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 | */ | |
72ef799b | 60 | static int 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 JA |
67 | |
68 | return 1; | |
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 | |
ffb86425 CH |
86 | static inline bool elv_support_features(struct request_queue *q, |
87 | const struct elevator_type *e) | |
8ac0d9a8 | 88 | { |
ffb86425 CH |
89 | return (q->required_elevator_features & e->elevator_features) == |
90 | q->required_elevator_features; | |
68c43f13 DLM |
91 | } |
92 | ||
93 | /** | |
94 | * elevator_match - Test an elevator name and features | |
95 | * @e: Scheduler to test | |
96 | * @name: Elevator name to test | |
68c43f13 | 97 | * |
5b8f65e1 RD |
98 | * Return true if the elevator @e name matches @name and if @e provides all |
99 | * the features specified by @required_features. | |
68c43f13 | 100 | */ |
ffb86425 | 101 | static bool elevator_match(const struct elevator_type *e, const char *name) |
68c43f13 | 102 | { |
ffb86425 CH |
103 | return !strcmp(e->elevator_name, name) || |
104 | (e->elevator_alias && !strcmp(e->elevator_alias, name)); | |
8ac0d9a8 JA |
105 | } |
106 | ||
ffb86425 | 107 | static struct elevator_type *__elevator_find(const char *name) |
1da177e4 | 108 | { |
a22b169d | 109 | struct elevator_type *e; |
1da177e4 | 110 | |
ffb86425 CH |
111 | list_for_each_entry(e, &elv_list, list) |
112 | if (elevator_match(e, name)) | |
a22b169d | 113 | return e; |
a22b169d | 114 | return NULL; |
1da177e4 LT |
115 | } |
116 | ||
81eaca44 CH |
117 | static struct elevator_type *elevator_find_get(struct request_queue *q, |
118 | const char *name) | |
1da177e4 | 119 | { |
2824bc93 | 120 | struct elevator_type *e; |
1da177e4 | 121 | |
2a12dcd7 | 122 | spin_lock(&elv_list_lock); |
ffb86425 CH |
123 | e = __elevator_find(name); |
124 | if (e && (!elv_support_features(q, e) || !elevator_tryget(e))) | |
2824bc93 | 125 | e = NULL; |
2a12dcd7 | 126 | spin_unlock(&elv_list_lock); |
1da177e4 LT |
127 | return e; |
128 | } | |
129 | ||
3d1ab40f AV |
130 | static struct kobj_type elv_ktype; |
131 | ||
d50235b7 | 132 | struct elevator_queue *elevator_alloc(struct request_queue *q, |
165125e1 | 133 | struct elevator_type *e) |
3d1ab40f | 134 | { |
b374d18a | 135 | struct elevator_queue *eq; |
9817064b | 136 | |
c1b511eb | 137 | eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node); |
9817064b | 138 | if (unlikely(!eq)) |
8406a4d5 | 139 | return NULL; |
9817064b | 140 | |
8ed40ee3 | 141 | __elevator_get(e); |
22f746e2 | 142 | eq->type = e; |
f9cb074b | 143 | kobject_init(&eq->kobj, &elv_ktype); |
9817064b | 144 | mutex_init(&eq->sysfs_lock); |
242d98f0 | 145 | hash_init(eq->hash); |
9817064b | 146 | |
3d1ab40f AV |
147 | return eq; |
148 | } | |
d50235b7 | 149 | EXPORT_SYMBOL(elevator_alloc); |
3d1ab40f AV |
150 | |
151 | static void elevator_release(struct kobject *kobj) | |
152 | { | |
b374d18a | 153 | struct elevator_queue *e; |
9817064b | 154 | |
b374d18a | 155 | e = container_of(kobj, struct elevator_queue, kobj); |
22f746e2 | 156 | elevator_put(e->type); |
3d1ab40f AV |
157 | kfree(e); |
158 | } | |
159 | ||
0c6cb3a2 | 160 | void elevator_exit(struct request_queue *q) |
1da177e4 | 161 | { |
0c6cb3a2 CH |
162 | struct elevator_queue *e = q->elevator; |
163 | ||
28883074 CH |
164 | ioc_clear_queue(q); |
165 | blk_mq_sched_free_rqs(q); | |
166 | ||
3d1ab40f | 167 | mutex_lock(&e->sysfs_lock); |
dd1c372d | 168 | blk_mq_exit_sched(q, e); |
3d1ab40f | 169 | mutex_unlock(&e->sysfs_lock); |
1da177e4 | 170 | |
3d1ab40f | 171 | kobject_put(&e->kobj); |
1da177e4 | 172 | } |
2e662b65 | 173 | |
9817064b JA |
174 | static inline void __elv_rqhash_del(struct request *rq) |
175 | { | |
242d98f0 | 176 | hash_del(&rq->hash); |
e8064021 | 177 | rq->rq_flags &= ~RQF_HASHED; |
9817064b JA |
178 | } |
179 | ||
70b3ea05 | 180 | void elv_rqhash_del(struct request_queue *q, struct request *rq) |
9817064b JA |
181 | { |
182 | if (ELV_ON_HASH(rq)) | |
183 | __elv_rqhash_del(rq); | |
184 | } | |
bd166ef1 | 185 | EXPORT_SYMBOL_GPL(elv_rqhash_del); |
9817064b | 186 | |
70b3ea05 | 187 | void elv_rqhash_add(struct request_queue *q, struct request *rq) |
9817064b | 188 | { |
b374d18a | 189 | struct elevator_queue *e = q->elevator; |
9817064b JA |
190 | |
191 | BUG_ON(ELV_ON_HASH(rq)); | |
242d98f0 | 192 | hash_add(e->hash, &rq->hash, rq_hash_key(rq)); |
e8064021 | 193 | rq->rq_flags |= RQF_HASHED; |
9817064b | 194 | } |
bd166ef1 | 195 | EXPORT_SYMBOL_GPL(elv_rqhash_add); |
9817064b | 196 | |
70b3ea05 | 197 | void elv_rqhash_reposition(struct request_queue *q, struct request *rq) |
9817064b JA |
198 | { |
199 | __elv_rqhash_del(rq); | |
200 | elv_rqhash_add(q, rq); | |
201 | } | |
202 | ||
70b3ea05 | 203 | struct request *elv_rqhash_find(struct request_queue *q, sector_t offset) |
9817064b | 204 | { |
b374d18a | 205 | struct elevator_queue *e = q->elevator; |
b67bfe0d | 206 | struct hlist_node *next; |
9817064b JA |
207 | struct request *rq; |
208 | ||
ee89f812 | 209 | hash_for_each_possible_safe(e->hash, rq, next, hash, offset) { |
9817064b JA |
210 | BUG_ON(!ELV_ON_HASH(rq)); |
211 | ||
212 | if (unlikely(!rq_mergeable(rq))) { | |
213 | __elv_rqhash_del(rq); | |
214 | continue; | |
215 | } | |
216 | ||
217 | if (rq_hash_key(rq) == offset) | |
218 | return rq; | |
219 | } | |
220 | ||
221 | return NULL; | |
222 | } | |
223 | ||
2e662b65 JA |
224 | /* |
225 | * RB-tree support functions for inserting/lookup/removal of requests | |
226 | * in a sorted RB tree. | |
227 | */ | |
796d5116 | 228 | void elv_rb_add(struct rb_root *root, struct request *rq) |
2e662b65 JA |
229 | { |
230 | struct rb_node **p = &root->rb_node; | |
231 | struct rb_node *parent = NULL; | |
232 | struct request *__rq; | |
233 | ||
234 | while (*p) { | |
235 | parent = *p; | |
236 | __rq = rb_entry(parent, struct request, rb_node); | |
237 | ||
83096ebf | 238 | if (blk_rq_pos(rq) < blk_rq_pos(__rq)) |
2e662b65 | 239 | p = &(*p)->rb_left; |
796d5116 | 240 | else if (blk_rq_pos(rq) >= blk_rq_pos(__rq)) |
2e662b65 | 241 | p = &(*p)->rb_right; |
2e662b65 JA |
242 | } |
243 | ||
244 | rb_link_node(&rq->rb_node, parent, p); | |
245 | rb_insert_color(&rq->rb_node, root); | |
2e662b65 | 246 | } |
2e662b65 JA |
247 | EXPORT_SYMBOL(elv_rb_add); |
248 | ||
249 | void elv_rb_del(struct rb_root *root, struct request *rq) | |
250 | { | |
251 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); | |
252 | rb_erase(&rq->rb_node, root); | |
253 | RB_CLEAR_NODE(&rq->rb_node); | |
254 | } | |
2e662b65 JA |
255 | EXPORT_SYMBOL(elv_rb_del); |
256 | ||
257 | struct request *elv_rb_find(struct rb_root *root, sector_t sector) | |
258 | { | |
259 | struct rb_node *n = root->rb_node; | |
260 | struct request *rq; | |
261 | ||
262 | while (n) { | |
263 | rq = rb_entry(n, struct request, rb_node); | |
264 | ||
83096ebf | 265 | if (sector < blk_rq_pos(rq)) |
2e662b65 | 266 | n = n->rb_left; |
83096ebf | 267 | else if (sector > blk_rq_pos(rq)) |
2e662b65 JA |
268 | n = n->rb_right; |
269 | else | |
270 | return rq; | |
271 | } | |
272 | ||
273 | return NULL; | |
274 | } | |
2e662b65 JA |
275 | EXPORT_SYMBOL(elv_rb_find); |
276 | ||
34fe7c05 CH |
277 | enum elv_merge elv_merge(struct request_queue *q, struct request **req, |
278 | struct bio *bio) | |
1da177e4 | 279 | { |
b374d18a | 280 | struct elevator_queue *e = q->elevator; |
9817064b | 281 | struct request *__rq; |
06b86245 | 282 | |
488991e2 AB |
283 | /* |
284 | * Levels of merges: | |
285 | * nomerges: No merges at all attempted | |
286 | * noxmerges: Only simple one-hit cache try | |
287 | * merges: All merge tries attempted | |
288 | */ | |
7460d389 | 289 | if (blk_queue_nomerges(q) || !bio_mergeable(bio)) |
488991e2 AB |
290 | return ELEVATOR_NO_MERGE; |
291 | ||
9817064b JA |
292 | /* |
293 | * First try one-hit cache. | |
294 | */ | |
72ef799b | 295 | if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) { |
34fe7c05 CH |
296 | enum elv_merge ret = blk_try_merge(q->last_merge, bio); |
297 | ||
06b86245 TH |
298 | if (ret != ELEVATOR_NO_MERGE) { |
299 | *req = q->last_merge; | |
300 | return ret; | |
301 | } | |
302 | } | |
1da177e4 | 303 | |
488991e2 | 304 | if (blk_queue_noxmerges(q)) |
ac9fafa1 AB |
305 | return ELEVATOR_NO_MERGE; |
306 | ||
9817064b JA |
307 | /* |
308 | * See if our hash lookup can find a potential backmerge. | |
309 | */ | |
4f024f37 | 310 | __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector); |
72ef799b | 311 | if (__rq && elv_bio_merge_ok(__rq, bio)) { |
9817064b | 312 | *req = __rq; |
866663b7 ML |
313 | |
314 | if (blk_discard_mergable(__rq)) | |
315 | return ELEVATOR_DISCARD_MERGE; | |
9817064b JA |
316 | return ELEVATOR_BACK_MERGE; |
317 | } | |
318 | ||
f9cd4bfe JA |
319 | if (e->type->ops.request_merge) |
320 | return e->type->ops.request_merge(q, req, bio); | |
1da177e4 LT |
321 | |
322 | return ELEVATOR_NO_MERGE; | |
323 | } | |
324 | ||
5e84ea3a JA |
325 | /* |
326 | * Attempt to do an insertion back merge. Only check for the case where | |
327 | * we can append 'rq' to an existing request, so we can throw 'rq' away | |
328 | * afterwards. | |
329 | * | |
fd2ef39c JK |
330 | * Returns true if we merged, false otherwise. 'free' will contain all |
331 | * requests that need to be freed. | |
5e84ea3a | 332 | */ |
fd2ef39c JK |
333 | bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq, |
334 | struct list_head *free) | |
5e84ea3a JA |
335 | { |
336 | struct request *__rq; | |
bee0393c | 337 | bool ret; |
5e84ea3a JA |
338 | |
339 | if (blk_queue_nomerges(q)) | |
340 | return false; | |
341 | ||
342 | /* | |
343 | * First try one-hit cache. | |
344 | */ | |
fd2ef39c JK |
345 | if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) { |
346 | list_add(&rq->queuelist, free); | |
5e84ea3a | 347 | return true; |
fd2ef39c | 348 | } |
5e84ea3a JA |
349 | |
350 | if (blk_queue_noxmerges(q)) | |
351 | return false; | |
352 | ||
bee0393c | 353 | ret = false; |
5e84ea3a JA |
354 | /* |
355 | * See if our hash lookup can find a potential backmerge. | |
356 | */ | |
bee0393c SL |
357 | while (1) { |
358 | __rq = elv_rqhash_find(q, blk_rq_pos(rq)); | |
359 | if (!__rq || !blk_attempt_req_merge(q, __rq, rq)) | |
360 | break; | |
361 | ||
fd2ef39c | 362 | list_add(&rq->queuelist, free); |
bee0393c SL |
363 | /* The merged request could be merged with others, try again */ |
364 | ret = true; | |
365 | rq = __rq; | |
366 | } | |
27419322 | 367 | |
bee0393c | 368 | return ret; |
5e84ea3a JA |
369 | } |
370 | ||
34fe7c05 CH |
371 | void elv_merged_request(struct request_queue *q, struct request *rq, |
372 | enum elv_merge type) | |
1da177e4 | 373 | { |
b374d18a | 374 | struct elevator_queue *e = q->elevator; |
1da177e4 | 375 | |
f9cd4bfe JA |
376 | if (e->type->ops.request_merged) |
377 | e->type->ops.request_merged(q, rq, type); | |
06b86245 | 378 | |
2e662b65 JA |
379 | if (type == ELEVATOR_BACK_MERGE) |
380 | elv_rqhash_reposition(q, rq); | |
9817064b | 381 | |
06b86245 | 382 | q->last_merge = rq; |
1da177e4 LT |
383 | } |
384 | ||
165125e1 | 385 | void elv_merge_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
386 | struct request *next) |
387 | { | |
b374d18a | 388 | struct elevator_queue *e = q->elevator; |
bd166ef1 | 389 | |
f9cd4bfe JA |
390 | if (e->type->ops.requests_merged) |
391 | e->type->ops.requests_merged(q, rq, next); | |
06b86245 | 392 | |
9817064b | 393 | elv_rqhash_reposition(q, rq); |
06b86245 | 394 | q->last_merge = rq; |
1da177e4 LT |
395 | } |
396 | ||
165125e1 | 397 | struct request *elv_latter_request(struct request_queue *q, struct request *rq) |
1da177e4 | 398 | { |
b374d18a | 399 | struct elevator_queue *e = q->elevator; |
1da177e4 | 400 | |
f9cd4bfe JA |
401 | if (e->type->ops.next_request) |
402 | return e->type->ops.next_request(q, rq); | |
bd166ef1 | 403 | |
1da177e4 LT |
404 | return NULL; |
405 | } | |
406 | ||
165125e1 | 407 | struct request *elv_former_request(struct request_queue *q, struct request *rq) |
1da177e4 | 408 | { |
b374d18a | 409 | struct elevator_queue *e = q->elevator; |
1da177e4 | 410 | |
f9cd4bfe JA |
411 | if (e->type->ops.former_request) |
412 | return e->type->ops.former_request(q, rq); | |
bd166ef1 | 413 | |
a1ce35fa | 414 | return NULL; |
1da177e4 LT |
415 | } |
416 | ||
3d1ab40f AV |
417 | #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr) |
418 | ||
419 | static ssize_t | |
420 | elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page) | |
1da177e4 | 421 | { |
3d1ab40f | 422 | struct elv_fs_entry *entry = to_elv(attr); |
b374d18a | 423 | struct elevator_queue *e; |
3d1ab40f AV |
424 | ssize_t error; |
425 | ||
426 | if (!entry->show) | |
427 | return -EIO; | |
428 | ||
b374d18a | 429 | e = container_of(kobj, struct elevator_queue, kobj); |
3d1ab40f | 430 | mutex_lock(&e->sysfs_lock); |
22f746e2 | 431 | error = e->type ? entry->show(e, page) : -ENOENT; |
3d1ab40f AV |
432 | mutex_unlock(&e->sysfs_lock); |
433 | return error; | |
434 | } | |
1da177e4 | 435 | |
3d1ab40f AV |
436 | static ssize_t |
437 | elv_attr_store(struct kobject *kobj, struct attribute *attr, | |
438 | const char *page, size_t length) | |
439 | { | |
3d1ab40f | 440 | struct elv_fs_entry *entry = to_elv(attr); |
b374d18a | 441 | struct elevator_queue *e; |
3d1ab40f | 442 | ssize_t error; |
1da177e4 | 443 | |
3d1ab40f AV |
444 | if (!entry->store) |
445 | return -EIO; | |
1da177e4 | 446 | |
b374d18a | 447 | e = container_of(kobj, struct elevator_queue, kobj); |
3d1ab40f | 448 | mutex_lock(&e->sysfs_lock); |
22f746e2 | 449 | error = e->type ? entry->store(e, page, length) : -ENOENT; |
3d1ab40f AV |
450 | mutex_unlock(&e->sysfs_lock); |
451 | return error; | |
452 | } | |
453 | ||
52cf25d0 | 454 | static const struct sysfs_ops elv_sysfs_ops = { |
3d1ab40f AV |
455 | .show = elv_attr_show, |
456 | .store = elv_attr_store, | |
457 | }; | |
458 | ||
459 | static struct kobj_type elv_ktype = { | |
460 | .sysfs_ops = &elv_sysfs_ops, | |
461 | .release = elevator_release, | |
462 | }; | |
463 | ||
cecf5d87 | 464 | int elv_register_queue(struct request_queue *q, bool uevent) |
3d1ab40f | 465 | { |
5a5bafdc | 466 | struct elevator_queue *e = q->elevator; |
3d1ab40f AV |
467 | int error; |
468 | ||
f0c6ae09 YY |
469 | lockdep_assert_held(&q->sysfs_lock); |
470 | ||
b2d6db58 | 471 | error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched"); |
3d1ab40f | 472 | if (!error) { |
22f746e2 | 473 | struct elv_fs_entry *attr = e->type->elevator_attrs; |
3d1ab40f | 474 | if (attr) { |
e572ec7e AV |
475 | while (attr->attr.name) { |
476 | if (sysfs_create_file(&e->kobj, &attr->attr)) | |
3d1ab40f | 477 | break; |
e572ec7e | 478 | attr++; |
3d1ab40f AV |
479 | } |
480 | } | |
cecf5d87 ML |
481 | if (uevent) |
482 | kobject_uevent(&e->kobj, KOBJ_ADD); | |
483 | ||
181d0663 | 484 | set_bit(ELEVATOR_FLAG_REGISTERED, &e->flags); |
3d1ab40f AV |
485 | } |
486 | return error; | |
1da177e4 | 487 | } |
bc1c1169 | 488 | |
1da177e4 LT |
489 | void elv_unregister_queue(struct request_queue *q) |
490 | { | |
f5ec592d EB |
491 | struct elevator_queue *e = q->elevator; |
492 | ||
f0c6ae09 YY |
493 | lockdep_assert_held(&q->sysfs_lock); |
494 | ||
181d0663 | 495 | if (e && test_and_clear_bit(ELEVATOR_FLAG_REGISTERED, &e->flags)) { |
f8fc877d TH |
496 | kobject_uevent(&e->kobj, KOBJ_REMOVE); |
497 | kobject_del(&e->kobj); | |
f8fc877d | 498 | } |
1da177e4 LT |
499 | } |
500 | ||
e567bf71 | 501 | int elv_register(struct elevator_type *e) |
1da177e4 | 502 | { |
e42cfb1d DLM |
503 | /* insert_requests and dispatch_request are mandatory */ |
504 | if (WARN_ON_ONCE(!e->ops.insert_requests || !e->ops.dispatch_request)) | |
505 | return -EINVAL; | |
506 | ||
3d3c2379 TH |
507 | /* create icq_cache if requested */ |
508 | if (e->icq_size) { | |
509 | if (WARN_ON(e->icq_size < sizeof(struct io_cq)) || | |
510 | WARN_ON(e->icq_align < __alignof__(struct io_cq))) | |
511 | return -EINVAL; | |
512 | ||
513 | snprintf(e->icq_cache_name, sizeof(e->icq_cache_name), | |
514 | "%s_io_cq", e->elevator_name); | |
515 | e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size, | |
516 | e->icq_align, 0, NULL); | |
517 | if (!e->icq_cache) | |
518 | return -ENOMEM; | |
519 | } | |
520 | ||
521 | /* register, don't allow duplicate names */ | |
2a12dcd7 | 522 | spin_lock(&elv_list_lock); |
ffb86425 | 523 | if (__elevator_find(e->elevator_name)) { |
3d3c2379 | 524 | spin_unlock(&elv_list_lock); |
62d2a194 | 525 | kmem_cache_destroy(e->icq_cache); |
3d3c2379 TH |
526 | return -EBUSY; |
527 | } | |
1da177e4 | 528 | list_add_tail(&e->list, &elv_list); |
2a12dcd7 | 529 | spin_unlock(&elv_list_lock); |
1da177e4 | 530 | |
d0b0a81a HT |
531 | printk(KERN_INFO "io scheduler %s registered\n", e->elevator_name); |
532 | ||
3d3c2379 | 533 | return 0; |
1da177e4 LT |
534 | } |
535 | EXPORT_SYMBOL_GPL(elv_register); | |
536 | ||
537 | void elv_unregister(struct elevator_type *e) | |
538 | { | |
3d3c2379 | 539 | /* unregister */ |
2a12dcd7 | 540 | spin_lock(&elv_list_lock); |
1da177e4 | 541 | list_del_init(&e->list); |
2a12dcd7 | 542 | spin_unlock(&elv_list_lock); |
3d3c2379 TH |
543 | |
544 | /* | |
545 | * Destroy icq_cache if it exists. icq's are RCU managed. Make | |
546 | * sure all RCU operations are complete before proceeding. | |
547 | */ | |
548 | if (e->icq_cache) { | |
549 | rcu_barrier(); | |
550 | kmem_cache_destroy(e->icq_cache); | |
551 | e->icq_cache = NULL; | |
552 | } | |
1da177e4 LT |
553 | } |
554 | EXPORT_SYMBOL_GPL(elv_unregister); | |
555 | ||
61db437d DLM |
556 | static inline bool elv_support_iosched(struct request_queue *q) |
557 | { | |
6251b754 | 558 | if (!queue_is_mq(q) || |
7a7c5e71 | 559 | (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED))) |
61db437d DLM |
560 | return false; |
561 | return true; | |
562 | } | |
563 | ||
131d08e1 | 564 | /* |
a0958ba7 DLM |
565 | * For single queue devices, default to using mq-deadline. If we have multiple |
566 | * queues or mq-deadline is not available, default to "none". | |
567 | */ | |
568 | static struct elevator_type *elevator_get_default(struct request_queue *q) | |
569 | { | |
90b71980 BVA |
570 | if (q->tag_set && q->tag_set->flags & BLK_MQ_F_NO_SCHED_BY_DEFAULT) |
571 | return NULL; | |
572 | ||
580dca81 | 573 | if (q->nr_hw_queues != 1 && |
079a2e3e | 574 | !blk_mq_is_shared_tags(q->tag_set->flags)) |
a0958ba7 DLM |
575 | return NULL; |
576 | ||
81eaca44 | 577 | return elevator_find_get(q, "mq-deadline"); |
a0958ba7 DLM |
578 | } |
579 | ||
580 | /* | |
581 | * Get the first elevator providing the features required by the request queue. | |
582 | * Default to "none" if no matching elevator is found. | |
583 | */ | |
584 | static struct elevator_type *elevator_get_by_features(struct request_queue *q) | |
585 | { | |
a2614255 | 586 | struct elevator_type *e, *found = NULL; |
a0958ba7 DLM |
587 | |
588 | spin_lock(&elv_list_lock); | |
589 | ||
590 | list_for_each_entry(e, &elv_list, list) { | |
ffb86425 | 591 | if (elv_support_features(q, e)) { |
a2614255 | 592 | found = e; |
a0958ba7 | 593 | break; |
a2614255 | 594 | } |
a0958ba7 DLM |
595 | } |
596 | ||
dd6f7f17 | 597 | if (found && !elevator_tryget(found)) |
a2614255 | 598 | found = NULL; |
a0958ba7 DLM |
599 | |
600 | spin_unlock(&elv_list_lock); | |
a2614255 | 601 | return found; |
a0958ba7 DLM |
602 | } |
603 | ||
604 | /* | |
605 | * For a device queue that has no required features, use the default elevator | |
606 | * settings. Otherwise, use the first elevator available matching the required | |
607 | * features. If no suitable elevator is find or if the chosen elevator | |
608 | * initialization fails, fall back to the "none" elevator (no elevator). | |
131d08e1 | 609 | */ |
954b4a5c | 610 | void elevator_init_mq(struct request_queue *q) |
131d08e1 CH |
611 | { |
612 | struct elevator_type *e; | |
954b4a5c | 613 | int err; |
131d08e1 | 614 | |
61db437d | 615 | if (!elv_support_iosched(q)) |
954b4a5c | 616 | return; |
61db437d | 617 | |
75e6c00f | 618 | WARN_ON_ONCE(blk_queue_registered(q)); |
c48dac13 | 619 | |
131d08e1 | 620 | if (unlikely(q->elevator)) |
954b4a5c | 621 | return; |
131d08e1 | 622 | |
a0958ba7 DLM |
623 | if (!q->required_elevator_features) |
624 | e = elevator_get_default(q); | |
625 | else | |
626 | e = elevator_get_by_features(q); | |
131d08e1 | 627 | if (!e) |
954b4a5c | 628 | return; |
131d08e1 | 629 | |
245a489e ML |
630 | /* |
631 | * We are called before adding disk, when there isn't any FS I/O, | |
632 | * so freezing queue plus canceling dispatch work is enough to | |
633 | * drain any dispatch activities originated from passthrough | |
634 | * requests, then no need to quiesce queue which may add long boot | |
635 | * latency, especially when lots of disks are involved. | |
636 | */ | |
737eb78e | 637 | blk_mq_freeze_queue(q); |
245a489e | 638 | blk_mq_cancel_work_sync(q); |
737eb78e | 639 | |
131d08e1 | 640 | err = blk_mq_init_sched(q, e); |
737eb78e | 641 | |
737eb78e DLM |
642 | blk_mq_unfreeze_queue(q); |
643 | ||
954b4a5c DLM |
644 | if (err) { |
645 | pr_warn("\"%s\" elevator initialization failed, " | |
646 | "falling back to \"none\"\n", e->elevator_name); | |
954b4a5c | 647 | } |
8ed40ee3 JC |
648 | |
649 | elevator_put(e); | |
131d08e1 | 650 | } |
131d08e1 | 651 | |
1da177e4 | 652 | /* |
ac1171bd JC |
653 | * Switch to new_e io scheduler. |
654 | * | |
655 | * If switching fails, we are most likely running out of memory and not able | |
656 | * to restore the old io scheduler, so leaving the io scheduler being none. | |
1da177e4 | 657 | */ |
8237c01f | 658 | int elevator_switch(struct request_queue *q, struct elevator_type *new_e) |
1da177e4 | 659 | { |
64b36075 | 660 | int ret; |
1da177e4 | 661 | |
14a23498 BVA |
662 | lockdep_assert_held(&q->sysfs_lock); |
663 | ||
a1ce35fa JA |
664 | blk_mq_freeze_queue(q); |
665 | blk_mq_quiesce_queue(q); | |
4722dc52 | 666 | |
64b36075 CH |
667 | if (q->elevator) { |
668 | elv_unregister_queue(q); | |
669 | elevator_exit(q); | |
670 | } | |
671 | ||
672 | ret = blk_mq_init_sched(q, new_e); | |
673 | if (ret) | |
674 | goto out_unfreeze; | |
675 | ||
676 | ret = elv_register_queue(q, true); | |
677 | if (ret) { | |
678 | elevator_exit(q); | |
679 | goto out_unfreeze; | |
680 | } | |
681 | blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name); | |
1da177e4 | 682 | |
64b36075 | 683 | out_unfreeze: |
a1ce35fa JA |
684 | blk_mq_unquiesce_queue(q); |
685 | blk_mq_unfreeze_queue(q); | |
64b36075 CH |
686 | return ret; |
687 | } | |
688 | ||
689 | void elevator_disable(struct request_queue *q) | |
690 | { | |
691 | lockdep_assert_held(&q->sysfs_lock); | |
75ad23bc | 692 | |
64b36075 CH |
693 | blk_mq_freeze_queue(q); |
694 | blk_mq_quiesce_queue(q); | |
695 | ||
696 | elv_unregister_queue(q); | |
697 | elevator_exit(q); | |
698 | blk_queue_flag_clear(QUEUE_FLAG_SQ_SCHED, q); | |
699 | q->elevator = NULL; | |
700 | q->nr_requests = q->tag_set->queue_depth; | |
701 | blk_add_trace_msg(q, "elv switch: none"); | |
702 | ||
703 | blk_mq_unquiesce_queue(q); | |
704 | blk_mq_unfreeze_queue(q); | |
1da177e4 LT |
705 | } |
706 | ||
5dd531a0 JA |
707 | /* |
708 | * Switch this queue to the given IO scheduler. | |
709 | */ | |
58367c8a | 710 | static int elevator_change(struct request_queue *q, const char *elevator_name) |
1da177e4 | 711 | { |
1da177e4 | 712 | struct elevator_type *e; |
8ed40ee3 | 713 | int ret; |
1da177e4 | 714 | |
e9a823fb | 715 | /* Make sure queue is not in the middle of being removed */ |
58c898ba | 716 | if (!blk_queue_registered(q)) |
e9a823fb DJ |
717 | return -ENOENT; |
718 | ||
bd166ef1 JA |
719 | /* |
720 | * Special case for mq, turn off scheduling | |
721 | */ | |
58367c8a | 722 | if (!strncmp(elevator_name, "none", 4)) { |
64b36075 CH |
723 | if (q->elevator) |
724 | elevator_disable(q); | |
725 | return 0; | |
fbd72127 | 726 | } |
cd43e26f | 727 | |
ffb86425 | 728 | if (q->elevator && elevator_match(q->elevator->type, elevator_name)) |
b54c2ad9 JC |
729 | return 0; |
730 | ||
81eaca44 CH |
731 | e = elevator_find_get(q, elevator_name); |
732 | if (!e) { | |
733 | request_module("%s-iosched", elevator_name); | |
734 | e = elevator_find_get(q, elevator_name); | |
735 | if (!e) | |
736 | return -EINVAL; | |
737 | } | |
8ed40ee3 JC |
738 | ret = elevator_switch(q, e); |
739 | elevator_put(e); | |
740 | return ret; | |
5dd531a0 | 741 | } |
7c8a3679 | 742 | |
58367c8a | 743 | ssize_t elv_iosched_store(struct request_queue *q, const char *buf, |
5dd531a0 JA |
744 | size_t count) |
745 | { | |
58367c8a | 746 | char elevator_name[ELV_NAME_MAX]; |
5dd531a0 JA |
747 | int ret; |
748 | ||
6251b754 | 749 | if (!elv_support_iosched(q)) |
5dd531a0 JA |
750 | return count; |
751 | ||
58367c8a CH |
752 | strlcpy(elevator_name, buf, sizeof(elevator_name)); |
753 | ret = elevator_change(q, strstrip(elevator_name)); | |
5dd531a0 JA |
754 | if (!ret) |
755 | return count; | |
5dd531a0 | 756 | return ret; |
1da177e4 LT |
757 | } |
758 | ||
165125e1 | 759 | ssize_t elv_iosched_show(struct request_queue *q, char *name) |
1da177e4 | 760 | { |
16095af2 CH |
761 | struct elevator_queue *eq = q->elevator; |
762 | struct elevator_type *cur = NULL, *e; | |
1da177e4 LT |
763 | int len = 0; |
764 | ||
aae2a643 | 765 | if (!elv_support_iosched(q)) |
cd43e26f MP |
766 | return sprintf(name, "none\n"); |
767 | ||
bd166ef1 JA |
768 | if (!q->elevator) |
769 | len += sprintf(name+len, "[none] "); | |
770 | else | |
16095af2 | 771 | cur = eq->type; |
cd43e26f | 772 | |
2a12dcd7 | 773 | spin_lock(&elv_list_lock); |
16095af2 | 774 | list_for_each_entry(e, &elv_list, list) { |
2eef17a2 | 775 | if (e == cur) { |
16095af2 | 776 | len += sprintf(name+len, "[%s] ", cur->elevator_name); |
bd166ef1 JA |
777 | continue; |
778 | } | |
ffb86425 | 779 | if (elv_support_features(q, e)) |
16095af2 | 780 | len += sprintf(name+len, "%s ", e->elevator_name); |
1da177e4 | 781 | } |
2a12dcd7 | 782 | spin_unlock(&elv_list_lock); |
1da177e4 | 783 | |
344e9ffc | 784 | if (q->elevator) |
bd166ef1 JA |
785 | len += sprintf(name+len, "none"); |
786 | ||
1da177e4 LT |
787 | len += sprintf(len+name, "\n"); |
788 | return len; | |
789 | } | |
790 | ||
165125e1 JA |
791 | struct request *elv_rb_former_request(struct request_queue *q, |
792 | struct request *rq) | |
2e662b65 JA |
793 | { |
794 | struct rb_node *rbprev = rb_prev(&rq->rb_node); | |
795 | ||
796 | if (rbprev) | |
797 | return rb_entry_rq(rbprev); | |
798 | ||
799 | return NULL; | |
800 | } | |
2e662b65 JA |
801 | EXPORT_SYMBOL(elv_rb_former_request); |
802 | ||
165125e1 JA |
803 | struct request *elv_rb_latter_request(struct request_queue *q, |
804 | struct request *rq) | |
2e662b65 JA |
805 | { |
806 | struct rb_node *rbnext = rb_next(&rq->rb_node); | |
807 | ||
808 | if (rbnext) | |
809 | return rb_entry_rq(rbnext); | |
810 | ||
811 | return NULL; | |
812 | } | |
2e662b65 | 813 | EXPORT_SYMBOL(elv_rb_latter_request); |
f8db3835 JK |
814 | |
815 | static int __init elevator_setup(char *str) | |
816 | { | |
817 | pr_warn("Kernel parameter elevator= does not have any effect anymore.\n" | |
818 | "Please use sysfs to set IO scheduler for individual devices.\n"); | |
819 | return 1; | |
820 | } | |
821 | ||
822 | __setup("elevator=", elevator_setup); |