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#include <linux/pm_runtime.h>
38#include <linux/blk-cgroup.h>
39
40#include <trace/events/block.h>
41
42#include "blk.h"
43
44static DEFINE_SPINLOCK(elv_list_lock);
45static LIST_HEAD(elv_list);
46
47/*
48 * Merge hash stuff.
49 */
50#define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq))
51
52/*
53 * Query io scheduler to see if the current process issuing bio may be
54 * merged with rq.
55 */
56static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
57{
58 struct request_queue *q = rq->q;
59 struct elevator_queue *e = q->elevator;
60
61 if (e->type->ops.elevator_allow_merge_fn)
62 return e->type->ops.elevator_allow_merge_fn(q, rq, bio);
63
64 return 1;
65}
66
67/*
68 * can we safely merge with this request?
69 */
70bool elv_rq_merge_ok(struct request *rq, struct bio *bio)
71{
72 if (!blk_rq_merge_ok(rq, bio))
73 return 0;
74
75 if (!elv_iosched_allow_merge(rq, bio))
76 return 0;
77
78 return 1;
79}
80EXPORT_SYMBOL(elv_rq_merge_ok);
81
82static struct elevator_type *elevator_find(const char *name)
83{
84 struct elevator_type *e;
85
86 list_for_each_entry(e, &elv_list, list) {
87 if (!strcmp(e->elevator_name, name))
88 return e;
89 }
90
91 return NULL;
92}
93
94static void elevator_put(struct elevator_type *e)
95{
96 module_put(e->elevator_owner);
97}
98
99static struct elevator_type *elevator_get(const char *name, bool try_loading)
100{
101 struct elevator_type *e;
102
103 spin_lock(&elv_list_lock);
104
105 e = elevator_find(name);
106 if (!e && try_loading) {
107 spin_unlock(&elv_list_lock);
108 request_module("%s-iosched", name);
109 spin_lock(&elv_list_lock);
110 e = elevator_find(name);
111 }
112
113 if (e && !try_module_get(e->elevator_owner))
114 e = NULL;
115
116 spin_unlock(&elv_list_lock);
117
118 return e;
119}
120
121static char chosen_elevator[ELV_NAME_MAX];
122
123static int __init elevator_setup(char *str)
124{
125 /*
126 * Be backwards-compatible with previous kernels, so users
127 * won't get the wrong elevator.
128 */
129 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
130 return 1;
131}
132
133__setup("elevator=", elevator_setup);
134
135/* called during boot to load the elevator chosen by the elevator param */
136void __init load_default_elevator_module(void)
137{
138 struct elevator_type *e;
139
140 if (!chosen_elevator[0])
141 return;
142
143 spin_lock(&elv_list_lock);
144 e = elevator_find(chosen_elevator);
145 spin_unlock(&elv_list_lock);
146
147 if (!e)
148 request_module("%s-iosched", chosen_elevator);
149}
150
151static struct kobj_type elv_ktype;
152
153struct elevator_queue *elevator_alloc(struct request_queue *q,
154 struct elevator_type *e)
155{
156 struct elevator_queue *eq;
157
158 eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
159 if (unlikely(!eq))
160 return NULL;
161
162 eq->type = e;
163 kobject_init(&eq->kobj, &elv_ktype);
164 mutex_init(&eq->sysfs_lock);
165 hash_init(eq->hash);
166
167 return eq;
168}
169EXPORT_SYMBOL(elevator_alloc);
170
171static void elevator_release(struct kobject *kobj)
172{
173 struct elevator_queue *e;
174
175 e = container_of(kobj, struct elevator_queue, kobj);
176 elevator_put(e->type);
177 kfree(e);
178}
179
180int elevator_init(struct request_queue *q, char *name)
181{
182 struct elevator_type *e = NULL;
183 int err;
184
185 /*
186 * q->sysfs_lock must be held to provide mutual exclusion between
187 * elevator_switch() and here.
188 */
189 lockdep_assert_held(&q->sysfs_lock);
190
191 if (unlikely(q->elevator))
192 return 0;
193
194 INIT_LIST_HEAD(&q->queue_head);
195 q->last_merge = NULL;
196 q->end_sector = 0;
197 q->boundary_rq = NULL;
198
199 if (name) {
200 e = elevator_get(name, true);
201 if (!e)
202 return -EINVAL;
203 }
204
205 /*
206 * Use the default elevator specified by config boot param or
207 * config option. Don't try to load modules as we could be running
208 * off async and request_module() isn't allowed from async.
209 */
210 if (!e && *chosen_elevator) {
211 e = elevator_get(chosen_elevator, false);
212 if (!e)
213 printk(KERN_ERR "I/O scheduler %s not found\n",
214 chosen_elevator);
215 }
216
217 if (!e) {
218 e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
219 if (!e) {
220 printk(KERN_ERR
221 "Default I/O scheduler not found. " \
222 "Using noop.\n");
223 e = elevator_get("noop", false);
224 }
225 }
226
227 err = e->ops.elevator_init_fn(q, e);
228 if (err)
229 elevator_put(e);
230 return err;
231}
232EXPORT_SYMBOL(elevator_init);
233
234void elevator_exit(struct elevator_queue *e)
235{
236 mutex_lock(&e->sysfs_lock);
237 if (e->type->ops.elevator_exit_fn)
238 e->type->ops.elevator_exit_fn(e);
239 mutex_unlock(&e->sysfs_lock);
240
241 kobject_put(&e->kobj);
242}
243EXPORT_SYMBOL(elevator_exit);
244
245static inline void __elv_rqhash_del(struct request *rq)
246{
247 hash_del(&rq->hash);
248 rq->cmd_flags &= ~REQ_HASHED;
249}
250
251static void elv_rqhash_del(struct request_queue *q, struct request *rq)
252{
253 if (ELV_ON_HASH(rq))
254 __elv_rqhash_del(rq);
255}
256
257static void elv_rqhash_add(struct request_queue *q, struct request *rq)
258{
259 struct elevator_queue *e = q->elevator;
260
261 BUG_ON(ELV_ON_HASH(rq));
262 hash_add(e->hash, &rq->hash, rq_hash_key(rq));
263 rq->cmd_flags |= REQ_HASHED;
264}
265
266static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
267{
268 __elv_rqhash_del(rq);
269 elv_rqhash_add(q, rq);
270}
271
272static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
273{
274 struct elevator_queue *e = q->elevator;
275 struct hlist_node *next;
276 struct request *rq;
277
278 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
279 BUG_ON(!ELV_ON_HASH(rq));
280
281 if (unlikely(!rq_mergeable(rq))) {
282 __elv_rqhash_del(rq);
283 continue;
284 }
285
286 if (rq_hash_key(rq) == offset)
287 return rq;
288 }
289
290 return NULL;
291}
292
293/*
294 * RB-tree support functions for inserting/lookup/removal of requests
295 * in a sorted RB tree.
296 */
297void elv_rb_add(struct rb_root *root, struct request *rq)
298{
299 struct rb_node **p = &root->rb_node;
300 struct rb_node *parent = NULL;
301 struct request *__rq;
302
303 while (*p) {
304 parent = *p;
305 __rq = rb_entry(parent, struct request, rb_node);
306
307 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
308 p = &(*p)->rb_left;
309 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
310 p = &(*p)->rb_right;
311 }
312
313 rb_link_node(&rq->rb_node, parent, p);
314 rb_insert_color(&rq->rb_node, root);
315}
316EXPORT_SYMBOL(elv_rb_add);
317
318void elv_rb_del(struct rb_root *root, struct request *rq)
319{
320 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
321 rb_erase(&rq->rb_node, root);
322 RB_CLEAR_NODE(&rq->rb_node);
323}
324EXPORT_SYMBOL(elv_rb_del);
325
326struct request *elv_rb_find(struct rb_root *root, sector_t sector)
327{
328 struct rb_node *n = root->rb_node;
329 struct request *rq;
330
331 while (n) {
332 rq = rb_entry(n, struct request, rb_node);
333
334 if (sector < blk_rq_pos(rq))
335 n = n->rb_left;
336 else if (sector > blk_rq_pos(rq))
337 n = n->rb_right;
338 else
339 return rq;
340 }
341
342 return NULL;
343}
344EXPORT_SYMBOL(elv_rb_find);
345
346/*
347 * Insert rq into dispatch queue of q. Queue lock must be held on
348 * entry. rq is sort instead into the dispatch queue. To be used by
349 * specific elevators.
350 */
351void elv_dispatch_sort(struct request_queue *q, struct request *rq)
352{
353 sector_t boundary;
354 struct list_head *entry;
355 int stop_flags;
356
357 if (q->last_merge == rq)
358 q->last_merge = NULL;
359
360 elv_rqhash_del(q, rq);
361
362 q->nr_sorted--;
363
364 boundary = q->end_sector;
365 stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
366 list_for_each_prev(entry, &q->queue_head) {
367 struct request *pos = list_entry_rq(entry);
368
369 if ((rq->cmd_flags & REQ_DISCARD) !=
370 (pos->cmd_flags & REQ_DISCARD))
371 break;
372 if (rq_data_dir(rq) != rq_data_dir(pos))
373 break;
374 if (pos->cmd_flags & stop_flags)
375 break;
376 if (blk_rq_pos(rq) >= boundary) {
377 if (blk_rq_pos(pos) < boundary)
378 continue;
379 } else {
380 if (blk_rq_pos(pos) >= boundary)
381 break;
382 }
383 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
384 break;
385 }
386
387 list_add(&rq->queuelist, entry);
388}
389EXPORT_SYMBOL(elv_dispatch_sort);
390
391/*
392 * Insert rq into dispatch queue of q. Queue lock must be held on
393 * entry. rq is added to the back of the dispatch queue. To be used by
394 * specific elevators.
395 */
396void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
397{
398 if (q->last_merge == rq)
399 q->last_merge = NULL;
400
401 elv_rqhash_del(q, rq);
402
403 q->nr_sorted--;
404
405 q->end_sector = rq_end_sector(rq);
406 q->boundary_rq = rq;
407 list_add_tail(&rq->queuelist, &q->queue_head);
408}
409EXPORT_SYMBOL(elv_dispatch_add_tail);
410
411int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
412{
413 struct elevator_queue *e = q->elevator;
414 struct request *__rq;
415 int ret;
416
417 /*
418 * Levels of merges:
419 * nomerges: No merges at all attempted
420 * noxmerges: Only simple one-hit cache try
421 * merges: All merge tries attempted
422 */
423 if (blk_queue_nomerges(q) || !bio_mergeable(bio))
424 return ELEVATOR_NO_MERGE;
425
426 /*
427 * First try one-hit cache.
428 */
429 if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) {
430 ret = blk_try_merge(q->last_merge, bio);
431 if (ret != ELEVATOR_NO_MERGE) {
432 *req = q->last_merge;
433 return ret;
434 }
435 }
436
437 if (blk_queue_noxmerges(q))
438 return ELEVATOR_NO_MERGE;
439
440 /*
441 * See if our hash lookup can find a potential backmerge.
442 */
443 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
444 if (__rq && elv_rq_merge_ok(__rq, bio)) {
445 *req = __rq;
446 return ELEVATOR_BACK_MERGE;
447 }
448
449 if (e->type->ops.elevator_merge_fn)
450 return e->type->ops.elevator_merge_fn(q, req, bio);
451
452 return ELEVATOR_NO_MERGE;
453}
454
455/*
456 * Attempt to do an insertion back merge. Only check for the case where
457 * we can append 'rq' to an existing request, so we can throw 'rq' away
458 * afterwards.
459 *
460 * Returns true if we merged, false otherwise
461 */
462static bool elv_attempt_insert_merge(struct request_queue *q,
463 struct request *rq)
464{
465 struct request *__rq;
466 bool ret;
467
468 if (blk_queue_nomerges(q))
469 return false;
470
471 /*
472 * First try one-hit cache.
473 */
474 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
475 return true;
476
477 if (blk_queue_noxmerges(q))
478 return false;
479
480 ret = false;
481 /*
482 * See if our hash lookup can find a potential backmerge.
483 */
484 while (1) {
485 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
486 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
487 break;
488
489 /* The merged request could be merged with others, try again */
490 ret = true;
491 rq = __rq;
492 }
493
494 return ret;
495}
496
497void elv_merged_request(struct request_queue *q, struct request *rq, int type)
498{
499 struct elevator_queue *e = q->elevator;
500
501 if (e->type->ops.elevator_merged_fn)
502 e->type->ops.elevator_merged_fn(q, rq, type);
503
504 if (type == ELEVATOR_BACK_MERGE)
505 elv_rqhash_reposition(q, rq);
506
507 q->last_merge = rq;
508}
509
510void elv_merge_requests(struct request_queue *q, struct request *rq,
511 struct request *next)
512{
513 struct elevator_queue *e = q->elevator;
514 const int next_sorted = next->cmd_flags & REQ_SORTED;
515
516 if (next_sorted && e->type->ops.elevator_merge_req_fn)
517 e->type->ops.elevator_merge_req_fn(q, rq, next);
518
519 elv_rqhash_reposition(q, rq);
520
521 if (next_sorted) {
522 elv_rqhash_del(q, next);
523 q->nr_sorted--;
524 }
525
526 q->last_merge = rq;
527}
528
529void elv_bio_merged(struct request_queue *q, struct request *rq,
530 struct bio *bio)
531{
532 struct elevator_queue *e = q->elevator;
533
534 if (e->type->ops.elevator_bio_merged_fn)
535 e->type->ops.elevator_bio_merged_fn(q, rq, bio);
536}
537
538#ifdef CONFIG_PM
539static void blk_pm_requeue_request(struct request *rq)
540{
541 if (rq->q->dev && !(rq->cmd_flags & REQ_PM))
542 rq->q->nr_pending--;
543}
544
545static void blk_pm_add_request(struct request_queue *q, struct request *rq)
546{
547 if (q->dev && !(rq->cmd_flags & REQ_PM) && q->nr_pending++ == 0 &&
548 (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
549 pm_request_resume(q->dev);
550}
551#else
552static inline void blk_pm_requeue_request(struct request *rq) {}
553static inline void blk_pm_add_request(struct request_queue *q,
554 struct request *rq)
555{
556}
557#endif
558
559void elv_requeue_request(struct request_queue *q, struct request *rq)
560{
561 /*
562 * it already went through dequeue, we need to decrement the
563 * in_flight count again
564 */
565 if (blk_account_rq(rq)) {
566 q->in_flight[rq_is_sync(rq)]--;
567 if (rq->cmd_flags & REQ_SORTED)
568 elv_deactivate_rq(q, rq);
569 }
570
571 rq->cmd_flags &= ~REQ_STARTED;
572
573 blk_pm_requeue_request(rq);
574
575 __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
576}
577
578void elv_drain_elevator(struct request_queue *q)
579{
580 static int printed;
581
582 lockdep_assert_held(q->queue_lock);
583
584 while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
585 ;
586 if (q->nr_sorted && printed++ < 10) {
587 printk(KERN_ERR "%s: forced dispatching is broken "
588 "(nr_sorted=%u), please report this\n",
589 q->elevator->type->elevator_name, q->nr_sorted);
590 }
591}
592
593void __elv_add_request(struct request_queue *q, struct request *rq, int where)
594{
595 trace_block_rq_insert(q, rq);
596
597 blk_pm_add_request(q, rq);
598
599 rq->q = q;
600
601 if (rq->cmd_flags & REQ_SOFTBARRIER) {
602 /* barriers are scheduling boundary, update end_sector */
603 if (rq->cmd_type == REQ_TYPE_FS) {
604 q->end_sector = rq_end_sector(rq);
605 q->boundary_rq = rq;
606 }
607 } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
608 (where == ELEVATOR_INSERT_SORT ||
609 where == ELEVATOR_INSERT_SORT_MERGE))
610 where = ELEVATOR_INSERT_BACK;
611
612 switch (where) {
613 case ELEVATOR_INSERT_REQUEUE:
614 case ELEVATOR_INSERT_FRONT:
615 rq->cmd_flags |= REQ_SOFTBARRIER;
616 list_add(&rq->queuelist, &q->queue_head);
617 break;
618
619 case ELEVATOR_INSERT_BACK:
620 rq->cmd_flags |= REQ_SOFTBARRIER;
621 elv_drain_elevator(q);
622 list_add_tail(&rq->queuelist, &q->queue_head);
623 /*
624 * We kick the queue here for the following reasons.
625 * - The elevator might have returned NULL previously
626 * to delay requests and returned them now. As the
627 * queue wasn't empty before this request, ll_rw_blk
628 * won't run the queue on return, resulting in hang.
629 * - Usually, back inserted requests won't be merged
630 * with anything. There's no point in delaying queue
631 * processing.
632 */
633 __blk_run_queue(q);
634 break;
635
636 case ELEVATOR_INSERT_SORT_MERGE:
637 /*
638 * If we succeed in merging this request with one in the
639 * queue already, we are done - rq has now been freed,
640 * so no need to do anything further.
641 */
642 if (elv_attempt_insert_merge(q, rq))
643 break;
644 case ELEVATOR_INSERT_SORT:
645 BUG_ON(rq->cmd_type != REQ_TYPE_FS);
646 rq->cmd_flags |= REQ_SORTED;
647 q->nr_sorted++;
648 if (rq_mergeable(rq)) {
649 elv_rqhash_add(q, rq);
650 if (!q->last_merge)
651 q->last_merge = rq;
652 }
653
654 /*
655 * Some ioscheds (cfq) run q->request_fn directly, so
656 * rq cannot be accessed after calling
657 * elevator_add_req_fn.
658 */
659 q->elevator->type->ops.elevator_add_req_fn(q, rq);
660 break;
661
662 case ELEVATOR_INSERT_FLUSH:
663 rq->cmd_flags |= REQ_SOFTBARRIER;
664 blk_insert_flush(rq);
665 break;
666 default:
667 printk(KERN_ERR "%s: bad insertion point %d\n",
668 __func__, where);
669 BUG();
670 }
671}
672EXPORT_SYMBOL(__elv_add_request);
673
674void elv_add_request(struct request_queue *q, struct request *rq, int where)
675{
676 unsigned long flags;
677
678 spin_lock_irqsave(q->queue_lock, flags);
679 __elv_add_request(q, rq, where);
680 spin_unlock_irqrestore(q->queue_lock, flags);
681}
682EXPORT_SYMBOL(elv_add_request);
683
684struct request *elv_latter_request(struct request_queue *q, struct request *rq)
685{
686 struct elevator_queue *e = q->elevator;
687
688 if (e->type->ops.elevator_latter_req_fn)
689 return e->type->ops.elevator_latter_req_fn(q, rq);
690 return NULL;
691}
692
693struct request *elv_former_request(struct request_queue *q, struct request *rq)
694{
695 struct elevator_queue *e = q->elevator;
696
697 if (e->type->ops.elevator_former_req_fn)
698 return e->type->ops.elevator_former_req_fn(q, rq);
699 return NULL;
700}
701
702int elv_set_request(struct request_queue *q, struct request *rq,
703 struct bio *bio, gfp_t gfp_mask)
704{
705 struct elevator_queue *e = q->elevator;
706
707 if (e->type->ops.elevator_set_req_fn)
708 return e->type->ops.elevator_set_req_fn(q, rq, bio, gfp_mask);
709 return 0;
710}
711
712void elv_put_request(struct request_queue *q, struct request *rq)
713{
714 struct elevator_queue *e = q->elevator;
715
716 if (e->type->ops.elevator_put_req_fn)
717 e->type->ops.elevator_put_req_fn(rq);
718}
719
720int elv_may_queue(struct request_queue *q, int rw)
721{
722 struct elevator_queue *e = q->elevator;
723
724 if (e->type->ops.elevator_may_queue_fn)
725 return e->type->ops.elevator_may_queue_fn(q, rw);
726
727 return ELV_MQUEUE_MAY;
728}
729
730void elv_completed_request(struct request_queue *q, struct request *rq)
731{
732 struct elevator_queue *e = q->elevator;
733
734 /*
735 * request is released from the driver, io must be done
736 */
737 if (blk_account_rq(rq)) {
738 q->in_flight[rq_is_sync(rq)]--;
739 if ((rq->cmd_flags & REQ_SORTED) &&
740 e->type->ops.elevator_completed_req_fn)
741 e->type->ops.elevator_completed_req_fn(q, rq);
742 }
743}
744
745#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
746
747static ssize_t
748elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
749{
750 struct elv_fs_entry *entry = to_elv(attr);
751 struct elevator_queue *e;
752 ssize_t error;
753
754 if (!entry->show)
755 return -EIO;
756
757 e = container_of(kobj, struct elevator_queue, kobj);
758 mutex_lock(&e->sysfs_lock);
759 error = e->type ? entry->show(e, page) : -ENOENT;
760 mutex_unlock(&e->sysfs_lock);
761 return error;
762}
763
764static ssize_t
765elv_attr_store(struct kobject *kobj, struct attribute *attr,
766 const char *page, size_t length)
767{
768 struct elv_fs_entry *entry = to_elv(attr);
769 struct elevator_queue *e;
770 ssize_t error;
771
772 if (!entry->store)
773 return -EIO;
774
775 e = container_of(kobj, struct elevator_queue, kobj);
776 mutex_lock(&e->sysfs_lock);
777 error = e->type ? entry->store(e, page, length) : -ENOENT;
778 mutex_unlock(&e->sysfs_lock);
779 return error;
780}
781
782static const struct sysfs_ops elv_sysfs_ops = {
783 .show = elv_attr_show,
784 .store = elv_attr_store,
785};
786
787static struct kobj_type elv_ktype = {
788 .sysfs_ops = &elv_sysfs_ops,
789 .release = elevator_release,
790};
791
792int elv_register_queue(struct request_queue *q)
793{
794 struct elevator_queue *e = q->elevator;
795 int error;
796
797 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
798 if (!error) {
799 struct elv_fs_entry *attr = e->type->elevator_attrs;
800 if (attr) {
801 while (attr->attr.name) {
802 if (sysfs_create_file(&e->kobj, &attr->attr))
803 break;
804 attr++;
805 }
806 }
807 kobject_uevent(&e->kobj, KOBJ_ADD);
808 e->registered = 1;
809 if (e->type->ops.elevator_registered_fn)
810 e->type->ops.elevator_registered_fn(q);
811 }
812 return error;
813}
814EXPORT_SYMBOL(elv_register_queue);
815
816void elv_unregister_queue(struct request_queue *q)
817{
818 if (q) {
819 struct elevator_queue *e = q->elevator;
820
821 kobject_uevent(&e->kobj, KOBJ_REMOVE);
822 kobject_del(&e->kobj);
823 e->registered = 0;
824 }
825}
826EXPORT_SYMBOL(elv_unregister_queue);
827
828int elv_register(struct elevator_type *e)
829{
830 char *def = "";
831
832 /* create icq_cache if requested */
833 if (e->icq_size) {
834 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
835 WARN_ON(e->icq_align < __alignof__(struct io_cq)))
836 return -EINVAL;
837
838 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
839 "%s_io_cq", e->elevator_name);
840 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
841 e->icq_align, 0, NULL);
842 if (!e->icq_cache)
843 return -ENOMEM;
844 }
845
846 /* register, don't allow duplicate names */
847 spin_lock(&elv_list_lock);
848 if (elevator_find(e->elevator_name)) {
849 spin_unlock(&elv_list_lock);
850 if (e->icq_cache)
851 kmem_cache_destroy(e->icq_cache);
852 return -EBUSY;
853 }
854 list_add_tail(&e->list, &elv_list);
855 spin_unlock(&elv_list_lock);
856
857 /* print pretty message */
858 if (!strcmp(e->elevator_name, chosen_elevator) ||
859 (!*chosen_elevator &&
860 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
861 def = " (default)";
862
863 printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
864 def);
865 return 0;
866}
867EXPORT_SYMBOL_GPL(elv_register);
868
869void elv_unregister(struct elevator_type *e)
870{
871 /* unregister */
872 spin_lock(&elv_list_lock);
873 list_del_init(&e->list);
874 spin_unlock(&elv_list_lock);
875
876 /*
877 * Destroy icq_cache if it exists. icq's are RCU managed. Make
878 * sure all RCU operations are complete before proceeding.
879 */
880 if (e->icq_cache) {
881 rcu_barrier();
882 kmem_cache_destroy(e->icq_cache);
883 e->icq_cache = NULL;
884 }
885}
886EXPORT_SYMBOL_GPL(elv_unregister);
887
888/*
889 * switch to new_e io scheduler. be careful not to introduce deadlocks -
890 * we don't free the old io scheduler, before we have allocated what we
891 * need for the new one. this way we have a chance of going back to the old
892 * one, if the new one fails init for some reason.
893 */
894static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
895{
896 struct elevator_queue *old = q->elevator;
897 bool registered = old->registered;
898 int err;
899
900 /*
901 * Turn on BYPASS and drain all requests w/ elevator private data.
902 * Block layer doesn't call into a quiesced elevator - all requests
903 * are directly put on the dispatch list without elevator data
904 * using INSERT_BACK. All requests have SOFTBARRIER set and no
905 * merge happens either.
906 */
907 blk_queue_bypass_start(q);
908
909 /* unregister and clear all auxiliary data of the old elevator */
910 if (registered)
911 elv_unregister_queue(q);
912
913 spin_lock_irq(q->queue_lock);
914 ioc_clear_queue(q);
915 spin_unlock_irq(q->queue_lock);
916
917 /* allocate, init and register new elevator */
918 err = new_e->ops.elevator_init_fn(q, new_e);
919 if (err)
920 goto fail_init;
921
922 if (registered) {
923 err = elv_register_queue(q);
924 if (err)
925 goto fail_register;
926 }
927
928 /* done, kill the old one and finish */
929 elevator_exit(old);
930 blk_queue_bypass_end(q);
931
932 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
933
934 return 0;
935
936fail_register:
937 elevator_exit(q->elevator);
938fail_init:
939 /* switch failed, restore and re-register old elevator */
940 q->elevator = old;
941 elv_register_queue(q);
942 blk_queue_bypass_end(q);
943
944 return err;
945}
946
947/*
948 * Switch this queue to the given IO scheduler.
949 */
950static int __elevator_change(struct request_queue *q, const char *name)
951{
952 char elevator_name[ELV_NAME_MAX];
953 struct elevator_type *e;
954
955 if (!q->elevator)
956 return -ENXIO;
957
958 strlcpy(elevator_name, name, sizeof(elevator_name));
959 e = elevator_get(strstrip(elevator_name), true);
960 if (!e) {
961 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
962 return -EINVAL;
963 }
964
965 if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
966 elevator_put(e);
967 return 0;
968 }
969
970 return elevator_switch(q, e);
971}
972
973int elevator_change(struct request_queue *q, const char *name)
974{
975 int ret;
976
977 /* Protect q->elevator from elevator_init() */
978 mutex_lock(&q->sysfs_lock);
979 ret = __elevator_change(q, name);
980 mutex_unlock(&q->sysfs_lock);
981
982 return ret;
983}
984EXPORT_SYMBOL(elevator_change);
985
986ssize_t elv_iosched_store(struct request_queue *q, const char *name,
987 size_t count)
988{
989 int ret;
990
991 if (!q->elevator)
992 return count;
993
994 ret = __elevator_change(q, name);
995 if (!ret)
996 return count;
997
998 printk(KERN_ERR "elevator: switch to %s failed\n", name);
999 return ret;
1000}
1001
1002ssize_t elv_iosched_show(struct request_queue *q, char *name)
1003{
1004 struct elevator_queue *e = q->elevator;
1005 struct elevator_type *elv;
1006 struct elevator_type *__e;
1007 int len = 0;
1008
1009 if (!q->elevator || !blk_queue_stackable(q))
1010 return sprintf(name, "none\n");
1011
1012 elv = e->type;
1013
1014 spin_lock(&elv_list_lock);
1015 list_for_each_entry(__e, &elv_list, list) {
1016 if (!strcmp(elv->elevator_name, __e->elevator_name))
1017 len += sprintf(name+len, "[%s] ", elv->elevator_name);
1018 else
1019 len += sprintf(name+len, "%s ", __e->elevator_name);
1020 }
1021 spin_unlock(&elv_list_lock);
1022
1023 len += sprintf(len+name, "\n");
1024 return len;
1025}
1026
1027struct request *elv_rb_former_request(struct request_queue *q,
1028 struct request *rq)
1029{
1030 struct rb_node *rbprev = rb_prev(&rq->rb_node);
1031
1032 if (rbprev)
1033 return rb_entry_rq(rbprev);
1034
1035 return NULL;
1036}
1037EXPORT_SYMBOL(elv_rb_former_request);
1038
1039struct request *elv_rb_latter_request(struct request_queue *q,
1040 struct request *rq)
1041{
1042 struct rb_node *rbnext = rb_next(&rq->rb_node);
1043
1044 if (rbnext)
1045 return rb_entry_rq(rbnext);
1046
1047 return NULL;
1048}
1049EXPORT_SYMBOL(elv_rb_latter_request);
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/delay.h>
35#include <linux/blktrace_api.h>
36#include <linux/hash.h>
37#include <linux/uaccess.h>
38
39#include <trace/events/block.h>
40
41#include "blk.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->ops->elevator_allow_merge_fn)
66 return e->ops->elevator_allow_merge_fn(q, rq, bio);
67
68 return 1;
69}
70
71/*
72 * can we safely merge with this request?
73 */
74int elv_rq_merge_ok(struct request *rq, struct bio *bio)
75{
76 if (!rq_mergeable(rq))
77 return 0;
78
79 /*
80 * Don't merge file system requests and discard requests
81 */
82 if ((bio->bi_rw & REQ_DISCARD) != (rq->bio->bi_rw & REQ_DISCARD))
83 return 0;
84
85 /*
86 * Don't merge discard requests and secure discard requests
87 */
88 if ((bio->bi_rw & REQ_SECURE) != (rq->bio->bi_rw & REQ_SECURE))
89 return 0;
90
91 /*
92 * different data direction or already started, don't merge
93 */
94 if (bio_data_dir(bio) != rq_data_dir(rq))
95 return 0;
96
97 /*
98 * must be same device and not a special request
99 */
100 if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
101 return 0;
102
103 /*
104 * only merge integrity protected bio into ditto rq
105 */
106 if (bio_integrity(bio) != blk_integrity_rq(rq))
107 return 0;
108
109 if (!elv_iosched_allow_merge(rq, bio))
110 return 0;
111
112 return 1;
113}
114EXPORT_SYMBOL(elv_rq_merge_ok);
115
116int elv_try_merge(struct request *__rq, struct bio *bio)
117{
118 int ret = ELEVATOR_NO_MERGE;
119
120 /*
121 * we can merge and sequence is ok, check if it's possible
122 */
123 if (elv_rq_merge_ok(__rq, bio)) {
124 if (blk_rq_pos(__rq) + blk_rq_sectors(__rq) == bio->bi_sector)
125 ret = ELEVATOR_BACK_MERGE;
126 else if (blk_rq_pos(__rq) - bio_sectors(bio) == bio->bi_sector)
127 ret = ELEVATOR_FRONT_MERGE;
128 }
129
130 return ret;
131}
132
133static struct elevator_type *elevator_find(const char *name)
134{
135 struct elevator_type *e;
136
137 list_for_each_entry(e, &elv_list, list) {
138 if (!strcmp(e->elevator_name, name))
139 return e;
140 }
141
142 return NULL;
143}
144
145static void elevator_put(struct elevator_type *e)
146{
147 module_put(e->elevator_owner);
148}
149
150static struct elevator_type *elevator_get(const char *name)
151{
152 struct elevator_type *e;
153
154 spin_lock(&elv_list_lock);
155
156 e = elevator_find(name);
157 if (!e) {
158 spin_unlock(&elv_list_lock);
159 request_module("%s-iosched", name);
160 spin_lock(&elv_list_lock);
161 e = elevator_find(name);
162 }
163
164 if (e && !try_module_get(e->elevator_owner))
165 e = NULL;
166
167 spin_unlock(&elv_list_lock);
168
169 return e;
170}
171
172static void *elevator_init_queue(struct request_queue *q,
173 struct elevator_queue *eq)
174{
175 return eq->ops->elevator_init_fn(q);
176}
177
178static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
179 void *data)
180{
181 q->elevator = eq;
182 eq->elevator_data = data;
183}
184
185static char chosen_elevator[16];
186
187static int __init elevator_setup(char *str)
188{
189 /*
190 * Be backwards-compatible with previous kernels, so users
191 * won't get the wrong elevator.
192 */
193 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
194 return 1;
195}
196
197__setup("elevator=", elevator_setup);
198
199static struct kobj_type elv_ktype;
200
201static struct elevator_queue *elevator_alloc(struct request_queue *q,
202 struct elevator_type *e)
203{
204 struct elevator_queue *eq;
205 int i;
206
207 eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
208 if (unlikely(!eq))
209 goto err;
210
211 eq->ops = &e->ops;
212 eq->elevator_type = e;
213 kobject_init(&eq->kobj, &elv_ktype);
214 mutex_init(&eq->sysfs_lock);
215
216 eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
217 GFP_KERNEL, q->node);
218 if (!eq->hash)
219 goto err;
220
221 for (i = 0; i < ELV_HASH_ENTRIES; i++)
222 INIT_HLIST_HEAD(&eq->hash[i]);
223
224 return eq;
225err:
226 kfree(eq);
227 elevator_put(e);
228 return NULL;
229}
230
231static void elevator_release(struct kobject *kobj)
232{
233 struct elevator_queue *e;
234
235 e = container_of(kobj, struct elevator_queue, kobj);
236 elevator_put(e->elevator_type);
237 kfree(e->hash);
238 kfree(e);
239}
240
241int elevator_init(struct request_queue *q, char *name)
242{
243 struct elevator_type *e = NULL;
244 struct elevator_queue *eq;
245 void *data;
246
247 if (unlikely(q->elevator))
248 return 0;
249
250 INIT_LIST_HEAD(&q->queue_head);
251 q->last_merge = NULL;
252 q->end_sector = 0;
253 q->boundary_rq = NULL;
254
255 if (name) {
256 e = elevator_get(name);
257 if (!e)
258 return -EINVAL;
259 }
260
261 if (!e && *chosen_elevator) {
262 e = elevator_get(chosen_elevator);
263 if (!e)
264 printk(KERN_ERR "I/O scheduler %s not found\n",
265 chosen_elevator);
266 }
267
268 if (!e) {
269 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
270 if (!e) {
271 printk(KERN_ERR
272 "Default I/O scheduler not found. " \
273 "Using noop.\n");
274 e = elevator_get("noop");
275 }
276 }
277
278 eq = elevator_alloc(q, e);
279 if (!eq)
280 return -ENOMEM;
281
282 data = elevator_init_queue(q, eq);
283 if (!data) {
284 kobject_put(&eq->kobj);
285 return -ENOMEM;
286 }
287
288 elevator_attach(q, eq, data);
289 return 0;
290}
291EXPORT_SYMBOL(elevator_init);
292
293void elevator_exit(struct elevator_queue *e)
294{
295 mutex_lock(&e->sysfs_lock);
296 if (e->ops->elevator_exit_fn)
297 e->ops->elevator_exit_fn(e);
298 e->ops = NULL;
299 mutex_unlock(&e->sysfs_lock);
300
301 kobject_put(&e->kobj);
302}
303EXPORT_SYMBOL(elevator_exit);
304
305static inline void __elv_rqhash_del(struct request *rq)
306{
307 hlist_del_init(&rq->hash);
308}
309
310static void elv_rqhash_del(struct request_queue *q, struct request *rq)
311{
312 if (ELV_ON_HASH(rq))
313 __elv_rqhash_del(rq);
314}
315
316static void elv_rqhash_add(struct request_queue *q, struct request *rq)
317{
318 struct elevator_queue *e = q->elevator;
319
320 BUG_ON(ELV_ON_HASH(rq));
321 hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
322}
323
324static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
325{
326 __elv_rqhash_del(rq);
327 elv_rqhash_add(q, rq);
328}
329
330static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
331{
332 struct elevator_queue *e = q->elevator;
333 struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
334 struct hlist_node *entry, *next;
335 struct request *rq;
336
337 hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
338 BUG_ON(!ELV_ON_HASH(rq));
339
340 if (unlikely(!rq_mergeable(rq))) {
341 __elv_rqhash_del(rq);
342 continue;
343 }
344
345 if (rq_hash_key(rq) == offset)
346 return rq;
347 }
348
349 return NULL;
350}
351
352/*
353 * RB-tree support functions for inserting/lookup/removal of requests
354 * in a sorted RB tree.
355 */
356void elv_rb_add(struct rb_root *root, struct request *rq)
357{
358 struct rb_node **p = &root->rb_node;
359 struct rb_node *parent = NULL;
360 struct request *__rq;
361
362 while (*p) {
363 parent = *p;
364 __rq = rb_entry(parent, struct request, rb_node);
365
366 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
367 p = &(*p)->rb_left;
368 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
369 p = &(*p)->rb_right;
370 }
371
372 rb_link_node(&rq->rb_node, parent, p);
373 rb_insert_color(&rq->rb_node, root);
374}
375EXPORT_SYMBOL(elv_rb_add);
376
377void elv_rb_del(struct rb_root *root, struct request *rq)
378{
379 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
380 rb_erase(&rq->rb_node, root);
381 RB_CLEAR_NODE(&rq->rb_node);
382}
383EXPORT_SYMBOL(elv_rb_del);
384
385struct request *elv_rb_find(struct rb_root *root, sector_t sector)
386{
387 struct rb_node *n = root->rb_node;
388 struct request *rq;
389
390 while (n) {
391 rq = rb_entry(n, struct request, rb_node);
392
393 if (sector < blk_rq_pos(rq))
394 n = n->rb_left;
395 else if (sector > blk_rq_pos(rq))
396 n = n->rb_right;
397 else
398 return rq;
399 }
400
401 return NULL;
402}
403EXPORT_SYMBOL(elv_rb_find);
404
405/*
406 * Insert rq into dispatch queue of q. Queue lock must be held on
407 * entry. rq is sort instead into the dispatch queue. To be used by
408 * specific elevators.
409 */
410void elv_dispatch_sort(struct request_queue *q, struct request *rq)
411{
412 sector_t boundary;
413 struct list_head *entry;
414 int stop_flags;
415
416 if (q->last_merge == rq)
417 q->last_merge = NULL;
418
419 elv_rqhash_del(q, rq);
420
421 q->nr_sorted--;
422
423 boundary = q->end_sector;
424 stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
425 list_for_each_prev(entry, &q->queue_head) {
426 struct request *pos = list_entry_rq(entry);
427
428 if ((rq->cmd_flags & REQ_DISCARD) !=
429 (pos->cmd_flags & REQ_DISCARD))
430 break;
431 if (rq_data_dir(rq) != rq_data_dir(pos))
432 break;
433 if (pos->cmd_flags & stop_flags)
434 break;
435 if (blk_rq_pos(rq) >= boundary) {
436 if (blk_rq_pos(pos) < boundary)
437 continue;
438 } else {
439 if (blk_rq_pos(pos) >= boundary)
440 break;
441 }
442 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
443 break;
444 }
445
446 list_add(&rq->queuelist, entry);
447}
448EXPORT_SYMBOL(elv_dispatch_sort);
449
450/*
451 * Insert rq into dispatch queue of q. Queue lock must be held on
452 * entry. rq is added to the back of the dispatch queue. To be used by
453 * specific elevators.
454 */
455void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
456{
457 if (q->last_merge == rq)
458 q->last_merge = NULL;
459
460 elv_rqhash_del(q, rq);
461
462 q->nr_sorted--;
463
464 q->end_sector = rq_end_sector(rq);
465 q->boundary_rq = rq;
466 list_add_tail(&rq->queuelist, &q->queue_head);
467}
468EXPORT_SYMBOL(elv_dispatch_add_tail);
469
470int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
471{
472 struct elevator_queue *e = q->elevator;
473 struct request *__rq;
474 int ret;
475
476 /*
477 * Levels of merges:
478 * nomerges: No merges at all attempted
479 * noxmerges: Only simple one-hit cache try
480 * merges: All merge tries attempted
481 */
482 if (blk_queue_nomerges(q))
483 return ELEVATOR_NO_MERGE;
484
485 /*
486 * First try one-hit cache.
487 */
488 if (q->last_merge) {
489 ret = elv_try_merge(q->last_merge, bio);
490 if (ret != ELEVATOR_NO_MERGE) {
491 *req = q->last_merge;
492 return ret;
493 }
494 }
495
496 if (blk_queue_noxmerges(q))
497 return ELEVATOR_NO_MERGE;
498
499 /*
500 * See if our hash lookup can find a potential backmerge.
501 */
502 __rq = elv_rqhash_find(q, bio->bi_sector);
503 if (__rq && elv_rq_merge_ok(__rq, bio)) {
504 *req = __rq;
505 return ELEVATOR_BACK_MERGE;
506 }
507
508 if (e->ops->elevator_merge_fn)
509 return e->ops->elevator_merge_fn(q, req, bio);
510
511 return ELEVATOR_NO_MERGE;
512}
513
514/*
515 * Attempt to do an insertion back merge. Only check for the case where
516 * we can append 'rq' to an existing request, so we can throw 'rq' away
517 * afterwards.
518 *
519 * Returns true if we merged, false otherwise
520 */
521static bool elv_attempt_insert_merge(struct request_queue *q,
522 struct request *rq)
523{
524 struct request *__rq;
525
526 if (blk_queue_nomerges(q))
527 return false;
528
529 /*
530 * First try one-hit cache.
531 */
532 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
533 return true;
534
535 if (blk_queue_noxmerges(q))
536 return false;
537
538 /*
539 * See if our hash lookup can find a potential backmerge.
540 */
541 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
542 if (__rq && blk_attempt_req_merge(q, __rq, rq))
543 return true;
544
545 return false;
546}
547
548void elv_merged_request(struct request_queue *q, struct request *rq, int type)
549{
550 struct elevator_queue *e = q->elevator;
551
552 if (e->ops->elevator_merged_fn)
553 e->ops->elevator_merged_fn(q, rq, type);
554
555 if (type == ELEVATOR_BACK_MERGE)
556 elv_rqhash_reposition(q, rq);
557
558 q->last_merge = rq;
559}
560
561void elv_merge_requests(struct request_queue *q, struct request *rq,
562 struct request *next)
563{
564 struct elevator_queue *e = q->elevator;
565 const int next_sorted = next->cmd_flags & REQ_SORTED;
566
567 if (next_sorted && e->ops->elevator_merge_req_fn)
568 e->ops->elevator_merge_req_fn(q, rq, next);
569
570 elv_rqhash_reposition(q, rq);
571
572 if (next_sorted) {
573 elv_rqhash_del(q, next);
574 q->nr_sorted--;
575 }
576
577 q->last_merge = rq;
578}
579
580void elv_bio_merged(struct request_queue *q, struct request *rq,
581 struct bio *bio)
582{
583 struct elevator_queue *e = q->elevator;
584
585 if (e->ops->elevator_bio_merged_fn)
586 e->ops->elevator_bio_merged_fn(q, rq, bio);
587}
588
589void elv_requeue_request(struct request_queue *q, struct request *rq)
590{
591 /*
592 * it already went through dequeue, we need to decrement the
593 * in_flight count again
594 */
595 if (blk_account_rq(rq)) {
596 q->in_flight[rq_is_sync(rq)]--;
597 if (rq->cmd_flags & REQ_SORTED)
598 elv_deactivate_rq(q, rq);
599 }
600
601 rq->cmd_flags &= ~REQ_STARTED;
602
603 __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
604}
605
606void elv_drain_elevator(struct request_queue *q)
607{
608 static int printed;
609 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
610 ;
611 if (q->nr_sorted == 0)
612 return;
613 if (printed++ < 10) {
614 printk(KERN_ERR "%s: forced dispatching is broken "
615 "(nr_sorted=%u), please report this\n",
616 q->elevator->elevator_type->elevator_name, q->nr_sorted);
617 }
618}
619
620/*
621 * Call with queue lock held, interrupts disabled
622 */
623void elv_quiesce_start(struct request_queue *q)
624{
625 if (!q->elevator)
626 return;
627
628 queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
629
630 /*
631 * make sure we don't have any requests in flight
632 */
633 elv_drain_elevator(q);
634 while (q->rq.elvpriv) {
635 __blk_run_queue(q);
636 spin_unlock_irq(q->queue_lock);
637 msleep(10);
638 spin_lock_irq(q->queue_lock);
639 elv_drain_elevator(q);
640 }
641}
642
643void elv_quiesce_end(struct request_queue *q)
644{
645 queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
646}
647
648void __elv_add_request(struct request_queue *q, struct request *rq, int where)
649{
650 trace_block_rq_insert(q, rq);
651
652 rq->q = q;
653
654 if (rq->cmd_flags & REQ_SOFTBARRIER) {
655 /* barriers are scheduling boundary, update end_sector */
656 if (rq->cmd_type == REQ_TYPE_FS ||
657 (rq->cmd_flags & REQ_DISCARD)) {
658 q->end_sector = rq_end_sector(rq);
659 q->boundary_rq = rq;
660 }
661 } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
662 (where == ELEVATOR_INSERT_SORT ||
663 where == ELEVATOR_INSERT_SORT_MERGE))
664 where = ELEVATOR_INSERT_BACK;
665
666 switch (where) {
667 case ELEVATOR_INSERT_REQUEUE:
668 case ELEVATOR_INSERT_FRONT:
669 rq->cmd_flags |= REQ_SOFTBARRIER;
670 list_add(&rq->queuelist, &q->queue_head);
671 break;
672
673 case ELEVATOR_INSERT_BACK:
674 rq->cmd_flags |= REQ_SOFTBARRIER;
675 elv_drain_elevator(q);
676 list_add_tail(&rq->queuelist, &q->queue_head);
677 /*
678 * We kick the queue here for the following reasons.
679 * - The elevator might have returned NULL previously
680 * to delay requests and returned them now. As the
681 * queue wasn't empty before this request, ll_rw_blk
682 * won't run the queue on return, resulting in hang.
683 * - Usually, back inserted requests won't be merged
684 * with anything. There's no point in delaying queue
685 * processing.
686 */
687 __blk_run_queue(q);
688 break;
689
690 case ELEVATOR_INSERT_SORT_MERGE:
691 /*
692 * If we succeed in merging this request with one in the
693 * queue already, we are done - rq has now been freed,
694 * so no need to do anything further.
695 */
696 if (elv_attempt_insert_merge(q, rq))
697 break;
698 case ELEVATOR_INSERT_SORT:
699 BUG_ON(rq->cmd_type != REQ_TYPE_FS &&
700 !(rq->cmd_flags & REQ_DISCARD));
701 rq->cmd_flags |= REQ_SORTED;
702 q->nr_sorted++;
703 if (rq_mergeable(rq)) {
704 elv_rqhash_add(q, rq);
705 if (!q->last_merge)
706 q->last_merge = rq;
707 }
708
709 /*
710 * Some ioscheds (cfq) run q->request_fn directly, so
711 * rq cannot be accessed after calling
712 * elevator_add_req_fn.
713 */
714 q->elevator->ops->elevator_add_req_fn(q, rq);
715 break;
716
717 case ELEVATOR_INSERT_FLUSH:
718 rq->cmd_flags |= REQ_SOFTBARRIER;
719 blk_insert_flush(rq);
720 break;
721 default:
722 printk(KERN_ERR "%s: bad insertion point %d\n",
723 __func__, where);
724 BUG();
725 }
726}
727EXPORT_SYMBOL(__elv_add_request);
728
729void elv_add_request(struct request_queue *q, struct request *rq, int where)
730{
731 unsigned long flags;
732
733 spin_lock_irqsave(q->queue_lock, flags);
734 __elv_add_request(q, rq, where);
735 spin_unlock_irqrestore(q->queue_lock, flags);
736}
737EXPORT_SYMBOL(elv_add_request);
738
739struct request *elv_latter_request(struct request_queue *q, struct request *rq)
740{
741 struct elevator_queue *e = q->elevator;
742
743 if (e->ops->elevator_latter_req_fn)
744 return e->ops->elevator_latter_req_fn(q, rq);
745 return NULL;
746}
747
748struct request *elv_former_request(struct request_queue *q, struct request *rq)
749{
750 struct elevator_queue *e = q->elevator;
751
752 if (e->ops->elevator_former_req_fn)
753 return e->ops->elevator_former_req_fn(q, rq);
754 return NULL;
755}
756
757int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
758{
759 struct elevator_queue *e = q->elevator;
760
761 if (e->ops->elevator_set_req_fn)
762 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
763
764 rq->elevator_private[0] = NULL;
765 return 0;
766}
767
768void elv_put_request(struct request_queue *q, struct request *rq)
769{
770 struct elevator_queue *e = q->elevator;
771
772 if (e->ops->elevator_put_req_fn)
773 e->ops->elevator_put_req_fn(rq);
774}
775
776int elv_may_queue(struct request_queue *q, int rw)
777{
778 struct elevator_queue *e = q->elevator;
779
780 if (e->ops->elevator_may_queue_fn)
781 return e->ops->elevator_may_queue_fn(q, rw);
782
783 return ELV_MQUEUE_MAY;
784}
785
786void elv_abort_queue(struct request_queue *q)
787{
788 struct request *rq;
789
790 blk_abort_flushes(q);
791
792 while (!list_empty(&q->queue_head)) {
793 rq = list_entry_rq(q->queue_head.next);
794 rq->cmd_flags |= REQ_QUIET;
795 trace_block_rq_abort(q, rq);
796 /*
797 * Mark this request as started so we don't trigger
798 * any debug logic in the end I/O path.
799 */
800 blk_start_request(rq);
801 __blk_end_request_all(rq, -EIO);
802 }
803}
804EXPORT_SYMBOL(elv_abort_queue);
805
806void elv_completed_request(struct request_queue *q, struct request *rq)
807{
808 struct elevator_queue *e = q->elevator;
809
810 /*
811 * request is released from the driver, io must be done
812 */
813 if (blk_account_rq(rq)) {
814 q->in_flight[rq_is_sync(rq)]--;
815 if ((rq->cmd_flags & REQ_SORTED) &&
816 e->ops->elevator_completed_req_fn)
817 e->ops->elevator_completed_req_fn(q, rq);
818 }
819}
820
821#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
822
823static ssize_t
824elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
825{
826 struct elv_fs_entry *entry = to_elv(attr);
827 struct elevator_queue *e;
828 ssize_t error;
829
830 if (!entry->show)
831 return -EIO;
832
833 e = container_of(kobj, struct elevator_queue, kobj);
834 mutex_lock(&e->sysfs_lock);
835 error = e->ops ? entry->show(e, page) : -ENOENT;
836 mutex_unlock(&e->sysfs_lock);
837 return error;
838}
839
840static ssize_t
841elv_attr_store(struct kobject *kobj, struct attribute *attr,
842 const char *page, size_t length)
843{
844 struct elv_fs_entry *entry = to_elv(attr);
845 struct elevator_queue *e;
846 ssize_t error;
847
848 if (!entry->store)
849 return -EIO;
850
851 e = container_of(kobj, struct elevator_queue, kobj);
852 mutex_lock(&e->sysfs_lock);
853 error = e->ops ? entry->store(e, page, length) : -ENOENT;
854 mutex_unlock(&e->sysfs_lock);
855 return error;
856}
857
858static const struct sysfs_ops elv_sysfs_ops = {
859 .show = elv_attr_show,
860 .store = elv_attr_store,
861};
862
863static struct kobj_type elv_ktype = {
864 .sysfs_ops = &elv_sysfs_ops,
865 .release = elevator_release,
866};
867
868int elv_register_queue(struct request_queue *q)
869{
870 struct elevator_queue *e = q->elevator;
871 int error;
872
873 error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
874 if (!error) {
875 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
876 if (attr) {
877 while (attr->attr.name) {
878 if (sysfs_create_file(&e->kobj, &attr->attr))
879 break;
880 attr++;
881 }
882 }
883 kobject_uevent(&e->kobj, KOBJ_ADD);
884 e->registered = 1;
885 }
886 return error;
887}
888EXPORT_SYMBOL(elv_register_queue);
889
890static void __elv_unregister_queue(struct elevator_queue *e)
891{
892 kobject_uevent(&e->kobj, KOBJ_REMOVE);
893 kobject_del(&e->kobj);
894 e->registered = 0;
895}
896
897void elv_unregister_queue(struct request_queue *q)
898{
899 if (q)
900 __elv_unregister_queue(q->elevator);
901}
902EXPORT_SYMBOL(elv_unregister_queue);
903
904void elv_register(struct elevator_type *e)
905{
906 char *def = "";
907
908 spin_lock(&elv_list_lock);
909 BUG_ON(elevator_find(e->elevator_name));
910 list_add_tail(&e->list, &elv_list);
911 spin_unlock(&elv_list_lock);
912
913 if (!strcmp(e->elevator_name, chosen_elevator) ||
914 (!*chosen_elevator &&
915 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
916 def = " (default)";
917
918 printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
919 def);
920}
921EXPORT_SYMBOL_GPL(elv_register);
922
923void elv_unregister(struct elevator_type *e)
924{
925 struct task_struct *g, *p;
926
927 /*
928 * Iterate every thread in the process to remove the io contexts.
929 */
930 if (e->ops.trim) {
931 read_lock(&tasklist_lock);
932 do_each_thread(g, p) {
933 task_lock(p);
934 if (p->io_context)
935 e->ops.trim(p->io_context);
936 task_unlock(p);
937 } while_each_thread(g, p);
938 read_unlock(&tasklist_lock);
939 }
940
941 spin_lock(&elv_list_lock);
942 list_del_init(&e->list);
943 spin_unlock(&elv_list_lock);
944}
945EXPORT_SYMBOL_GPL(elv_unregister);
946
947/*
948 * switch to new_e io scheduler. be careful not to introduce deadlocks -
949 * we don't free the old io scheduler, before we have allocated what we
950 * need for the new one. this way we have a chance of going back to the old
951 * one, if the new one fails init for some reason.
952 */
953static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
954{
955 struct elevator_queue *old_elevator, *e;
956 void *data;
957 int err;
958
959 /*
960 * Allocate new elevator
961 */
962 e = elevator_alloc(q, new_e);
963 if (!e)
964 return -ENOMEM;
965
966 data = elevator_init_queue(q, e);
967 if (!data) {
968 kobject_put(&e->kobj);
969 return -ENOMEM;
970 }
971
972 /*
973 * Turn on BYPASS and drain all requests w/ elevator private data
974 */
975 spin_lock_irq(q->queue_lock);
976 elv_quiesce_start(q);
977
978 /*
979 * Remember old elevator.
980 */
981 old_elevator = q->elevator;
982
983 /*
984 * attach and start new elevator
985 */
986 elevator_attach(q, e, data);
987
988 spin_unlock_irq(q->queue_lock);
989
990 if (old_elevator->registered) {
991 __elv_unregister_queue(old_elevator);
992
993 err = elv_register_queue(q);
994 if (err)
995 goto fail_register;
996 }
997
998 /*
999 * finally exit old elevator and turn off BYPASS.
1000 */
1001 elevator_exit(old_elevator);
1002 spin_lock_irq(q->queue_lock);
1003 elv_quiesce_end(q);
1004 spin_unlock_irq(q->queue_lock);
1005
1006 blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1007
1008 return 0;
1009
1010fail_register:
1011 /*
1012 * switch failed, exit the new io scheduler and reattach the old
1013 * one again (along with re-adding the sysfs dir)
1014 */
1015 elevator_exit(e);
1016 q->elevator = old_elevator;
1017 elv_register_queue(q);
1018
1019 spin_lock_irq(q->queue_lock);
1020 queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1021 spin_unlock_irq(q->queue_lock);
1022
1023 return err;
1024}
1025
1026/*
1027 * Switch this queue to the given IO scheduler.
1028 */
1029int elevator_change(struct request_queue *q, const char *name)
1030{
1031 char elevator_name[ELV_NAME_MAX];
1032 struct elevator_type *e;
1033
1034 if (!q->elevator)
1035 return -ENXIO;
1036
1037 strlcpy(elevator_name, name, sizeof(elevator_name));
1038 e = elevator_get(strstrip(elevator_name));
1039 if (!e) {
1040 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1041 return -EINVAL;
1042 }
1043
1044 if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1045 elevator_put(e);
1046 return 0;
1047 }
1048
1049 return elevator_switch(q, e);
1050}
1051EXPORT_SYMBOL(elevator_change);
1052
1053ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1054 size_t count)
1055{
1056 int ret;
1057
1058 if (!q->elevator)
1059 return count;
1060
1061 ret = elevator_change(q, name);
1062 if (!ret)
1063 return count;
1064
1065 printk(KERN_ERR "elevator: switch to %s failed\n", name);
1066 return ret;
1067}
1068
1069ssize_t elv_iosched_show(struct request_queue *q, char *name)
1070{
1071 struct elevator_queue *e = q->elevator;
1072 struct elevator_type *elv;
1073 struct elevator_type *__e;
1074 int len = 0;
1075
1076 if (!q->elevator || !blk_queue_stackable(q))
1077 return sprintf(name, "none\n");
1078
1079 elv = e->elevator_type;
1080
1081 spin_lock(&elv_list_lock);
1082 list_for_each_entry(__e, &elv_list, list) {
1083 if (!strcmp(elv->elevator_name, __e->elevator_name))
1084 len += sprintf(name+len, "[%s] ", elv->elevator_name);
1085 else
1086 len += sprintf(name+len, "%s ", __e->elevator_name);
1087 }
1088 spin_unlock(&elv_list_lock);
1089
1090 len += sprintf(len+name, "\n");
1091 return len;
1092}
1093
1094struct request *elv_rb_former_request(struct request_queue *q,
1095 struct request *rq)
1096{
1097 struct rb_node *rbprev = rb_prev(&rq->rb_node);
1098
1099 if (rbprev)
1100 return rb_entry_rq(rbprev);
1101
1102 return NULL;
1103}
1104EXPORT_SYMBOL(elv_rb_former_request);
1105
1106struct request *elv_rb_latter_request(struct request_queue *q,
1107 struct request *rq)
1108{
1109 struct rb_node *rbnext = rb_next(&rq->rb_node);
1110
1111 if (rbnext)
1112 return rb_entry_rq(rbnext);
1113
1114 return NULL;
1115}
1116EXPORT_SYMBOL(elv_rb_latter_request);