Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 | 2 | /* |
1da177e4 LT |
3 | * Block device elevator/IO-scheduler. |
4 | * | |
5 | * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE | |
6 | * | |
0fe23479 | 7 | * 30042000 Jens Axboe <axboe@kernel.dk> : |
1da177e4 LT |
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> | |
1da177e4 | 29 | #include <linux/bio.h> |
1da177e4 LT |
30 | #include <linux/module.h> |
31 | #include <linux/slab.h> | |
32 | #include <linux/init.h> | |
33 | #include <linux/compiler.h> | |
2056a782 | 34 | #include <linux/blktrace_api.h> |
9817064b | 35 | #include <linux/hash.h> |
0835da67 | 36 | #include <linux/uaccess.h> |
c8158819 | 37 | #include <linux/pm_runtime.h> |
1da177e4 | 38 | |
55782138 LZ |
39 | #include <trace/events/block.h> |
40 | ||
2e9bc346 | 41 | #include "elevator.h" |
242f9dcb | 42 | #include "blk.h" |
bd166ef1 | 43 | #include "blk-mq-sched.h" |
bca6b067 | 44 | #include "blk-pm.h" |
8330cdb0 | 45 | #include "blk-wbt.h" |
672fdcf0 | 46 | #include "blk-cgroup.h" |
242f9dcb | 47 | |
1da177e4 LT |
48 | static DEFINE_SPINLOCK(elv_list_lock); |
49 | static LIST_HEAD(elv_list); | |
50 | ||
9817064b JA |
51 | /* |
52 | * Merge hash stuff. | |
53 | */ | |
83096ebf | 54 | #define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq)) |
9817064b | 55 | |
da775265 JA |
56 | /* |
57 | * Query io scheduler to see if the current process issuing bio may be | |
58 | * merged with rq. | |
59 | */ | |
8d283ee6 | 60 | static bool elv_iosched_allow_bio_merge(struct request *rq, struct bio *bio) |
da775265 | 61 | { |
165125e1 | 62 | struct request_queue *q = rq->q; |
b374d18a | 63 | struct elevator_queue *e = q->elevator; |
da775265 | 64 | |
f9cd4bfe JA |
65 | if (e->type->ops.allow_merge) |
66 | return e->type->ops.allow_merge(q, rq, bio); | |
da775265 | 67 | |
8d283ee6 | 68 | return true; |
da775265 JA |
69 | } |
70 | ||
1da177e4 LT |
71 | /* |
72 | * can we safely merge with this request? | |
73 | */ | |
72ef799b | 74 | bool elv_bio_merge_ok(struct request *rq, struct bio *bio) |
1da177e4 | 75 | { |
050c8ea8 | 76 | if (!blk_rq_merge_ok(rq, bio)) |
72ef799b | 77 | return false; |
7ba1ba12 | 78 | |
72ef799b TE |
79 | if (!elv_iosched_allow_bio_merge(rq, bio)) |
80 | return false; | |
1da177e4 | 81 | |
72ef799b | 82 | return true; |
1da177e4 | 83 | } |
72ef799b | 84 | EXPORT_SYMBOL(elv_bio_merge_ok); |
1da177e4 | 85 | |
68c43f13 | 86 | /** |
f69b5e8f | 87 | * elevator_match - Check whether @e's name or alias matches @name |
68c43f13 DLM |
88 | * @e: Scheduler to test |
89 | * @name: Elevator name to test | |
68c43f13 | 90 | * |
f69b5e8f | 91 | * Return true if the elevator @e's name or alias matches @name. |
68c43f13 | 92 | */ |
ffb86425 | 93 | static bool elevator_match(const struct elevator_type *e, const char *name) |
68c43f13 | 94 | { |
ffb86425 CH |
95 | return !strcmp(e->elevator_name, name) || |
96 | (e->elevator_alias && !strcmp(e->elevator_alias, name)); | |
8ac0d9a8 JA |
97 | } |
98 | ||
ffb86425 | 99 | static struct elevator_type *__elevator_find(const char *name) |
1da177e4 | 100 | { |
a22b169d | 101 | struct elevator_type *e; |
1da177e4 | 102 | |
ffb86425 CH |
103 | list_for_each_entry(e, &elv_list, list) |
104 | if (elevator_match(e, name)) | |
a22b169d | 105 | return e; |
a22b169d | 106 | return NULL; |
1da177e4 LT |
107 | } |
108 | ||
81eaca44 CH |
109 | static struct elevator_type *elevator_find_get(struct request_queue *q, |
110 | const char *name) | |
1da177e4 | 111 | { |
2824bc93 | 112 | struct elevator_type *e; |
1da177e4 | 113 | |
2a12dcd7 | 114 | spin_lock(&elv_list_lock); |
ffb86425 | 115 | e = __elevator_find(name); |
e4eb37cc | 116 | if (e && (!elevator_tryget(e))) |
2824bc93 | 117 | e = NULL; |
2a12dcd7 | 118 | spin_unlock(&elv_list_lock); |
1da177e4 LT |
119 | return e; |
120 | } | |
121 | ||
5f622417 | 122 | static const struct kobj_type elv_ktype; |
3d1ab40f | 123 | |
d50235b7 | 124 | struct elevator_queue *elevator_alloc(struct request_queue *q, |
165125e1 | 125 | struct elevator_type *e) |
3d1ab40f | 126 | { |
b374d18a | 127 | struct elevator_queue *eq; |
9817064b | 128 | |
c1b511eb | 129 | eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node); |
9817064b | 130 | if (unlikely(!eq)) |
8406a4d5 | 131 | return NULL; |
9817064b | 132 | |
8ed40ee3 | 133 | __elevator_get(e); |
22f746e2 | 134 | eq->type = e; |
f9cb074b | 135 | kobject_init(&eq->kobj, &elv_ktype); |
9817064b | 136 | mutex_init(&eq->sysfs_lock); |
242d98f0 | 137 | hash_init(eq->hash); |
9817064b | 138 | |
3d1ab40f AV |
139 | return eq; |
140 | } | |
d50235b7 | 141 | EXPORT_SYMBOL(elevator_alloc); |
3d1ab40f AV |
142 | |
143 | static void elevator_release(struct kobject *kobj) | |
144 | { | |
b374d18a | 145 | struct elevator_queue *e; |
9817064b | 146 | |
b374d18a | 147 | e = container_of(kobj, struct elevator_queue, kobj); |
22f746e2 | 148 | elevator_put(e->type); |
3d1ab40f AV |
149 | kfree(e); |
150 | } | |
151 | ||
0c6cb3a2 | 152 | void elevator_exit(struct request_queue *q) |
1da177e4 | 153 | { |
0c6cb3a2 CH |
154 | struct elevator_queue *e = q->elevator; |
155 | ||
28883074 CH |
156 | ioc_clear_queue(q); |
157 | blk_mq_sched_free_rqs(q); | |
158 | ||
3d1ab40f | 159 | mutex_lock(&e->sysfs_lock); |
dd1c372d | 160 | blk_mq_exit_sched(q, e); |
3d1ab40f | 161 | mutex_unlock(&e->sysfs_lock); |
1da177e4 | 162 | |
3d1ab40f | 163 | kobject_put(&e->kobj); |
1da177e4 | 164 | } |
2e662b65 | 165 | |
9817064b JA |
166 | static inline void __elv_rqhash_del(struct request *rq) |
167 | { | |
242d98f0 | 168 | hash_del(&rq->hash); |
e8064021 | 169 | rq->rq_flags &= ~RQF_HASHED; |
9817064b JA |
170 | } |
171 | ||
70b3ea05 | 172 | void elv_rqhash_del(struct request_queue *q, struct request *rq) |
9817064b JA |
173 | { |
174 | if (ELV_ON_HASH(rq)) | |
175 | __elv_rqhash_del(rq); | |
176 | } | |
bd166ef1 | 177 | EXPORT_SYMBOL_GPL(elv_rqhash_del); |
9817064b | 178 | |
70b3ea05 | 179 | void elv_rqhash_add(struct request_queue *q, struct request *rq) |
9817064b | 180 | { |
b374d18a | 181 | struct elevator_queue *e = q->elevator; |
9817064b JA |
182 | |
183 | BUG_ON(ELV_ON_HASH(rq)); | |
242d98f0 | 184 | hash_add(e->hash, &rq->hash, rq_hash_key(rq)); |
e8064021 | 185 | rq->rq_flags |= RQF_HASHED; |
9817064b | 186 | } |
bd166ef1 | 187 | EXPORT_SYMBOL_GPL(elv_rqhash_add); |
9817064b | 188 | |
70b3ea05 | 189 | void elv_rqhash_reposition(struct request_queue *q, struct request *rq) |
9817064b JA |
190 | { |
191 | __elv_rqhash_del(rq); | |
192 | elv_rqhash_add(q, rq); | |
193 | } | |
194 | ||
70b3ea05 | 195 | struct request *elv_rqhash_find(struct request_queue *q, sector_t offset) |
9817064b | 196 | { |
b374d18a | 197 | struct elevator_queue *e = q->elevator; |
b67bfe0d | 198 | struct hlist_node *next; |
9817064b JA |
199 | struct request *rq; |
200 | ||
ee89f812 | 201 | hash_for_each_possible_safe(e->hash, rq, next, hash, offset) { |
9817064b JA |
202 | BUG_ON(!ELV_ON_HASH(rq)); |
203 | ||
204 | if (unlikely(!rq_mergeable(rq))) { | |
205 | __elv_rqhash_del(rq); | |
206 | continue; | |
207 | } | |
208 | ||
209 | if (rq_hash_key(rq) == offset) | |
210 | return rq; | |
211 | } | |
212 | ||
213 | return NULL; | |
214 | } | |
215 | ||
2e662b65 JA |
216 | /* |
217 | * RB-tree support functions for inserting/lookup/removal of requests | |
218 | * in a sorted RB tree. | |
219 | */ | |
796d5116 | 220 | void elv_rb_add(struct rb_root *root, struct request *rq) |
2e662b65 JA |
221 | { |
222 | struct rb_node **p = &root->rb_node; | |
223 | struct rb_node *parent = NULL; | |
224 | struct request *__rq; | |
225 | ||
226 | while (*p) { | |
227 | parent = *p; | |
228 | __rq = rb_entry(parent, struct request, rb_node); | |
229 | ||
83096ebf | 230 | if (blk_rq_pos(rq) < blk_rq_pos(__rq)) |
2e662b65 | 231 | p = &(*p)->rb_left; |
796d5116 | 232 | else if (blk_rq_pos(rq) >= blk_rq_pos(__rq)) |
2e662b65 | 233 | p = &(*p)->rb_right; |
2e662b65 JA |
234 | } |
235 | ||
236 | rb_link_node(&rq->rb_node, parent, p); | |
237 | rb_insert_color(&rq->rb_node, root); | |
2e662b65 | 238 | } |
2e662b65 JA |
239 | EXPORT_SYMBOL(elv_rb_add); |
240 | ||
241 | void elv_rb_del(struct rb_root *root, struct request *rq) | |
242 | { | |
243 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); | |
244 | rb_erase(&rq->rb_node, root); | |
245 | RB_CLEAR_NODE(&rq->rb_node); | |
246 | } | |
2e662b65 JA |
247 | EXPORT_SYMBOL(elv_rb_del); |
248 | ||
249 | struct request *elv_rb_find(struct rb_root *root, sector_t sector) | |
250 | { | |
251 | struct rb_node *n = root->rb_node; | |
252 | struct request *rq; | |
253 | ||
254 | while (n) { | |
255 | rq = rb_entry(n, struct request, rb_node); | |
256 | ||
83096ebf | 257 | if (sector < blk_rq_pos(rq)) |
2e662b65 | 258 | n = n->rb_left; |
83096ebf | 259 | else if (sector > blk_rq_pos(rq)) |
2e662b65 JA |
260 | n = n->rb_right; |
261 | else | |
262 | return rq; | |
263 | } | |
264 | ||
265 | return NULL; | |
266 | } | |
2e662b65 JA |
267 | EXPORT_SYMBOL(elv_rb_find); |
268 | ||
34fe7c05 CH |
269 | enum elv_merge elv_merge(struct request_queue *q, struct request **req, |
270 | struct bio *bio) | |
1da177e4 | 271 | { |
b374d18a | 272 | struct elevator_queue *e = q->elevator; |
9817064b | 273 | struct request *__rq; |
06b86245 | 274 | |
488991e2 AB |
275 | /* |
276 | * Levels of merges: | |
277 | * nomerges: No merges at all attempted | |
278 | * noxmerges: Only simple one-hit cache try | |
279 | * merges: All merge tries attempted | |
280 | */ | |
7460d389 | 281 | if (blk_queue_nomerges(q) || !bio_mergeable(bio)) |
488991e2 AB |
282 | return ELEVATOR_NO_MERGE; |
283 | ||
9817064b JA |
284 | /* |
285 | * First try one-hit cache. | |
286 | */ | |
72ef799b | 287 | if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) { |
34fe7c05 CH |
288 | enum elv_merge ret = blk_try_merge(q->last_merge, bio); |
289 | ||
06b86245 TH |
290 | if (ret != ELEVATOR_NO_MERGE) { |
291 | *req = q->last_merge; | |
292 | return ret; | |
293 | } | |
294 | } | |
1da177e4 | 295 | |
488991e2 | 296 | if (blk_queue_noxmerges(q)) |
ac9fafa1 AB |
297 | return ELEVATOR_NO_MERGE; |
298 | ||
9817064b JA |
299 | /* |
300 | * See if our hash lookup can find a potential backmerge. | |
301 | */ | |
4f024f37 | 302 | __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector); |
72ef799b | 303 | if (__rq && elv_bio_merge_ok(__rq, bio)) { |
9817064b | 304 | *req = __rq; |
866663b7 ML |
305 | |
306 | if (blk_discard_mergable(__rq)) | |
307 | return ELEVATOR_DISCARD_MERGE; | |
9817064b JA |
308 | return ELEVATOR_BACK_MERGE; |
309 | } | |
310 | ||
f9cd4bfe JA |
311 | if (e->type->ops.request_merge) |
312 | return e->type->ops.request_merge(q, req, bio); | |
1da177e4 LT |
313 | |
314 | return ELEVATOR_NO_MERGE; | |
315 | } | |
316 | ||
5e84ea3a JA |
317 | /* |
318 | * Attempt to do an insertion back merge. Only check for the case where | |
319 | * we can append 'rq' to an existing request, so we can throw 'rq' away | |
320 | * afterwards. | |
321 | * | |
fd2ef39c JK |
322 | * Returns true if we merged, false otherwise. 'free' will contain all |
323 | * requests that need to be freed. | |
5e84ea3a | 324 | */ |
fd2ef39c JK |
325 | bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq, |
326 | struct list_head *free) | |
5e84ea3a JA |
327 | { |
328 | struct request *__rq; | |
bee0393c | 329 | bool ret; |
5e84ea3a JA |
330 | |
331 | if (blk_queue_nomerges(q)) | |
332 | return false; | |
333 | ||
334 | /* | |
335 | * First try one-hit cache. | |
336 | */ | |
fd2ef39c JK |
337 | if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) { |
338 | list_add(&rq->queuelist, free); | |
5e84ea3a | 339 | return true; |
fd2ef39c | 340 | } |
5e84ea3a JA |
341 | |
342 | if (blk_queue_noxmerges(q)) | |
343 | return false; | |
344 | ||
bee0393c | 345 | ret = false; |
5e84ea3a JA |
346 | /* |
347 | * See if our hash lookup can find a potential backmerge. | |
348 | */ | |
bee0393c SL |
349 | while (1) { |
350 | __rq = elv_rqhash_find(q, blk_rq_pos(rq)); | |
351 | if (!__rq || !blk_attempt_req_merge(q, __rq, rq)) | |
352 | break; | |
353 | ||
fd2ef39c | 354 | list_add(&rq->queuelist, free); |
bee0393c SL |
355 | /* The merged request could be merged with others, try again */ |
356 | ret = true; | |
357 | rq = __rq; | |
358 | } | |
27419322 | 359 | |
bee0393c | 360 | return ret; |
5e84ea3a JA |
361 | } |
362 | ||
34fe7c05 CH |
363 | void elv_merged_request(struct request_queue *q, struct request *rq, |
364 | enum elv_merge type) | |
1da177e4 | 365 | { |
b374d18a | 366 | struct elevator_queue *e = q->elevator; |
1da177e4 | 367 | |
f9cd4bfe JA |
368 | if (e->type->ops.request_merged) |
369 | e->type->ops.request_merged(q, rq, type); | |
06b86245 | 370 | |
2e662b65 JA |
371 | if (type == ELEVATOR_BACK_MERGE) |
372 | elv_rqhash_reposition(q, rq); | |
9817064b | 373 | |
06b86245 | 374 | q->last_merge = rq; |
1da177e4 LT |
375 | } |
376 | ||
165125e1 | 377 | void elv_merge_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
378 | struct request *next) |
379 | { | |
b374d18a | 380 | struct elevator_queue *e = q->elevator; |
bd166ef1 | 381 | |
f9cd4bfe JA |
382 | if (e->type->ops.requests_merged) |
383 | e->type->ops.requests_merged(q, rq, next); | |
06b86245 | 384 | |
9817064b | 385 | elv_rqhash_reposition(q, rq); |
06b86245 | 386 | q->last_merge = rq; |
1da177e4 LT |
387 | } |
388 | ||
165125e1 | 389 | struct request *elv_latter_request(struct request_queue *q, struct request *rq) |
1da177e4 | 390 | { |
b374d18a | 391 | struct elevator_queue *e = q->elevator; |
1da177e4 | 392 | |
f9cd4bfe JA |
393 | if (e->type->ops.next_request) |
394 | return e->type->ops.next_request(q, rq); | |
bd166ef1 | 395 | |
1da177e4 LT |
396 | return NULL; |
397 | } | |
398 | ||
165125e1 | 399 | struct request *elv_former_request(struct request_queue *q, struct request *rq) |
1da177e4 | 400 | { |
b374d18a | 401 | struct elevator_queue *e = q->elevator; |
1da177e4 | 402 | |
f9cd4bfe JA |
403 | if (e->type->ops.former_request) |
404 | return e->type->ops.former_request(q, rq); | |
bd166ef1 | 405 | |
a1ce35fa | 406 | return NULL; |
1da177e4 LT |
407 | } |
408 | ||
3d1ab40f AV |
409 | #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr) |
410 | ||
411 | static ssize_t | |
412 | elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page) | |
1da177e4 | 413 | { |
3d1ab40f | 414 | struct elv_fs_entry *entry = to_elv(attr); |
b374d18a | 415 | struct elevator_queue *e; |
3d1ab40f AV |
416 | ssize_t error; |
417 | ||
418 | if (!entry->show) | |
419 | return -EIO; | |
420 | ||
b374d18a | 421 | e = container_of(kobj, struct elevator_queue, kobj); |
3d1ab40f | 422 | mutex_lock(&e->sysfs_lock); |
22f746e2 | 423 | error = e->type ? entry->show(e, page) : -ENOENT; |
3d1ab40f AV |
424 | mutex_unlock(&e->sysfs_lock); |
425 | return error; | |
426 | } | |
1da177e4 | 427 | |
3d1ab40f AV |
428 | static ssize_t |
429 | elv_attr_store(struct kobject *kobj, struct attribute *attr, | |
430 | const char *page, size_t length) | |
431 | { | |
3d1ab40f | 432 | struct elv_fs_entry *entry = to_elv(attr); |
b374d18a | 433 | struct elevator_queue *e; |
3d1ab40f | 434 | ssize_t error; |
1da177e4 | 435 | |
3d1ab40f AV |
436 | if (!entry->store) |
437 | return -EIO; | |
1da177e4 | 438 | |
b374d18a | 439 | e = container_of(kobj, struct elevator_queue, kobj); |
3d1ab40f | 440 | mutex_lock(&e->sysfs_lock); |
22f746e2 | 441 | error = e->type ? entry->store(e, page, length) : -ENOENT; |
3d1ab40f AV |
442 | mutex_unlock(&e->sysfs_lock); |
443 | return error; | |
444 | } | |
445 | ||
52cf25d0 | 446 | static const struct sysfs_ops elv_sysfs_ops = { |
3d1ab40f AV |
447 | .show = elv_attr_show, |
448 | .store = elv_attr_store, | |
449 | }; | |
450 | ||
5f622417 | 451 | static const struct kobj_type elv_ktype = { |
3d1ab40f AV |
452 | .sysfs_ops = &elv_sysfs_ops, |
453 | .release = elevator_release, | |
454 | }; | |
455 | ||
cecf5d87 | 456 | int elv_register_queue(struct request_queue *q, bool uevent) |
3d1ab40f | 457 | { |
5a5bafdc | 458 | struct elevator_queue *e = q->elevator; |
3d1ab40f AV |
459 | int error; |
460 | ||
f0c6ae09 YY |
461 | lockdep_assert_held(&q->sysfs_lock); |
462 | ||
2bd85221 | 463 | error = kobject_add(&e->kobj, &q->disk->queue_kobj, "iosched"); |
3d1ab40f | 464 | if (!error) { |
22f746e2 | 465 | struct elv_fs_entry *attr = e->type->elevator_attrs; |
3d1ab40f | 466 | if (attr) { |
e572ec7e AV |
467 | while (attr->attr.name) { |
468 | if (sysfs_create_file(&e->kobj, &attr->attr)) | |
3d1ab40f | 469 | break; |
e572ec7e | 470 | attr++; |
3d1ab40f AV |
471 | } |
472 | } | |
cecf5d87 ML |
473 | if (uevent) |
474 | kobject_uevent(&e->kobj, KOBJ_ADD); | |
475 | ||
181d0663 | 476 | set_bit(ELEVATOR_FLAG_REGISTERED, &e->flags); |
3d1ab40f AV |
477 | } |
478 | return error; | |
1da177e4 | 479 | } |
bc1c1169 | 480 | |
1da177e4 LT |
481 | void elv_unregister_queue(struct request_queue *q) |
482 | { | |
f5ec592d EB |
483 | struct elevator_queue *e = q->elevator; |
484 | ||
f0c6ae09 YY |
485 | lockdep_assert_held(&q->sysfs_lock); |
486 | ||
181d0663 | 487 | if (e && test_and_clear_bit(ELEVATOR_FLAG_REGISTERED, &e->flags)) { |
f8fc877d TH |
488 | kobject_uevent(&e->kobj, KOBJ_REMOVE); |
489 | kobject_del(&e->kobj); | |
f8fc877d | 490 | } |
1da177e4 LT |
491 | } |
492 | ||
e567bf71 | 493 | int elv_register(struct elevator_type *e) |
1da177e4 | 494 | { |
e5c0ca13 CZ |
495 | /* finish request is mandatory */ |
496 | if (WARN_ON_ONCE(!e->ops.finish_request)) | |
497 | return -EINVAL; | |
e42cfb1d DLM |
498 | /* insert_requests and dispatch_request are mandatory */ |
499 | if (WARN_ON_ONCE(!e->ops.insert_requests || !e->ops.dispatch_request)) | |
500 | return -EINVAL; | |
501 | ||
3d3c2379 TH |
502 | /* create icq_cache if requested */ |
503 | if (e->icq_size) { | |
504 | if (WARN_ON(e->icq_size < sizeof(struct io_cq)) || | |
505 | WARN_ON(e->icq_align < __alignof__(struct io_cq))) | |
506 | return -EINVAL; | |
507 | ||
508 | snprintf(e->icq_cache_name, sizeof(e->icq_cache_name), | |
509 | "%s_io_cq", e->elevator_name); | |
510 | e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size, | |
511 | e->icq_align, 0, NULL); | |
512 | if (!e->icq_cache) | |
513 | return -ENOMEM; | |
514 | } | |
515 | ||
516 | /* register, don't allow duplicate names */ | |
2a12dcd7 | 517 | spin_lock(&elv_list_lock); |
ffb86425 | 518 | if (__elevator_find(e->elevator_name)) { |
3d3c2379 | 519 | spin_unlock(&elv_list_lock); |
62d2a194 | 520 | kmem_cache_destroy(e->icq_cache); |
3d3c2379 TH |
521 | return -EBUSY; |
522 | } | |
1da177e4 | 523 | list_add_tail(&e->list, &elv_list); |
2a12dcd7 | 524 | spin_unlock(&elv_list_lock); |
1da177e4 | 525 | |
d0b0a81a HT |
526 | printk(KERN_INFO "io scheduler %s registered\n", e->elevator_name); |
527 | ||
3d3c2379 | 528 | return 0; |
1da177e4 LT |
529 | } |
530 | EXPORT_SYMBOL_GPL(elv_register); | |
531 | ||
532 | void elv_unregister(struct elevator_type *e) | |
533 | { | |
3d3c2379 | 534 | /* unregister */ |
2a12dcd7 | 535 | spin_lock(&elv_list_lock); |
1da177e4 | 536 | list_del_init(&e->list); |
2a12dcd7 | 537 | spin_unlock(&elv_list_lock); |
3d3c2379 TH |
538 | |
539 | /* | |
540 | * Destroy icq_cache if it exists. icq's are RCU managed. Make | |
541 | * sure all RCU operations are complete before proceeding. | |
542 | */ | |
543 | if (e->icq_cache) { | |
544 | rcu_barrier(); | |
545 | kmem_cache_destroy(e->icq_cache); | |
546 | e->icq_cache = NULL; | |
547 | } | |
1da177e4 LT |
548 | } |
549 | EXPORT_SYMBOL_GPL(elv_unregister); | |
550 | ||
61db437d DLM |
551 | static inline bool elv_support_iosched(struct request_queue *q) |
552 | { | |
6251b754 | 553 | if (!queue_is_mq(q) || |
7a7c5e71 | 554 | (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED))) |
61db437d DLM |
555 | return false; |
556 | return true; | |
557 | } | |
558 | ||
131d08e1 | 559 | /* |
a0958ba7 DLM |
560 | * For single queue devices, default to using mq-deadline. If we have multiple |
561 | * queues or mq-deadline is not available, default to "none". | |
562 | */ | |
563 | static struct elevator_type *elevator_get_default(struct request_queue *q) | |
564 | { | |
90b71980 BVA |
565 | if (q->tag_set && q->tag_set->flags & BLK_MQ_F_NO_SCHED_BY_DEFAULT) |
566 | return NULL; | |
567 | ||
580dca81 | 568 | if (q->nr_hw_queues != 1 && |
079a2e3e | 569 | !blk_mq_is_shared_tags(q->tag_set->flags)) |
a0958ba7 DLM |
570 | return NULL; |
571 | ||
81eaca44 | 572 | return elevator_find_get(q, "mq-deadline"); |
a0958ba7 DLM |
573 | } |
574 | ||
575 | /* | |
e4eb37cc DLM |
576 | * Use the default elevator settings. If the chosen elevator initialization |
577 | * fails, fall back to the "none" elevator (no elevator). | |
131d08e1 | 578 | */ |
954b4a5c | 579 | void elevator_init_mq(struct request_queue *q) |
131d08e1 CH |
580 | { |
581 | struct elevator_type *e; | |
954b4a5c | 582 | int err; |
131d08e1 | 583 | |
61db437d | 584 | if (!elv_support_iosched(q)) |
954b4a5c | 585 | return; |
61db437d | 586 | |
75e6c00f | 587 | WARN_ON_ONCE(blk_queue_registered(q)); |
c48dac13 | 588 | |
131d08e1 | 589 | if (unlikely(q->elevator)) |
954b4a5c | 590 | return; |
131d08e1 | 591 | |
e4eb37cc | 592 | e = elevator_get_default(q); |
131d08e1 | 593 | if (!e) |
954b4a5c | 594 | return; |
131d08e1 | 595 | |
245a489e ML |
596 | /* |
597 | * We are called before adding disk, when there isn't any FS I/O, | |
598 | * so freezing queue plus canceling dispatch work is enough to | |
599 | * drain any dispatch activities originated from passthrough | |
600 | * requests, then no need to quiesce queue which may add long boot | |
601 | * latency, especially when lots of disks are involved. | |
602 | */ | |
737eb78e | 603 | blk_mq_freeze_queue(q); |
245a489e | 604 | blk_mq_cancel_work_sync(q); |
737eb78e | 605 | |
131d08e1 | 606 | err = blk_mq_init_sched(q, e); |
737eb78e | 607 | |
737eb78e DLM |
608 | blk_mq_unfreeze_queue(q); |
609 | ||
954b4a5c DLM |
610 | if (err) { |
611 | pr_warn("\"%s\" elevator initialization failed, " | |
612 | "falling back to \"none\"\n", e->elevator_name); | |
954b4a5c | 613 | } |
8ed40ee3 JC |
614 | |
615 | elevator_put(e); | |
131d08e1 | 616 | } |
131d08e1 | 617 | |
1da177e4 | 618 | /* |
ac1171bd JC |
619 | * Switch to new_e io scheduler. |
620 | * | |
621 | * If switching fails, we are most likely running out of memory and not able | |
622 | * to restore the old io scheduler, so leaving the io scheduler being none. | |
1da177e4 | 623 | */ |
8237c01f | 624 | int elevator_switch(struct request_queue *q, struct elevator_type *new_e) |
1da177e4 | 625 | { |
64b36075 | 626 | int ret; |
1da177e4 | 627 | |
14a23498 BVA |
628 | lockdep_assert_held(&q->sysfs_lock); |
629 | ||
a1ce35fa JA |
630 | blk_mq_freeze_queue(q); |
631 | blk_mq_quiesce_queue(q); | |
4722dc52 | 632 | |
64b36075 CH |
633 | if (q->elevator) { |
634 | elv_unregister_queue(q); | |
635 | elevator_exit(q); | |
636 | } | |
637 | ||
638 | ret = blk_mq_init_sched(q, new_e); | |
639 | if (ret) | |
640 | goto out_unfreeze; | |
641 | ||
642 | ret = elv_register_queue(q, true); | |
643 | if (ret) { | |
644 | elevator_exit(q); | |
645 | goto out_unfreeze; | |
646 | } | |
647 | blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name); | |
1da177e4 | 648 | |
64b36075 | 649 | out_unfreeze: |
a1ce35fa JA |
650 | blk_mq_unquiesce_queue(q); |
651 | blk_mq_unfreeze_queue(q); | |
e0cca8bc JC |
652 | |
653 | if (ret) { | |
654 | pr_warn("elv: switch to \"%s\" failed, falling back to \"none\"\n", | |
655 | new_e->elevator_name); | |
656 | } | |
657 | ||
64b36075 CH |
658 | return ret; |
659 | } | |
660 | ||
661 | void elevator_disable(struct request_queue *q) | |
662 | { | |
663 | lockdep_assert_held(&q->sysfs_lock); | |
75ad23bc | 664 | |
64b36075 CH |
665 | blk_mq_freeze_queue(q); |
666 | blk_mq_quiesce_queue(q); | |
667 | ||
668 | elv_unregister_queue(q); | |
669 | elevator_exit(q); | |
670 | blk_queue_flag_clear(QUEUE_FLAG_SQ_SCHED, q); | |
671 | q->elevator = NULL; | |
672 | q->nr_requests = q->tag_set->queue_depth; | |
673 | blk_add_trace_msg(q, "elv switch: none"); | |
674 | ||
675 | blk_mq_unquiesce_queue(q); | |
676 | blk_mq_unfreeze_queue(q); | |
1da177e4 LT |
677 | } |
678 | ||
5dd531a0 JA |
679 | /* |
680 | * Switch this queue to the given IO scheduler. | |
681 | */ | |
58367c8a | 682 | static int elevator_change(struct request_queue *q, const char *elevator_name) |
1da177e4 | 683 | { |
1da177e4 | 684 | struct elevator_type *e; |
8ed40ee3 | 685 | int ret; |
1da177e4 | 686 | |
e9a823fb | 687 | /* Make sure queue is not in the middle of being removed */ |
58c898ba | 688 | if (!blk_queue_registered(q)) |
e9a823fb DJ |
689 | return -ENOENT; |
690 | ||
58367c8a | 691 | if (!strncmp(elevator_name, "none", 4)) { |
64b36075 CH |
692 | if (q->elevator) |
693 | elevator_disable(q); | |
694 | return 0; | |
fbd72127 | 695 | } |
cd43e26f | 696 | |
ffb86425 | 697 | if (q->elevator && elevator_match(q->elevator->type, elevator_name)) |
b54c2ad9 JC |
698 | return 0; |
699 | ||
81eaca44 CH |
700 | e = elevator_find_get(q, elevator_name); |
701 | if (!e) { | |
702 | request_module("%s-iosched", elevator_name); | |
703 | e = elevator_find_get(q, elevator_name); | |
704 | if (!e) | |
705 | return -EINVAL; | |
706 | } | |
8ed40ee3 JC |
707 | ret = elevator_switch(q, e); |
708 | elevator_put(e); | |
709 | return ret; | |
5dd531a0 | 710 | } |
7c8a3679 | 711 | |
58367c8a | 712 | ssize_t elv_iosched_store(struct request_queue *q, const char *buf, |
5dd531a0 JA |
713 | size_t count) |
714 | { | |
58367c8a | 715 | char elevator_name[ELV_NAME_MAX]; |
5dd531a0 JA |
716 | int ret; |
717 | ||
6251b754 | 718 | if (!elv_support_iosched(q)) |
5dd531a0 JA |
719 | return count; |
720 | ||
20d09975 | 721 | strscpy(elevator_name, buf, sizeof(elevator_name)); |
58367c8a | 722 | ret = elevator_change(q, strstrip(elevator_name)); |
5dd531a0 JA |
723 | if (!ret) |
724 | return count; | |
5dd531a0 | 725 | return ret; |
1da177e4 LT |
726 | } |
727 | ||
165125e1 | 728 | ssize_t elv_iosched_show(struct request_queue *q, char *name) |
1da177e4 | 729 | { |
16095af2 CH |
730 | struct elevator_queue *eq = q->elevator; |
731 | struct elevator_type *cur = NULL, *e; | |
1da177e4 LT |
732 | int len = 0; |
733 | ||
aae2a643 | 734 | if (!elv_support_iosched(q)) |
cd43e26f MP |
735 | return sprintf(name, "none\n"); |
736 | ||
7919d679 | 737 | if (!q->elevator) { |
bd166ef1 | 738 | len += sprintf(name+len, "[none] "); |
7919d679 JC |
739 | } else { |
740 | len += sprintf(name+len, "none "); | |
16095af2 | 741 | cur = eq->type; |
7919d679 | 742 | } |
cd43e26f | 743 | |
2a12dcd7 | 744 | spin_lock(&elv_list_lock); |
16095af2 | 745 | list_for_each_entry(e, &elv_list, list) { |
5998249e | 746 | if (e == cur) |
7a3b3660 | 747 | len += sprintf(name+len, "[%s] ", e->elevator_name); |
e4eb37cc | 748 | else |
16095af2 | 749 | len += sprintf(name+len, "%s ", e->elevator_name); |
1da177e4 | 750 | } |
2a12dcd7 | 751 | spin_unlock(&elv_list_lock); |
1da177e4 | 752 | |
c6451ede | 753 | len += sprintf(name+len, "\n"); |
1da177e4 LT |
754 | return len; |
755 | } | |
756 | ||
165125e1 JA |
757 | struct request *elv_rb_former_request(struct request_queue *q, |
758 | struct request *rq) | |
2e662b65 JA |
759 | { |
760 | struct rb_node *rbprev = rb_prev(&rq->rb_node); | |
761 | ||
762 | if (rbprev) | |
763 | return rb_entry_rq(rbprev); | |
764 | ||
765 | return NULL; | |
766 | } | |
2e662b65 JA |
767 | EXPORT_SYMBOL(elv_rb_former_request); |
768 | ||
165125e1 JA |
769 | struct request *elv_rb_latter_request(struct request_queue *q, |
770 | struct request *rq) | |
2e662b65 JA |
771 | { |
772 | struct rb_node *rbnext = rb_next(&rq->rb_node); | |
773 | ||
774 | if (rbnext) | |
775 | return rb_entry_rq(rbnext); | |
776 | ||
777 | return NULL; | |
778 | } | |
2e662b65 | 779 | EXPORT_SYMBOL(elv_rb_latter_request); |
f8db3835 JK |
780 | |
781 | static int __init elevator_setup(char *str) | |
782 | { | |
783 | pr_warn("Kernel parameter elevator= does not have any effect anymore.\n" | |
784 | "Please use sysfs to set IO scheduler for individual devices.\n"); | |
785 | return 1; | |
786 | } | |
787 | ||
788 | __setup("elevator=", elevator_setup); |