data of the old elevator */
930 if (registered)
931 elv_unregister_queue(q);
932
933 spin_lock_irq(q->queue_lock);
934 ioc_clear_queue(q);
935 spin_unlock_irq(q->queue_lock);
936
937 /* allocate, init and register new elevator */
938 err = new_e->ops.elevator_init_fn(q, new_e);
939 if (err)
940 goto fail_init;
941
942 if (registered) {
943 err = elv_register_queue(q);
944 if (err)
945 goto fail_register;
946 }
947
948 /* done, kill the old one and finish */
949 elevator_exit(old);
950 blk_queue_bypass_end(q);
951
952 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
953
954 return 0;
955
956fail_register:
957 elevator_exit(q->elevator);
958fail_init:
959 /* switch failed, restore and re-register old elevator */
960 q->elevator = old;
961 elv_register_queue(q);
962 blk_queue_bypass_end(q);
963
964 return err;
965}
966
967/*
968 * Switch this queue to the given IO scheduler.
969 */
970static int __elevator_change(struct request_queue *q, const char *name)
971{
972 char elevator_name[ELV_NAME_MAX];
973 struct elevator_type *e;
974
975 if (!q->elevator)
976 return -ENXIO;
977
978 strlcpy(elevator_name, name, sizeof(elevator_name));
979 e = elevator_get(strstrip(elevator_name), true);
980 if (!e) {
981 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
982 return -EINVAL;
983 }
984
985 if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
986 elevator_put(e);
987 return 0;
988 }
989
990 return elevator_switch(q, e);
991}
992
993int elevator_change(struct request_queue *q, const char *name)
994{
995 int ret;
996
997 /* Protect q->elevator from elevator_init() */
998 mutex_lock(&q->sysfs_lock);
999 ret = __elevator_change(q, name);
1000 mutex_unlock(&q->sysfs_lock);
1001
1002 return ret;
1003}
1004EXPORT_SYMBOL(elevator_change);
1005
1006ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1007 size_t count)
1008{
1009 int ret;
1010
1011 if (!q->elevator)
1012 return count;
1013
1014 ret = __elevator_change(q, name);
1015 if (!ret)
1016 return count;
1017
1018 printk(KERN_ERR "elevator: switch to %s failed\n", name);
1019 return ret;
1020}
1021
1022ssize_t elv_iosched_show(struct request_queue *q, char *name)
1023{
1024 struct elevator_queue *e = q->elevator;
1025 struct elevator_type *elv;
1026 struct elevator_type *__e;
1027 int len = 0;
1028
1029 if (!q->elevator || !blk_queue_stackable(q))
1030 return sprintf(name, "none\n");
1031
1032 elv = e->type;
1033
1034 spin_lock(&elv_list_lock);
1035 list_for_each_entry(__e, &elv_list, list) {
1036 if (!strcmp(elv->elevator_name, __e->elevator_name))
1037 len += sprintf(name+len, "[%s] ", elv->elevator_name);
1038 else
1039 len += sprintf(name+len, "%s ", __e->elevator_name);
1040 }
1041 spin_unlock(&elv_list_lock);
1042
1043 len += sprintf(len+name, "\n");
1044 return len;
1045}
1046
1047struct request *elv_rb_former_request(struct request_queue *q,
1048 struct request *rq)
1049{
1050 struct rb_node *rbprev = rb_prev(&rq->rb_node);
1051
1052 if (rbprev)
1053 return rb_entry_rq(rbprev);
1054
1055 return NULL;
1056}
1057EXPORT_SYMBOL(elv_rb_former_request);
1058
1059struct request *elv_rb_latter_request(struct request_queue *q,
1060 struct request *rq)
1061{
1062 struct rb_node *rbnext = rb_next(&rq->rb_node);
1063
1064 if (rbnext)
1065 return rb_entry_rq(rbnext);
1066
1067 return NULL;
1068}
1069EXPORT_SYMBOL(elv_rb_latter_request);