xref: /openbmc/linux/block/elevator.c (revision 9d56dd3b083a3bec56e9da35ce07baca81030b03)
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 
43 static DEFINE_SPINLOCK(elv_list_lock);
44 static LIST_HEAD(elv_list);
45 
46 /*
47  * Merge hash stuff.
48  */
49 static 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  */
60 static 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  */
74 int 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_rw_flagged(bio, BIO_RW_DISCARD) !=
83 	    bio_rw_flagged(rq->bio, BIO_RW_DISCARD))
84 		return 0;
85 
86 	/*
87 	 * different data direction or already started, don't merge
88 	 */
89 	if (bio_data_dir(bio) != rq_data_dir(rq))
90 		return 0;
91 
92 	/*
93 	 * must be same device and not a special request
94 	 */
95 	if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
96 		return 0;
97 
98 	/*
99 	 * only merge integrity protected bio into ditto rq
100 	 */
101 	if (bio_integrity(bio) != blk_integrity_rq(rq))
102 		return 0;
103 
104 	if (!elv_iosched_allow_merge(rq, bio))
105 		return 0;
106 
107 	return 1;
108 }
109 EXPORT_SYMBOL(elv_rq_merge_ok);
110 
111 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
112 {
113 	int ret = ELEVATOR_NO_MERGE;
114 
115 	/*
116 	 * we can merge and sequence is ok, check if it's possible
117 	 */
118 	if (elv_rq_merge_ok(__rq, bio)) {
119 		if (blk_rq_pos(__rq) + blk_rq_sectors(__rq) == bio->bi_sector)
120 			ret = ELEVATOR_BACK_MERGE;
121 		else if (blk_rq_pos(__rq) - bio_sectors(bio) == bio->bi_sector)
122 			ret = ELEVATOR_FRONT_MERGE;
123 	}
124 
125 	return ret;
126 }
127 
128 static struct elevator_type *elevator_find(const char *name)
129 {
130 	struct elevator_type *e;
131 
132 	list_for_each_entry(e, &elv_list, list) {
133 		if (!strcmp(e->elevator_name, name))
134 			return e;
135 	}
136 
137 	return NULL;
138 }
139 
140 static void elevator_put(struct elevator_type *e)
141 {
142 	module_put(e->elevator_owner);
143 }
144 
145 static struct elevator_type *elevator_get(const char *name)
146 {
147 	struct elevator_type *e;
148 
149 	spin_lock(&elv_list_lock);
150 
151 	e = elevator_find(name);
152 	if (!e) {
153 		char elv[ELV_NAME_MAX + strlen("-iosched")];
154 
155 		spin_unlock(&elv_list_lock);
156 
157 		sprintf(elv, "%s-iosched", name);
158 
159 		request_module("%s", elv);
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 
172 static void *elevator_init_queue(struct request_queue *q,
173 				 struct elevator_queue *eq)
174 {
175 	return eq->ops->elevator_init_fn(q);
176 }
177 
178 static 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 
185 static char chosen_elevator[16];
186 
187 static 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 
199 static struct kobj_type elv_ktype;
200 
201 static 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;
225 err:
226 	kfree(eq);
227 	elevator_put(e);
228 	return NULL;
229 }
230 
231 static 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 
241 int elevator_init(struct request_queue *q, char *name)
242 {
243 	struct elevator_type *e = NULL;
244 	struct elevator_queue *eq;
245 	int ret = 0;
246 	void *data;
247 
248 	INIT_LIST_HEAD(&q->queue_head);
249 	q->last_merge = NULL;
250 	q->end_sector = 0;
251 	q->boundary_rq = NULL;
252 
253 	if (name) {
254 		e = elevator_get(name);
255 		if (!e)
256 			return -EINVAL;
257 	}
258 
259 	if (!e && *chosen_elevator) {
260 		e = elevator_get(chosen_elevator);
261 		if (!e)
262 			printk(KERN_ERR "I/O scheduler %s not found\n",
263 							chosen_elevator);
264 	}
265 
266 	if (!e) {
267 		e = elevator_get(CONFIG_DEFAULT_IOSCHED);
268 		if (!e) {
269 			printk(KERN_ERR
270 				"Default I/O scheduler not found. " \
271 				"Using noop.\n");
272 			e = elevator_get("noop");
273 		}
274 	}
275 
276 	eq = elevator_alloc(q, e);
277 	if (!eq)
278 		return -ENOMEM;
279 
280 	data = elevator_init_queue(q, eq);
281 	if (!data) {
282 		kobject_put(&eq->kobj);
283 		return -ENOMEM;
284 	}
285 
286 	elevator_attach(q, eq, data);
287 	return ret;
288 }
289 EXPORT_SYMBOL(elevator_init);
290 
291 void elevator_exit(struct elevator_queue *e)
292 {
293 	mutex_lock(&e->sysfs_lock);
294 	if (e->ops->elevator_exit_fn)
295 		e->ops->elevator_exit_fn(e);
296 	e->ops = NULL;
297 	mutex_unlock(&e->sysfs_lock);
298 
299 	kobject_put(&e->kobj);
300 }
301 EXPORT_SYMBOL(elevator_exit);
302 
303 static inline void __elv_rqhash_del(struct request *rq)
304 {
305 	hlist_del_init(&rq->hash);
306 }
307 
308 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
309 {
310 	if (ELV_ON_HASH(rq))
311 		__elv_rqhash_del(rq);
312 }
313 
314 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
315 {
316 	struct elevator_queue *e = q->elevator;
317 
318 	BUG_ON(ELV_ON_HASH(rq));
319 	hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
320 }
321 
322 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
323 {
324 	__elv_rqhash_del(rq);
325 	elv_rqhash_add(q, rq);
326 }
327 
328 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
329 {
330 	struct elevator_queue *e = q->elevator;
331 	struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
332 	struct hlist_node *entry, *next;
333 	struct request *rq;
334 
335 	hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
336 		BUG_ON(!ELV_ON_HASH(rq));
337 
338 		if (unlikely(!rq_mergeable(rq))) {
339 			__elv_rqhash_del(rq);
340 			continue;
341 		}
342 
343 		if (rq_hash_key(rq) == offset)
344 			return rq;
345 	}
346 
347 	return NULL;
348 }
349 
350 /*
351  * RB-tree support functions for inserting/lookup/removal of requests
352  * in a sorted RB tree.
353  */
354 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
355 {
356 	struct rb_node **p = &root->rb_node;
357 	struct rb_node *parent = NULL;
358 	struct request *__rq;
359 
360 	while (*p) {
361 		parent = *p;
362 		__rq = rb_entry(parent, struct request, rb_node);
363 
364 		if (blk_rq_pos(rq) < blk_rq_pos(__rq))
365 			p = &(*p)->rb_left;
366 		else if (blk_rq_pos(rq) > blk_rq_pos(__rq))
367 			p = &(*p)->rb_right;
368 		else
369 			return __rq;
370 	}
371 
372 	rb_link_node(&rq->rb_node, parent, p);
373 	rb_insert_color(&rq->rb_node, root);
374 	return NULL;
375 }
376 EXPORT_SYMBOL(elv_rb_add);
377 
378 void elv_rb_del(struct rb_root *root, struct request *rq)
379 {
380 	BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
381 	rb_erase(&rq->rb_node, root);
382 	RB_CLEAR_NODE(&rq->rb_node);
383 }
384 EXPORT_SYMBOL(elv_rb_del);
385 
386 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
387 {
388 	struct rb_node *n = root->rb_node;
389 	struct request *rq;
390 
391 	while (n) {
392 		rq = rb_entry(n, struct request, rb_node);
393 
394 		if (sector < blk_rq_pos(rq))
395 			n = n->rb_left;
396 		else if (sector > blk_rq_pos(rq))
397 			n = n->rb_right;
398 		else
399 			return rq;
400 	}
401 
402 	return NULL;
403 }
404 EXPORT_SYMBOL(elv_rb_find);
405 
406 /*
407  * Insert rq into dispatch queue of q.  Queue lock must be held on
408  * entry.  rq is sort instead into the dispatch queue. To be used by
409  * specific elevators.
410  */
411 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
412 {
413 	sector_t boundary;
414 	struct list_head *entry;
415 	int stop_flags;
416 
417 	if (q->last_merge == rq)
418 		q->last_merge = NULL;
419 
420 	elv_rqhash_del(q, rq);
421 
422 	q->nr_sorted--;
423 
424 	boundary = q->end_sector;
425 	stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
426 	list_for_each_prev(entry, &q->queue_head) {
427 		struct request *pos = list_entry_rq(entry);
428 
429 		if (blk_discard_rq(rq) != blk_discard_rq(pos))
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 }
448 EXPORT_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  */
455 void 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 }
468 EXPORT_SYMBOL(elv_dispatch_add_tail);
469 
470 int 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 	 * First try one-hit cache.
478 	 */
479 	if (q->last_merge) {
480 		ret = elv_try_merge(q->last_merge, bio);
481 		if (ret != ELEVATOR_NO_MERGE) {
482 			*req = q->last_merge;
483 			return ret;
484 		}
485 	}
486 
487 	if (blk_queue_nomerges(q))
488 		return ELEVATOR_NO_MERGE;
489 
490 	/*
491 	 * See if our hash lookup can find a potential backmerge.
492 	 */
493 	__rq = elv_rqhash_find(q, bio->bi_sector);
494 	if (__rq && elv_rq_merge_ok(__rq, bio)) {
495 		*req = __rq;
496 		return ELEVATOR_BACK_MERGE;
497 	}
498 
499 	if (e->ops->elevator_merge_fn)
500 		return e->ops->elevator_merge_fn(q, req, bio);
501 
502 	return ELEVATOR_NO_MERGE;
503 }
504 
505 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
506 {
507 	struct elevator_queue *e = q->elevator;
508 
509 	if (e->ops->elevator_merged_fn)
510 		e->ops->elevator_merged_fn(q, rq, type);
511 
512 	if (type == ELEVATOR_BACK_MERGE)
513 		elv_rqhash_reposition(q, rq);
514 
515 	q->last_merge = rq;
516 }
517 
518 void elv_merge_requests(struct request_queue *q, struct request *rq,
519 			     struct request *next)
520 {
521 	struct elevator_queue *e = q->elevator;
522 
523 	if (e->ops->elevator_merge_req_fn)
524 		e->ops->elevator_merge_req_fn(q, rq, next);
525 
526 	elv_rqhash_reposition(q, rq);
527 	elv_rqhash_del(q, next);
528 
529 	q->nr_sorted--;
530 	q->last_merge = rq;
531 }
532 
533 void elv_requeue_request(struct request_queue *q, struct request *rq)
534 {
535 	/*
536 	 * it already went through dequeue, we need to decrement the
537 	 * in_flight count again
538 	 */
539 	if (blk_account_rq(rq)) {
540 		q->in_flight[rq_is_sync(rq)]--;
541 		if (blk_sorted_rq(rq))
542 			elv_deactivate_rq(q, rq);
543 	}
544 
545 	rq->cmd_flags &= ~REQ_STARTED;
546 
547 	elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
548 }
549 
550 void elv_drain_elevator(struct request_queue *q)
551 {
552 	static int printed;
553 	while (q->elevator->ops->elevator_dispatch_fn(q, 1))
554 		;
555 	if (q->nr_sorted == 0)
556 		return;
557 	if (printed++ < 10) {
558 		printk(KERN_ERR "%s: forced dispatching is broken "
559 		       "(nr_sorted=%u), please report this\n",
560 		       q->elevator->elevator_type->elevator_name, q->nr_sorted);
561 	}
562 }
563 
564 /*
565  * Call with queue lock held, interrupts disabled
566  */
567 void elv_quiesce_start(struct request_queue *q)
568 {
569 	if (!q->elevator)
570 		return;
571 
572 	queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
573 
574 	/*
575 	 * make sure we don't have any requests in flight
576 	 */
577 	elv_drain_elevator(q);
578 	while (q->rq.elvpriv) {
579 		__blk_run_queue(q);
580 		spin_unlock_irq(q->queue_lock);
581 		msleep(10);
582 		spin_lock_irq(q->queue_lock);
583 		elv_drain_elevator(q);
584 	}
585 }
586 
587 void elv_quiesce_end(struct request_queue *q)
588 {
589 	queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
590 }
591 
592 void elv_insert(struct request_queue *q, struct request *rq, int where)
593 {
594 	struct list_head *pos;
595 	unsigned ordseq;
596 	int unplug_it = 1;
597 
598 	trace_block_rq_insert(q, rq);
599 
600 	rq->q = q;
601 
602 	switch (where) {
603 	case ELEVATOR_INSERT_FRONT:
604 		rq->cmd_flags |= REQ_SOFTBARRIER;
605 
606 		list_add(&rq->queuelist, &q->queue_head);
607 		break;
608 
609 	case ELEVATOR_INSERT_BACK:
610 		rq->cmd_flags |= REQ_SOFTBARRIER;
611 		elv_drain_elevator(q);
612 		list_add_tail(&rq->queuelist, &q->queue_head);
613 		/*
614 		 * We kick the queue here for the following reasons.
615 		 * - The elevator might have returned NULL previously
616 		 *   to delay requests and returned them now.  As the
617 		 *   queue wasn't empty before this request, ll_rw_blk
618 		 *   won't run the queue on return, resulting in hang.
619 		 * - Usually, back inserted requests won't be merged
620 		 *   with anything.  There's no point in delaying queue
621 		 *   processing.
622 		 */
623 		__blk_run_queue(q);
624 		break;
625 
626 	case ELEVATOR_INSERT_SORT:
627 		BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
628 		rq->cmd_flags |= REQ_SORTED;
629 		q->nr_sorted++;
630 		if (rq_mergeable(rq)) {
631 			elv_rqhash_add(q, rq);
632 			if (!q->last_merge)
633 				q->last_merge = rq;
634 		}
635 
636 		/*
637 		 * Some ioscheds (cfq) run q->request_fn directly, so
638 		 * rq cannot be accessed after calling
639 		 * elevator_add_req_fn.
640 		 */
641 		q->elevator->ops->elevator_add_req_fn(q, rq);
642 		break;
643 
644 	case ELEVATOR_INSERT_REQUEUE:
645 		/*
646 		 * If ordered flush isn't in progress, we do front
647 		 * insertion; otherwise, requests should be requeued
648 		 * in ordseq order.
649 		 */
650 		rq->cmd_flags |= REQ_SOFTBARRIER;
651 
652 		/*
653 		 * Most requeues happen because of a busy condition,
654 		 * don't force unplug of the queue for that case.
655 		 */
656 		unplug_it = 0;
657 
658 		if (q->ordseq == 0) {
659 			list_add(&rq->queuelist, &q->queue_head);
660 			break;
661 		}
662 
663 		ordseq = blk_ordered_req_seq(rq);
664 
665 		list_for_each(pos, &q->queue_head) {
666 			struct request *pos_rq = list_entry_rq(pos);
667 			if (ordseq <= blk_ordered_req_seq(pos_rq))
668 				break;
669 		}
670 
671 		list_add_tail(&rq->queuelist, pos);
672 		break;
673 
674 	default:
675 		printk(KERN_ERR "%s: bad insertion point %d\n",
676 		       __func__, where);
677 		BUG();
678 	}
679 
680 	if (unplug_it && blk_queue_plugged(q)) {
681 		int nrq = q->rq.count[BLK_RW_SYNC] + q->rq.count[BLK_RW_ASYNC]
682 				- queue_in_flight(q);
683 
684 		if (nrq >= q->unplug_thresh)
685 			__generic_unplug_device(q);
686 	}
687 }
688 
689 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
690 		       int plug)
691 {
692 	if (q->ordcolor)
693 		rq->cmd_flags |= REQ_ORDERED_COLOR;
694 
695 	if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
696 		/*
697 		 * toggle ordered color
698 		 */
699 		if (blk_barrier_rq(rq))
700 			q->ordcolor ^= 1;
701 
702 		/*
703 		 * barriers implicitly indicate back insertion
704 		 */
705 		if (where == ELEVATOR_INSERT_SORT)
706 			where = ELEVATOR_INSERT_BACK;
707 
708 		/*
709 		 * this request is scheduling boundary, update
710 		 * end_sector
711 		 */
712 		if (blk_fs_request(rq) || blk_discard_rq(rq)) {
713 			q->end_sector = rq_end_sector(rq);
714 			q->boundary_rq = rq;
715 		}
716 	} else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
717 		    where == ELEVATOR_INSERT_SORT)
718 		where = ELEVATOR_INSERT_BACK;
719 
720 	if (plug)
721 		blk_plug_device(q);
722 
723 	elv_insert(q, rq, where);
724 }
725 EXPORT_SYMBOL(__elv_add_request);
726 
727 void elv_add_request(struct request_queue *q, struct request *rq, int where,
728 		     int plug)
729 {
730 	unsigned long flags;
731 
732 	spin_lock_irqsave(q->queue_lock, flags);
733 	__elv_add_request(q, rq, where, plug);
734 	spin_unlock_irqrestore(q->queue_lock, flags);
735 }
736 EXPORT_SYMBOL(elv_add_request);
737 
738 int elv_queue_empty(struct request_queue *q)
739 {
740 	struct elevator_queue *e = q->elevator;
741 
742 	if (!list_empty(&q->queue_head))
743 		return 0;
744 
745 	if (e->ops->elevator_queue_empty_fn)
746 		return e->ops->elevator_queue_empty_fn(q);
747 
748 	return 1;
749 }
750 EXPORT_SYMBOL(elv_queue_empty);
751 
752 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
753 {
754 	struct elevator_queue *e = q->elevator;
755 
756 	if (e->ops->elevator_latter_req_fn)
757 		return e->ops->elevator_latter_req_fn(q, rq);
758 	return NULL;
759 }
760 
761 struct request *elv_former_request(struct request_queue *q, struct request *rq)
762 {
763 	struct elevator_queue *e = q->elevator;
764 
765 	if (e->ops->elevator_former_req_fn)
766 		return e->ops->elevator_former_req_fn(q, rq);
767 	return NULL;
768 }
769 
770 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
771 {
772 	struct elevator_queue *e = q->elevator;
773 
774 	if (e->ops->elevator_set_req_fn)
775 		return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
776 
777 	rq->elevator_private = NULL;
778 	return 0;
779 }
780 
781 void elv_put_request(struct request_queue *q, struct request *rq)
782 {
783 	struct elevator_queue *e = q->elevator;
784 
785 	if (e->ops->elevator_put_req_fn)
786 		e->ops->elevator_put_req_fn(rq);
787 }
788 
789 int elv_may_queue(struct request_queue *q, int rw)
790 {
791 	struct elevator_queue *e = q->elevator;
792 
793 	if (e->ops->elevator_may_queue_fn)
794 		return e->ops->elevator_may_queue_fn(q, rw);
795 
796 	return ELV_MQUEUE_MAY;
797 }
798 
799 void elv_abort_queue(struct request_queue *q)
800 {
801 	struct request *rq;
802 
803 	while (!list_empty(&q->queue_head)) {
804 		rq = list_entry_rq(q->queue_head.next);
805 		rq->cmd_flags |= REQ_QUIET;
806 		trace_block_rq_abort(q, rq);
807 		/*
808 		 * Mark this request as started so we don't trigger
809 		 * any debug logic in the end I/O path.
810 		 */
811 		blk_start_request(rq);
812 		__blk_end_request_all(rq, -EIO);
813 	}
814 }
815 EXPORT_SYMBOL(elv_abort_queue);
816 
817 void elv_completed_request(struct request_queue *q, struct request *rq)
818 {
819 	struct elevator_queue *e = q->elevator;
820 
821 	/*
822 	 * request is released from the driver, io must be done
823 	 */
824 	if (blk_account_rq(rq)) {
825 		q->in_flight[rq_is_sync(rq)]--;
826 		if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
827 			e->ops->elevator_completed_req_fn(q, rq);
828 	}
829 
830 	/*
831 	 * Check if the queue is waiting for fs requests to be
832 	 * drained for flush sequence.
833 	 */
834 	if (unlikely(q->ordseq)) {
835 		struct request *next = NULL;
836 
837 		if (!list_empty(&q->queue_head))
838 			next = list_entry_rq(q->queue_head.next);
839 
840 		if (!queue_in_flight(q) &&
841 		    blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
842 		    (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
843 			blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
844 			__blk_run_queue(q);
845 		}
846 	}
847 }
848 
849 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
850 
851 static ssize_t
852 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
853 {
854 	struct elv_fs_entry *entry = to_elv(attr);
855 	struct elevator_queue *e;
856 	ssize_t error;
857 
858 	if (!entry->show)
859 		return -EIO;
860 
861 	e = container_of(kobj, struct elevator_queue, kobj);
862 	mutex_lock(&e->sysfs_lock);
863 	error = e->ops ? entry->show(e, page) : -ENOENT;
864 	mutex_unlock(&e->sysfs_lock);
865 	return error;
866 }
867 
868 static ssize_t
869 elv_attr_store(struct kobject *kobj, struct attribute *attr,
870 	       const char *page, size_t length)
871 {
872 	struct elv_fs_entry *entry = to_elv(attr);
873 	struct elevator_queue *e;
874 	ssize_t error;
875 
876 	if (!entry->store)
877 		return -EIO;
878 
879 	e = container_of(kobj, struct elevator_queue, kobj);
880 	mutex_lock(&e->sysfs_lock);
881 	error = e->ops ? entry->store(e, page, length) : -ENOENT;
882 	mutex_unlock(&e->sysfs_lock);
883 	return error;
884 }
885 
886 static struct sysfs_ops elv_sysfs_ops = {
887 	.show	= elv_attr_show,
888 	.store	= elv_attr_store,
889 };
890 
891 static struct kobj_type elv_ktype = {
892 	.sysfs_ops	= &elv_sysfs_ops,
893 	.release	= elevator_release,
894 };
895 
896 int elv_register_queue(struct request_queue *q)
897 {
898 	struct elevator_queue *e = q->elevator;
899 	int error;
900 
901 	error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
902 	if (!error) {
903 		struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
904 		if (attr) {
905 			while (attr->attr.name) {
906 				if (sysfs_create_file(&e->kobj, &attr->attr))
907 					break;
908 				attr++;
909 			}
910 		}
911 		kobject_uevent(&e->kobj, KOBJ_ADD);
912 	}
913 	return error;
914 }
915 
916 static void __elv_unregister_queue(struct elevator_queue *e)
917 {
918 	kobject_uevent(&e->kobj, KOBJ_REMOVE);
919 	kobject_del(&e->kobj);
920 }
921 
922 void elv_unregister_queue(struct request_queue *q)
923 {
924 	if (q)
925 		__elv_unregister_queue(q->elevator);
926 }
927 
928 void elv_register(struct elevator_type *e)
929 {
930 	char *def = "";
931 
932 	spin_lock(&elv_list_lock);
933 	BUG_ON(elevator_find(e->elevator_name));
934 	list_add_tail(&e->list, &elv_list);
935 	spin_unlock(&elv_list_lock);
936 
937 	if (!strcmp(e->elevator_name, chosen_elevator) ||
938 			(!*chosen_elevator &&
939 			 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
940 				def = " (default)";
941 
942 	printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
943 								def);
944 }
945 EXPORT_SYMBOL_GPL(elv_register);
946 
947 void elv_unregister(struct elevator_type *e)
948 {
949 	struct task_struct *g, *p;
950 
951 	/*
952 	 * Iterate every thread in the process to remove the io contexts.
953 	 */
954 	if (e->ops.trim) {
955 		read_lock(&tasklist_lock);
956 		do_each_thread(g, p) {
957 			task_lock(p);
958 			if (p->io_context)
959 				e->ops.trim(p->io_context);
960 			task_unlock(p);
961 		} while_each_thread(g, p);
962 		read_unlock(&tasklist_lock);
963 	}
964 
965 	spin_lock(&elv_list_lock);
966 	list_del_init(&e->list);
967 	spin_unlock(&elv_list_lock);
968 }
969 EXPORT_SYMBOL_GPL(elv_unregister);
970 
971 /*
972  * switch to new_e io scheduler. be careful not to introduce deadlocks -
973  * we don't free the old io scheduler, before we have allocated what we
974  * need for the new one. this way we have a chance of going back to the old
975  * one, if the new one fails init for some reason.
976  */
977 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
978 {
979 	struct elevator_queue *old_elevator, *e;
980 	void *data;
981 
982 	/*
983 	 * Allocate new elevator
984 	 */
985 	e = elevator_alloc(q, new_e);
986 	if (!e)
987 		return 0;
988 
989 	data = elevator_init_queue(q, e);
990 	if (!data) {
991 		kobject_put(&e->kobj);
992 		return 0;
993 	}
994 
995 	/*
996 	 * Turn on BYPASS and drain all requests w/ elevator private data
997 	 */
998 	spin_lock_irq(q->queue_lock);
999 	elv_quiesce_start(q);
1000 
1001 	/*
1002 	 * Remember old elevator.
1003 	 */
1004 	old_elevator = q->elevator;
1005 
1006 	/*
1007 	 * attach and start new elevator
1008 	 */
1009 	elevator_attach(q, e, data);
1010 
1011 	spin_unlock_irq(q->queue_lock);
1012 
1013 	__elv_unregister_queue(old_elevator);
1014 
1015 	if (elv_register_queue(q))
1016 		goto fail_register;
1017 
1018 	/*
1019 	 * finally exit old elevator and turn off BYPASS.
1020 	 */
1021 	elevator_exit(old_elevator);
1022 	spin_lock_irq(q->queue_lock);
1023 	elv_quiesce_end(q);
1024 	spin_unlock_irq(q->queue_lock);
1025 
1026 	blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1027 
1028 	return 1;
1029 
1030 fail_register:
1031 	/*
1032 	 * switch failed, exit the new io scheduler and reattach the old
1033 	 * one again (along with re-adding the sysfs dir)
1034 	 */
1035 	elevator_exit(e);
1036 	q->elevator = old_elevator;
1037 	elv_register_queue(q);
1038 
1039 	spin_lock_irq(q->queue_lock);
1040 	queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1041 	spin_unlock_irq(q->queue_lock);
1042 
1043 	return 0;
1044 }
1045 
1046 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1047 			  size_t count)
1048 {
1049 	char elevator_name[ELV_NAME_MAX];
1050 	struct elevator_type *e;
1051 
1052 	if (!q->elevator)
1053 		return count;
1054 
1055 	strlcpy(elevator_name, name, sizeof(elevator_name));
1056 	e = elevator_get(strstrip(elevator_name));
1057 	if (!e) {
1058 		printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1059 		return -EINVAL;
1060 	}
1061 
1062 	if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1063 		elevator_put(e);
1064 		return count;
1065 	}
1066 
1067 	if (!elevator_switch(q, e))
1068 		printk(KERN_ERR "elevator: switch to %s failed\n",
1069 							elevator_name);
1070 	return count;
1071 }
1072 
1073 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1074 {
1075 	struct elevator_queue *e = q->elevator;
1076 	struct elevator_type *elv;
1077 	struct elevator_type *__e;
1078 	int len = 0;
1079 
1080 	if (!q->elevator)
1081 		return sprintf(name, "none\n");
1082 
1083 	elv = e->elevator_type;
1084 
1085 	spin_lock(&elv_list_lock);
1086 	list_for_each_entry(__e, &elv_list, list) {
1087 		if (!strcmp(elv->elevator_name, __e->elevator_name))
1088 			len += sprintf(name+len, "[%s] ", elv->elevator_name);
1089 		else
1090 			len += sprintf(name+len, "%s ", __e->elevator_name);
1091 	}
1092 	spin_unlock(&elv_list_lock);
1093 
1094 	len += sprintf(len+name, "\n");
1095 	return len;
1096 }
1097 
1098 struct request *elv_rb_former_request(struct request_queue *q,
1099 				      struct request *rq)
1100 {
1101 	struct rb_node *rbprev = rb_prev(&rq->rb_node);
1102 
1103 	if (rbprev)
1104 		return rb_entry_rq(rbprev);
1105 
1106 	return NULL;
1107 }
1108 EXPORT_SYMBOL(elv_rb_former_request);
1109 
1110 struct request *elv_rb_latter_request(struct request_queue *q,
1111 				      struct request *rq)
1112 {
1113 	struct rb_node *rbnext = rb_next(&rq->rb_node);
1114 
1115 	if (rbnext)
1116 		return rb_entry_rq(rbnext);
1117 
1118 	return NULL;
1119 }
1120 EXPORT_SYMBOL(elv_rb_latter_request);
1121