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 | ||
135 | static void elevator_put(struct elevator_type *e) | |
136 | { | |
137 | module_put(e->elevator_owner); | |
138 | } | |
139 | ||
2527d997 JA |
140 | static struct elevator_type *elevator_get(struct request_queue *q, |
141 | const char *name, bool try_loading) | |
1da177e4 | 142 | { |
2824bc93 | 143 | struct elevator_type *e; |
1da177e4 | 144 | |
2a12dcd7 | 145 | spin_lock(&elv_list_lock); |
2824bc93 | 146 | |
68c43f13 | 147 | e = elevator_find(name, q->required_elevator_features); |
21c3c5d2 | 148 | if (!e && try_loading) { |
e1640949 | 149 | spin_unlock(&elv_list_lock); |
490b94be | 150 | request_module("%s-iosched", name); |
e1640949 | 151 | spin_lock(&elv_list_lock); |
68c43f13 | 152 | e = elevator_find(name, q->required_elevator_features); |
e1640949 JA |
153 | } |
154 | ||
2824bc93 TH |
155 | if (e && !try_module_get(e->elevator_owner)) |
156 | e = NULL; | |
157 | ||
2a12dcd7 | 158 | spin_unlock(&elv_list_lock); |
1da177e4 LT |
159 | return e; |
160 | } | |
161 | ||
3d1ab40f AV |
162 | static struct kobj_type elv_ktype; |
163 | ||
d50235b7 | 164 | struct elevator_queue *elevator_alloc(struct request_queue *q, |
165125e1 | 165 | struct elevator_type *e) |
3d1ab40f | 166 | { |
b374d18a | 167 | struct elevator_queue *eq; |
9817064b | 168 | |
c1b511eb | 169 | eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node); |
9817064b | 170 | if (unlikely(!eq)) |
8406a4d5 | 171 | return NULL; |
9817064b | 172 | |
22f746e2 | 173 | eq->type = e; |
f9cb074b | 174 | kobject_init(&eq->kobj, &elv_ktype); |
9817064b | 175 | mutex_init(&eq->sysfs_lock); |
242d98f0 | 176 | hash_init(eq->hash); |
9817064b | 177 | |
3d1ab40f AV |
178 | return eq; |
179 | } | |
d50235b7 | 180 | EXPORT_SYMBOL(elevator_alloc); |
3d1ab40f AV |
181 | |
182 | static void elevator_release(struct kobject *kobj) | |
183 | { | |
b374d18a | 184 | struct elevator_queue *e; |
9817064b | 185 | |
b374d18a | 186 | e = container_of(kobj, struct elevator_queue, kobj); |
22f746e2 | 187 | elevator_put(e->type); |
3d1ab40f AV |
188 | kfree(e); |
189 | } | |
190 | ||
0c6cb3a2 | 191 | void elevator_exit(struct request_queue *q) |
1da177e4 | 192 | { |
0c6cb3a2 CH |
193 | struct elevator_queue *e = q->elevator; |
194 | ||
28883074 CH |
195 | ioc_clear_queue(q); |
196 | blk_mq_sched_free_rqs(q); | |
197 | ||
3d1ab40f | 198 | mutex_lock(&e->sysfs_lock); |
dd1c372d | 199 | blk_mq_exit_sched(q, e); |
3d1ab40f | 200 | mutex_unlock(&e->sysfs_lock); |
1da177e4 | 201 | |
3d1ab40f | 202 | kobject_put(&e->kobj); |
1da177e4 | 203 | } |
2e662b65 | 204 | |
9817064b JA |
205 | static inline void __elv_rqhash_del(struct request *rq) |
206 | { | |
242d98f0 | 207 | hash_del(&rq->hash); |
e8064021 | 208 | rq->rq_flags &= ~RQF_HASHED; |
9817064b JA |
209 | } |
210 | ||
70b3ea05 | 211 | void elv_rqhash_del(struct request_queue *q, struct request *rq) |
9817064b JA |
212 | { |
213 | if (ELV_ON_HASH(rq)) | |
214 | __elv_rqhash_del(rq); | |
215 | } | |
bd166ef1 | 216 | EXPORT_SYMBOL_GPL(elv_rqhash_del); |
9817064b | 217 | |
70b3ea05 | 218 | void elv_rqhash_add(struct request_queue *q, struct request *rq) |
9817064b | 219 | { |
b374d18a | 220 | struct elevator_queue *e = q->elevator; |
9817064b JA |
221 | |
222 | BUG_ON(ELV_ON_HASH(rq)); | |
242d98f0 | 223 | hash_add(e->hash, &rq->hash, rq_hash_key(rq)); |
e8064021 | 224 | rq->rq_flags |= RQF_HASHED; |
9817064b | 225 | } |
bd166ef1 | 226 | EXPORT_SYMBOL_GPL(elv_rqhash_add); |
9817064b | 227 | |
70b3ea05 | 228 | void elv_rqhash_reposition(struct request_queue *q, struct request *rq) |
9817064b JA |
229 | { |
230 | __elv_rqhash_del(rq); | |
231 | elv_rqhash_add(q, rq); | |
232 | } | |
233 | ||
70b3ea05 | 234 | struct request *elv_rqhash_find(struct request_queue *q, sector_t offset) |
9817064b | 235 | { |
b374d18a | 236 | struct elevator_queue *e = q->elevator; |
b67bfe0d | 237 | struct hlist_node *next; |
9817064b JA |
238 | struct request *rq; |
239 | ||
ee89f812 | 240 | hash_for_each_possible_safe(e->hash, rq, next, hash, offset) { |
9817064b JA |
241 | BUG_ON(!ELV_ON_HASH(rq)); |
242 | ||
243 | if (unlikely(!rq_mergeable(rq))) { | |
244 | __elv_rqhash_del(rq); | |
245 | continue; | |
246 | } | |
247 | ||
248 | if (rq_hash_key(rq) == offset) | |
249 | return rq; | |
250 | } | |
251 | ||
252 | return NULL; | |
253 | } | |
254 | ||
2e662b65 JA |
255 | /* |
256 | * RB-tree support functions for inserting/lookup/removal of requests | |
257 | * in a sorted RB tree. | |
258 | */ | |
796d5116 | 259 | void elv_rb_add(struct rb_root *root, struct request *rq) |
2e662b65 JA |
260 | { |
261 | struct rb_node **p = &root->rb_node; | |
262 | struct rb_node *parent = NULL; | |
263 | struct request *__rq; | |
264 | ||
265 | while (*p) { | |
266 | parent = *p; | |
267 | __rq = rb_entry(parent, struct request, rb_node); | |
268 | ||
83096ebf | 269 | if (blk_rq_pos(rq) < blk_rq_pos(__rq)) |
2e662b65 | 270 | p = &(*p)->rb_left; |
796d5116 | 271 | else if (blk_rq_pos(rq) >= blk_rq_pos(__rq)) |
2e662b65 | 272 | p = &(*p)->rb_right; |
2e662b65 JA |
273 | } |
274 | ||
275 | rb_link_node(&rq->rb_node, parent, p); | |
276 | rb_insert_color(&rq->rb_node, root); | |
2e662b65 | 277 | } |
2e662b65 JA |
278 | EXPORT_SYMBOL(elv_rb_add); |
279 | ||
280 | void elv_rb_del(struct rb_root *root, struct request *rq) | |
281 | { | |
282 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); | |
283 | rb_erase(&rq->rb_node, root); | |
284 | RB_CLEAR_NODE(&rq->rb_node); | |
285 | } | |
2e662b65 JA |
286 | EXPORT_SYMBOL(elv_rb_del); |
287 | ||
288 | struct request *elv_rb_find(struct rb_root *root, sector_t sector) | |
289 | { | |
290 | struct rb_node *n = root->rb_node; | |
291 | struct request *rq; | |
292 | ||
293 | while (n) { | |
294 | rq = rb_entry(n, struct request, rb_node); | |
295 | ||
83096ebf | 296 | if (sector < blk_rq_pos(rq)) |
2e662b65 | 297 | n = n->rb_left; |
83096ebf | 298 | else if (sector > blk_rq_pos(rq)) |
2e662b65 JA |
299 | n = n->rb_right; |
300 | else | |
301 | return rq; | |
302 | } | |
303 | ||
304 | return NULL; | |
305 | } | |
2e662b65 JA |
306 | EXPORT_SYMBOL(elv_rb_find); |
307 | ||
34fe7c05 CH |
308 | enum elv_merge elv_merge(struct request_queue *q, struct request **req, |
309 | struct bio *bio) | |
1da177e4 | 310 | { |
b374d18a | 311 | struct elevator_queue *e = q->elevator; |
9817064b | 312 | struct request *__rq; |
06b86245 | 313 | |
488991e2 AB |
314 | /* |
315 | * Levels of merges: | |
316 | * nomerges: No merges at all attempted | |
317 | * noxmerges: Only simple one-hit cache try | |
318 | * merges: All merge tries attempted | |
319 | */ | |
7460d389 | 320 | if (blk_queue_nomerges(q) || !bio_mergeable(bio)) |
488991e2 AB |
321 | return ELEVATOR_NO_MERGE; |
322 | ||
9817064b JA |
323 | /* |
324 | * First try one-hit cache. | |
325 | */ | |
72ef799b | 326 | if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) { |
34fe7c05 CH |
327 | enum elv_merge ret = blk_try_merge(q->last_merge, bio); |
328 | ||
06b86245 TH |
329 | if (ret != ELEVATOR_NO_MERGE) { |
330 | *req = q->last_merge; | |
331 | return ret; | |
332 | } | |
333 | } | |
1da177e4 | 334 | |
488991e2 | 335 | if (blk_queue_noxmerges(q)) |
ac9fafa1 AB |
336 | return ELEVATOR_NO_MERGE; |
337 | ||
9817064b JA |
338 | /* |
339 | * See if our hash lookup can find a potential backmerge. | |
340 | */ | |
4f024f37 | 341 | __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector); |
72ef799b | 342 | if (__rq && elv_bio_merge_ok(__rq, bio)) { |
9817064b | 343 | *req = __rq; |
866663b7 ML |
344 | |
345 | if (blk_discard_mergable(__rq)) | |
346 | return ELEVATOR_DISCARD_MERGE; | |
9817064b JA |
347 | return ELEVATOR_BACK_MERGE; |
348 | } | |
349 | ||
f9cd4bfe JA |
350 | if (e->type->ops.request_merge) |
351 | return e->type->ops.request_merge(q, req, bio); | |
1da177e4 LT |
352 | |
353 | return ELEVATOR_NO_MERGE; | |
354 | } | |
355 | ||
5e84ea3a JA |
356 | /* |
357 | * Attempt to do an insertion back merge. Only check for the case where | |
358 | * we can append 'rq' to an existing request, so we can throw 'rq' away | |
359 | * afterwards. | |
360 | * | |
fd2ef39c JK |
361 | * Returns true if we merged, false otherwise. 'free' will contain all |
362 | * requests that need to be freed. | |
5e84ea3a | 363 | */ |
fd2ef39c JK |
364 | bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq, |
365 | struct list_head *free) | |
5e84ea3a JA |
366 | { |
367 | struct request *__rq; | |
bee0393c | 368 | bool ret; |
5e84ea3a JA |
369 | |
370 | if (blk_queue_nomerges(q)) | |
371 | return false; | |
372 | ||
373 | /* | |
374 | * First try one-hit cache. | |
375 | */ | |
fd2ef39c JK |
376 | if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) { |
377 | list_add(&rq->queuelist, free); | |
5e84ea3a | 378 | return true; |
fd2ef39c | 379 | } |
5e84ea3a JA |
380 | |
381 | if (blk_queue_noxmerges(q)) | |
382 | return false; | |
383 | ||
bee0393c | 384 | ret = false; |
5e84ea3a JA |
385 | /* |
386 | * See if our hash lookup can find a potential backmerge. | |
387 | */ | |
bee0393c SL |
388 | while (1) { |
389 | __rq = elv_rqhash_find(q, blk_rq_pos(rq)); | |
390 | if (!__rq || !blk_attempt_req_merge(q, __rq, rq)) | |
391 | break; | |
392 | ||
fd2ef39c | 393 | list_add(&rq->queuelist, free); |
bee0393c SL |
394 | /* The merged request could be merged with others, try again */ |
395 | ret = true; | |
396 | rq = __rq; | |
397 | } | |
27419322 | 398 | |
bee0393c | 399 | return ret; |
5e84ea3a JA |
400 | } |
401 | ||
34fe7c05 CH |
402 | void elv_merged_request(struct request_queue *q, struct request *rq, |
403 | enum elv_merge type) | |
1da177e4 | 404 | { |
b374d18a | 405 | struct elevator_queue *e = q->elevator; |
1da177e4 | 406 | |
f9cd4bfe JA |
407 | if (e->type->ops.request_merged) |
408 | e->type->ops.request_merged(q, rq, type); | |
06b86245 | 409 | |
2e662b65 JA |
410 | if (type == ELEVATOR_BACK_MERGE) |
411 | elv_rqhash_reposition(q, rq); | |
9817064b | 412 | |
06b86245 | 413 | q->last_merge = rq; |
1da177e4 LT |
414 | } |
415 | ||
165125e1 | 416 | void elv_merge_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
417 | struct request *next) |
418 | { | |
b374d18a | 419 | struct elevator_queue *e = q->elevator; |
bd166ef1 | 420 | |
f9cd4bfe JA |
421 | if (e->type->ops.requests_merged) |
422 | e->type->ops.requests_merged(q, rq, next); | |
06b86245 | 423 | |
9817064b | 424 | elv_rqhash_reposition(q, rq); |
06b86245 | 425 | q->last_merge = rq; |
1da177e4 LT |
426 | } |
427 | ||
165125e1 | 428 | struct request *elv_latter_request(struct request_queue *q, struct request *rq) |
1da177e4 | 429 | { |
b374d18a | 430 | struct elevator_queue *e = q->elevator; |
1da177e4 | 431 | |
f9cd4bfe JA |
432 | if (e->type->ops.next_request) |
433 | return e->type->ops.next_request(q, rq); | |
bd166ef1 | 434 | |
1da177e4 LT |
435 | return NULL; |
436 | } | |
437 | ||
165125e1 | 438 | struct request *elv_former_request(struct request_queue *q, struct request *rq) |
1da177e4 | 439 | { |
b374d18a | 440 | struct elevator_queue *e = q->elevator; |
1da177e4 | 441 | |
f9cd4bfe JA |
442 | if (e->type->ops.former_request) |
443 | return e->type->ops.former_request(q, rq); | |
bd166ef1 | 444 | |
a1ce35fa | 445 | return NULL; |
1da177e4 LT |
446 | } |
447 | ||
3d1ab40f AV |
448 | #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr) |
449 | ||
450 | static ssize_t | |
451 | elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page) | |
1da177e4 | 452 | { |
3d1ab40f | 453 | struct elv_fs_entry *entry = to_elv(attr); |
b374d18a | 454 | struct elevator_queue *e; |
3d1ab40f AV |
455 | ssize_t error; |
456 | ||
457 | if (!entry->show) | |
458 | return -EIO; | |
459 | ||
b374d18a | 460 | e = container_of(kobj, struct elevator_queue, kobj); |
3d1ab40f | 461 | mutex_lock(&e->sysfs_lock); |
22f746e2 | 462 | error = e->type ? entry->show(e, page) : -ENOENT; |
3d1ab40f AV |
463 | mutex_unlock(&e->sysfs_lock); |
464 | return error; | |
465 | } | |
1da177e4 | 466 | |
3d1ab40f AV |
467 | static ssize_t |
468 | elv_attr_store(struct kobject *kobj, struct attribute *attr, | |
469 | const char *page, size_t length) | |
470 | { | |
3d1ab40f | 471 | struct elv_fs_entry *entry = to_elv(attr); |
b374d18a | 472 | struct elevator_queue *e; |
3d1ab40f | 473 | ssize_t error; |
1da177e4 | 474 | |
3d1ab40f AV |
475 | if (!entry->store) |
476 | return -EIO; | |
1da177e4 | 477 | |
b374d18a | 478 | e = container_of(kobj, struct elevator_queue, kobj); |
3d1ab40f | 479 | mutex_lock(&e->sysfs_lock); |
22f746e2 | 480 | error = e->type ? entry->store(e, page, length) : -ENOENT; |
3d1ab40f AV |
481 | mutex_unlock(&e->sysfs_lock); |
482 | return error; | |
483 | } | |
484 | ||
52cf25d0 | 485 | static const struct sysfs_ops elv_sysfs_ops = { |
3d1ab40f AV |
486 | .show = elv_attr_show, |
487 | .store = elv_attr_store, | |
488 | }; | |
489 | ||
490 | static struct kobj_type elv_ktype = { | |
491 | .sysfs_ops = &elv_sysfs_ops, | |
492 | .release = elevator_release, | |
493 | }; | |
494 | ||
cecf5d87 | 495 | int elv_register_queue(struct request_queue *q, bool uevent) |
3d1ab40f | 496 | { |
5a5bafdc | 497 | struct elevator_queue *e = q->elevator; |
3d1ab40f AV |
498 | int error; |
499 | ||
f0c6ae09 YY |
500 | lockdep_assert_held(&q->sysfs_lock); |
501 | ||
b2d6db58 | 502 | error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched"); |
3d1ab40f | 503 | if (!error) { |
22f746e2 | 504 | struct elv_fs_entry *attr = e->type->elevator_attrs; |
3d1ab40f | 505 | if (attr) { |
e572ec7e AV |
506 | while (attr->attr.name) { |
507 | if (sysfs_create_file(&e->kobj, &attr->attr)) | |
3d1ab40f | 508 | break; |
e572ec7e | 509 | attr++; |
3d1ab40f AV |
510 | } |
511 | } | |
cecf5d87 ML |
512 | if (uevent) |
513 | kobject_uevent(&e->kobj, KOBJ_ADD); | |
514 | ||
430c62fb | 515 | e->registered = 1; |
3d1ab40f AV |
516 | } |
517 | return error; | |
1da177e4 | 518 | } |
bc1c1169 | 519 | |
1da177e4 LT |
520 | void elv_unregister_queue(struct request_queue *q) |
521 | { | |
f5ec592d EB |
522 | struct elevator_queue *e = q->elevator; |
523 | ||
f0c6ae09 YY |
524 | lockdep_assert_held(&q->sysfs_lock); |
525 | ||
f5ec592d | 526 | if (e && e->registered) { |
f8fc877d TH |
527 | struct elevator_queue *e = q->elevator; |
528 | ||
529 | kobject_uevent(&e->kobj, KOBJ_REMOVE); | |
530 | kobject_del(&e->kobj); | |
cecf5d87 | 531 | |
f8fc877d TH |
532 | e->registered = 0; |
533 | } | |
1da177e4 LT |
534 | } |
535 | ||
e567bf71 | 536 | int elv_register(struct elevator_type *e) |
1da177e4 | 537 | { |
e42cfb1d DLM |
538 | /* insert_requests and dispatch_request are mandatory */ |
539 | if (WARN_ON_ONCE(!e->ops.insert_requests || !e->ops.dispatch_request)) | |
540 | return -EINVAL; | |
541 | ||
3d3c2379 TH |
542 | /* create icq_cache if requested */ |
543 | if (e->icq_size) { | |
544 | if (WARN_ON(e->icq_size < sizeof(struct io_cq)) || | |
545 | WARN_ON(e->icq_align < __alignof__(struct io_cq))) | |
546 | return -EINVAL; | |
547 | ||
548 | snprintf(e->icq_cache_name, sizeof(e->icq_cache_name), | |
549 | "%s_io_cq", e->elevator_name); | |
550 | e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size, | |
551 | e->icq_align, 0, NULL); | |
552 | if (!e->icq_cache) | |
553 | return -ENOMEM; | |
554 | } | |
555 | ||
556 | /* register, don't allow duplicate names */ | |
2a12dcd7 | 557 | spin_lock(&elv_list_lock); |
68c43f13 | 558 | if (elevator_find(e->elevator_name, 0)) { |
3d3c2379 | 559 | spin_unlock(&elv_list_lock); |
62d2a194 | 560 | kmem_cache_destroy(e->icq_cache); |
3d3c2379 TH |
561 | return -EBUSY; |
562 | } | |
1da177e4 | 563 | list_add_tail(&e->list, &elv_list); |
2a12dcd7 | 564 | spin_unlock(&elv_list_lock); |
1da177e4 | 565 | |
d0b0a81a HT |
566 | printk(KERN_INFO "io scheduler %s registered\n", e->elevator_name); |
567 | ||
3d3c2379 | 568 | return 0; |
1da177e4 LT |
569 | } |
570 | EXPORT_SYMBOL_GPL(elv_register); | |
571 | ||
572 | void elv_unregister(struct elevator_type *e) | |
573 | { | |
3d3c2379 | 574 | /* unregister */ |
2a12dcd7 | 575 | spin_lock(&elv_list_lock); |
1da177e4 | 576 | list_del_init(&e->list); |
2a12dcd7 | 577 | spin_unlock(&elv_list_lock); |
3d3c2379 TH |
578 | |
579 | /* | |
580 | * Destroy icq_cache if it exists. icq's are RCU managed. Make | |
581 | * sure all RCU operations are complete before proceeding. | |
582 | */ | |
583 | if (e->icq_cache) { | |
584 | rcu_barrier(); | |
585 | kmem_cache_destroy(e->icq_cache); | |
586 | e->icq_cache = NULL; | |
587 | } | |
1da177e4 LT |
588 | } |
589 | EXPORT_SYMBOL_GPL(elv_unregister); | |
590 | ||
d48ece20 | 591 | int elevator_switch_mq(struct request_queue *q, |
54d5329d OS |
592 | struct elevator_type *new_e) |
593 | { | |
594 | int ret; | |
595 | ||
14a23498 BVA |
596 | lockdep_assert_held(&q->sysfs_lock); |
597 | ||
54d5329d | 598 | if (q->elevator) { |
f5ec592d | 599 | elv_unregister_queue(q); |
0c6cb3a2 | 600 | elevator_exit(q); |
54d5329d OS |
601 | } |
602 | ||
603 | ret = blk_mq_init_sched(q, new_e); | |
604 | if (ret) | |
605 | goto out; | |
606 | ||
607 | if (new_e) { | |
cecf5d87 | 608 | ret = elv_register_queue(q, true); |
54d5329d | 609 | if (ret) { |
0c6cb3a2 | 610 | elevator_exit(q); |
54d5329d OS |
611 | goto out; |
612 | } | |
613 | } | |
614 | ||
615 | if (new_e) | |
616 | blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name); | |
617 | else | |
618 | blk_add_trace_msg(q, "elv switch: none"); | |
619 | ||
620 | out: | |
54d5329d | 621 | return ret; |
54d5329d OS |
622 | } |
623 | ||
61db437d DLM |
624 | static inline bool elv_support_iosched(struct request_queue *q) |
625 | { | |
6251b754 | 626 | if (!queue_is_mq(q) || |
7a7c5e71 | 627 | (q->tag_set && (q->tag_set->flags & BLK_MQ_F_NO_SCHED))) |
61db437d DLM |
628 | return false; |
629 | return true; | |
630 | } | |
631 | ||
131d08e1 | 632 | /* |
a0958ba7 DLM |
633 | * For single queue devices, default to using mq-deadline. If we have multiple |
634 | * queues or mq-deadline is not available, default to "none". | |
635 | */ | |
636 | static struct elevator_type *elevator_get_default(struct request_queue *q) | |
637 | { | |
90b71980 BVA |
638 | if (q->tag_set && q->tag_set->flags & BLK_MQ_F_NO_SCHED_BY_DEFAULT) |
639 | return NULL; | |
640 | ||
580dca81 | 641 | if (q->nr_hw_queues != 1 && |
079a2e3e | 642 | !blk_mq_is_shared_tags(q->tag_set->flags)) |
a0958ba7 DLM |
643 | return NULL; |
644 | ||
645 | return elevator_get(q, "mq-deadline", false); | |
646 | } | |
647 | ||
648 | /* | |
649 | * Get the first elevator providing the features required by the request queue. | |
650 | * Default to "none" if no matching elevator is found. | |
651 | */ | |
652 | static struct elevator_type *elevator_get_by_features(struct request_queue *q) | |
653 | { | |
a2614255 | 654 | struct elevator_type *e, *found = NULL; |
a0958ba7 DLM |
655 | |
656 | spin_lock(&elv_list_lock); | |
657 | ||
658 | list_for_each_entry(e, &elv_list, list) { | |
659 | if (elv_support_features(e->elevator_features, | |
a2614255 JA |
660 | q->required_elevator_features)) { |
661 | found = e; | |
a0958ba7 | 662 | break; |
a2614255 | 663 | } |
a0958ba7 DLM |
664 | } |
665 | ||
a2614255 JA |
666 | if (found && !try_module_get(found->elevator_owner)) |
667 | found = NULL; | |
a0958ba7 DLM |
668 | |
669 | spin_unlock(&elv_list_lock); | |
a2614255 | 670 | return found; |
a0958ba7 DLM |
671 | } |
672 | ||
673 | /* | |
674 | * For a device queue that has no required features, use the default elevator | |
675 | * settings. Otherwise, use the first elevator available matching the required | |
676 | * features. If no suitable elevator is find or if the chosen elevator | |
677 | * initialization fails, fall back to the "none" elevator (no elevator). | |
131d08e1 | 678 | */ |
954b4a5c | 679 | void elevator_init_mq(struct request_queue *q) |
131d08e1 CH |
680 | { |
681 | struct elevator_type *e; | |
954b4a5c | 682 | int err; |
131d08e1 | 683 | |
61db437d | 684 | if (!elv_support_iosched(q)) |
954b4a5c | 685 | return; |
61db437d | 686 | |
75e6c00f | 687 | WARN_ON_ONCE(blk_queue_registered(q)); |
c48dac13 | 688 | |
131d08e1 | 689 | if (unlikely(q->elevator)) |
954b4a5c | 690 | return; |
131d08e1 | 691 | |
a0958ba7 DLM |
692 | if (!q->required_elevator_features) |
693 | e = elevator_get_default(q); | |
694 | else | |
695 | e = elevator_get_by_features(q); | |
131d08e1 | 696 | if (!e) |
954b4a5c | 697 | return; |
131d08e1 | 698 | |
245a489e ML |
699 | /* |
700 | * We are called before adding disk, when there isn't any FS I/O, | |
701 | * so freezing queue plus canceling dispatch work is enough to | |
702 | * drain any dispatch activities originated from passthrough | |
703 | * requests, then no need to quiesce queue which may add long boot | |
704 | * latency, especially when lots of disks are involved. | |
705 | */ | |
737eb78e | 706 | blk_mq_freeze_queue(q); |
245a489e | 707 | blk_mq_cancel_work_sync(q); |
737eb78e | 708 | |
131d08e1 | 709 | err = blk_mq_init_sched(q, e); |
737eb78e | 710 | |
737eb78e DLM |
711 | blk_mq_unfreeze_queue(q); |
712 | ||
954b4a5c DLM |
713 | if (err) { |
714 | pr_warn("\"%s\" elevator initialization failed, " | |
715 | "falling back to \"none\"\n", e->elevator_name); | |
131d08e1 | 716 | elevator_put(e); |
954b4a5c | 717 | } |
131d08e1 | 718 | } |
131d08e1 | 719 | |
1da177e4 LT |
720 | /* |
721 | * switch to new_e io scheduler. be careful not to introduce deadlocks - | |
722 | * we don't free the old io scheduler, before we have allocated what we | |
723 | * need for the new one. this way we have a chance of going back to the old | |
cb98fc8b | 724 | * one, if the new one fails init for some reason. |
1da177e4 | 725 | */ |
165125e1 | 726 | static int elevator_switch(struct request_queue *q, struct elevator_type *new_e) |
1da177e4 | 727 | { |
e8989fae | 728 | int err; |
1da177e4 | 729 | |
14a23498 BVA |
730 | lockdep_assert_held(&q->sysfs_lock); |
731 | ||
a1ce35fa JA |
732 | blk_mq_freeze_queue(q); |
733 | blk_mq_quiesce_queue(q); | |
4722dc52 | 734 | |
a1ce35fa | 735 | err = elevator_switch_mq(q, new_e); |
1da177e4 | 736 | |
a1ce35fa JA |
737 | blk_mq_unquiesce_queue(q); |
738 | blk_mq_unfreeze_queue(q); | |
75ad23bc | 739 | |
5dd531a0 | 740 | return err; |
1da177e4 LT |
741 | } |
742 | ||
5dd531a0 JA |
743 | /* |
744 | * Switch this queue to the given IO scheduler. | |
745 | */ | |
7c8a3679 | 746 | static int __elevator_change(struct request_queue *q, const char *name) |
1da177e4 LT |
747 | { |
748 | char elevator_name[ELV_NAME_MAX]; | |
749 | struct elevator_type *e; | |
750 | ||
e9a823fb | 751 | /* Make sure queue is not in the middle of being removed */ |
58c898ba | 752 | if (!blk_queue_registered(q)) |
e9a823fb DJ |
753 | return -ENOENT; |
754 | ||
bd166ef1 JA |
755 | /* |
756 | * Special case for mq, turn off scheduling | |
757 | */ | |
fbd72127 AZ |
758 | if (!strncmp(name, "none", 4)) { |
759 | if (!q->elevator) | |
760 | return 0; | |
bd166ef1 | 761 | return elevator_switch(q, NULL); |
fbd72127 | 762 | } |
cd43e26f | 763 | |
ee2e992c | 764 | strlcpy(elevator_name, name, sizeof(elevator_name)); |
2527d997 | 765 | e = elevator_get(q, strstrip(elevator_name), true); |
340ff321 | 766 | if (!e) |
1da177e4 | 767 | return -EINVAL; |
1da177e4 | 768 | |
68c43f13 DLM |
769 | if (q->elevator && |
770 | elevator_match(q->elevator->type, elevator_name, 0)) { | |
2ca7d93b | 771 | elevator_put(e); |
5dd531a0 | 772 | return 0; |
2ca7d93b | 773 | } |
1da177e4 | 774 | |
5dd531a0 JA |
775 | return elevator_switch(q, e); |
776 | } | |
7c8a3679 | 777 | |
5dd531a0 JA |
778 | ssize_t elv_iosched_store(struct request_queue *q, const char *name, |
779 | size_t count) | |
780 | { | |
781 | int ret; | |
782 | ||
6251b754 | 783 | if (!elv_support_iosched(q)) |
5dd531a0 JA |
784 | return count; |
785 | ||
7c8a3679 | 786 | ret = __elevator_change(q, name); |
5dd531a0 JA |
787 | if (!ret) |
788 | return count; | |
789 | ||
5dd531a0 | 790 | return ret; |
1da177e4 LT |
791 | } |
792 | ||
165125e1 | 793 | ssize_t elv_iosched_show(struct request_queue *q, char *name) |
1da177e4 | 794 | { |
b374d18a | 795 | struct elevator_queue *e = q->elevator; |
bd166ef1 | 796 | struct elevator_type *elv = NULL; |
70cee26e | 797 | struct elevator_type *__e; |
1da177e4 LT |
798 | int len = 0; |
799 | ||
344e9ffc | 800 | if (!queue_is_mq(q)) |
cd43e26f MP |
801 | return sprintf(name, "none\n"); |
802 | ||
bd166ef1 JA |
803 | if (!q->elevator) |
804 | len += sprintf(name+len, "[none] "); | |
805 | else | |
806 | elv = e->type; | |
cd43e26f | 807 | |
2a12dcd7 | 808 | spin_lock(&elv_list_lock); |
70cee26e | 809 | list_for_each_entry(__e, &elv_list, list) { |
68c43f13 | 810 | if (elv && elevator_match(elv, __e->elevator_name, 0)) { |
1da177e4 | 811 | len += sprintf(name+len, "[%s] ", elv->elevator_name); |
bd166ef1 JA |
812 | continue; |
813 | } | |
68c43f13 DLM |
814 | if (elv_support_iosched(q) && |
815 | elevator_match(__e, __e->elevator_name, | |
816 | q->required_elevator_features)) | |
1da177e4 LT |
817 | len += sprintf(name+len, "%s ", __e->elevator_name); |
818 | } | |
2a12dcd7 | 819 | spin_unlock(&elv_list_lock); |
1da177e4 | 820 | |
344e9ffc | 821 | if (q->elevator) |
bd166ef1 JA |
822 | len += sprintf(name+len, "none"); |
823 | ||
1da177e4 LT |
824 | len += sprintf(len+name, "\n"); |
825 | return len; | |
826 | } | |
827 | ||
165125e1 JA |
828 | struct request *elv_rb_former_request(struct request_queue *q, |
829 | struct request *rq) | |
2e662b65 JA |
830 | { |
831 | struct rb_node *rbprev = rb_prev(&rq->rb_node); | |
832 | ||
833 | if (rbprev) | |
834 | return rb_entry_rq(rbprev); | |
835 | ||
836 | return NULL; | |
837 | } | |
2e662b65 JA |
838 | EXPORT_SYMBOL(elv_rb_former_request); |
839 | ||
165125e1 JA |
840 | struct request *elv_rb_latter_request(struct request_queue *q, |
841 | struct request *rq) | |
2e662b65 JA |
842 | { |
843 | struct rb_node *rbnext = rb_next(&rq->rb_node); | |
844 | ||
845 | if (rbnext) | |
846 | return rb_entry_rq(rbnext); | |
847 | ||
848 | return NULL; | |
849 | } | |
2e662b65 | 850 | EXPORT_SYMBOL(elv_rb_latter_request); |
f8db3835 JK |
851 | |
852 | static int __init elevator_setup(char *str) | |
853 | { | |
854 | pr_warn("Kernel parameter elevator= does not have any effect anymore.\n" | |
855 | "Please use sysfs to set IO scheduler for individual devices.\n"); | |
856 | return 1; | |
857 | } | |
858 | ||
859 | __setup("elevator=", elevator_setup); |