]>
Commit | Line | Data |
---|---|---|
1 | // SPDX-License-Identifier: GPL-2.0 | |
2 | /* | |
3 | * Block device elevator/IO-scheduler. | |
4 | * | |
5 | * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE | |
6 | * | |
7 | * 30042000 Jens Axboe <axboe@kernel.dk> : | |
8 | * | |
9 | * Split the elevator a bit so that it is possible to choose a different | |
10 | * one or even write a new "plug in". There are three pieces: | |
11 | * - elevator_fn, inserts a new request in the queue list | |
12 | * - elevator_merge_fn, decides whether a new buffer can be merged with | |
13 | * an existing request | |
14 | * - elevator_dequeue_fn, called when a request is taken off the active list | |
15 | * | |
16 | * 20082000 Dave Jones <davej@suse.de> : | |
17 | * Removed tests for max-bomb-segments, which was breaking elvtune | |
18 | * when run without -bN | |
19 | * | |
20 | * Jens: | |
21 | * - Rework again to work with bio instead of buffer_heads | |
22 | * - loose bi_dev comparisons, partition handling is right now | |
23 | * - completely modularize elevator setup and teardown | |
24 | * | |
25 | */ | |
26 | #include <linux/kernel.h> | |
27 | #include <linux/fs.h> | |
28 | #include <linux/blkdev.h> | |
29 | #include <linux/elevator.h> | |
30 | #include <linux/bio.h> | |
31 | #include <linux/module.h> | |
32 | #include <linux/slab.h> | |
33 | #include <linux/init.h> | |
34 | #include <linux/compiler.h> | |
35 | #include <linux/blktrace_api.h> | |
36 | #include <linux/hash.h> | |
37 | #include <linux/uaccess.h> | |
38 | #include <linux/pm_runtime.h> | |
39 | #include <linux/blk-cgroup.h> | |
40 | ||
41 | #include <trace/events/block.h> | |
42 | ||
43 | #include "blk.h" | |
44 | #include "blk-mq-sched.h" | |
45 | #include "blk-pm.h" | |
46 | #include "blk-wbt.h" | |
47 | ||
48 | static DEFINE_SPINLOCK(elv_list_lock); | |
49 | static LIST_HEAD(elv_list); | |
50 | ||
51 | /* | |
52 | * Merge hash stuff. | |
53 | */ | |
54 | #define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq)) | |
55 | ||
56 | /* | |
57 | * Query io scheduler to see if the current process issuing bio may be | |
58 | * merged with rq. | |
59 | */ | |
60 | static int elv_iosched_allow_bio_merge(struct request *rq, struct bio *bio) | |
61 | { | |
62 | struct request_queue *q = rq->q; | |
63 | struct elevator_queue *e = q->elevator; | |
64 | ||
65 | if (e->type->ops.allow_merge) | |
66 | return e->type->ops.allow_merge(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 inline bool elv_support_features(unsigned int elv_features, | |
87 | unsigned int required_features) | |
88 | { | |
89 | return (required_features & elv_features) == required_features; | |
90 | } | |
91 | ||
92 | /** | |
93 | * elevator_match - Test an elevator name and features | |
94 | * @e: Scheduler to test | |
95 | * @name: Elevator name to test | |
96 | * @required_features: Features that the elevator must provide | |
97 | * | |
98 | * Return true is the elevator @e name matches @name and if @e provides all the | |
99 | * the feratures spcified by @required_features. | |
100 | */ | |
101 | static bool elevator_match(const struct elevator_type *e, const char *name, | |
102 | unsigned int required_features) | |
103 | { | |
104 | if (!elv_support_features(e->elevator_features, required_features)) | |
105 | return false; | |
106 | if (!strcmp(e->elevator_name, name)) | |
107 | return true; | |
108 | if (e->elevator_alias && !strcmp(e->elevator_alias, name)) | |
109 | return true; | |
110 | ||
111 | return false; | |
112 | } | |
113 | ||
114 | /** | |
115 | * elevator_find - Find an elevator | |
116 | * @name: Name of the elevator to find | |
117 | * @required_features: Features that the elevator must provide | |
118 | * | |
119 | * Return the first registered scheduler with name @name and supporting the | |
120 | * features @required_features and NULL otherwise. | |
121 | */ | |
122 | static struct elevator_type *elevator_find(const char *name, | |
123 | unsigned int required_features) | |
124 | { | |
125 | struct elevator_type *e; | |
126 | ||
127 | list_for_each_entry(e, &elv_list, list) { | |
128 | if (elevator_match(e, name, required_features)) | |
129 | return e; | |
130 | } | |
131 | ||
132 | return NULL; | |
133 | } | |
134 | ||
135 | static void elevator_put(struct elevator_type *e) | |
136 | { | |
137 | module_put(e->elevator_owner); | |
138 | } | |
139 | ||
140 | static struct elevator_type *elevator_get(struct request_queue *q, | |
141 | const char *name, bool try_loading) | |
142 | { | |
143 | struct elevator_type *e; | |
144 | ||
145 | spin_lock(&elv_list_lock); | |
146 | ||
147 | e = elevator_find(name, q->required_elevator_features); | |
148 | if (!e && try_loading) { | |
149 | spin_unlock(&elv_list_lock); | |
150 | request_module("%s-iosched", name); | |
151 | spin_lock(&elv_list_lock); | |
152 | e = elevator_find(name, q->required_elevator_features); | |
153 | } | |
154 | ||
155 | if (e && !try_module_get(e->elevator_owner)) | |
156 | e = NULL; | |
157 | ||
158 | spin_unlock(&elv_list_lock); | |
159 | return e; | |
160 | } | |
161 | ||
162 | static struct kobj_type elv_ktype; | |
163 | ||
164 | struct elevator_queue *elevator_alloc(struct request_queue *q, | |
165 | struct elevator_type *e) | |
166 | { | |
167 | struct elevator_queue *eq; | |
168 | ||
169 | eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node); | |
170 | if (unlikely(!eq)) | |
171 | return NULL; | |
172 | ||
173 | eq->type = e; | |
174 | kobject_init(&eq->kobj, &elv_ktype); | |
175 | mutex_init(&eq->sysfs_lock); | |
176 | hash_init(eq->hash); | |
177 | ||
178 | return eq; | |
179 | } | |
180 | EXPORT_SYMBOL(elevator_alloc); | |
181 | ||
182 | static void elevator_release(struct kobject *kobj) | |
183 | { | |
184 | struct elevator_queue *e; | |
185 | ||
186 | e = container_of(kobj, struct elevator_queue, kobj); | |
187 | elevator_put(e->type); | |
188 | kfree(e); | |
189 | } | |
190 | ||
191 | void __elevator_exit(struct request_queue *q, struct elevator_queue *e) | |
192 | { | |
193 | mutex_lock(&e->sysfs_lock); | |
194 | if (e->type->ops.exit_sched) | |
195 | blk_mq_exit_sched(q, e); | |
196 | mutex_unlock(&e->sysfs_lock); | |
197 | ||
198 | kobject_put(&e->kobj); | |
199 | } | |
200 | ||
201 | static inline void __elv_rqhash_del(struct request *rq) | |
202 | { | |
203 | hash_del(&rq->hash); | |
204 | rq->rq_flags &= ~RQF_HASHED; | |
205 | } | |
206 | ||
207 | void elv_rqhash_del(struct request_queue *q, struct request *rq) | |
208 | { | |
209 | if (ELV_ON_HASH(rq)) | |
210 | __elv_rqhash_del(rq); | |
211 | } | |
212 | EXPORT_SYMBOL_GPL(elv_rqhash_del); | |
213 | ||
214 | void elv_rqhash_add(struct request_queue *q, struct request *rq) | |
215 | { | |
216 | struct elevator_queue *e = q->elevator; | |
217 | ||
218 | BUG_ON(ELV_ON_HASH(rq)); | |
219 | hash_add(e->hash, &rq->hash, rq_hash_key(rq)); | |
220 | rq->rq_flags |= RQF_HASHED; | |
221 | } | |
222 | EXPORT_SYMBOL_GPL(elv_rqhash_add); | |
223 | ||
224 | void elv_rqhash_reposition(struct request_queue *q, struct request *rq) | |
225 | { | |
226 | __elv_rqhash_del(rq); | |
227 | elv_rqhash_add(q, rq); | |
228 | } | |
229 | ||
230 | struct request *elv_rqhash_find(struct request_queue *q, sector_t offset) | |
231 | { | |
232 | struct elevator_queue *e = q->elevator; | |
233 | struct hlist_node *next; | |
234 | struct request *rq; | |
235 | ||
236 | hash_for_each_possible_safe(e->hash, rq, next, hash, offset) { | |
237 | BUG_ON(!ELV_ON_HASH(rq)); | |
238 | ||
239 | if (unlikely(!rq_mergeable(rq))) { | |
240 | __elv_rqhash_del(rq); | |
241 | continue; | |
242 | } | |
243 | ||
244 | if (rq_hash_key(rq) == offset) | |
245 | return rq; | |
246 | } | |
247 | ||
248 | return NULL; | |
249 | } | |
250 | ||
251 | /* | |
252 | * RB-tree support functions for inserting/lookup/removal of requests | |
253 | * in a sorted RB tree. | |
254 | */ | |
255 | void elv_rb_add(struct rb_root *root, struct request *rq) | |
256 | { | |
257 | struct rb_node **p = &root->rb_node; | |
258 | struct rb_node *parent = NULL; | |
259 | struct request *__rq; | |
260 | ||
261 | while (*p) { | |
262 | parent = *p; | |
263 | __rq = rb_entry(parent, struct request, rb_node); | |
264 | ||
265 | if (blk_rq_pos(rq) < blk_rq_pos(__rq)) | |
266 | p = &(*p)->rb_left; | |
267 | else if (blk_rq_pos(rq) >= blk_rq_pos(__rq)) | |
268 | p = &(*p)->rb_right; | |
269 | } | |
270 | ||
271 | rb_link_node(&rq->rb_node, parent, p); | |
272 | rb_insert_color(&rq->rb_node, root); | |
273 | } | |
274 | EXPORT_SYMBOL(elv_rb_add); | |
275 | ||
276 | void elv_rb_del(struct rb_root *root, struct request *rq) | |
277 | { | |
278 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); | |
279 | rb_erase(&rq->rb_node, root); | |
280 | RB_CLEAR_NODE(&rq->rb_node); | |
281 | } | |
282 | EXPORT_SYMBOL(elv_rb_del); | |
283 | ||
284 | struct request *elv_rb_find(struct rb_root *root, sector_t sector) | |
285 | { | |
286 | struct rb_node *n = root->rb_node; | |
287 | struct request *rq; | |
288 | ||
289 | while (n) { | |
290 | rq = rb_entry(n, struct request, rb_node); | |
291 | ||
292 | if (sector < blk_rq_pos(rq)) | |
293 | n = n->rb_left; | |
294 | else if (sector > blk_rq_pos(rq)) | |
295 | n = n->rb_right; | |
296 | else | |
297 | return rq; | |
298 | } | |
299 | ||
300 | return NULL; | |
301 | } | |
302 | EXPORT_SYMBOL(elv_rb_find); | |
303 | ||
304 | enum elv_merge elv_merge(struct request_queue *q, struct request **req, | |
305 | struct bio *bio) | |
306 | { | |
307 | struct elevator_queue *e = q->elevator; | |
308 | struct request *__rq; | |
309 | ||
310 | /* | |
311 | * Levels of merges: | |
312 | * nomerges: No merges at all attempted | |
313 | * noxmerges: Only simple one-hit cache try | |
314 | * merges: All merge tries attempted | |
315 | */ | |
316 | if (blk_queue_nomerges(q) || !bio_mergeable(bio)) | |
317 | return ELEVATOR_NO_MERGE; | |
318 | ||
319 | /* | |
320 | * First try one-hit cache. | |
321 | */ | |
322 | if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) { | |
323 | enum elv_merge ret = blk_try_merge(q->last_merge, bio); | |
324 | ||
325 | if (ret != ELEVATOR_NO_MERGE) { | |
326 | *req = q->last_merge; | |
327 | return ret; | |
328 | } | |
329 | } | |
330 | ||
331 | if (blk_queue_noxmerges(q)) | |
332 | return ELEVATOR_NO_MERGE; | |
333 | ||
334 | /* | |
335 | * See if our hash lookup can find a potential backmerge. | |
336 | */ | |
337 | __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector); | |
338 | if (__rq && elv_bio_merge_ok(__rq, bio)) { | |
339 | *req = __rq; | |
340 | return ELEVATOR_BACK_MERGE; | |
341 | } | |
342 | ||
343 | if (e->type->ops.request_merge) | |
344 | return e->type->ops.request_merge(q, req, bio); | |
345 | ||
346 | return ELEVATOR_NO_MERGE; | |
347 | } | |
348 | ||
349 | /* | |
350 | * Attempt to do an insertion back merge. Only check for the case where | |
351 | * we can append 'rq' to an existing request, so we can throw 'rq' away | |
352 | * afterwards. | |
353 | * | |
354 | * Returns true if we merged, false otherwise | |
355 | */ | |
356 | bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq) | |
357 | { | |
358 | struct request *__rq; | |
359 | bool ret; | |
360 | ||
361 | if (blk_queue_nomerges(q)) | |
362 | return false; | |
363 | ||
364 | /* | |
365 | * First try one-hit cache. | |
366 | */ | |
367 | if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) | |
368 | return true; | |
369 | ||
370 | if (blk_queue_noxmerges(q)) | |
371 | return false; | |
372 | ||
373 | ret = false; | |
374 | /* | |
375 | * See if our hash lookup can find a potential backmerge. | |
376 | */ | |
377 | while (1) { | |
378 | __rq = elv_rqhash_find(q, blk_rq_pos(rq)); | |
379 | if (!__rq || !blk_attempt_req_merge(q, __rq, rq)) | |
380 | break; | |
381 | ||
382 | /* The merged request could be merged with others, try again */ | |
383 | ret = true; | |
384 | rq = __rq; | |
385 | } | |
386 | ||
387 | return ret; | |
388 | } | |
389 | ||
390 | void elv_merged_request(struct request_queue *q, struct request *rq, | |
391 | enum elv_merge type) | |
392 | { | |
393 | struct elevator_queue *e = q->elevator; | |
394 | ||
395 | if (e->type->ops.request_merged) | |
396 | e->type->ops.request_merged(q, rq, type); | |
397 | ||
398 | if (type == ELEVATOR_BACK_MERGE) | |
399 | elv_rqhash_reposition(q, rq); | |
400 | ||
401 | q->last_merge = rq; | |
402 | } | |
403 | ||
404 | void elv_merge_requests(struct request_queue *q, struct request *rq, | |
405 | struct request *next) | |
406 | { | |
407 | struct elevator_queue *e = q->elevator; | |
408 | ||
409 | if (e->type->ops.requests_merged) | |
410 | e->type->ops.requests_merged(q, rq, next); | |
411 | ||
412 | elv_rqhash_reposition(q, rq); | |
413 | q->last_merge = rq; | |
414 | } | |
415 | ||
416 | struct request *elv_latter_request(struct request_queue *q, struct request *rq) | |
417 | { | |
418 | struct elevator_queue *e = q->elevator; | |
419 | ||
420 | if (e->type->ops.next_request) | |
421 | return e->type->ops.next_request(q, rq); | |
422 | ||
423 | return NULL; | |
424 | } | |
425 | ||
426 | struct request *elv_former_request(struct request_queue *q, struct request *rq) | |
427 | { | |
428 | struct elevator_queue *e = q->elevator; | |
429 | ||
430 | if (e->type->ops.former_request) | |
431 | return e->type->ops.former_request(q, rq); | |
432 | ||
433 | return NULL; | |
434 | } | |
435 | ||
436 | #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr) | |
437 | ||
438 | static ssize_t | |
439 | elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page) | |
440 | { | |
441 | struct elv_fs_entry *entry = to_elv(attr); | |
442 | struct elevator_queue *e; | |
443 | ssize_t error; | |
444 | ||
445 | if (!entry->show) | |
446 | return -EIO; | |
447 | ||
448 | e = container_of(kobj, struct elevator_queue, kobj); | |
449 | mutex_lock(&e->sysfs_lock); | |
450 | error = e->type ? entry->show(e, page) : -ENOENT; | |
451 | mutex_unlock(&e->sysfs_lock); | |
452 | return error; | |
453 | } | |
454 | ||
455 | static ssize_t | |
456 | elv_attr_store(struct kobject *kobj, struct attribute *attr, | |
457 | const char *page, size_t length) | |
458 | { | |
459 | struct elv_fs_entry *entry = to_elv(attr); | |
460 | struct elevator_queue *e; | |
461 | ssize_t error; | |
462 | ||
463 | if (!entry->store) | |
464 | return -EIO; | |
465 | ||
466 | e = container_of(kobj, struct elevator_queue, kobj); | |
467 | mutex_lock(&e->sysfs_lock); | |
468 | error = e->type ? entry->store(e, page, length) : -ENOENT; | |
469 | mutex_unlock(&e->sysfs_lock); | |
470 | return error; | |
471 | } | |
472 | ||
473 | static const struct sysfs_ops elv_sysfs_ops = { | |
474 | .show = elv_attr_show, | |
475 | .store = elv_attr_store, | |
476 | }; | |
477 | ||
478 | static struct kobj_type elv_ktype = { | |
479 | .sysfs_ops = &elv_sysfs_ops, | |
480 | .release = elevator_release, | |
481 | }; | |
482 | ||
483 | /* | |
484 | * elv_register_queue is called from either blk_register_queue or | |
485 | * elevator_switch, elevator switch is prevented from being happen | |
486 | * in the two paths, so it is safe to not hold q->sysfs_lock. | |
487 | */ | |
488 | int elv_register_queue(struct request_queue *q, bool uevent) | |
489 | { | |
490 | struct elevator_queue *e = q->elevator; | |
491 | int error; | |
492 | ||
493 | error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched"); | |
494 | if (!error) { | |
495 | struct elv_fs_entry *attr = e->type->elevator_attrs; | |
496 | if (attr) { | |
497 | while (attr->attr.name) { | |
498 | if (sysfs_create_file(&e->kobj, &attr->attr)) | |
499 | break; | |
500 | attr++; | |
501 | } | |
502 | } | |
503 | if (uevent) | |
504 | kobject_uevent(&e->kobj, KOBJ_ADD); | |
505 | ||
506 | e->registered = 1; | |
507 | } | |
508 | return error; | |
509 | } | |
510 | ||
511 | /* | |
512 | * elv_unregister_queue is called from either blk_unregister_queue or | |
513 | * elevator_switch, elevator switch is prevented from being happen | |
514 | * in the two paths, so it is safe to not hold q->sysfs_lock. | |
515 | */ | |
516 | void elv_unregister_queue(struct request_queue *q) | |
517 | { | |
518 | if (q) { | |
519 | struct elevator_queue *e = q->elevator; | |
520 | ||
521 | kobject_uevent(&e->kobj, KOBJ_REMOVE); | |
522 | kobject_del(&e->kobj); | |
523 | ||
524 | e->registered = 0; | |
525 | /* Re-enable throttling in case elevator disabled it */ | |
526 | wbt_enable_default(q); | |
527 | } | |
528 | } | |
529 | ||
530 | int elv_register(struct elevator_type *e) | |
531 | { | |
532 | /* create icq_cache if requested */ | |
533 | if (e->icq_size) { | |
534 | if (WARN_ON(e->icq_size < sizeof(struct io_cq)) || | |
535 | WARN_ON(e->icq_align < __alignof__(struct io_cq))) | |
536 | return -EINVAL; | |
537 | ||
538 | snprintf(e->icq_cache_name, sizeof(e->icq_cache_name), | |
539 | "%s_io_cq", e->elevator_name); | |
540 | e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size, | |
541 | e->icq_align, 0, NULL); | |
542 | if (!e->icq_cache) | |
543 | return -ENOMEM; | |
544 | } | |
545 | ||
546 | /* register, don't allow duplicate names */ | |
547 | spin_lock(&elv_list_lock); | |
548 | if (elevator_find(e->elevator_name, 0)) { | |
549 | spin_unlock(&elv_list_lock); | |
550 | kmem_cache_destroy(e->icq_cache); | |
551 | return -EBUSY; | |
552 | } | |
553 | list_add_tail(&e->list, &elv_list); | |
554 | spin_unlock(&elv_list_lock); | |
555 | ||
556 | printk(KERN_INFO "io scheduler %s registered\n", e->elevator_name); | |
557 | ||
558 | return 0; | |
559 | } | |
560 | EXPORT_SYMBOL_GPL(elv_register); | |
561 | ||
562 | void elv_unregister(struct elevator_type *e) | |
563 | { | |
564 | /* unregister */ | |
565 | spin_lock(&elv_list_lock); | |
566 | list_del_init(&e->list); | |
567 | spin_unlock(&elv_list_lock); | |
568 | ||
569 | /* | |
570 | * Destroy icq_cache if it exists. icq's are RCU managed. Make | |
571 | * sure all RCU operations are complete before proceeding. | |
572 | */ | |
573 | if (e->icq_cache) { | |
574 | rcu_barrier(); | |
575 | kmem_cache_destroy(e->icq_cache); | |
576 | e->icq_cache = NULL; | |
577 | } | |
578 | } | |
579 | EXPORT_SYMBOL_GPL(elv_unregister); | |
580 | ||
581 | int elevator_switch_mq(struct request_queue *q, | |
582 | struct elevator_type *new_e) | |
583 | { | |
584 | int ret; | |
585 | ||
586 | lockdep_assert_held(&q->sysfs_lock); | |
587 | ||
588 | if (q->elevator) { | |
589 | if (q->elevator->registered) | |
590 | elv_unregister_queue(q); | |
591 | ||
592 | ioc_clear_queue(q); | |
593 | elevator_exit(q, q->elevator); | |
594 | } | |
595 | ||
596 | ret = blk_mq_init_sched(q, new_e); | |
597 | if (ret) | |
598 | goto out; | |
599 | ||
600 | if (new_e) { | |
601 | ret = elv_register_queue(q, true); | |
602 | if (ret) { | |
603 | elevator_exit(q, q->elevator); | |
604 | goto out; | |
605 | } | |
606 | } | |
607 | ||
608 | if (new_e) | |
609 | blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name); | |
610 | else | |
611 | blk_add_trace_msg(q, "elv switch: none"); | |
612 | ||
613 | out: | |
614 | return ret; | |
615 | } | |
616 | ||
617 | static inline bool elv_support_iosched(struct request_queue *q) | |
618 | { | |
619 | if (!q->mq_ops || | |
620 | (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED))) | |
621 | return false; | |
622 | return true; | |
623 | } | |
624 | ||
625 | /* | |
626 | * For single queue devices, default to using mq-deadline. If we have multiple | |
627 | * queues or mq-deadline is not available, default to "none". | |
628 | */ | |
629 | static struct elevator_type *elevator_get_default(struct request_queue *q) | |
630 | { | |
631 | if (q->nr_hw_queues != 1) | |
632 | return NULL; | |
633 | ||
634 | return elevator_get(q, "mq-deadline", false); | |
635 | } | |
636 | ||
637 | /* | |
638 | * Get the first elevator providing the features required by the request queue. | |
639 | * Default to "none" if no matching elevator is found. | |
640 | */ | |
641 | static struct elevator_type *elevator_get_by_features(struct request_queue *q) | |
642 | { | |
643 | struct elevator_type *e, *found = NULL; | |
644 | ||
645 | spin_lock(&elv_list_lock); | |
646 | ||
647 | list_for_each_entry(e, &elv_list, list) { | |
648 | if (elv_support_features(e->elevator_features, | |
649 | q->required_elevator_features)) { | |
650 | found = e; | |
651 | break; | |
652 | } | |
653 | } | |
654 | ||
655 | if (found && !try_module_get(found->elevator_owner)) | |
656 | found = NULL; | |
657 | ||
658 | spin_unlock(&elv_list_lock); | |
659 | return found; | |
660 | } | |
661 | ||
662 | /* | |
663 | * For a device queue that has no required features, use the default elevator | |
664 | * settings. Otherwise, use the first elevator available matching the required | |
665 | * features. If no suitable elevator is find or if the chosen elevator | |
666 | * initialization fails, fall back to the "none" elevator (no elevator). | |
667 | */ | |
668 | void elevator_init_mq(struct request_queue *q) | |
669 | { | |
670 | struct elevator_type *e; | |
671 | int err; | |
672 | ||
673 | if (!elv_support_iosched(q)) | |
674 | return; | |
675 | ||
676 | WARN_ON_ONCE(test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags)); | |
677 | ||
678 | if (unlikely(q->elevator)) | |
679 | return; | |
680 | ||
681 | if (!q->required_elevator_features) | |
682 | e = elevator_get_default(q); | |
683 | else | |
684 | e = elevator_get_by_features(q); | |
685 | if (!e) | |
686 | return; | |
687 | ||
688 | blk_mq_freeze_queue(q); | |
689 | blk_mq_quiesce_queue(q); | |
690 | ||
691 | err = blk_mq_init_sched(q, e); | |
692 | ||
693 | blk_mq_unquiesce_queue(q); | |
694 | blk_mq_unfreeze_queue(q); | |
695 | ||
696 | if (err) { | |
697 | pr_warn("\"%s\" elevator initialization failed, " | |
698 | "falling back to \"none\"\n", e->elevator_name); | |
699 | elevator_put(e); | |
700 | } | |
701 | } | |
702 | ||
703 | ||
704 | /* | |
705 | * switch to new_e io scheduler. be careful not to introduce deadlocks - | |
706 | * we don't free the old io scheduler, before we have allocated what we | |
707 | * need for the new one. this way we have a chance of going back to the old | |
708 | * one, if the new one fails init for some reason. | |
709 | */ | |
710 | static int elevator_switch(struct request_queue *q, struct elevator_type *new_e) | |
711 | { | |
712 | int err; | |
713 | ||
714 | lockdep_assert_held(&q->sysfs_lock); | |
715 | ||
716 | blk_mq_freeze_queue(q); | |
717 | blk_mq_quiesce_queue(q); | |
718 | ||
719 | err = elevator_switch_mq(q, new_e); | |
720 | ||
721 | blk_mq_unquiesce_queue(q); | |
722 | blk_mq_unfreeze_queue(q); | |
723 | ||
724 | return err; | |
725 | } | |
726 | ||
727 | /* | |
728 | * Switch this queue to the given IO scheduler. | |
729 | */ | |
730 | static int __elevator_change(struct request_queue *q, const char *name) | |
731 | { | |
732 | char elevator_name[ELV_NAME_MAX]; | |
733 | struct elevator_type *e; | |
734 | ||
735 | /* Make sure queue is not in the middle of being removed */ | |
736 | if (!blk_queue_registered(q)) | |
737 | return -ENOENT; | |
738 | ||
739 | /* | |
740 | * Special case for mq, turn off scheduling | |
741 | */ | |
742 | if (!strncmp(name, "none", 4)) { | |
743 | if (!q->elevator) | |
744 | return 0; | |
745 | return elevator_switch(q, NULL); | |
746 | } | |
747 | ||
748 | strlcpy(elevator_name, name, sizeof(elevator_name)); | |
749 | e = elevator_get(q, strstrip(elevator_name), true); | |
750 | if (!e) | |
751 | return -EINVAL; | |
752 | ||
753 | if (q->elevator && | |
754 | elevator_match(q->elevator->type, elevator_name, 0)) { | |
755 | elevator_put(e); | |
756 | return 0; | |
757 | } | |
758 | ||
759 | return elevator_switch(q, e); | |
760 | } | |
761 | ||
762 | ssize_t elv_iosched_store(struct request_queue *q, const char *name, | |
763 | size_t count) | |
764 | { | |
765 | int ret; | |
766 | ||
767 | if (!queue_is_mq(q) || !elv_support_iosched(q)) | |
768 | return count; | |
769 | ||
770 | ret = __elevator_change(q, name); | |
771 | if (!ret) | |
772 | return count; | |
773 | ||
774 | return ret; | |
775 | } | |
776 | ||
777 | ssize_t elv_iosched_show(struct request_queue *q, char *name) | |
778 | { | |
779 | struct elevator_queue *e = q->elevator; | |
780 | struct elevator_type *elv = NULL; | |
781 | struct elevator_type *__e; | |
782 | int len = 0; | |
783 | ||
784 | if (!queue_is_mq(q)) | |
785 | return sprintf(name, "none\n"); | |
786 | ||
787 | if (!q->elevator) | |
788 | len += sprintf(name+len, "[none] "); | |
789 | else | |
790 | elv = e->type; | |
791 | ||
792 | spin_lock(&elv_list_lock); | |
793 | list_for_each_entry(__e, &elv_list, list) { | |
794 | if (elv && elevator_match(elv, __e->elevator_name, 0)) { | |
795 | len += sprintf(name+len, "[%s] ", elv->elevator_name); | |
796 | continue; | |
797 | } | |
798 | if (elv_support_iosched(q) && | |
799 | elevator_match(__e, __e->elevator_name, | |
800 | q->required_elevator_features)) | |
801 | len += sprintf(name+len, "%s ", __e->elevator_name); | |
802 | } | |
803 | spin_unlock(&elv_list_lock); | |
804 | ||
805 | if (q->elevator) | |
806 | len += sprintf(name+len, "none"); | |
807 | ||
808 | len += sprintf(len+name, "\n"); | |
809 | return len; | |
810 | } | |
811 | ||
812 | struct request *elv_rb_former_request(struct request_queue *q, | |
813 | struct request *rq) | |
814 | { | |
815 | struct rb_node *rbprev = rb_prev(&rq->rb_node); | |
816 | ||
817 | if (rbprev) | |
818 | return rb_entry_rq(rbprev); | |
819 | ||
820 | return NULL; | |
821 | } | |
822 | EXPORT_SYMBOL(elv_rb_former_request); | |
823 | ||
824 | struct request *elv_rb_latter_request(struct request_queue *q, | |
825 | struct request *rq) | |
826 | { | |
827 | struct rb_node *rbnext = rb_next(&rq->rb_node); | |
828 | ||
829 | if (rbnext) | |
830 | return rb_entry_rq(rbnext); | |
831 | ||
832 | return NULL; | |
833 | } | |
834 | EXPORT_SYMBOL(elv_rb_latter_request); | |
835 | ||
836 | static int __init elevator_setup(char *str) | |
837 | { | |
838 | pr_warn("Kernel parameter elevator= does not have any effect anymore.\n" | |
839 | "Please use sysfs to set IO scheduler for individual devices.\n"); | |
840 | return 1; | |
841 | } | |
842 | ||
843 | __setup("elevator=", elevator_setup); |