xref: /openbmc/linux/block/elevator.c (revision 5d0e4d78)
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 
46 static DEFINE_SPINLOCK(elv_list_lock);
47 static 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  */
58 static 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  */
74 bool 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 }
84 EXPORT_SYMBOL(elv_bio_merge_ok);
85 
86 static struct elevator_type *elevator_find(const char *name)
87 {
88 	struct elevator_type *e;
89 
90 	list_for_each_entry(e, &elv_list, list) {
91 		if (!strcmp(e->elevator_name, name))
92 			return e;
93 	}
94 
95 	return NULL;
96 }
97 
98 static void elevator_put(struct elevator_type *e)
99 {
100 	module_put(e->elevator_owner);
101 }
102 
103 static struct elevator_type *elevator_get(const char *name, bool try_loading)
104 {
105 	struct elevator_type *e;
106 
107 	spin_lock(&elv_list_lock);
108 
109 	e = elevator_find(name);
110 	if (!e && try_loading) {
111 		spin_unlock(&elv_list_lock);
112 		request_module("%s-iosched", name);
113 		spin_lock(&elv_list_lock);
114 		e = elevator_find(name);
115 	}
116 
117 	if (e && !try_module_get(e->elevator_owner))
118 		e = NULL;
119 
120 	spin_unlock(&elv_list_lock);
121 
122 	return e;
123 }
124 
125 static char chosen_elevator[ELV_NAME_MAX];
126 
127 static int __init elevator_setup(char *str)
128 {
129 	/*
130 	 * Be backwards-compatible with previous kernels, so users
131 	 * won't get the wrong elevator.
132 	 */
133 	strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
134 	return 1;
135 }
136 
137 __setup("elevator=", elevator_setup);
138 
139 /* called during boot to load the elevator chosen by the elevator param */
140 void __init load_default_elevator_module(void)
141 {
142 	struct elevator_type *e;
143 
144 	if (!chosen_elevator[0])
145 		return;
146 
147 	spin_lock(&elv_list_lock);
148 	e = elevator_find(chosen_elevator);
149 	spin_unlock(&elv_list_lock);
150 
151 	if (!e)
152 		request_module("%s-iosched", chosen_elevator);
153 }
154 
155 static struct kobj_type elv_ktype;
156 
157 struct elevator_queue *elevator_alloc(struct request_queue *q,
158 				  struct elevator_type *e)
159 {
160 	struct elevator_queue *eq;
161 
162 	eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
163 	if (unlikely(!eq))
164 		return NULL;
165 
166 	eq->type = e;
167 	kobject_init(&eq->kobj, &elv_ktype);
168 	mutex_init(&eq->sysfs_lock);
169 	hash_init(eq->hash);
170 	eq->uses_mq = e->uses_mq;
171 
172 	return eq;
173 }
174 EXPORT_SYMBOL(elevator_alloc);
175 
176 static void elevator_release(struct kobject *kobj)
177 {
178 	struct elevator_queue *e;
179 
180 	e = container_of(kobj, struct elevator_queue, kobj);
181 	elevator_put(e->type);
182 	kfree(e);
183 }
184 
185 int elevator_init(struct request_queue *q, char *name)
186 {
187 	struct elevator_type *e = NULL;
188 	int err;
189 
190 	/*
191 	 * q->sysfs_lock must be held to provide mutual exclusion between
192 	 * elevator_switch() and here.
193 	 */
194 	lockdep_assert_held(&q->sysfs_lock);
195 
196 	if (unlikely(q->elevator))
197 		return 0;
198 
199 	INIT_LIST_HEAD(&q->queue_head);
200 	q->last_merge = NULL;
201 	q->end_sector = 0;
202 	q->boundary_rq = NULL;
203 
204 	if (name) {
205 		e = elevator_get(name, true);
206 		if (!e)
207 			return -EINVAL;
208 	}
209 
210 	/*
211 	 * Use the default elevator specified by config boot param for
212 	 * non-mq devices, or by config option. Don't try to load modules
213 	 * as we could be running off async and request_module() isn't
214 	 * allowed from async.
215 	 */
216 	if (!e && !q->mq_ops && *chosen_elevator) {
217 		e = elevator_get(chosen_elevator, false);
218 		if (!e)
219 			printk(KERN_ERR "I/O scheduler %s not found\n",
220 							chosen_elevator);
221 	}
222 
223 	if (!e) {
224 		/*
225 		 * For blk-mq devices, we default to using mq-deadline,
226 		 * if available, for single queue devices. If deadline
227 		 * isn't available OR we have multiple queues, default
228 		 * to "none".
229 		 */
230 		if (q->mq_ops) {
231 			if (q->nr_hw_queues == 1)
232 				e = elevator_get("mq-deadline", false);
233 			if (!e)
234 				return 0;
235 		} else
236 			e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
237 
238 		if (!e) {
239 			printk(KERN_ERR
240 				"Default I/O scheduler not found. " \
241 				"Using noop.\n");
242 			e = elevator_get("noop", false);
243 		}
244 	}
245 
246 	if (e->uses_mq)
247 		err = blk_mq_init_sched(q, e);
248 	else
249 		err = e->ops.sq.elevator_init_fn(q, e);
250 	if (err)
251 		elevator_put(e);
252 	return err;
253 }
254 EXPORT_SYMBOL(elevator_init);
255 
256 void elevator_exit(struct request_queue *q, struct elevator_queue *e)
257 {
258 	mutex_lock(&e->sysfs_lock);
259 	if (e->uses_mq && e->type->ops.mq.exit_sched)
260 		blk_mq_exit_sched(q, e);
261 	else if (!e->uses_mq && e->type->ops.sq.elevator_exit_fn)
262 		e->type->ops.sq.elevator_exit_fn(e);
263 	mutex_unlock(&e->sysfs_lock);
264 
265 	kobject_put(&e->kobj);
266 }
267 EXPORT_SYMBOL(elevator_exit);
268 
269 static inline void __elv_rqhash_del(struct request *rq)
270 {
271 	hash_del(&rq->hash);
272 	rq->rq_flags &= ~RQF_HASHED;
273 }
274 
275 void elv_rqhash_del(struct request_queue *q, struct request *rq)
276 {
277 	if (ELV_ON_HASH(rq))
278 		__elv_rqhash_del(rq);
279 }
280 EXPORT_SYMBOL_GPL(elv_rqhash_del);
281 
282 void elv_rqhash_add(struct request_queue *q, struct request *rq)
283 {
284 	struct elevator_queue *e = q->elevator;
285 
286 	BUG_ON(ELV_ON_HASH(rq));
287 	hash_add(e->hash, &rq->hash, rq_hash_key(rq));
288 	rq->rq_flags |= RQF_HASHED;
289 }
290 EXPORT_SYMBOL_GPL(elv_rqhash_add);
291 
292 void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
293 {
294 	__elv_rqhash_del(rq);
295 	elv_rqhash_add(q, rq);
296 }
297 
298 struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
299 {
300 	struct elevator_queue *e = q->elevator;
301 	struct hlist_node *next;
302 	struct request *rq;
303 
304 	hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
305 		BUG_ON(!ELV_ON_HASH(rq));
306 
307 		if (unlikely(!rq_mergeable(rq))) {
308 			__elv_rqhash_del(rq);
309 			continue;
310 		}
311 
312 		if (rq_hash_key(rq) == offset)
313 			return rq;
314 	}
315 
316 	return NULL;
317 }
318 
319 /*
320  * RB-tree support functions for inserting/lookup/removal of requests
321  * in a sorted RB tree.
322  */
323 void elv_rb_add(struct rb_root *root, struct request *rq)
324 {
325 	struct rb_node **p = &root->rb_node;
326 	struct rb_node *parent = NULL;
327 	struct request *__rq;
328 
329 	while (*p) {
330 		parent = *p;
331 		__rq = rb_entry(parent, struct request, rb_node);
332 
333 		if (blk_rq_pos(rq) < blk_rq_pos(__rq))
334 			p = &(*p)->rb_left;
335 		else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
336 			p = &(*p)->rb_right;
337 	}
338 
339 	rb_link_node(&rq->rb_node, parent, p);
340 	rb_insert_color(&rq->rb_node, root);
341 }
342 EXPORT_SYMBOL(elv_rb_add);
343 
344 void elv_rb_del(struct rb_root *root, struct request *rq)
345 {
346 	BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
347 	rb_erase(&rq->rb_node, root);
348 	RB_CLEAR_NODE(&rq->rb_node);
349 }
350 EXPORT_SYMBOL(elv_rb_del);
351 
352 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
353 {
354 	struct rb_node *n = root->rb_node;
355 	struct request *rq;
356 
357 	while (n) {
358 		rq = rb_entry(n, struct request, rb_node);
359 
360 		if (sector < blk_rq_pos(rq))
361 			n = n->rb_left;
362 		else if (sector > blk_rq_pos(rq))
363 			n = n->rb_right;
364 		else
365 			return rq;
366 	}
367 
368 	return NULL;
369 }
370 EXPORT_SYMBOL(elv_rb_find);
371 
372 /*
373  * Insert rq into dispatch queue of q.  Queue lock must be held on
374  * entry.  rq is sort instead into the dispatch queue. To be used by
375  * specific elevators.
376  */
377 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
378 {
379 	sector_t boundary;
380 	struct list_head *entry;
381 
382 	if (q->last_merge == rq)
383 		q->last_merge = NULL;
384 
385 	elv_rqhash_del(q, rq);
386 
387 	q->nr_sorted--;
388 
389 	boundary = q->end_sector;
390 	list_for_each_prev(entry, &q->queue_head) {
391 		struct request *pos = list_entry_rq(entry);
392 
393 		if (req_op(rq) != req_op(pos))
394 			break;
395 		if (rq_data_dir(rq) != rq_data_dir(pos))
396 			break;
397 		if (pos->rq_flags & (RQF_STARTED | RQF_SOFTBARRIER))
398 			break;
399 		if (blk_rq_pos(rq) >= boundary) {
400 			if (blk_rq_pos(pos) < boundary)
401 				continue;
402 		} else {
403 			if (blk_rq_pos(pos) >= boundary)
404 				break;
405 		}
406 		if (blk_rq_pos(rq) >= blk_rq_pos(pos))
407 			break;
408 	}
409 
410 	list_add(&rq->queuelist, entry);
411 }
412 EXPORT_SYMBOL(elv_dispatch_sort);
413 
414 /*
415  * Insert rq into dispatch queue of q.  Queue lock must be held on
416  * entry.  rq is added to the back of the dispatch queue. To be used by
417  * specific elevators.
418  */
419 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
420 {
421 	if (q->last_merge == rq)
422 		q->last_merge = NULL;
423 
424 	elv_rqhash_del(q, rq);
425 
426 	q->nr_sorted--;
427 
428 	q->end_sector = rq_end_sector(rq);
429 	q->boundary_rq = rq;
430 	list_add_tail(&rq->queuelist, &q->queue_head);
431 }
432 EXPORT_SYMBOL(elv_dispatch_add_tail);
433 
434 enum elv_merge elv_merge(struct request_queue *q, struct request **req,
435 		struct bio *bio)
436 {
437 	struct elevator_queue *e = q->elevator;
438 	struct request *__rq;
439 
440 	/*
441 	 * Levels of merges:
442 	 * 	nomerges:  No merges at all attempted
443 	 * 	noxmerges: Only simple one-hit cache try
444 	 * 	merges:	   All merge tries attempted
445 	 */
446 	if (blk_queue_nomerges(q) || !bio_mergeable(bio))
447 		return ELEVATOR_NO_MERGE;
448 
449 	/*
450 	 * First try one-hit cache.
451 	 */
452 	if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
453 		enum elv_merge ret = blk_try_merge(q->last_merge, bio);
454 
455 		if (ret != ELEVATOR_NO_MERGE) {
456 			*req = q->last_merge;
457 			return ret;
458 		}
459 	}
460 
461 	if (blk_queue_noxmerges(q))
462 		return ELEVATOR_NO_MERGE;
463 
464 	/*
465 	 * See if our hash lookup can find a potential backmerge.
466 	 */
467 	__rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
468 	if (__rq && elv_bio_merge_ok(__rq, bio)) {
469 		*req = __rq;
470 		return ELEVATOR_BACK_MERGE;
471 	}
472 
473 	if (e->uses_mq && e->type->ops.mq.request_merge)
474 		return e->type->ops.mq.request_merge(q, req, bio);
475 	else if (!e->uses_mq && e->type->ops.sq.elevator_merge_fn)
476 		return e->type->ops.sq.elevator_merge_fn(q, req, bio);
477 
478 	return ELEVATOR_NO_MERGE;
479 }
480 
481 /*
482  * Attempt to do an insertion back merge. Only check for the case where
483  * we can append 'rq' to an existing request, so we can throw 'rq' away
484  * afterwards.
485  *
486  * Returns true if we merged, false otherwise
487  */
488 bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq)
489 {
490 	struct request *__rq;
491 	bool ret;
492 
493 	if (blk_queue_nomerges(q))
494 		return false;
495 
496 	/*
497 	 * First try one-hit cache.
498 	 */
499 	if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
500 		return true;
501 
502 	if (blk_queue_noxmerges(q))
503 		return false;
504 
505 	ret = false;
506 	/*
507 	 * See if our hash lookup can find a potential backmerge.
508 	 */
509 	while (1) {
510 		__rq = elv_rqhash_find(q, blk_rq_pos(rq));
511 		if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
512 			break;
513 
514 		/* The merged request could be merged with others, try again */
515 		ret = true;
516 		rq = __rq;
517 	}
518 
519 	return ret;
520 }
521 
522 void elv_merged_request(struct request_queue *q, struct request *rq,
523 		enum elv_merge type)
524 {
525 	struct elevator_queue *e = q->elevator;
526 
527 	if (e->uses_mq && e->type->ops.mq.request_merged)
528 		e->type->ops.mq.request_merged(q, rq, type);
529 	else if (!e->uses_mq && e->type->ops.sq.elevator_merged_fn)
530 		e->type->ops.sq.elevator_merged_fn(q, rq, type);
531 
532 	if (type == ELEVATOR_BACK_MERGE)
533 		elv_rqhash_reposition(q, rq);
534 
535 	q->last_merge = rq;
536 }
537 
538 void elv_merge_requests(struct request_queue *q, struct request *rq,
539 			     struct request *next)
540 {
541 	struct elevator_queue *e = q->elevator;
542 	bool next_sorted = false;
543 
544 	if (e->uses_mq && e->type->ops.mq.requests_merged)
545 		e->type->ops.mq.requests_merged(q, rq, next);
546 	else if (e->type->ops.sq.elevator_merge_req_fn) {
547 		next_sorted = (__force bool)(next->rq_flags & RQF_SORTED);
548 		if (next_sorted)
549 			e->type->ops.sq.elevator_merge_req_fn(q, rq, next);
550 	}
551 
552 	elv_rqhash_reposition(q, rq);
553 
554 	if (next_sorted) {
555 		elv_rqhash_del(q, next);
556 		q->nr_sorted--;
557 	}
558 
559 	q->last_merge = rq;
560 }
561 
562 void elv_bio_merged(struct request_queue *q, struct request *rq,
563 			struct bio *bio)
564 {
565 	struct elevator_queue *e = q->elevator;
566 
567 	if (WARN_ON_ONCE(e->uses_mq))
568 		return;
569 
570 	if (e->type->ops.sq.elevator_bio_merged_fn)
571 		e->type->ops.sq.elevator_bio_merged_fn(q, rq, bio);
572 }
573 
574 #ifdef CONFIG_PM
575 static void blk_pm_requeue_request(struct request *rq)
576 {
577 	if (rq->q->dev && !(rq->rq_flags & RQF_PM))
578 		rq->q->nr_pending--;
579 }
580 
581 static void blk_pm_add_request(struct request_queue *q, struct request *rq)
582 {
583 	if (q->dev && !(rq->rq_flags & RQF_PM) && q->nr_pending++ == 0 &&
584 	    (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
585 		pm_request_resume(q->dev);
586 }
587 #else
588 static inline void blk_pm_requeue_request(struct request *rq) {}
589 static inline void blk_pm_add_request(struct request_queue *q,
590 				      struct request *rq)
591 {
592 }
593 #endif
594 
595 void elv_requeue_request(struct request_queue *q, struct request *rq)
596 {
597 	/*
598 	 * it already went through dequeue, we need to decrement the
599 	 * in_flight count again
600 	 */
601 	if (blk_account_rq(rq)) {
602 		q->in_flight[rq_is_sync(rq)]--;
603 		if (rq->rq_flags & RQF_SORTED)
604 			elv_deactivate_rq(q, rq);
605 	}
606 
607 	rq->rq_flags &= ~RQF_STARTED;
608 
609 	blk_pm_requeue_request(rq);
610 
611 	__elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
612 }
613 
614 void elv_drain_elevator(struct request_queue *q)
615 {
616 	struct elevator_queue *e = q->elevator;
617 	static int printed;
618 
619 	if (WARN_ON_ONCE(e->uses_mq))
620 		return;
621 
622 	lockdep_assert_held(q->queue_lock);
623 
624 	while (e->type->ops.sq.elevator_dispatch_fn(q, 1))
625 		;
626 	if (q->nr_sorted && printed++ < 10) {
627 		printk(KERN_ERR "%s: forced dispatching is broken "
628 		       "(nr_sorted=%u), please report this\n",
629 		       q->elevator->type->elevator_name, q->nr_sorted);
630 	}
631 }
632 
633 void __elv_add_request(struct request_queue *q, struct request *rq, int where)
634 {
635 	trace_block_rq_insert(q, rq);
636 
637 	blk_pm_add_request(q, rq);
638 
639 	rq->q = q;
640 
641 	if (rq->rq_flags & RQF_SOFTBARRIER) {
642 		/* barriers are scheduling boundary, update end_sector */
643 		if (!blk_rq_is_passthrough(rq)) {
644 			q->end_sector = rq_end_sector(rq);
645 			q->boundary_rq = rq;
646 		}
647 	} else if (!(rq->rq_flags & RQF_ELVPRIV) &&
648 		    (where == ELEVATOR_INSERT_SORT ||
649 		     where == ELEVATOR_INSERT_SORT_MERGE))
650 		where = ELEVATOR_INSERT_BACK;
651 
652 	switch (where) {
653 	case ELEVATOR_INSERT_REQUEUE:
654 	case ELEVATOR_INSERT_FRONT:
655 		rq->rq_flags |= RQF_SOFTBARRIER;
656 		list_add(&rq->queuelist, &q->queue_head);
657 		break;
658 
659 	case ELEVATOR_INSERT_BACK:
660 		rq->rq_flags |= RQF_SOFTBARRIER;
661 		elv_drain_elevator(q);
662 		list_add_tail(&rq->queuelist, &q->queue_head);
663 		/*
664 		 * We kick the queue here for the following reasons.
665 		 * - The elevator might have returned NULL previously
666 		 *   to delay requests and returned them now.  As the
667 		 *   queue wasn't empty before this request, ll_rw_blk
668 		 *   won't run the queue on return, resulting in hang.
669 		 * - Usually, back inserted requests won't be merged
670 		 *   with anything.  There's no point in delaying queue
671 		 *   processing.
672 		 */
673 		__blk_run_queue(q);
674 		break;
675 
676 	case ELEVATOR_INSERT_SORT_MERGE:
677 		/*
678 		 * If we succeed in merging this request with one in the
679 		 * queue already, we are done - rq has now been freed,
680 		 * so no need to do anything further.
681 		 */
682 		if (elv_attempt_insert_merge(q, rq))
683 			break;
684 		/* fall through */
685 	case ELEVATOR_INSERT_SORT:
686 		BUG_ON(blk_rq_is_passthrough(rq));
687 		rq->rq_flags |= RQF_SORTED;
688 		q->nr_sorted++;
689 		if (rq_mergeable(rq)) {
690 			elv_rqhash_add(q, rq);
691 			if (!q->last_merge)
692 				q->last_merge = rq;
693 		}
694 
695 		/*
696 		 * Some ioscheds (cfq) run q->request_fn directly, so
697 		 * rq cannot be accessed after calling
698 		 * elevator_add_req_fn.
699 		 */
700 		q->elevator->type->ops.sq.elevator_add_req_fn(q, rq);
701 		break;
702 
703 	case ELEVATOR_INSERT_FLUSH:
704 		rq->rq_flags |= RQF_SOFTBARRIER;
705 		blk_insert_flush(rq);
706 		break;
707 	default:
708 		printk(KERN_ERR "%s: bad insertion point %d\n",
709 		       __func__, where);
710 		BUG();
711 	}
712 }
713 EXPORT_SYMBOL(__elv_add_request);
714 
715 void elv_add_request(struct request_queue *q, struct request *rq, int where)
716 {
717 	unsigned long flags;
718 
719 	spin_lock_irqsave(q->queue_lock, flags);
720 	__elv_add_request(q, rq, where);
721 	spin_unlock_irqrestore(q->queue_lock, flags);
722 }
723 EXPORT_SYMBOL(elv_add_request);
724 
725 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
726 {
727 	struct elevator_queue *e = q->elevator;
728 
729 	if (e->uses_mq && e->type->ops.mq.next_request)
730 		return e->type->ops.mq.next_request(q, rq);
731 	else if (!e->uses_mq && e->type->ops.sq.elevator_latter_req_fn)
732 		return e->type->ops.sq.elevator_latter_req_fn(q, rq);
733 
734 	return NULL;
735 }
736 
737 struct request *elv_former_request(struct request_queue *q, struct request *rq)
738 {
739 	struct elevator_queue *e = q->elevator;
740 
741 	if (e->uses_mq && e->type->ops.mq.former_request)
742 		return e->type->ops.mq.former_request(q, rq);
743 	if (!e->uses_mq && e->type->ops.sq.elevator_former_req_fn)
744 		return e->type->ops.sq.elevator_former_req_fn(q, rq);
745 	return NULL;
746 }
747 
748 int elv_set_request(struct request_queue *q, struct request *rq,
749 		    struct bio *bio, gfp_t gfp_mask)
750 {
751 	struct elevator_queue *e = q->elevator;
752 
753 	if (WARN_ON_ONCE(e->uses_mq))
754 		return 0;
755 
756 	if (e->type->ops.sq.elevator_set_req_fn)
757 		return e->type->ops.sq.elevator_set_req_fn(q, rq, bio, gfp_mask);
758 	return 0;
759 }
760 
761 void elv_put_request(struct request_queue *q, struct request *rq)
762 {
763 	struct elevator_queue *e = q->elevator;
764 
765 	if (WARN_ON_ONCE(e->uses_mq))
766 		return;
767 
768 	if (e->type->ops.sq.elevator_put_req_fn)
769 		e->type->ops.sq.elevator_put_req_fn(rq);
770 }
771 
772 int elv_may_queue(struct request_queue *q, unsigned int op)
773 {
774 	struct elevator_queue *e = q->elevator;
775 
776 	if (WARN_ON_ONCE(e->uses_mq))
777 		return 0;
778 
779 	if (e->type->ops.sq.elevator_may_queue_fn)
780 		return e->type->ops.sq.elevator_may_queue_fn(q, op);
781 
782 	return ELV_MQUEUE_MAY;
783 }
784 
785 void elv_completed_request(struct request_queue *q, struct request *rq)
786 {
787 	struct elevator_queue *e = q->elevator;
788 
789 	if (WARN_ON_ONCE(e->uses_mq))
790 		return;
791 
792 	/*
793 	 * request is released from the driver, io must be done
794 	 */
795 	if (blk_account_rq(rq)) {
796 		q->in_flight[rq_is_sync(rq)]--;
797 		if ((rq->rq_flags & RQF_SORTED) &&
798 		    e->type->ops.sq.elevator_completed_req_fn)
799 			e->type->ops.sq.elevator_completed_req_fn(q, rq);
800 	}
801 }
802 
803 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
804 
805 static ssize_t
806 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
807 {
808 	struct elv_fs_entry *entry = to_elv(attr);
809 	struct elevator_queue *e;
810 	ssize_t error;
811 
812 	if (!entry->show)
813 		return -EIO;
814 
815 	e = container_of(kobj, struct elevator_queue, kobj);
816 	mutex_lock(&e->sysfs_lock);
817 	error = e->type ? entry->show(e, page) : -ENOENT;
818 	mutex_unlock(&e->sysfs_lock);
819 	return error;
820 }
821 
822 static ssize_t
823 elv_attr_store(struct kobject *kobj, struct attribute *attr,
824 	       const char *page, size_t length)
825 {
826 	struct elv_fs_entry *entry = to_elv(attr);
827 	struct elevator_queue *e;
828 	ssize_t error;
829 
830 	if (!entry->store)
831 		return -EIO;
832 
833 	e = container_of(kobj, struct elevator_queue, kobj);
834 	mutex_lock(&e->sysfs_lock);
835 	error = e->type ? entry->store(e, page, length) : -ENOENT;
836 	mutex_unlock(&e->sysfs_lock);
837 	return error;
838 }
839 
840 static const struct sysfs_ops elv_sysfs_ops = {
841 	.show	= elv_attr_show,
842 	.store	= elv_attr_store,
843 };
844 
845 static struct kobj_type elv_ktype = {
846 	.sysfs_ops	= &elv_sysfs_ops,
847 	.release	= elevator_release,
848 };
849 
850 int elv_register_queue(struct request_queue *q)
851 {
852 	struct elevator_queue *e = q->elevator;
853 	int error;
854 
855 	error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
856 	if (!error) {
857 		struct elv_fs_entry *attr = e->type->elevator_attrs;
858 		if (attr) {
859 			while (attr->attr.name) {
860 				if (sysfs_create_file(&e->kobj, &attr->attr))
861 					break;
862 				attr++;
863 			}
864 		}
865 		kobject_uevent(&e->kobj, KOBJ_ADD);
866 		e->registered = 1;
867 		if (!e->uses_mq && e->type->ops.sq.elevator_registered_fn)
868 			e->type->ops.sq.elevator_registered_fn(q);
869 	}
870 	return error;
871 }
872 EXPORT_SYMBOL(elv_register_queue);
873 
874 void elv_unregister_queue(struct request_queue *q)
875 {
876 	if (q) {
877 		struct elevator_queue *e = q->elevator;
878 
879 		kobject_uevent(&e->kobj, KOBJ_REMOVE);
880 		kobject_del(&e->kobj);
881 		e->registered = 0;
882 		/* Re-enable throttling in case elevator disabled it */
883 		wbt_enable_default(q);
884 	}
885 }
886 EXPORT_SYMBOL(elv_unregister_queue);
887 
888 int elv_register(struct elevator_type *e)
889 {
890 	char *def = "";
891 
892 	/* create icq_cache if requested */
893 	if (e->icq_size) {
894 		if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
895 		    WARN_ON(e->icq_align < __alignof__(struct io_cq)))
896 			return -EINVAL;
897 
898 		snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
899 			 "%s_io_cq", e->elevator_name);
900 		e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
901 						 e->icq_align, 0, NULL);
902 		if (!e->icq_cache)
903 			return -ENOMEM;
904 	}
905 
906 	/* register, don't allow duplicate names */
907 	spin_lock(&elv_list_lock);
908 	if (elevator_find(e->elevator_name)) {
909 		spin_unlock(&elv_list_lock);
910 		if (e->icq_cache)
911 			kmem_cache_destroy(e->icq_cache);
912 		return -EBUSY;
913 	}
914 	list_add_tail(&e->list, &elv_list);
915 	spin_unlock(&elv_list_lock);
916 
917 	/* print pretty message */
918 	if (!strcmp(e->elevator_name, chosen_elevator) ||
919 			(!*chosen_elevator &&
920 			 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
921 				def = " (default)";
922 
923 	printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
924 								def);
925 	return 0;
926 }
927 EXPORT_SYMBOL_GPL(elv_register);
928 
929 void elv_unregister(struct elevator_type *e)
930 {
931 	/* unregister */
932 	spin_lock(&elv_list_lock);
933 	list_del_init(&e->list);
934 	spin_unlock(&elv_list_lock);
935 
936 	/*
937 	 * Destroy icq_cache if it exists.  icq's are RCU managed.  Make
938 	 * sure all RCU operations are complete before proceeding.
939 	 */
940 	if (e->icq_cache) {
941 		rcu_barrier();
942 		kmem_cache_destroy(e->icq_cache);
943 		e->icq_cache = NULL;
944 	}
945 }
946 EXPORT_SYMBOL_GPL(elv_unregister);
947 
948 static int elevator_switch_mq(struct request_queue *q,
949 			      struct elevator_type *new_e)
950 {
951 	int ret;
952 
953 	blk_mq_freeze_queue(q);
954 
955 	if (q->elevator) {
956 		if (q->elevator->registered)
957 			elv_unregister_queue(q);
958 		ioc_clear_queue(q);
959 		elevator_exit(q, q->elevator);
960 	}
961 
962 	ret = blk_mq_init_sched(q, new_e);
963 	if (ret)
964 		goto out;
965 
966 	if (new_e) {
967 		ret = elv_register_queue(q);
968 		if (ret) {
969 			elevator_exit(q, q->elevator);
970 			goto out;
971 		}
972 	}
973 
974 	if (new_e)
975 		blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
976 	else
977 		blk_add_trace_msg(q, "elv switch: none");
978 
979 out:
980 	blk_mq_unfreeze_queue(q);
981 	return ret;
982 }
983 
984 /*
985  * switch to new_e io scheduler. be careful not to introduce deadlocks -
986  * we don't free the old io scheduler, before we have allocated what we
987  * need for the new one. this way we have a chance of going back to the old
988  * one, if the new one fails init for some reason.
989  */
990 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
991 {
992 	struct elevator_queue *old = q->elevator;
993 	bool old_registered = false;
994 	int err;
995 
996 	if (q->mq_ops)
997 		return elevator_switch_mq(q, new_e);
998 
999 	/*
1000 	 * Turn on BYPASS and drain all requests w/ elevator private data.
1001 	 * Block layer doesn't call into a quiesced elevator - all requests
1002 	 * are directly put on the dispatch list without elevator data
1003 	 * using INSERT_BACK.  All requests have SOFTBARRIER set and no
1004 	 * merge happens either.
1005 	 */
1006 	if (old) {
1007 		old_registered = old->registered;
1008 
1009 		blk_queue_bypass_start(q);
1010 
1011 		/* unregister and clear all auxiliary data of the old elevator */
1012 		if (old_registered)
1013 			elv_unregister_queue(q);
1014 
1015 		ioc_clear_queue(q);
1016 	}
1017 
1018 	/* allocate, init and register new elevator */
1019 	err = new_e->ops.sq.elevator_init_fn(q, new_e);
1020 	if (err)
1021 		goto fail_init;
1022 
1023 	err = elv_register_queue(q);
1024 	if (err)
1025 		goto fail_register;
1026 
1027 	/* done, kill the old one and finish */
1028 	if (old) {
1029 		elevator_exit(q, old);
1030 		blk_queue_bypass_end(q);
1031 	}
1032 
1033 	blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
1034 
1035 	return 0;
1036 
1037 fail_register:
1038 	elevator_exit(q, q->elevator);
1039 fail_init:
1040 	/* switch failed, restore and re-register old elevator */
1041 	if (old) {
1042 		q->elevator = old;
1043 		elv_register_queue(q);
1044 		blk_queue_bypass_end(q);
1045 	}
1046 
1047 	return err;
1048 }
1049 
1050 /*
1051  * Switch this queue to the given IO scheduler.
1052  */
1053 static int __elevator_change(struct request_queue *q, const char *name)
1054 {
1055 	char elevator_name[ELV_NAME_MAX];
1056 	struct elevator_type *e;
1057 
1058 	/*
1059 	 * Special case for mq, turn off scheduling
1060 	 */
1061 	if (q->mq_ops && !strncmp(name, "none", 4))
1062 		return elevator_switch(q, NULL);
1063 
1064 	strlcpy(elevator_name, name, sizeof(elevator_name));
1065 	e = elevator_get(strstrip(elevator_name), true);
1066 	if (!e)
1067 		return -EINVAL;
1068 
1069 	if (q->elevator &&
1070 	    !strcmp(elevator_name, q->elevator->type->elevator_name)) {
1071 		elevator_put(e);
1072 		return 0;
1073 	}
1074 
1075 	if (!e->uses_mq && q->mq_ops) {
1076 		elevator_put(e);
1077 		return -EINVAL;
1078 	}
1079 	if (e->uses_mq && !q->mq_ops) {
1080 		elevator_put(e);
1081 		return -EINVAL;
1082 	}
1083 
1084 	return elevator_switch(q, e);
1085 }
1086 
1087 static inline bool elv_support_iosched(struct request_queue *q)
1088 {
1089 	if (q->mq_ops && q->tag_set && (q->tag_set->flags &
1090 				BLK_MQ_F_NO_SCHED))
1091 		return false;
1092 	return true;
1093 }
1094 
1095 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1096 			  size_t count)
1097 {
1098 	int ret;
1099 
1100 	if (!(q->mq_ops || q->request_fn) || !elv_support_iosched(q))
1101 		return count;
1102 
1103 	ret = __elevator_change(q, name);
1104 	if (!ret)
1105 		return count;
1106 
1107 	return ret;
1108 }
1109 
1110 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1111 {
1112 	struct elevator_queue *e = q->elevator;
1113 	struct elevator_type *elv = NULL;
1114 	struct elevator_type *__e;
1115 	int len = 0;
1116 
1117 	if (!blk_queue_stackable(q))
1118 		return sprintf(name, "none\n");
1119 
1120 	if (!q->elevator)
1121 		len += sprintf(name+len, "[none] ");
1122 	else
1123 		elv = e->type;
1124 
1125 	spin_lock(&elv_list_lock);
1126 	list_for_each_entry(__e, &elv_list, list) {
1127 		if (elv && !strcmp(elv->elevator_name, __e->elevator_name)) {
1128 			len += sprintf(name+len, "[%s] ", elv->elevator_name);
1129 			continue;
1130 		}
1131 		if (__e->uses_mq && q->mq_ops && elv_support_iosched(q))
1132 			len += sprintf(name+len, "%s ", __e->elevator_name);
1133 		else if (!__e->uses_mq && !q->mq_ops)
1134 			len += sprintf(name+len, "%s ", __e->elevator_name);
1135 	}
1136 	spin_unlock(&elv_list_lock);
1137 
1138 	if (q->mq_ops && q->elevator)
1139 		len += sprintf(name+len, "none");
1140 
1141 	len += sprintf(len+name, "\n");
1142 	return len;
1143 }
1144 
1145 struct request *elv_rb_former_request(struct request_queue *q,
1146 				      struct request *rq)
1147 {
1148 	struct rb_node *rbprev = rb_prev(&rq->rb_node);
1149 
1150 	if (rbprev)
1151 		return rb_entry_rq(rbprev);
1152 
1153 	return NULL;
1154 }
1155 EXPORT_SYMBOL(elv_rb_former_request);
1156 
1157 struct request *elv_rb_latter_request(struct request_queue *q,
1158 				      struct request *rq)
1159 {
1160 	struct rb_node *rbnext = rb_next(&rq->rb_node);
1161 
1162 	if (rbnext)
1163 		return rb_entry_rq(rbnext);
1164 
1165 	return NULL;
1166 }
1167 EXPORT_SYMBOL(elv_rb_latter_request);
1168