xref: /openbmc/linux/block/elevator.c (revision 9ac8d3fb)
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 "blk.h"
40 
41 static DEFINE_SPINLOCK(elv_list_lock);
42 static LIST_HEAD(elv_list);
43 
44 /*
45  * Merge hash stuff.
46  */
47 static const int elv_hash_shift = 6;
48 #define ELV_HASH_BLOCK(sec)	((sec) >> 3)
49 #define ELV_HASH_FN(sec)	\
50 		(hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
51 #define ELV_HASH_ENTRIES	(1 << elv_hash_shift)
52 #define rq_hash_key(rq)		((rq)->sector + (rq)->nr_sectors)
53 #define ELV_ON_HASH(rq)		(!hlist_unhashed(&(rq)->hash))
54 
55 /*
56  * Query io scheduler to see if the current process issuing bio may be
57  * merged with rq.
58  */
59 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
60 {
61 	struct request_queue *q = rq->q;
62 	elevator_t *e = q->elevator;
63 
64 	if (e->ops->elevator_allow_merge_fn)
65 		return e->ops->elevator_allow_merge_fn(q, rq, bio);
66 
67 	return 1;
68 }
69 
70 /*
71  * can we safely merge with this request?
72  */
73 int elv_rq_merge_ok(struct request *rq, struct bio *bio)
74 {
75 	if (!rq_mergeable(rq))
76 		return 0;
77 
78 	/*
79 	 * Don't merge file system requests and discard requests
80 	 */
81 	if (bio_discard(bio) != bio_discard(rq->bio))
82 		return 0;
83 
84 	/*
85 	 * different data direction or already started, don't merge
86 	 */
87 	if (bio_data_dir(bio) != rq_data_dir(rq))
88 		return 0;
89 
90 	/*
91 	 * must be same device and not a special request
92 	 */
93 	if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
94 		return 0;
95 
96 	/*
97 	 * only merge integrity protected bio into ditto rq
98 	 */
99 	if (bio_integrity(bio) != blk_integrity_rq(rq))
100 		return 0;
101 
102 	if (!elv_iosched_allow_merge(rq, bio))
103 		return 0;
104 
105 	return 1;
106 }
107 EXPORT_SYMBOL(elv_rq_merge_ok);
108 
109 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
110 {
111 	int ret = ELEVATOR_NO_MERGE;
112 
113 	/*
114 	 * we can merge and sequence is ok, check if it's possible
115 	 */
116 	if (elv_rq_merge_ok(__rq, bio)) {
117 		if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
118 			ret = ELEVATOR_BACK_MERGE;
119 		else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
120 			ret = ELEVATOR_FRONT_MERGE;
121 	}
122 
123 	return ret;
124 }
125 
126 static struct elevator_type *elevator_find(const char *name)
127 {
128 	struct elevator_type *e;
129 
130 	list_for_each_entry(e, &elv_list, list) {
131 		if (!strcmp(e->elevator_name, name))
132 			return e;
133 	}
134 
135 	return NULL;
136 }
137 
138 static void elevator_put(struct elevator_type *e)
139 {
140 	module_put(e->elevator_owner);
141 }
142 
143 static struct elevator_type *elevator_get(const char *name)
144 {
145 	struct elevator_type *e;
146 
147 	spin_lock(&elv_list_lock);
148 
149 	e = elevator_find(name);
150 	if (!e) {
151 		char elv[ELV_NAME_MAX + strlen("-iosched")];
152 
153 		spin_unlock(&elv_list_lock);
154 
155 		if (!strcmp(name, "anticipatory"))
156 			sprintf(elv, "as-iosched");
157 		else
158 			sprintf(elv, "%s-iosched", name);
159 
160 		request_module("%s", elv);
161 		spin_lock(&elv_list_lock);
162 		e = elevator_find(name);
163 	}
164 
165 	if (e && !try_module_get(e->elevator_owner))
166 		e = NULL;
167 
168 	spin_unlock(&elv_list_lock);
169 
170 	return e;
171 }
172 
173 static void *elevator_init_queue(struct request_queue *q,
174 				 struct elevator_queue *eq)
175 {
176 	return eq->ops->elevator_init_fn(q);
177 }
178 
179 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
180 			   void *data)
181 {
182 	q->elevator = eq;
183 	eq->elevator_data = data;
184 }
185 
186 static char chosen_elevator[16];
187 
188 static int __init elevator_setup(char *str)
189 {
190 	/*
191 	 * Be backwards-compatible with previous kernels, so users
192 	 * won't get the wrong elevator.
193 	 */
194 	if (!strcmp(str, "as"))
195 		strcpy(chosen_elevator, "anticipatory");
196 	else
197 		strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
198 	return 1;
199 }
200 
201 __setup("elevator=", elevator_setup);
202 
203 static struct kobj_type elv_ktype;
204 
205 static elevator_t *elevator_alloc(struct request_queue *q,
206 				  struct elevator_type *e)
207 {
208 	elevator_t *eq;
209 	int i;
210 
211 	eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL | __GFP_ZERO, q->node);
212 	if (unlikely(!eq))
213 		goto err;
214 
215 	eq->ops = &e->ops;
216 	eq->elevator_type = e;
217 	kobject_init(&eq->kobj, &elv_ktype);
218 	mutex_init(&eq->sysfs_lock);
219 
220 	eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
221 					GFP_KERNEL, q->node);
222 	if (!eq->hash)
223 		goto err;
224 
225 	for (i = 0; i < ELV_HASH_ENTRIES; i++)
226 		INIT_HLIST_HEAD(&eq->hash[i]);
227 
228 	return eq;
229 err:
230 	kfree(eq);
231 	elevator_put(e);
232 	return NULL;
233 }
234 
235 static void elevator_release(struct kobject *kobj)
236 {
237 	elevator_t *e = container_of(kobj, elevator_t, kobj);
238 
239 	elevator_put(e->elevator_type);
240 	kfree(e->hash);
241 	kfree(e);
242 }
243 
244 int elevator_init(struct request_queue *q, char *name)
245 {
246 	struct elevator_type *e = NULL;
247 	struct elevator_queue *eq;
248 	int ret = 0;
249 	void *data;
250 
251 	INIT_LIST_HEAD(&q->queue_head);
252 	q->last_merge = NULL;
253 	q->end_sector = 0;
254 	q->boundary_rq = NULL;
255 
256 	if (name) {
257 		e = elevator_get(name);
258 		if (!e)
259 			return -EINVAL;
260 	}
261 
262 	if (!e && *chosen_elevator) {
263 		e = elevator_get(chosen_elevator);
264 		if (!e)
265 			printk(KERN_ERR "I/O scheduler %s not found\n",
266 							chosen_elevator);
267 	}
268 
269 	if (!e) {
270 		e = elevator_get(CONFIG_DEFAULT_IOSCHED);
271 		if (!e) {
272 			printk(KERN_ERR
273 				"Default I/O scheduler not found. " \
274 				"Using noop.\n");
275 			e = elevator_get("noop");
276 		}
277 	}
278 
279 	eq = elevator_alloc(q, e);
280 	if (!eq)
281 		return -ENOMEM;
282 
283 	data = elevator_init_queue(q, eq);
284 	if (!data) {
285 		kobject_put(&eq->kobj);
286 		return -ENOMEM;
287 	}
288 
289 	elevator_attach(q, eq, data);
290 	return ret;
291 }
292 EXPORT_SYMBOL(elevator_init);
293 
294 void elevator_exit(elevator_t *e)
295 {
296 	mutex_lock(&e->sysfs_lock);
297 	if (e->ops->elevator_exit_fn)
298 		e->ops->elevator_exit_fn(e);
299 	e->ops = NULL;
300 	mutex_unlock(&e->sysfs_lock);
301 
302 	kobject_put(&e->kobj);
303 }
304 EXPORT_SYMBOL(elevator_exit);
305 
306 static void elv_activate_rq(struct request_queue *q, struct request *rq)
307 {
308 	elevator_t *e = q->elevator;
309 
310 	if (e->ops->elevator_activate_req_fn)
311 		e->ops->elevator_activate_req_fn(q, rq);
312 }
313 
314 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
315 {
316 	elevator_t *e = q->elevator;
317 
318 	if (e->ops->elevator_deactivate_req_fn)
319 		e->ops->elevator_deactivate_req_fn(q, rq);
320 }
321 
322 static inline void __elv_rqhash_del(struct request *rq)
323 {
324 	hlist_del_init(&rq->hash);
325 }
326 
327 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
328 {
329 	if (ELV_ON_HASH(rq))
330 		__elv_rqhash_del(rq);
331 }
332 
333 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
334 {
335 	elevator_t *e = q->elevator;
336 
337 	BUG_ON(ELV_ON_HASH(rq));
338 	hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
339 }
340 
341 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
342 {
343 	__elv_rqhash_del(rq);
344 	elv_rqhash_add(q, rq);
345 }
346 
347 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
348 {
349 	elevator_t *e = q->elevator;
350 	struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
351 	struct hlist_node *entry, *next;
352 	struct request *rq;
353 
354 	hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
355 		BUG_ON(!ELV_ON_HASH(rq));
356 
357 		if (unlikely(!rq_mergeable(rq))) {
358 			__elv_rqhash_del(rq);
359 			continue;
360 		}
361 
362 		if (rq_hash_key(rq) == offset)
363 			return rq;
364 	}
365 
366 	return NULL;
367 }
368 
369 /*
370  * RB-tree support functions for inserting/lookup/removal of requests
371  * in a sorted RB tree.
372  */
373 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
374 {
375 	struct rb_node **p = &root->rb_node;
376 	struct rb_node *parent = NULL;
377 	struct request *__rq;
378 
379 	while (*p) {
380 		parent = *p;
381 		__rq = rb_entry(parent, struct request, rb_node);
382 
383 		if (rq->sector < __rq->sector)
384 			p = &(*p)->rb_left;
385 		else if (rq->sector > __rq->sector)
386 			p = &(*p)->rb_right;
387 		else
388 			return __rq;
389 	}
390 
391 	rb_link_node(&rq->rb_node, parent, p);
392 	rb_insert_color(&rq->rb_node, root);
393 	return NULL;
394 }
395 EXPORT_SYMBOL(elv_rb_add);
396 
397 void elv_rb_del(struct rb_root *root, struct request *rq)
398 {
399 	BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
400 	rb_erase(&rq->rb_node, root);
401 	RB_CLEAR_NODE(&rq->rb_node);
402 }
403 EXPORT_SYMBOL(elv_rb_del);
404 
405 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
406 {
407 	struct rb_node *n = root->rb_node;
408 	struct request *rq;
409 
410 	while (n) {
411 		rq = rb_entry(n, struct request, rb_node);
412 
413 		if (sector < rq->sector)
414 			n = n->rb_left;
415 		else if (sector > rq->sector)
416 			n = n->rb_right;
417 		else
418 			return rq;
419 	}
420 
421 	return NULL;
422 }
423 EXPORT_SYMBOL(elv_rb_find);
424 
425 /*
426  * Insert rq into dispatch queue of q.  Queue lock must be held on
427  * entry.  rq is sort instead into the dispatch queue. To be used by
428  * specific elevators.
429  */
430 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
431 {
432 	sector_t boundary;
433 	struct list_head *entry;
434 	int stop_flags;
435 
436 	if (q->last_merge == rq)
437 		q->last_merge = NULL;
438 
439 	elv_rqhash_del(q, rq);
440 
441 	q->nr_sorted--;
442 
443 	boundary = q->end_sector;
444 	stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
445 	list_for_each_prev(entry, &q->queue_head) {
446 		struct request *pos = list_entry_rq(entry);
447 
448 		if (blk_discard_rq(rq) != blk_discard_rq(pos))
449 			break;
450 		if (rq_data_dir(rq) != rq_data_dir(pos))
451 			break;
452 		if (pos->cmd_flags & stop_flags)
453 			break;
454 		if (rq->sector >= boundary) {
455 			if (pos->sector < boundary)
456 				continue;
457 		} else {
458 			if (pos->sector >= boundary)
459 				break;
460 		}
461 		if (rq->sector >= pos->sector)
462 			break;
463 	}
464 
465 	list_add(&rq->queuelist, entry);
466 }
467 EXPORT_SYMBOL(elv_dispatch_sort);
468 
469 /*
470  * Insert rq into dispatch queue of q.  Queue lock must be held on
471  * entry.  rq is added to the back of the dispatch queue. To be used by
472  * specific elevators.
473  */
474 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
475 {
476 	if (q->last_merge == rq)
477 		q->last_merge = NULL;
478 
479 	elv_rqhash_del(q, rq);
480 
481 	q->nr_sorted--;
482 
483 	q->end_sector = rq_end_sector(rq);
484 	q->boundary_rq = rq;
485 	list_add_tail(&rq->queuelist, &q->queue_head);
486 }
487 EXPORT_SYMBOL(elv_dispatch_add_tail);
488 
489 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
490 {
491 	elevator_t *e = q->elevator;
492 	struct request *__rq;
493 	int ret;
494 
495 	/*
496 	 * First try one-hit cache.
497 	 */
498 	if (q->last_merge) {
499 		ret = elv_try_merge(q->last_merge, bio);
500 		if (ret != ELEVATOR_NO_MERGE) {
501 			*req = q->last_merge;
502 			return ret;
503 		}
504 	}
505 
506 	if (blk_queue_nomerges(q))
507 		return ELEVATOR_NO_MERGE;
508 
509 	/*
510 	 * See if our hash lookup can find a potential backmerge.
511 	 */
512 	__rq = elv_rqhash_find(q, bio->bi_sector);
513 	if (__rq && elv_rq_merge_ok(__rq, bio)) {
514 		*req = __rq;
515 		return ELEVATOR_BACK_MERGE;
516 	}
517 
518 	if (e->ops->elevator_merge_fn)
519 		return e->ops->elevator_merge_fn(q, req, bio);
520 
521 	return ELEVATOR_NO_MERGE;
522 }
523 
524 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
525 {
526 	elevator_t *e = q->elevator;
527 
528 	if (e->ops->elevator_merged_fn)
529 		e->ops->elevator_merged_fn(q, rq, type);
530 
531 	if (type == ELEVATOR_BACK_MERGE)
532 		elv_rqhash_reposition(q, rq);
533 
534 	q->last_merge = rq;
535 }
536 
537 void elv_merge_requests(struct request_queue *q, struct request *rq,
538 			     struct request *next)
539 {
540 	elevator_t *e = q->elevator;
541 
542 	if (e->ops->elevator_merge_req_fn)
543 		e->ops->elevator_merge_req_fn(q, rq, next);
544 
545 	elv_rqhash_reposition(q, rq);
546 	elv_rqhash_del(q, next);
547 
548 	q->nr_sorted--;
549 	q->last_merge = rq;
550 }
551 
552 void elv_requeue_request(struct request_queue *q, struct request *rq)
553 {
554 	/*
555 	 * it already went through dequeue, we need to decrement the
556 	 * in_flight count again
557 	 */
558 	if (blk_account_rq(rq)) {
559 		q->in_flight--;
560 		if (blk_sorted_rq(rq))
561 			elv_deactivate_rq(q, rq);
562 	}
563 
564 	rq->cmd_flags &= ~REQ_STARTED;
565 
566 	elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
567 }
568 
569 static void elv_drain_elevator(struct request_queue *q)
570 {
571 	static int printed;
572 	while (q->elevator->ops->elevator_dispatch_fn(q, 1))
573 		;
574 	if (q->nr_sorted == 0)
575 		return;
576 	if (printed++ < 10) {
577 		printk(KERN_ERR "%s: forced dispatching is broken "
578 		       "(nr_sorted=%u), please report this\n",
579 		       q->elevator->elevator_type->elevator_name, q->nr_sorted);
580 	}
581 }
582 
583 void elv_insert(struct request_queue *q, struct request *rq, int where)
584 {
585 	struct list_head *pos;
586 	unsigned ordseq;
587 	int unplug_it = 1;
588 
589 	blk_add_trace_rq(q, rq, BLK_TA_INSERT);
590 
591 	rq->q = q;
592 
593 	switch (where) {
594 	case ELEVATOR_INSERT_FRONT:
595 		rq->cmd_flags |= REQ_SOFTBARRIER;
596 
597 		list_add(&rq->queuelist, &q->queue_head);
598 		break;
599 
600 	case ELEVATOR_INSERT_BACK:
601 		rq->cmd_flags |= REQ_SOFTBARRIER;
602 		elv_drain_elevator(q);
603 		list_add_tail(&rq->queuelist, &q->queue_head);
604 		/*
605 		 * We kick the queue here for the following reasons.
606 		 * - The elevator might have returned NULL previously
607 		 *   to delay requests and returned them now.  As the
608 		 *   queue wasn't empty before this request, ll_rw_blk
609 		 *   won't run the queue on return, resulting in hang.
610 		 * - Usually, back inserted requests won't be merged
611 		 *   with anything.  There's no point in delaying queue
612 		 *   processing.
613 		 */
614 		blk_remove_plug(q);
615 		blk_start_queueing(q);
616 		break;
617 
618 	case ELEVATOR_INSERT_SORT:
619 		BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
620 		rq->cmd_flags |= REQ_SORTED;
621 		q->nr_sorted++;
622 		if (rq_mergeable(rq)) {
623 			elv_rqhash_add(q, rq);
624 			if (!q->last_merge)
625 				q->last_merge = rq;
626 		}
627 
628 		/*
629 		 * Some ioscheds (cfq) run q->request_fn directly, so
630 		 * rq cannot be accessed after calling
631 		 * elevator_add_req_fn.
632 		 */
633 		q->elevator->ops->elevator_add_req_fn(q, rq);
634 		break;
635 
636 	case ELEVATOR_INSERT_REQUEUE:
637 		/*
638 		 * If ordered flush isn't in progress, we do front
639 		 * insertion; otherwise, requests should be requeued
640 		 * in ordseq order.
641 		 */
642 		rq->cmd_flags |= REQ_SOFTBARRIER;
643 
644 		/*
645 		 * Most requeues happen because of a busy condition,
646 		 * don't force unplug of the queue for that case.
647 		 */
648 		unplug_it = 0;
649 
650 		if (q->ordseq == 0) {
651 			list_add(&rq->queuelist, &q->queue_head);
652 			break;
653 		}
654 
655 		ordseq = blk_ordered_req_seq(rq);
656 
657 		list_for_each(pos, &q->queue_head) {
658 			struct request *pos_rq = list_entry_rq(pos);
659 			if (ordseq <= blk_ordered_req_seq(pos_rq))
660 				break;
661 		}
662 
663 		list_add_tail(&rq->queuelist, pos);
664 		break;
665 
666 	default:
667 		printk(KERN_ERR "%s: bad insertion point %d\n",
668 		       __func__, where);
669 		BUG();
670 	}
671 
672 	if (unplug_it && blk_queue_plugged(q)) {
673 		int nrq = q->rq.count[READ] + q->rq.count[WRITE]
674 			- q->in_flight;
675 
676 		if (nrq >= q->unplug_thresh)
677 			__generic_unplug_device(q);
678 	}
679 }
680 
681 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
682 		       int plug)
683 {
684 	if (q->ordcolor)
685 		rq->cmd_flags |= REQ_ORDERED_COLOR;
686 
687 	if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
688 		/*
689 		 * toggle ordered color
690 		 */
691 		if (blk_barrier_rq(rq))
692 			q->ordcolor ^= 1;
693 
694 		/*
695 		 * barriers implicitly indicate back insertion
696 		 */
697 		if (where == ELEVATOR_INSERT_SORT)
698 			where = ELEVATOR_INSERT_BACK;
699 
700 		/*
701 		 * this request is scheduling boundary, update
702 		 * end_sector
703 		 */
704 		if (blk_fs_request(rq) || blk_discard_rq(rq)) {
705 			q->end_sector = rq_end_sector(rq);
706 			q->boundary_rq = rq;
707 		}
708 	} else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
709 		    where == ELEVATOR_INSERT_SORT)
710 		where = ELEVATOR_INSERT_BACK;
711 
712 	if (plug)
713 		blk_plug_device(q);
714 
715 	elv_insert(q, rq, where);
716 }
717 EXPORT_SYMBOL(__elv_add_request);
718 
719 void elv_add_request(struct request_queue *q, struct request *rq, int where,
720 		     int plug)
721 {
722 	unsigned long flags;
723 
724 	spin_lock_irqsave(q->queue_lock, flags);
725 	__elv_add_request(q, rq, where, plug);
726 	spin_unlock_irqrestore(q->queue_lock, flags);
727 }
728 EXPORT_SYMBOL(elv_add_request);
729 
730 static inline struct request *__elv_next_request(struct request_queue *q)
731 {
732 	struct request *rq;
733 
734 	while (1) {
735 		while (!list_empty(&q->queue_head)) {
736 			rq = list_entry_rq(q->queue_head.next);
737 			if (blk_do_ordered(q, &rq))
738 				return rq;
739 		}
740 
741 		if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
742 			return NULL;
743 	}
744 }
745 
746 struct request *elv_next_request(struct request_queue *q)
747 {
748 	struct request *rq;
749 	int ret;
750 
751 	while ((rq = __elv_next_request(q)) != NULL) {
752 		/*
753 		 * Kill the empty barrier place holder, the driver must
754 		 * not ever see it.
755 		 */
756 		if (blk_empty_barrier(rq)) {
757 			__blk_end_request(rq, 0, blk_rq_bytes(rq));
758 			continue;
759 		}
760 		if (!(rq->cmd_flags & REQ_STARTED)) {
761 			/*
762 			 * This is the first time the device driver
763 			 * sees this request (possibly after
764 			 * requeueing).  Notify IO scheduler.
765 			 */
766 			if (blk_sorted_rq(rq))
767 				elv_activate_rq(q, rq);
768 
769 			/*
770 			 * just mark as started even if we don't start
771 			 * it, a request that has been delayed should
772 			 * not be passed by new incoming requests
773 			 */
774 			rq->cmd_flags |= REQ_STARTED;
775 			blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
776 
777 			/*
778 			 * We are now handing the request to the hardware,
779 			 * add the timeout handler
780 			 */
781 			blk_add_timer(rq);
782 		}
783 
784 		if (!q->boundary_rq || q->boundary_rq == rq) {
785 			q->end_sector = rq_end_sector(rq);
786 			q->boundary_rq = NULL;
787 		}
788 
789 		if (rq->cmd_flags & REQ_DONTPREP)
790 			break;
791 
792 		if (q->dma_drain_size && rq->data_len) {
793 			/*
794 			 * make sure space for the drain appears we
795 			 * know we can do this because max_hw_segments
796 			 * has been adjusted to be one fewer than the
797 			 * device can handle
798 			 */
799 			rq->nr_phys_segments++;
800 		}
801 
802 		if (!q->prep_rq_fn)
803 			break;
804 
805 		ret = q->prep_rq_fn(q, rq);
806 		if (ret == BLKPREP_OK) {
807 			break;
808 		} else if (ret == BLKPREP_DEFER) {
809 			/*
810 			 * the request may have been (partially) prepped.
811 			 * we need to keep this request in the front to
812 			 * avoid resource deadlock.  REQ_STARTED will
813 			 * prevent other fs requests from passing this one.
814 			 */
815 			if (q->dma_drain_size && rq->data_len &&
816 			    !(rq->cmd_flags & REQ_DONTPREP)) {
817 				/*
818 				 * remove the space for the drain we added
819 				 * so that we don't add it again
820 				 */
821 				--rq->nr_phys_segments;
822 			}
823 
824 			rq = NULL;
825 			break;
826 		} else if (ret == BLKPREP_KILL) {
827 			rq->cmd_flags |= REQ_QUIET;
828 			__blk_end_request(rq, -EIO, blk_rq_bytes(rq));
829 		} else {
830 			printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
831 			break;
832 		}
833 	}
834 
835 	return rq;
836 }
837 EXPORT_SYMBOL(elv_next_request);
838 
839 void elv_dequeue_request(struct request_queue *q, struct request *rq)
840 {
841 	BUG_ON(list_empty(&rq->queuelist));
842 	BUG_ON(ELV_ON_HASH(rq));
843 
844 	list_del_init(&rq->queuelist);
845 
846 	/*
847 	 * the time frame between a request being removed from the lists
848 	 * and to it is freed is accounted as io that is in progress at
849 	 * the driver side.
850 	 */
851 	if (blk_account_rq(rq))
852 		q->in_flight++;
853 }
854 EXPORT_SYMBOL(elv_dequeue_request);
855 
856 int elv_queue_empty(struct request_queue *q)
857 {
858 	elevator_t *e = q->elevator;
859 
860 	if (!list_empty(&q->queue_head))
861 		return 0;
862 
863 	if (e->ops->elevator_queue_empty_fn)
864 		return e->ops->elevator_queue_empty_fn(q);
865 
866 	return 1;
867 }
868 EXPORT_SYMBOL(elv_queue_empty);
869 
870 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
871 {
872 	elevator_t *e = q->elevator;
873 
874 	if (e->ops->elevator_latter_req_fn)
875 		return e->ops->elevator_latter_req_fn(q, rq);
876 	return NULL;
877 }
878 
879 struct request *elv_former_request(struct request_queue *q, struct request *rq)
880 {
881 	elevator_t *e = q->elevator;
882 
883 	if (e->ops->elevator_former_req_fn)
884 		return e->ops->elevator_former_req_fn(q, rq);
885 	return NULL;
886 }
887 
888 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
889 {
890 	elevator_t *e = q->elevator;
891 
892 	if (e->ops->elevator_set_req_fn)
893 		return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
894 
895 	rq->elevator_private = NULL;
896 	return 0;
897 }
898 
899 void elv_put_request(struct request_queue *q, struct request *rq)
900 {
901 	elevator_t *e = q->elevator;
902 
903 	if (e->ops->elevator_put_req_fn)
904 		e->ops->elevator_put_req_fn(rq);
905 }
906 
907 int elv_may_queue(struct request_queue *q, int rw)
908 {
909 	elevator_t *e = q->elevator;
910 
911 	if (e->ops->elevator_may_queue_fn)
912 		return e->ops->elevator_may_queue_fn(q, rw);
913 
914 	return ELV_MQUEUE_MAY;
915 }
916 
917 void elv_abort_queue(struct request_queue *q)
918 {
919 	struct request *rq;
920 
921 	while (!list_empty(&q->queue_head)) {
922 		rq = list_entry_rq(q->queue_head.next);
923 		rq->cmd_flags |= REQ_QUIET;
924 		blk_add_trace_rq(q, rq, BLK_TA_ABORT);
925 		__blk_end_request(rq, -EIO, blk_rq_bytes(rq));
926 	}
927 }
928 EXPORT_SYMBOL(elv_abort_queue);
929 
930 void elv_completed_request(struct request_queue *q, struct request *rq)
931 {
932 	elevator_t *e = q->elevator;
933 
934 	/*
935 	 * request is released from the driver, io must be done
936 	 */
937 	if (blk_account_rq(rq)) {
938 		q->in_flight--;
939 		if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
940 			e->ops->elevator_completed_req_fn(q, rq);
941 	}
942 
943 	/*
944 	 * Check if the queue is waiting for fs requests to be
945 	 * drained for flush sequence.
946 	 */
947 	if (unlikely(q->ordseq)) {
948 		struct request *first_rq = list_entry_rq(q->queue_head.next);
949 		if (q->in_flight == 0 &&
950 		    blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
951 		    blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
952 			blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
953 			blk_start_queueing(q);
954 		}
955 	}
956 }
957 
958 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
959 
960 static ssize_t
961 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
962 {
963 	elevator_t *e = container_of(kobj, elevator_t, kobj);
964 	struct elv_fs_entry *entry = to_elv(attr);
965 	ssize_t error;
966 
967 	if (!entry->show)
968 		return -EIO;
969 
970 	mutex_lock(&e->sysfs_lock);
971 	error = e->ops ? entry->show(e, page) : -ENOENT;
972 	mutex_unlock(&e->sysfs_lock);
973 	return error;
974 }
975 
976 static ssize_t
977 elv_attr_store(struct kobject *kobj, struct attribute *attr,
978 	       const char *page, size_t length)
979 {
980 	elevator_t *e = container_of(kobj, elevator_t, kobj);
981 	struct elv_fs_entry *entry = to_elv(attr);
982 	ssize_t error;
983 
984 	if (!entry->store)
985 		return -EIO;
986 
987 	mutex_lock(&e->sysfs_lock);
988 	error = e->ops ? entry->store(e, page, length) : -ENOENT;
989 	mutex_unlock(&e->sysfs_lock);
990 	return error;
991 }
992 
993 static struct sysfs_ops elv_sysfs_ops = {
994 	.show	= elv_attr_show,
995 	.store	= elv_attr_store,
996 };
997 
998 static struct kobj_type elv_ktype = {
999 	.sysfs_ops	= &elv_sysfs_ops,
1000 	.release	= elevator_release,
1001 };
1002 
1003 int elv_register_queue(struct request_queue *q)
1004 {
1005 	elevator_t *e = q->elevator;
1006 	int error;
1007 
1008 	error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
1009 	if (!error) {
1010 		struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
1011 		if (attr) {
1012 			while (attr->attr.name) {
1013 				if (sysfs_create_file(&e->kobj, &attr->attr))
1014 					break;
1015 				attr++;
1016 			}
1017 		}
1018 		kobject_uevent(&e->kobj, KOBJ_ADD);
1019 	}
1020 	return error;
1021 }
1022 
1023 static void __elv_unregister_queue(elevator_t *e)
1024 {
1025 	kobject_uevent(&e->kobj, KOBJ_REMOVE);
1026 	kobject_del(&e->kobj);
1027 }
1028 
1029 void elv_unregister_queue(struct request_queue *q)
1030 {
1031 	if (q)
1032 		__elv_unregister_queue(q->elevator);
1033 }
1034 
1035 void elv_register(struct elevator_type *e)
1036 {
1037 	char *def = "";
1038 
1039 	spin_lock(&elv_list_lock);
1040 	BUG_ON(elevator_find(e->elevator_name));
1041 	list_add_tail(&e->list, &elv_list);
1042 	spin_unlock(&elv_list_lock);
1043 
1044 	if (!strcmp(e->elevator_name, chosen_elevator) ||
1045 			(!*chosen_elevator &&
1046 			 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
1047 				def = " (default)";
1048 
1049 	printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
1050 								def);
1051 }
1052 EXPORT_SYMBOL_GPL(elv_register);
1053 
1054 void elv_unregister(struct elevator_type *e)
1055 {
1056 	struct task_struct *g, *p;
1057 
1058 	/*
1059 	 * Iterate every thread in the process to remove the io contexts.
1060 	 */
1061 	if (e->ops.trim) {
1062 		read_lock(&tasklist_lock);
1063 		do_each_thread(g, p) {
1064 			task_lock(p);
1065 			if (p->io_context)
1066 				e->ops.trim(p->io_context);
1067 			task_unlock(p);
1068 		} while_each_thread(g, p);
1069 		read_unlock(&tasklist_lock);
1070 	}
1071 
1072 	spin_lock(&elv_list_lock);
1073 	list_del_init(&e->list);
1074 	spin_unlock(&elv_list_lock);
1075 }
1076 EXPORT_SYMBOL_GPL(elv_unregister);
1077 
1078 /*
1079  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1080  * we don't free the old io scheduler, before we have allocated what we
1081  * need for the new one. this way we have a chance of going back to the old
1082  * one, if the new one fails init for some reason.
1083  */
1084 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1085 {
1086 	elevator_t *old_elevator, *e;
1087 	void *data;
1088 
1089 	/*
1090 	 * Allocate new elevator
1091 	 */
1092 	e = elevator_alloc(q, new_e);
1093 	if (!e)
1094 		return 0;
1095 
1096 	data = elevator_init_queue(q, e);
1097 	if (!data) {
1098 		kobject_put(&e->kobj);
1099 		return 0;
1100 	}
1101 
1102 	/*
1103 	 * Turn on BYPASS and drain all requests w/ elevator private data
1104 	 */
1105 	spin_lock_irq(q->queue_lock);
1106 
1107 	queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
1108 
1109 	elv_drain_elevator(q);
1110 
1111 	while (q->rq.elvpriv) {
1112 		blk_start_queueing(q);
1113 		spin_unlock_irq(q->queue_lock);
1114 		msleep(10);
1115 		spin_lock_irq(q->queue_lock);
1116 		elv_drain_elevator(q);
1117 	}
1118 
1119 	/*
1120 	 * Remember old elevator.
1121 	 */
1122 	old_elevator = q->elevator;
1123 
1124 	/*
1125 	 * attach and start new elevator
1126 	 */
1127 	elevator_attach(q, e, data);
1128 
1129 	spin_unlock_irq(q->queue_lock);
1130 
1131 	__elv_unregister_queue(old_elevator);
1132 
1133 	if (elv_register_queue(q))
1134 		goto fail_register;
1135 
1136 	/*
1137 	 * finally exit old elevator and turn off BYPASS.
1138 	 */
1139 	elevator_exit(old_elevator);
1140 	spin_lock_irq(q->queue_lock);
1141 	queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1142 	spin_unlock_irq(q->queue_lock);
1143 
1144 	blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1145 
1146 	return 1;
1147 
1148 fail_register:
1149 	/*
1150 	 * switch failed, exit the new io scheduler and reattach the old
1151 	 * one again (along with re-adding the sysfs dir)
1152 	 */
1153 	elevator_exit(e);
1154 	q->elevator = old_elevator;
1155 	elv_register_queue(q);
1156 
1157 	spin_lock_irq(q->queue_lock);
1158 	queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1159 	spin_unlock_irq(q->queue_lock);
1160 
1161 	return 0;
1162 }
1163 
1164 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1165 			  size_t count)
1166 {
1167 	char elevator_name[ELV_NAME_MAX];
1168 	struct elevator_type *e;
1169 
1170 	strlcpy(elevator_name, name, sizeof(elevator_name));
1171 	strstrip(elevator_name);
1172 
1173 	e = elevator_get(elevator_name);
1174 	if (!e) {
1175 		printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1176 		return -EINVAL;
1177 	}
1178 
1179 	if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1180 		elevator_put(e);
1181 		return count;
1182 	}
1183 
1184 	if (!elevator_switch(q, e))
1185 		printk(KERN_ERR "elevator: switch to %s failed\n",
1186 							elevator_name);
1187 	return count;
1188 }
1189 
1190 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1191 {
1192 	elevator_t *e = q->elevator;
1193 	struct elevator_type *elv = e->elevator_type;
1194 	struct elevator_type *__e;
1195 	int len = 0;
1196 
1197 	spin_lock(&elv_list_lock);
1198 	list_for_each_entry(__e, &elv_list, list) {
1199 		if (!strcmp(elv->elevator_name, __e->elevator_name))
1200 			len += sprintf(name+len, "[%s] ", elv->elevator_name);
1201 		else
1202 			len += sprintf(name+len, "%s ", __e->elevator_name);
1203 	}
1204 	spin_unlock(&elv_list_lock);
1205 
1206 	len += sprintf(len+name, "\n");
1207 	return len;
1208 }
1209 
1210 struct request *elv_rb_former_request(struct request_queue *q,
1211 				      struct request *rq)
1212 {
1213 	struct rb_node *rbprev = rb_prev(&rq->rb_node);
1214 
1215 	if (rbprev)
1216 		return rb_entry_rq(rbprev);
1217 
1218 	return NULL;
1219 }
1220 EXPORT_SYMBOL(elv_rb_former_request);
1221 
1222 struct request *elv_rb_latter_request(struct request_queue *q,
1223 				      struct request *rq)
1224 {
1225 	struct rb_node *rbnext = rb_next(&rq->rb_node);
1226 
1227 	if (rbnext)
1228 		return rb_entry_rq(rbnext);
1229 
1230 	return NULL;
1231 }
1232 EXPORT_SYMBOL(elv_rb_latter_request);
1233