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 | |
68c43f13 DLM |
86 | static inline bool elv_support_features(unsigned int elv_features, |
87 | unsigned int required_features) | |
8ac0d9a8 | 88 | { |
68c43f13 DLM |
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 | * | |
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 DLM |
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; | |
8ac0d9a8 JA |
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 | ||
68c43f13 DLM |
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. | |
2527d997 | 121 | */ |
68c43f13 DLM |
122 | static struct elevator_type *elevator_find(const char *name, |
123 | unsigned int required_features) | |
1da177e4 | 124 | { |
a22b169d | 125 | struct elevator_type *e; |
1da177e4 | 126 | |
70cee26e | 127 | list_for_each_entry(e, &elv_list, list) { |
68c43f13 | 128 | if (elevator_match(e, name, required_features)) |
a22b169d | 129 | return e; |
1da177e4 | 130 | } |
1da177e4 | 131 | |
a22b169d | 132 | return NULL; |
1da177e4 LT |
133 | } |
134 | ||
81eaca44 CH |
135 | static struct elevator_type *elevator_find_get(struct request_queue *q, |
136 | const char *name) | |
1da177e4 | 137 | { |
2824bc93 | 138 | struct elevator_type *e; |
1da177e4 | 139 | |
2a12dcd7 | 140 | spin_lock(&elv_list_lock); |
68c43f13 | 141 | e = elevator_find(name, q->required_elevator_features); |
dd6f7f17 | 142 | if (e && !elevator_tryget(e)) |
2824bc93 | 143 | e = NULL; |
2a12dcd7 | 144 | spin_unlock(&elv_list_lock); |
1da177e4 LT |
145 | return e; |
146 | } | |
147 | ||
3d1ab40f AV |
148 | static struct kobj_type elv_ktype; |
149 | ||
d50235b7 | 150 | struct elevator_queue *elevator_alloc(struct request_queue *q, |
165125e1 | 151 | struct elevator_type *e) |
3d1ab40f | 152 | { |
b374d18a | 153 | struct elevator_queue *eq; |
9817064b | 154 | |
c1b511eb | 155 | eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node); |
9817064b | 156 | if (unlikely(!eq)) |
8406a4d5 | 157 | return NULL; |
9817064b | 158 | |
8ed40ee3 | 159 | __elevator_get(e); |
22f746e2 | 160 | eq->type = e; |
f9cb074b | 161 | kobject_init(&eq->kobj, &elv_ktype); |
9817064b | 162 | mutex_init(&eq->sysfs_lock); |
242d98f0 | 163 | hash_init(eq->hash); |
9817064b | 164 | |
3d1ab40f AV |
165 | return eq; |
166 | } | |
d50235b7 | 167 | EXPORT_SYMBOL(elevator_alloc); |
3d1ab40f AV |
168 | |
169 | static void elevator_release(struct kobject *kobj) | |
170 | { | |
b374d18a | 171 | struct elevator_queue *e; |
9817064b | 172 | |
b374d18a | 173 | e = container_of(kobj, struct elevator_queue, kobj); |
22f746e2 | 174 | elevator_put(e->type); |
3d1ab40f AV |
175 | kfree(e); |
176 | } | |
177 | ||
0c6cb3a2 | 178 | void elevator_exit(struct request_queue *q) |
1da177e4 | 179 | { |
0c6cb3a2 CH |
180 | struct elevator_queue *e = q->elevator; |
181 | ||
28883074 CH |
182 | ioc_clear_queue(q); |
183 | blk_mq_sched_free_rqs(q); | |
184 | ||
3d1ab40f | 185 | mutex_lock(&e->sysfs_lock); |
dd1c372d | 186 | blk_mq_exit_sched(q, e); |
3d1ab40f | 187 | mutex_unlock(&e->sysfs_lock); |
1da177e4 | 188 | |
3d1ab40f | 189 | kobject_put(&e->kobj); |
1da177e4 | 190 | } |
2e662b65 | 191 | |
9817064b JA |
192 | static inline void __elv_rqhash_del(struct request *rq) |
193 | { | |
242d98f0 | 194 | hash_del(&rq->hash); |
e8064021 | 195 | rq->rq_flags &= ~RQF_HASHED; |
9817064b JA |
196 | } |
197 | ||
70b3ea05 | 198 | void elv_rqhash_del(struct request_queue *q, struct request *rq) |
9817064b JA |
199 | { |
200 | if (ELV_ON_HASH(rq)) | |
201 | __elv_rqhash_del(rq); | |
202 | } | |
bd166ef1 | 203 | EXPORT_SYMBOL_GPL(elv_rqhash_del); |
9817064b | 204 | |
70b3ea05 | 205 | void elv_rqhash_add(struct request_queue *q, struct request *rq) |
9817064b | 206 | { |
b374d18a | 207 | struct elevator_queue *e = q->elevator; |
9817064b JA |
208 | |
209 | BUG_ON(ELV_ON_HASH(rq)); | |
242d98f0 | 210 | hash_add(e->hash, &rq->hash, rq_hash_key(rq)); |
e8064021 | 211 | rq->rq_flags |= RQF_HASHED; |
9817064b | 212 | } |
bd166ef1 | 213 | EXPORT_SYMBOL_GPL(elv_rqhash_add); |
9817064b | 214 | |
70b3ea05 | 215 | void elv_rqhash_reposition(struct request_queue *q, struct request *rq) |
9817064b JA |
216 | { |
217 | __elv_rqhash_del(rq); | |
218 | elv_rqhash_add(q, rq); | |
219 | } | |
220 | ||
70b3ea05 | 221 | struct request *elv_rqhash_find(struct request_queue *q, sector_t offset) |
9817064b | 222 | { |
b374d18a | 223 | struct elevator_queue *e = q->elevator; |
b67bfe0d | 224 | struct hlist_node *next; |
9817064b JA |
225 | struct request *rq; |
226 | ||
ee89f812 | 227 | hash_for_each_possible_safe(e->hash, rq, next, hash, offset) { |
9817064b JA |
228 | BUG_ON(!ELV_ON_HASH(rq)); |
229 | ||
230 | if (unlikely(!rq_mergeable(rq))) { | |
231 | __elv_rqhash_del(rq); | |
232 | continue; | |
233 | } | |
234 | ||
235 | if (rq_hash_key(rq) == offset) | |
236 | return rq; | |
237 | } | |
238 | ||
239 | return NULL; | |
240 | } | |
241 | ||
2e662b65 JA |
242 | /* |
243 | * RB-tree support functions for inserting/lookup/removal of requests | |
244 | * in a sorted RB tree. | |
245 | */ | |
796d5116 | 246 | void elv_rb_add(struct rb_root *root, struct request *rq) |
2e662b65 JA |
247 | { |
248 | struct rb_node **p = &root->rb_node; | |
249 | struct rb_node *parent = NULL; | |
250 | struct request *__rq; | |
251 | ||
252 | while (*p) { | |
253 | parent = *p; | |
254 | __rq = rb_entry(parent, struct request, rb_node); | |
255 | ||
83096ebf | 256 | if (blk_rq_pos(rq) < blk_rq_pos(__rq)) |
2e662b65 | 257 | p = &(*p)->rb_left; |
796d5116 | 258 | else if (blk_rq_pos(rq) >= blk_rq_pos(__rq)) |
2e662b65 | 259 | p = &(*p)->rb_right; |
2e662b65 JA |
260 | } |
261 | ||
262 | rb_link_node(&rq->rb_node, parent, p); | |
263 | rb_insert_color(&rq->rb_node, root); | |
2e662b65 | 264 | } |
2e662b65 JA |
265 | EXPORT_SYMBOL(elv_rb_add); |
266 | ||
267 | void elv_rb_del(struct rb_root *root, struct request *rq) | |
268 | { | |
269 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); | |
270 | rb_erase(&rq->rb_node, root); | |
271 | RB_CLEAR_NODE(&rq->rb_node); | |
272 | } | |
2e662b65 JA |
273 | EXPORT_SYMBOL(elv_rb_del); |
274 | ||
275 | struct request *elv_rb_find(struct rb_root *root, sector_t sector) | |
276 | { | |
277 | struct rb_node *n = root->rb_node; | |
278 | struct request *rq; | |
279 | ||
280 | while (n) { | |
281 | rq = rb_entry(n, struct request, rb_node); | |
282 | ||
83096ebf | 283 | if (sector < blk_rq_pos(rq)) |
2e662b65 | 284 | n = n->rb_left; |
83096ebf | 285 | else if (sector > blk_rq_pos(rq)) |
2e662b65 JA |
286 | n = n->rb_right; |
287 | else | |
288 | return rq; | |
289 | } | |
290 | ||
291 | return NULL; | |
292 | } | |
2e662b65 JA |
293 | EXPORT_SYMBOL(elv_rb_find); |
294 | ||
34fe7c05 CH |
295 | enum elv_merge elv_merge(struct request_queue *q, struct request **req, |
296 | struct bio *bio) | |
1da177e4 | 297 | { |
b374d18a | 298 | struct elevator_queue *e = q->elevator; |
9817064b | 299 | struct request *__rq; |
06b86245 | 300 | |
488991e2 AB |
301 | /* |
302 | * Levels of merges: | |
303 | * nomerges: No merges at all attempted | |
304 | * noxmerges: Only simple one-hit cache try | |
305 | * merges: All merge tries attempted | |
306 | */ | |
7460d389 | 307 | if (blk_queue_nomerges(q) || !bio_mergeable(bio)) |
488991e2 AB |
308 | return ELEVATOR_NO_MERGE; |
309 | ||
9817064b JA |
310 | /* |
311 | * First try one-hit cache. | |
312 | */ | |
72ef799b | 313 | if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) { |
34fe7c05 CH |
314 | enum elv_merge ret = blk_try_merge(q->last_merge, bio); |
315 | ||
06b86245 TH |
316 | if (ret != ELEVATOR_NO_MERGE) { |
317 | *req = q->last_merge; | |
318 | return ret; | |
319 | } | |
320 | } | |
1da177e4 | 321 | |
488991e2 | 322 | if (blk_queue_noxmerges(q)) |
ac9fafa1 AB |
323 | return ELEVATOR_NO_MERGE; |
324 | ||
9817064b JA |
325 | /* |
326 | * See if our hash lookup can find a potential backmerge. | |
327 | */ | |
4f024f37 | 328 | __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector); |
72ef799b | 329 | if (__rq && elv_bio_merge_ok(__rq, bio)) { |
9817064b | 330 | *req = __rq; |
866663b7 ML |
331 | |
332 | if (blk_discard_mergable(__rq)) | |
333 | return ELEVATOR_DISCARD_MERGE; | |
9817064b JA |
334 | return ELEVATOR_BACK_MERGE; |
335 | } | |
336 | ||
f9cd4bfe JA |
337 | if (e->type->ops.request_merge) |
338 | return e->type->ops.request_merge(q, req, bio); | |
1da177e4 LT |
339 | |
340 | return ELEVATOR_NO_MERGE; | |
341 | } | |
342 | ||
5e84ea3a JA |
343 | /* |
344 | * Attempt to do an insertion back merge. Only check for the case where | |
345 | * we can append 'rq' to an existing request, so we can throw 'rq' away | |
346 | * afterwards. | |
347 | * | |
fd2ef39c JK |
348 | * Returns true if we merged, false otherwise. 'free' will contain all |
349 | * requests that need to be freed. | |
5e84ea3a | 350 | */ |
fd2ef39c JK |
351 | bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq, |
352 | struct list_head *free) | |
5e84ea3a JA |
353 | { |
354 | struct request *__rq; | |
bee0393c | 355 | bool ret; |
5e84ea3a JA |
356 | |
357 | if (blk_queue_nomerges(q)) | |
358 | return false; | |
359 | ||
360 | /* | |
361 | * First try one-hit cache. | |
362 | */ | |
fd2ef39c JK |
363 | if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) { |
364 | list_add(&rq->queuelist, free); | |
5e84ea3a | 365 | return true; |
fd2ef39c | 366 | } |
5e84ea3a JA |
367 | |
368 | if (blk_queue_noxmerges(q)) | |
369 | return false; | |
370 | ||
bee0393c | 371 | ret = false; |
5e84ea3a JA |
372 | /* |
373 | * See if our hash lookup can find a potential backmerge. | |
374 | */ | |
bee0393c SL |
375 | while (1) { |
376 | __rq = elv_rqhash_find(q, blk_rq_pos(rq)); | |
377 | if (!__rq || !blk_attempt_req_merge(q, __rq, rq)) | |
378 | break; | |
379 | ||
fd2ef39c | 380 | list_add(&rq->queuelist, free); |
bee0393c SL |
381 | /* The merged request could be merged with others, try again */ |
382 | ret = true; | |
383 | rq = __rq; | |
384 | } | |
27419322 | 385 | |
bee0393c | 386 | return ret; |
5e84ea3a JA |
387 | } |
388 | ||
34fe7c05 CH |
389 | void elv_merged_request(struct request_queue *q, struct request *rq, |
390 | enum elv_merge type) | |
1da177e4 | 391 | { |
b374d18a | 392 | struct elevator_queue *e = q->elevator; |
1da177e4 | 393 | |
f9cd4bfe JA |
394 | if (e->type->ops.request_merged) |
395 | e->type->ops.request_merged(q, rq, type); | |
06b86245 | 396 | |
2e662b65 JA |
397 | if (type == ELEVATOR_BACK_MERGE) |
398 | elv_rqhash_reposition(q, rq); | |
9817064b | 399 | |
06b86245 | 400 | q->last_merge = rq; |
1da177e4 LT |
401 | } |
402 | ||
165125e1 | 403 | void elv_merge_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
404 | struct request *next) |
405 | { | |
b374d18a | 406 | struct elevator_queue *e = q->elevator; |
bd166ef1 | 407 | |
f9cd4bfe JA |
408 | if (e->type->ops.requests_merged) |
409 | e->type->ops.requests_merged(q, rq, next); | |
06b86245 | 410 | |
9817064b | 411 | elv_rqhash_reposition(q, rq); |
06b86245 | 412 | q->last_merge = rq; |
1da177e4 LT |
413 | } |
414 | ||
165125e1 | 415 | struct request *elv_latter_request(struct request_queue *q, struct request *rq) |
1da177e4 | 416 | { |
b374d18a | 417 | struct elevator_queue *e = q->elevator; |
1da177e4 | 418 | |
f9cd4bfe JA |
419 | if (e->type->ops.next_request) |
420 | return e->type->ops.next_request(q, rq); | |
bd166ef1 | 421 | |
1da177e4 LT |
422 | return NULL; |
423 | } | |
424 | ||
165125e1 | 425 | struct request *elv_former_request(struct request_queue *q, struct request *rq) |
1da177e4 | 426 | { |
b374d18a | 427 | struct elevator_queue *e = q->elevator; |
1da177e4 | 428 | |
f9cd4bfe JA |
429 | if (e->type->ops.former_request) |
430 | return e->type->ops.former_request(q, rq); | |
bd166ef1 | 431 | |
a1ce35fa | 432 | return NULL; |
1da177e4 LT |
433 | } |
434 | ||
3d1ab40f AV |
435 | #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr) |
436 | ||
437 | static ssize_t | |
438 | elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page) | |
1da177e4 | 439 | { |
3d1ab40f | 440 | struct elv_fs_entry *entry = to_elv(attr); |
b374d18a | 441 | struct elevator_queue *e; |
3d1ab40f AV |
442 | ssize_t error; |
443 | ||
444 | if (!entry->show) | |
445 | return -EIO; | |
446 | ||
b374d18a | 447 | e = container_of(kobj, struct elevator_queue, kobj); |
3d1ab40f | 448 | mutex_lock(&e->sysfs_lock); |
22f746e2 | 449 | error = e->type ? entry->show(e, page) : -ENOENT; |
3d1ab40f AV |
450 | mutex_unlock(&e->sysfs_lock); |
451 | return error; | |
452 | } | |
1da177e4 | 453 | |
3d1ab40f AV |
454 | static ssize_t |
455 | elv_attr_store(struct kobject *kobj, struct attribute *attr, | |
456 | const char *page, size_t length) | |
457 | { | |
3d1ab40f | 458 | struct elv_fs_entry *entry = to_elv(attr); |
b374d18a | 459 | struct elevator_queue *e; |
3d1ab40f | 460 | ssize_t error; |
1da177e4 | 461 | |
3d1ab40f AV |
462 | if (!entry->store) |
463 | return -EIO; | |
1da177e4 | 464 | |
b374d18a | 465 | e = container_of(kobj, struct elevator_queue, kobj); |
3d1ab40f | 466 | mutex_lock(&e->sysfs_lock); |
22f746e2 | 467 | error = e->type ? entry->store(e, page, length) : -ENOENT; |
3d1ab40f AV |
468 | mutex_unlock(&e->sysfs_lock); |
469 | return error; | |
470 | } | |
471 | ||
52cf25d0 | 472 | static const struct sysfs_ops elv_sysfs_ops = { |
3d1ab40f AV |
473 | .show = elv_attr_show, |
474 | .store = elv_attr_store, | |
475 | }; | |
476 | ||
477 | static struct kobj_type elv_ktype = { | |
478 | .sysfs_ops = &elv_sysfs_ops, | |
479 | .release = elevator_release, | |
480 | }; | |
481 | ||
cecf5d87 | 482 | int elv_register_queue(struct request_queue *q, bool uevent) |
3d1ab40f | 483 | { |
5a5bafdc | 484 | struct elevator_queue *e = q->elevator; |
3d1ab40f AV |
485 | int error; |
486 | ||
f0c6ae09 YY |
487 | lockdep_assert_held(&q->sysfs_lock); |
488 | ||
b2d6db58 | 489 | error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched"); |
3d1ab40f | 490 | if (!error) { |
22f746e2 | 491 | struct elv_fs_entry *attr = e->type->elevator_attrs; |
3d1ab40f | 492 | if (attr) { |
e572ec7e AV |
493 | while (attr->attr.name) { |
494 | if (sysfs_create_file(&e->kobj, &attr->attr)) | |
3d1ab40f | 495 | break; |
e572ec7e | 496 | attr++; |
3d1ab40f AV |
497 | } |
498 | } | |
cecf5d87 ML |
499 | if (uevent) |
500 | kobject_uevent(&e->kobj, KOBJ_ADD); | |
501 | ||
181d0663 | 502 | set_bit(ELEVATOR_FLAG_REGISTERED, &e->flags); |
3d1ab40f AV |
503 | } |
504 | return error; | |
1da177e4 | 505 | } |
bc1c1169 | 506 | |
1da177e4 LT |
507 | void elv_unregister_queue(struct request_queue *q) |
508 | { | |
f5ec592d EB |
509 | struct elevator_queue *e = q->elevator; |
510 | ||
f0c6ae09 YY |
511 | lockdep_assert_held(&q->sysfs_lock); |
512 | ||
181d0663 | 513 | if (e && test_and_clear_bit(ELEVATOR_FLAG_REGISTERED, &e->flags)) { |
f8fc877d TH |
514 | kobject_uevent(&e->kobj, KOBJ_REMOVE); |
515 | kobject_del(&e->kobj); | |
f8fc877d | 516 | } |
1da177e4 LT |
517 | } |
518 | ||
e567bf71 | 519 | int elv_register(struct elevator_type *e) |
1da177e4 | 520 | { |
e42cfb1d DLM |
521 | /* insert_requests and dispatch_request are mandatory */ |
522 | if (WARN_ON_ONCE(!e->ops.insert_requests || !e->ops.dispatch_request)) | |
523 | return -EINVAL; | |
524 | ||
3d3c2379 TH |
525 | /* create icq_cache if requested */ |
526 | if (e->icq_size) { | |
527 | if (WARN_ON(e->icq_size < sizeof(struct io_cq)) || | |
528 | WARN_ON(e->icq_align < __alignof__(struct io_cq))) | |
529 | return -EINVAL; | |
530 | ||
531 | snprintf(e->icq_cache_name, sizeof(e->icq_cache_name), | |
532 | "%s_io_cq", e->elevator_name); | |
533 | e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size, | |
534 | e->icq_align, 0, NULL); | |
535 | if (!e->icq_cache) | |
536 | return -ENOMEM; | |
537 | } | |
538 | ||
539 | /* register, don't allow duplicate names */ | |
2a12dcd7 | 540 | spin_lock(&elv_list_lock); |
68c43f13 | 541 | if (elevator_find(e->elevator_name, 0)) { |
3d3c2379 | 542 | spin_unlock(&elv_list_lock); |
62d2a194 | 543 | kmem_cache_destroy(e->icq_cache); |
3d3c2379 TH |
544 | return -EBUSY; |
545 | } | |
1da177e4 | 546 | list_add_tail(&e->list, &elv_list); |
2a12dcd7 | 547 | spin_unlock(&elv_list_lock); |
1da177e4 | 548 | |
d0b0a81a HT |
549 | printk(KERN_INFO "io scheduler %s registered\n", e->elevator_name); |
550 | ||
3d3c2379 | 551 | return 0; |
1da177e4 LT |
552 | } |
553 | EXPORT_SYMBOL_GPL(elv_register); | |
554 | ||
555 | void elv_unregister(struct elevator_type *e) | |
556 | { | |
3d3c2379 | 557 | /* unregister */ |
2a12dcd7 | 558 | spin_lock(&elv_list_lock); |
1da177e4 | 559 | list_del_init(&e->list); |
2a12dcd7 | 560 | spin_unlock(&elv_list_lock); |
3d3c2379 TH |
561 | |
562 | /* | |
563 | * Destroy icq_cache if it exists. icq's are RCU managed. Make | |
564 | * sure all RCU operations are complete before proceeding. | |
565 | */ | |
566 | if (e->icq_cache) { | |
567 | rcu_barrier(); | |
568 | kmem_cache_destroy(e->icq_cache); | |
569 | e->icq_cache = NULL; | |
570 | } | |
1da177e4 LT |
571 | } |
572 | EXPORT_SYMBOL_GPL(elv_unregister); | |
573 | ||
8237c01f | 574 | static int elevator_switch_mq(struct request_queue *q, |
54d5329d OS |
575 | struct elevator_type *new_e) |
576 | { | |
577 | int ret; | |
578 | ||
14a23498 BVA |
579 | lockdep_assert_held(&q->sysfs_lock); |
580 | ||
54d5329d | 581 | if (q->elevator) { |
f5ec592d | 582 | elv_unregister_queue(q); |
0c6cb3a2 | 583 | elevator_exit(q); |
54d5329d OS |
584 | } |
585 | ||
586 | ret = blk_mq_init_sched(q, new_e); | |
587 | if (ret) | |
588 | goto out; | |
589 | ||
590 | if (new_e) { | |
cecf5d87 | 591 | ret = elv_register_queue(q, true); |
54d5329d | 592 | if (ret) { |
0c6cb3a2 | 593 | elevator_exit(q); |
54d5329d OS |
594 | goto out; |
595 | } | |
596 | } | |
597 | ||
598 | if (new_e) | |
599 | blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name); | |
600 | else | |
601 | blk_add_trace_msg(q, "elv switch: none"); | |
602 | ||
603 | out: | |
54d5329d | 604 | return ret; |
54d5329d OS |
605 | } |
606 | ||
61db437d DLM |
607 | static inline bool elv_support_iosched(struct request_queue *q) |
608 | { | |
6251b754 | 609 | if (!queue_is_mq(q) || |
7a7c5e71 | 610 | (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED))) |
61db437d DLM |
611 | return false; |
612 | return true; | |
613 | } | |
614 | ||
131d08e1 | 615 | /* |
a0958ba7 DLM |
616 | * For single queue devices, default to using mq-deadline. If we have multiple |
617 | * queues or mq-deadline is not available, default to "none". | |
618 | */ | |
619 | static struct elevator_type *elevator_get_default(struct request_queue *q) | |
620 | { | |
90b71980 BVA |
621 | if (q->tag_set && q->tag_set->flags & BLK_MQ_F_NO_SCHED_BY_DEFAULT) |
622 | return NULL; | |
623 | ||
580dca81 | 624 | if (q->nr_hw_queues != 1 && |
079a2e3e | 625 | !blk_mq_is_shared_tags(q->tag_set->flags)) |
a0958ba7 DLM |
626 | return NULL; |
627 | ||
81eaca44 | 628 | return elevator_find_get(q, "mq-deadline"); |
a0958ba7 DLM |
629 | } |
630 | ||
631 | /* | |
632 | * Get the first elevator providing the features required by the request queue. | |
633 | * Default to "none" if no matching elevator is found. | |
634 | */ | |
635 | static struct elevator_type *elevator_get_by_features(struct request_queue *q) | |
636 | { | |
a2614255 | 637 | struct elevator_type *e, *found = NULL; |
a0958ba7 DLM |
638 | |
639 | spin_lock(&elv_list_lock); | |
640 | ||
641 | list_for_each_entry(e, &elv_list, list) { | |
642 | if (elv_support_features(e->elevator_features, | |
a2614255 JA |
643 | q->required_elevator_features)) { |
644 | found = e; | |
a0958ba7 | 645 | break; |
a2614255 | 646 | } |
a0958ba7 DLM |
647 | } |
648 | ||
dd6f7f17 | 649 | if (found && !elevator_tryget(found)) |
a2614255 | 650 | found = NULL; |
a0958ba7 DLM |
651 | |
652 | spin_unlock(&elv_list_lock); | |
a2614255 | 653 | return found; |
a0958ba7 DLM |
654 | } |
655 | ||
656 | /* | |
657 | * For a device queue that has no required features, use the default elevator | |
658 | * settings. Otherwise, use the first elevator available matching the required | |
659 | * features. If no suitable elevator is find or if the chosen elevator | |
660 | * initialization fails, fall back to the "none" elevator (no elevator). | |
131d08e1 | 661 | */ |
954b4a5c | 662 | void elevator_init_mq(struct request_queue *q) |
131d08e1 CH |
663 | { |
664 | struct elevator_type *e; | |
954b4a5c | 665 | int err; |
131d08e1 | 666 | |
61db437d | 667 | if (!elv_support_iosched(q)) |
954b4a5c | 668 | return; |
61db437d | 669 | |
75e6c00f | 670 | WARN_ON_ONCE(blk_queue_registered(q)); |
c48dac13 | 671 | |
131d08e1 | 672 | if (unlikely(q->elevator)) |
954b4a5c | 673 | return; |
131d08e1 | 674 | |
a0958ba7 DLM |
675 | if (!q->required_elevator_features) |
676 | e = elevator_get_default(q); | |
677 | else | |
678 | e = elevator_get_by_features(q); | |
131d08e1 | 679 | if (!e) |
954b4a5c | 680 | return; |
131d08e1 | 681 | |
245a489e ML |
682 | /* |
683 | * We are called before adding disk, when there isn't any FS I/O, | |
684 | * so freezing queue plus canceling dispatch work is enough to | |
685 | * drain any dispatch activities originated from passthrough | |
686 | * requests, then no need to quiesce queue which may add long boot | |
687 | * latency, especially when lots of disks are involved. | |
688 | */ | |
737eb78e | 689 | blk_mq_freeze_queue(q); |
245a489e | 690 | blk_mq_cancel_work_sync(q); |
737eb78e | 691 | |
131d08e1 | 692 | err = blk_mq_init_sched(q, e); |
737eb78e | 693 | |
737eb78e DLM |
694 | blk_mq_unfreeze_queue(q); |
695 | ||
954b4a5c DLM |
696 | if (err) { |
697 | pr_warn("\"%s\" elevator initialization failed, " | |
698 | "falling back to \"none\"\n", e->elevator_name); | |
954b4a5c | 699 | } |
8ed40ee3 JC |
700 | |
701 | elevator_put(e); | |
131d08e1 | 702 | } |
131d08e1 | 703 | |
1da177e4 LT |
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 | |
cb98fc8b | 708 | * one, if the new one fails init for some reason. |
1da177e4 | 709 | */ |
8237c01f | 710 | int elevator_switch(struct request_queue *q, struct elevator_type *new_e) |
1da177e4 | 711 | { |
e8989fae | 712 | int err; |
1da177e4 | 713 | |
14a23498 BVA |
714 | lockdep_assert_held(&q->sysfs_lock); |
715 | ||
a1ce35fa JA |
716 | blk_mq_freeze_queue(q); |
717 | blk_mq_quiesce_queue(q); | |
4722dc52 | 718 | |
a1ce35fa | 719 | err = elevator_switch_mq(q, new_e); |
1da177e4 | 720 | |
a1ce35fa JA |
721 | blk_mq_unquiesce_queue(q); |
722 | blk_mq_unfreeze_queue(q); | |
75ad23bc | 723 | |
5dd531a0 | 724 | return err; |
1da177e4 LT |
725 | } |
726 | ||
5dd531a0 JA |
727 | /* |
728 | * Switch this queue to the given IO scheduler. | |
729 | */ | |
58367c8a | 730 | static int elevator_change(struct request_queue *q, const char *elevator_name) |
1da177e4 | 731 | { |
1da177e4 | 732 | struct elevator_type *e; |
8ed40ee3 | 733 | int ret; |
1da177e4 | 734 | |
e9a823fb | 735 | /* Make sure queue is not in the middle of being removed */ |
58c898ba | 736 | if (!blk_queue_registered(q)) |
e9a823fb DJ |
737 | return -ENOENT; |
738 | ||
bd166ef1 JA |
739 | /* |
740 | * Special case for mq, turn off scheduling | |
741 | */ | |
58367c8a | 742 | if (!strncmp(elevator_name, "none", 4)) { |
fbd72127 AZ |
743 | if (!q->elevator) |
744 | return 0; | |
bd166ef1 | 745 | return elevator_switch(q, NULL); |
fbd72127 | 746 | } |
cd43e26f | 747 | |
b54c2ad9 JC |
748 | if (q->elevator && elevator_match(q->elevator->type, elevator_name, 0)) |
749 | return 0; | |
750 | ||
81eaca44 CH |
751 | e = elevator_find_get(q, elevator_name); |
752 | if (!e) { | |
753 | request_module("%s-iosched", elevator_name); | |
754 | e = elevator_find_get(q, elevator_name); | |
755 | if (!e) | |
756 | return -EINVAL; | |
757 | } | |
8ed40ee3 JC |
758 | ret = elevator_switch(q, e); |
759 | elevator_put(e); | |
760 | return ret; | |
5dd531a0 | 761 | } |
7c8a3679 | 762 | |
58367c8a | 763 | ssize_t elv_iosched_store(struct request_queue *q, const char *buf, |
5dd531a0 JA |
764 | size_t count) |
765 | { | |
58367c8a | 766 | char elevator_name[ELV_NAME_MAX]; |
5dd531a0 JA |
767 | int ret; |
768 | ||
6251b754 | 769 | if (!elv_support_iosched(q)) |
5dd531a0 JA |
770 | return count; |
771 | ||
58367c8a CH |
772 | strlcpy(elevator_name, buf, sizeof(elevator_name)); |
773 | ret = elevator_change(q, strstrip(elevator_name)); | |
5dd531a0 JA |
774 | if (!ret) |
775 | return count; | |
5dd531a0 | 776 | return ret; |
1da177e4 LT |
777 | } |
778 | ||
165125e1 | 779 | ssize_t elv_iosched_show(struct request_queue *q, char *name) |
1da177e4 | 780 | { |
16095af2 CH |
781 | struct elevator_queue *eq = q->elevator; |
782 | struct elevator_type *cur = NULL, *e; | |
1da177e4 LT |
783 | int len = 0; |
784 | ||
aae2a643 | 785 | if (!elv_support_iosched(q)) |
cd43e26f MP |
786 | return sprintf(name, "none\n"); |
787 | ||
bd166ef1 JA |
788 | if (!q->elevator) |
789 | len += sprintf(name+len, "[none] "); | |
790 | else | |
16095af2 | 791 | cur = eq->type; |
cd43e26f | 792 | |
2a12dcd7 | 793 | spin_lock(&elv_list_lock); |
16095af2 CH |
794 | list_for_each_entry(e, &elv_list, list) { |
795 | if (cur && elevator_match(cur, e->elevator_name, 0)) { | |
796 | len += sprintf(name+len, "[%s] ", cur->elevator_name); | |
bd166ef1 JA |
797 | continue; |
798 | } | |
16095af2 | 799 | if (elevator_match(e, e->elevator_name, |
68c43f13 | 800 | q->required_elevator_features)) |
16095af2 | 801 | len += sprintf(name+len, "%s ", e->elevator_name); |
1da177e4 | 802 | } |
2a12dcd7 | 803 | spin_unlock(&elv_list_lock); |
1da177e4 | 804 | |
344e9ffc | 805 | if (q->elevator) |
bd166ef1 JA |
806 | len += sprintf(name+len, "none"); |
807 | ||
1da177e4 LT |
808 | len += sprintf(len+name, "\n"); |
809 | return len; | |
810 | } | |
811 | ||
165125e1 JA |
812 | struct request *elv_rb_former_request(struct request_queue *q, |
813 | struct request *rq) | |
2e662b65 JA |
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 | } | |
2e662b65 JA |
822 | EXPORT_SYMBOL(elv_rb_former_request); |
823 | ||
165125e1 JA |
824 | struct request *elv_rb_latter_request(struct request_queue *q, |
825 | struct request *rq) | |
2e662b65 JA |
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 | } | |
2e662b65 | 834 | EXPORT_SYMBOL(elv_rb_latter_request); |
f8db3835 JK |
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); |