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