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