Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
1da177e4 LT |
2 | * Block device elevator/IO-scheduler. |
3 | * | |
4 | * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE | |
5 | * | |
0fe23479 | 6 | * 30042000 Jens Axboe <axboe@kernel.dk> : |
1da177e4 LT |
7 | * |
8 | * Split the elevator a bit so that it is possible to choose a different | |
9 | * one or even write a new "plug in". There are three pieces: | |
10 | * - elevator_fn, inserts a new request in the queue list | |
11 | * - elevator_merge_fn, decides whether a new buffer can be merged with | |
12 | * an existing request | |
13 | * - elevator_dequeue_fn, called when a request is taken off the active list | |
14 | * | |
15 | * 20082000 Dave Jones <davej@suse.de> : | |
16 | * Removed tests for max-bomb-segments, which was breaking elvtune | |
17 | * when run without -bN | |
18 | * | |
19 | * Jens: | |
20 | * - Rework again to work with bio instead of buffer_heads | |
21 | * - loose bi_dev comparisons, partition handling is right now | |
22 | * - completely modularize elevator setup and teardown | |
23 | * | |
24 | */ | |
25 | #include <linux/kernel.h> | |
26 | #include <linux/fs.h> | |
27 | #include <linux/blkdev.h> | |
28 | #include <linux/elevator.h> | |
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> |
1da177e4 | 37 | |
55782138 LZ |
38 | #include <trace/events/block.h> |
39 | ||
242f9dcb | 40 | #include "blk.h" |
72e06c25 | 41 | #include "blk-cgroup.h" |
242f9dcb | 42 | |
1da177e4 LT |
43 | static DEFINE_SPINLOCK(elv_list_lock); |
44 | static LIST_HEAD(elv_list); | |
45 | ||
9817064b JA |
46 | /* |
47 | * Merge hash stuff. | |
48 | */ | |
49 | static const int elv_hash_shift = 6; | |
50 | #define ELV_HASH_BLOCK(sec) ((sec) >> 3) | |
4eb166d9 JA |
51 | #define ELV_HASH_FN(sec) \ |
52 | (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift)) | |
9817064b | 53 | #define ELV_HASH_ENTRIES (1 << elv_hash_shift) |
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 | */ | |
60 | static int elv_iosched_allow_merge(struct request *rq, struct bio *bio) | |
61 | { | |
165125e1 | 62 | struct request_queue *q = rq->q; |
b374d18a | 63 | struct elevator_queue *e = q->elevator; |
da775265 | 64 | |
22f746e2 TH |
65 | if (e->type->ops.elevator_allow_merge_fn) |
66 | return e->type->ops.elevator_allow_merge_fn(q, rq, bio); | |
da775265 JA |
67 | |
68 | return 1; | |
69 | } | |
70 | ||
1da177e4 LT |
71 | /* |
72 | * can we safely merge with this request? | |
73 | */ | |
050c8ea8 | 74 | bool elv_rq_merge_ok(struct request *rq, struct bio *bio) |
1da177e4 | 75 | { |
050c8ea8 | 76 | if (!blk_rq_merge_ok(rq, bio)) |
7ba1ba12 MP |
77 | return 0; |
78 | ||
da775265 JA |
79 | if (!elv_iosched_allow_merge(rq, bio)) |
80 | return 0; | |
1da177e4 | 81 | |
da775265 | 82 | return 1; |
1da177e4 LT |
83 | } |
84 | EXPORT_SYMBOL(elv_rq_merge_ok); | |
85 | ||
1da177e4 LT |
86 | static struct elevator_type *elevator_find(const char *name) |
87 | { | |
a22b169d | 88 | struct elevator_type *e; |
1da177e4 | 89 | |
70cee26e | 90 | list_for_each_entry(e, &elv_list, list) { |
a22b169d VT |
91 | if (!strcmp(e->elevator_name, name)) |
92 | return e; | |
1da177e4 | 93 | } |
1da177e4 | 94 | |
a22b169d | 95 | return NULL; |
1da177e4 LT |
96 | } |
97 | ||
98 | static void elevator_put(struct elevator_type *e) | |
99 | { | |
100 | module_put(e->elevator_owner); | |
101 | } | |
102 | ||
103 | static struct elevator_type *elevator_get(const char *name) | |
104 | { | |
2824bc93 | 105 | struct elevator_type *e; |
1da177e4 | 106 | |
2a12dcd7 | 107 | spin_lock(&elv_list_lock); |
2824bc93 TH |
108 | |
109 | e = elevator_find(name); | |
e1640949 | 110 | if (!e) { |
e1640949 | 111 | spin_unlock(&elv_list_lock); |
490b94be | 112 | request_module("%s-iosched", name); |
e1640949 JA |
113 | spin_lock(&elv_list_lock); |
114 | e = elevator_find(name); | |
115 | } | |
116 | ||
2824bc93 TH |
117 | if (e && !try_module_get(e->elevator_owner)) |
118 | e = NULL; | |
119 | ||
2a12dcd7 | 120 | spin_unlock(&elv_list_lock); |
1da177e4 LT |
121 | |
122 | return e; | |
123 | } | |
124 | ||
484fc254 | 125 | static char chosen_elevator[ELV_NAME_MAX]; |
1da177e4 | 126 | |
5f003976 | 127 | static int __init elevator_setup(char *str) |
1da177e4 | 128 | { |
752a3b79 CE |
129 | /* |
130 | * Be backwards-compatible with previous kernels, so users | |
131 | * won't get the wrong elevator. | |
132 | */ | |
492af635 | 133 | strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1); |
9b41046c | 134 | return 1; |
1da177e4 LT |
135 | } |
136 | ||
137 | __setup("elevator=", elevator_setup); | |
138 | ||
3d1ab40f AV |
139 | static struct kobj_type elv_ktype; |
140 | ||
b374d18a | 141 | static struct elevator_queue *elevator_alloc(struct request_queue *q, |
165125e1 | 142 | struct elevator_type *e) |
3d1ab40f | 143 | { |
b374d18a | 144 | struct elevator_queue *eq; |
9817064b JA |
145 | int i; |
146 | ||
b374d18a | 147 | eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node); |
9817064b JA |
148 | if (unlikely(!eq)) |
149 | goto err; | |
150 | ||
22f746e2 | 151 | eq->type = e; |
f9cb074b | 152 | kobject_init(&eq->kobj, &elv_ktype); |
9817064b JA |
153 | mutex_init(&eq->sysfs_lock); |
154 | ||
b5deef90 JA |
155 | eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES, |
156 | GFP_KERNEL, q->node); | |
9817064b JA |
157 | if (!eq->hash) |
158 | goto err; | |
159 | ||
160 | for (i = 0; i < ELV_HASH_ENTRIES; i++) | |
161 | INIT_HLIST_HEAD(&eq->hash[i]); | |
162 | ||
3d1ab40f | 163 | return eq; |
9817064b JA |
164 | err: |
165 | kfree(eq); | |
166 | elevator_put(e); | |
167 | return NULL; | |
3d1ab40f AV |
168 | } |
169 | ||
170 | static void elevator_release(struct kobject *kobj) | |
171 | { | |
b374d18a | 172 | struct elevator_queue *e; |
9817064b | 173 | |
b374d18a | 174 | e = container_of(kobj, struct elevator_queue, kobj); |
22f746e2 | 175 | elevator_put(e->type); |
9817064b | 176 | kfree(e->hash); |
3d1ab40f AV |
177 | kfree(e); |
178 | } | |
179 | ||
165125e1 | 180 | int elevator_init(struct request_queue *q, char *name) |
1da177e4 LT |
181 | { |
182 | struct elevator_type *e = NULL; | |
f8fc877d | 183 | int err; |
1da177e4 | 184 | |
1abec4fd MS |
185 | if (unlikely(q->elevator)) |
186 | return 0; | |
187 | ||
cb98fc8b TH |
188 | INIT_LIST_HEAD(&q->queue_head); |
189 | q->last_merge = NULL; | |
190 | q->end_sector = 0; | |
191 | q->boundary_rq = NULL; | |
cb98fc8b | 192 | |
4eb166d9 JA |
193 | if (name) { |
194 | e = elevator_get(name); | |
195 | if (!e) | |
196 | return -EINVAL; | |
197 | } | |
1da177e4 | 198 | |
4eb166d9 JA |
199 | if (!e && *chosen_elevator) { |
200 | e = elevator_get(chosen_elevator); | |
201 | if (!e) | |
202 | printk(KERN_ERR "I/O scheduler %s not found\n", | |
203 | chosen_elevator); | |
204 | } | |
248d5ca5 | 205 | |
4eb166d9 JA |
206 | if (!e) { |
207 | e = elevator_get(CONFIG_DEFAULT_IOSCHED); | |
208 | if (!e) { | |
209 | printk(KERN_ERR | |
210 | "Default I/O scheduler not found. " \ | |
211 | "Using noop.\n"); | |
212 | e = elevator_get("noop"); | |
213 | } | |
5f003976 ND |
214 | } |
215 | ||
5a5bafdc TH |
216 | q->elevator = elevator_alloc(q, e); |
217 | if (!q->elevator) | |
1da177e4 | 218 | return -ENOMEM; |
1da177e4 | 219 | |
b2fab5ac | 220 | err = e->ops.elevator_init_fn(q); |
f8fc877d | 221 | if (err) { |
5a5bafdc | 222 | kobject_put(&q->elevator->kobj); |
f8fc877d | 223 | return err; |
bc1c1169 | 224 | } |
1da177e4 | 225 | |
1abec4fd | 226 | return 0; |
1da177e4 | 227 | } |
2e662b65 JA |
228 | EXPORT_SYMBOL(elevator_init); |
229 | ||
b374d18a | 230 | void elevator_exit(struct elevator_queue *e) |
1da177e4 | 231 | { |
3d1ab40f | 232 | mutex_lock(&e->sysfs_lock); |
22f746e2 TH |
233 | if (e->type->ops.elevator_exit_fn) |
234 | e->type->ops.elevator_exit_fn(e); | |
3d1ab40f | 235 | mutex_unlock(&e->sysfs_lock); |
1da177e4 | 236 | |
3d1ab40f | 237 | kobject_put(&e->kobj); |
1da177e4 | 238 | } |
2e662b65 JA |
239 | EXPORT_SYMBOL(elevator_exit); |
240 | ||
9817064b JA |
241 | static inline void __elv_rqhash_del(struct request *rq) |
242 | { | |
243 | hlist_del_init(&rq->hash); | |
244 | } | |
245 | ||
165125e1 | 246 | static void elv_rqhash_del(struct request_queue *q, struct request *rq) |
9817064b JA |
247 | { |
248 | if (ELV_ON_HASH(rq)) | |
249 | __elv_rqhash_del(rq); | |
250 | } | |
251 | ||
165125e1 | 252 | static void elv_rqhash_add(struct request_queue *q, struct request *rq) |
9817064b | 253 | { |
b374d18a | 254 | struct elevator_queue *e = q->elevator; |
9817064b JA |
255 | |
256 | BUG_ON(ELV_ON_HASH(rq)); | |
257 | hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]); | |
258 | } | |
259 | ||
165125e1 | 260 | static void elv_rqhash_reposition(struct request_queue *q, struct request *rq) |
9817064b JA |
261 | { |
262 | __elv_rqhash_del(rq); | |
263 | elv_rqhash_add(q, rq); | |
264 | } | |
265 | ||
165125e1 | 266 | static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset) |
9817064b | 267 | { |
b374d18a | 268 | struct elevator_queue *e = q->elevator; |
9817064b JA |
269 | struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)]; |
270 | struct hlist_node *entry, *next; | |
271 | struct request *rq; | |
272 | ||
273 | hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) { | |
274 | BUG_ON(!ELV_ON_HASH(rq)); | |
275 | ||
276 | if (unlikely(!rq_mergeable(rq))) { | |
277 | __elv_rqhash_del(rq); | |
278 | continue; | |
279 | } | |
280 | ||
281 | if (rq_hash_key(rq) == offset) | |
282 | return rq; | |
283 | } | |
284 | ||
285 | return NULL; | |
286 | } | |
287 | ||
2e662b65 JA |
288 | /* |
289 | * RB-tree support functions for inserting/lookup/removal of requests | |
290 | * in a sorted RB tree. | |
291 | */ | |
796d5116 | 292 | void elv_rb_add(struct rb_root *root, struct request *rq) |
2e662b65 JA |
293 | { |
294 | struct rb_node **p = &root->rb_node; | |
295 | struct rb_node *parent = NULL; | |
296 | struct request *__rq; | |
297 | ||
298 | while (*p) { | |
299 | parent = *p; | |
300 | __rq = rb_entry(parent, struct request, rb_node); | |
301 | ||
83096ebf | 302 | if (blk_rq_pos(rq) < blk_rq_pos(__rq)) |
2e662b65 | 303 | p = &(*p)->rb_left; |
796d5116 | 304 | else if (blk_rq_pos(rq) >= blk_rq_pos(__rq)) |
2e662b65 | 305 | p = &(*p)->rb_right; |
2e662b65 JA |
306 | } |
307 | ||
308 | rb_link_node(&rq->rb_node, parent, p); | |
309 | rb_insert_color(&rq->rb_node, root); | |
2e662b65 | 310 | } |
2e662b65 JA |
311 | EXPORT_SYMBOL(elv_rb_add); |
312 | ||
313 | void elv_rb_del(struct rb_root *root, struct request *rq) | |
314 | { | |
315 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); | |
316 | rb_erase(&rq->rb_node, root); | |
317 | RB_CLEAR_NODE(&rq->rb_node); | |
318 | } | |
2e662b65 JA |
319 | EXPORT_SYMBOL(elv_rb_del); |
320 | ||
321 | struct request *elv_rb_find(struct rb_root *root, sector_t sector) | |
322 | { | |
323 | struct rb_node *n = root->rb_node; | |
324 | struct request *rq; | |
325 | ||
326 | while (n) { | |
327 | rq = rb_entry(n, struct request, rb_node); | |
328 | ||
83096ebf | 329 | if (sector < blk_rq_pos(rq)) |
2e662b65 | 330 | n = n->rb_left; |
83096ebf | 331 | else if (sector > blk_rq_pos(rq)) |
2e662b65 JA |
332 | n = n->rb_right; |
333 | else | |
334 | return rq; | |
335 | } | |
336 | ||
337 | return NULL; | |
338 | } | |
2e662b65 JA |
339 | EXPORT_SYMBOL(elv_rb_find); |
340 | ||
8922e16c TH |
341 | /* |
342 | * Insert rq into dispatch queue of q. Queue lock must be held on | |
dbe7f76d | 343 | * entry. rq is sort instead into the dispatch queue. To be used by |
2e662b65 | 344 | * specific elevators. |
8922e16c | 345 | */ |
165125e1 | 346 | void elv_dispatch_sort(struct request_queue *q, struct request *rq) |
8922e16c TH |
347 | { |
348 | sector_t boundary; | |
8922e16c | 349 | struct list_head *entry; |
4eb166d9 | 350 | int stop_flags; |
8922e16c | 351 | |
06b86245 TH |
352 | if (q->last_merge == rq) |
353 | q->last_merge = NULL; | |
9817064b JA |
354 | |
355 | elv_rqhash_del(q, rq); | |
356 | ||
15853af9 | 357 | q->nr_sorted--; |
06b86245 | 358 | |
1b47f531 | 359 | boundary = q->end_sector; |
02e031cb | 360 | stop_flags = REQ_SOFTBARRIER | REQ_STARTED; |
8922e16c TH |
361 | list_for_each_prev(entry, &q->queue_head) { |
362 | struct request *pos = list_entry_rq(entry); | |
363 | ||
33659ebb CH |
364 | if ((rq->cmd_flags & REQ_DISCARD) != |
365 | (pos->cmd_flags & REQ_DISCARD)) | |
e17fc0a1 | 366 | break; |
783660b2 JA |
367 | if (rq_data_dir(rq) != rq_data_dir(pos)) |
368 | break; | |
4eb166d9 | 369 | if (pos->cmd_flags & stop_flags) |
8922e16c | 370 | break; |
83096ebf TH |
371 | if (blk_rq_pos(rq) >= boundary) { |
372 | if (blk_rq_pos(pos) < boundary) | |
8922e16c TH |
373 | continue; |
374 | } else { | |
83096ebf | 375 | if (blk_rq_pos(pos) >= boundary) |
8922e16c TH |
376 | break; |
377 | } | |
83096ebf | 378 | if (blk_rq_pos(rq) >= blk_rq_pos(pos)) |
8922e16c TH |
379 | break; |
380 | } | |
381 | ||
382 | list_add(&rq->queuelist, entry); | |
383 | } | |
2e662b65 JA |
384 | EXPORT_SYMBOL(elv_dispatch_sort); |
385 | ||
9817064b | 386 | /* |
2e662b65 JA |
387 | * Insert rq into dispatch queue of q. Queue lock must be held on |
388 | * entry. rq is added to the back of the dispatch queue. To be used by | |
389 | * specific elevators. | |
9817064b JA |
390 | */ |
391 | void elv_dispatch_add_tail(struct request_queue *q, struct request *rq) | |
392 | { | |
393 | if (q->last_merge == rq) | |
394 | q->last_merge = NULL; | |
395 | ||
396 | elv_rqhash_del(q, rq); | |
397 | ||
398 | q->nr_sorted--; | |
399 | ||
400 | q->end_sector = rq_end_sector(rq); | |
401 | q->boundary_rq = rq; | |
402 | list_add_tail(&rq->queuelist, &q->queue_head); | |
403 | } | |
2e662b65 JA |
404 | EXPORT_SYMBOL(elv_dispatch_add_tail); |
405 | ||
165125e1 | 406 | int elv_merge(struct request_queue *q, struct request **req, struct bio *bio) |
1da177e4 | 407 | { |
b374d18a | 408 | struct elevator_queue *e = q->elevator; |
9817064b | 409 | struct request *__rq; |
06b86245 TH |
410 | int ret; |
411 | ||
488991e2 AB |
412 | /* |
413 | * Levels of merges: | |
414 | * nomerges: No merges at all attempted | |
415 | * noxmerges: Only simple one-hit cache try | |
416 | * merges: All merge tries attempted | |
417 | */ | |
418 | if (blk_queue_nomerges(q)) | |
419 | return ELEVATOR_NO_MERGE; | |
420 | ||
9817064b JA |
421 | /* |
422 | * First try one-hit cache. | |
423 | */ | |
050c8ea8 TH |
424 | if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) { |
425 | ret = blk_try_merge(q->last_merge, bio); | |
06b86245 TH |
426 | if (ret != ELEVATOR_NO_MERGE) { |
427 | *req = q->last_merge; | |
428 | return ret; | |
429 | } | |
430 | } | |
1da177e4 | 431 | |
488991e2 | 432 | if (blk_queue_noxmerges(q)) |
ac9fafa1 AB |
433 | return ELEVATOR_NO_MERGE; |
434 | ||
9817064b JA |
435 | /* |
436 | * See if our hash lookup can find a potential backmerge. | |
437 | */ | |
438 | __rq = elv_rqhash_find(q, bio->bi_sector); | |
439 | if (__rq && elv_rq_merge_ok(__rq, bio)) { | |
440 | *req = __rq; | |
441 | return ELEVATOR_BACK_MERGE; | |
442 | } | |
443 | ||
22f746e2 TH |
444 | if (e->type->ops.elevator_merge_fn) |
445 | return e->type->ops.elevator_merge_fn(q, req, bio); | |
1da177e4 LT |
446 | |
447 | return ELEVATOR_NO_MERGE; | |
448 | } | |
449 | ||
5e84ea3a JA |
450 | /* |
451 | * Attempt to do an insertion back merge. Only check for the case where | |
452 | * we can append 'rq' to an existing request, so we can throw 'rq' away | |
453 | * afterwards. | |
454 | * | |
455 | * Returns true if we merged, false otherwise | |
456 | */ | |
457 | static bool elv_attempt_insert_merge(struct request_queue *q, | |
458 | struct request *rq) | |
459 | { | |
460 | struct request *__rq; | |
bee0393c | 461 | bool ret; |
5e84ea3a JA |
462 | |
463 | if (blk_queue_nomerges(q)) | |
464 | return false; | |
465 | ||
466 | /* | |
467 | * First try one-hit cache. | |
468 | */ | |
469 | if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) | |
470 | return true; | |
471 | ||
472 | if (blk_queue_noxmerges(q)) | |
473 | return false; | |
474 | ||
bee0393c | 475 | ret = false; |
5e84ea3a JA |
476 | /* |
477 | * See if our hash lookup can find a potential backmerge. | |
478 | */ | |
bee0393c SL |
479 | while (1) { |
480 | __rq = elv_rqhash_find(q, blk_rq_pos(rq)); | |
481 | if (!__rq || !blk_attempt_req_merge(q, __rq, rq)) | |
482 | break; | |
483 | ||
484 | /* The merged request could be merged with others, try again */ | |
485 | ret = true; | |
486 | rq = __rq; | |
487 | } | |
27419322 | 488 | |
bee0393c | 489 | return ret; |
5e84ea3a JA |
490 | } |
491 | ||
165125e1 | 492 | void elv_merged_request(struct request_queue *q, struct request *rq, int type) |
1da177e4 | 493 | { |
b374d18a | 494 | struct elevator_queue *e = q->elevator; |
1da177e4 | 495 | |
22f746e2 TH |
496 | if (e->type->ops.elevator_merged_fn) |
497 | e->type->ops.elevator_merged_fn(q, rq, type); | |
06b86245 | 498 | |
2e662b65 JA |
499 | if (type == ELEVATOR_BACK_MERGE) |
500 | elv_rqhash_reposition(q, rq); | |
9817064b | 501 | |
06b86245 | 502 | q->last_merge = rq; |
1da177e4 LT |
503 | } |
504 | ||
165125e1 | 505 | void elv_merge_requests(struct request_queue *q, struct request *rq, |
1da177e4 LT |
506 | struct request *next) |
507 | { | |
b374d18a | 508 | struct elevator_queue *e = q->elevator; |
5e84ea3a | 509 | const int next_sorted = next->cmd_flags & REQ_SORTED; |
1da177e4 | 510 | |
22f746e2 TH |
511 | if (next_sorted && e->type->ops.elevator_merge_req_fn) |
512 | e->type->ops.elevator_merge_req_fn(q, rq, next); | |
06b86245 | 513 | |
9817064b | 514 | elv_rqhash_reposition(q, rq); |
9817064b | 515 | |
5e84ea3a JA |
516 | if (next_sorted) { |
517 | elv_rqhash_del(q, next); | |
518 | q->nr_sorted--; | |
519 | } | |
520 | ||
06b86245 | 521 | q->last_merge = rq; |
1da177e4 LT |
522 | } |
523 | ||
812d4026 DS |
524 | void elv_bio_merged(struct request_queue *q, struct request *rq, |
525 | struct bio *bio) | |
526 | { | |
527 | struct elevator_queue *e = q->elevator; | |
528 | ||
22f746e2 TH |
529 | if (e->type->ops.elevator_bio_merged_fn) |
530 | e->type->ops.elevator_bio_merged_fn(q, rq, bio); | |
812d4026 DS |
531 | } |
532 | ||
165125e1 | 533 | void elv_requeue_request(struct request_queue *q, struct request *rq) |
1da177e4 | 534 | { |
1da177e4 LT |
535 | /* |
536 | * it already went through dequeue, we need to decrement the | |
537 | * in_flight count again | |
538 | */ | |
8922e16c | 539 | if (blk_account_rq(rq)) { |
0a7ae2ff | 540 | q->in_flight[rq_is_sync(rq)]--; |
33659ebb | 541 | if (rq->cmd_flags & REQ_SORTED) |
cad97516 | 542 | elv_deactivate_rq(q, rq); |
8922e16c | 543 | } |
1da177e4 | 544 | |
4aff5e23 | 545 | rq->cmd_flags &= ~REQ_STARTED; |
1da177e4 | 546 | |
b710a480 | 547 | __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE); |
1da177e4 LT |
548 | } |
549 | ||
26308eab | 550 | void elv_drain_elevator(struct request_queue *q) |
15853af9 TH |
551 | { |
552 | static int printed; | |
e3c78ca5 TH |
553 | |
554 | lockdep_assert_held(q->queue_lock); | |
555 | ||
22f746e2 | 556 | while (q->elevator->type->ops.elevator_dispatch_fn(q, 1)) |
15853af9 | 557 | ; |
e3c78ca5 | 558 | if (q->nr_sorted && printed++ < 10) { |
15853af9 TH |
559 | printk(KERN_ERR "%s: forced dispatching is broken " |
560 | "(nr_sorted=%u), please report this\n", | |
22f746e2 | 561 | q->elevator->type->elevator_name, q->nr_sorted); |
15853af9 TH |
562 | } |
563 | } | |
564 | ||
b710a480 | 565 | void __elv_add_request(struct request_queue *q, struct request *rq, int where) |
1da177e4 | 566 | { |
5f3ea37c | 567 | trace_block_rq_insert(q, rq); |
2056a782 | 568 | |
1da177e4 LT |
569 | rq->q = q; |
570 | ||
b710a480 JA |
571 | if (rq->cmd_flags & REQ_SOFTBARRIER) { |
572 | /* barriers are scheduling boundary, update end_sector */ | |
e2a60da7 | 573 | if (rq->cmd_type == REQ_TYPE_FS) { |
b710a480 JA |
574 | q->end_sector = rq_end_sector(rq); |
575 | q->boundary_rq = rq; | |
576 | } | |
577 | } else if (!(rq->cmd_flags & REQ_ELVPRIV) && | |
3aa72873 JA |
578 | (where == ELEVATOR_INSERT_SORT || |
579 | where == ELEVATOR_INSERT_SORT_MERGE)) | |
b710a480 JA |
580 | where = ELEVATOR_INSERT_BACK; |
581 | ||
8922e16c | 582 | switch (where) { |
28e7d184 | 583 | case ELEVATOR_INSERT_REQUEUE: |
8922e16c | 584 | case ELEVATOR_INSERT_FRONT: |
4aff5e23 | 585 | rq->cmd_flags |= REQ_SOFTBARRIER; |
8922e16c TH |
586 | list_add(&rq->queuelist, &q->queue_head); |
587 | break; | |
588 | ||
589 | case ELEVATOR_INSERT_BACK: | |
4aff5e23 | 590 | rq->cmd_flags |= REQ_SOFTBARRIER; |
15853af9 | 591 | elv_drain_elevator(q); |
8922e16c TH |
592 | list_add_tail(&rq->queuelist, &q->queue_head); |
593 | /* | |
594 | * We kick the queue here for the following reasons. | |
595 | * - The elevator might have returned NULL previously | |
596 | * to delay requests and returned them now. As the | |
597 | * queue wasn't empty before this request, ll_rw_blk | |
598 | * won't run the queue on return, resulting in hang. | |
599 | * - Usually, back inserted requests won't be merged | |
600 | * with anything. There's no point in delaying queue | |
601 | * processing. | |
602 | */ | |
24ecfbe2 | 603 | __blk_run_queue(q); |
8922e16c TH |
604 | break; |
605 | ||
5e84ea3a JA |
606 | case ELEVATOR_INSERT_SORT_MERGE: |
607 | /* | |
608 | * If we succeed in merging this request with one in the | |
609 | * queue already, we are done - rq has now been freed, | |
610 | * so no need to do anything further. | |
611 | */ | |
612 | if (elv_attempt_insert_merge(q, rq)) | |
613 | break; | |
8922e16c | 614 | case ELEVATOR_INSERT_SORT: |
e2a60da7 | 615 | BUG_ON(rq->cmd_type != REQ_TYPE_FS); |
4aff5e23 | 616 | rq->cmd_flags |= REQ_SORTED; |
15853af9 | 617 | q->nr_sorted++; |
9817064b JA |
618 | if (rq_mergeable(rq)) { |
619 | elv_rqhash_add(q, rq); | |
620 | if (!q->last_merge) | |
621 | q->last_merge = rq; | |
622 | } | |
623 | ||
ca23509f TH |
624 | /* |
625 | * Some ioscheds (cfq) run q->request_fn directly, so | |
626 | * rq cannot be accessed after calling | |
627 | * elevator_add_req_fn. | |
628 | */ | |
22f746e2 | 629 | q->elevator->type->ops.elevator_add_req_fn(q, rq); |
8922e16c TH |
630 | break; |
631 | ||
ae1b1539 TH |
632 | case ELEVATOR_INSERT_FLUSH: |
633 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
634 | blk_insert_flush(rq); | |
635 | break; | |
8922e16c TH |
636 | default: |
637 | printk(KERN_ERR "%s: bad insertion point %d\n", | |
24c03d47 | 638 | __func__, where); |
8922e16c TH |
639 | BUG(); |
640 | } | |
1da177e4 | 641 | } |
2e662b65 JA |
642 | EXPORT_SYMBOL(__elv_add_request); |
643 | ||
7eaceacc | 644 | void elv_add_request(struct request_queue *q, struct request *rq, int where) |
1da177e4 LT |
645 | { |
646 | unsigned long flags; | |
647 | ||
648 | spin_lock_irqsave(q->queue_lock, flags); | |
7eaceacc | 649 | __elv_add_request(q, rq, where); |
1da177e4 LT |
650 | spin_unlock_irqrestore(q->queue_lock, flags); |
651 | } | |
2e662b65 JA |
652 | EXPORT_SYMBOL(elv_add_request); |
653 | ||
165125e1 | 654 | struct request *elv_latter_request(struct request_queue *q, struct request *rq) |
1da177e4 | 655 | { |
b374d18a | 656 | struct elevator_queue *e = q->elevator; |
1da177e4 | 657 | |
22f746e2 TH |
658 | if (e->type->ops.elevator_latter_req_fn) |
659 | return e->type->ops.elevator_latter_req_fn(q, rq); | |
1da177e4 LT |
660 | return NULL; |
661 | } | |
662 | ||
165125e1 | 663 | struct request *elv_former_request(struct request_queue *q, struct request *rq) |
1da177e4 | 664 | { |
b374d18a | 665 | struct elevator_queue *e = q->elevator; |
1da177e4 | 666 | |
22f746e2 TH |
667 | if (e->type->ops.elevator_former_req_fn) |
668 | return e->type->ops.elevator_former_req_fn(q, rq); | |
1da177e4 LT |
669 | return NULL; |
670 | } | |
671 | ||
852c788f TH |
672 | int elv_set_request(struct request_queue *q, struct request *rq, |
673 | struct bio *bio, gfp_t gfp_mask) | |
1da177e4 | 674 | { |
b374d18a | 675 | struct elevator_queue *e = q->elevator; |
1da177e4 | 676 | |
22f746e2 | 677 | if (e->type->ops.elevator_set_req_fn) |
852c788f | 678 | return e->type->ops.elevator_set_req_fn(q, rq, bio, gfp_mask); |
1da177e4 LT |
679 | return 0; |
680 | } | |
681 | ||
165125e1 | 682 | void elv_put_request(struct request_queue *q, struct request *rq) |
1da177e4 | 683 | { |
b374d18a | 684 | struct elevator_queue *e = q->elevator; |
1da177e4 | 685 | |
22f746e2 TH |
686 | if (e->type->ops.elevator_put_req_fn) |
687 | e->type->ops.elevator_put_req_fn(rq); | |
1da177e4 LT |
688 | } |
689 | ||
165125e1 | 690 | int elv_may_queue(struct request_queue *q, int rw) |
1da177e4 | 691 | { |
b374d18a | 692 | struct elevator_queue *e = q->elevator; |
1da177e4 | 693 | |
22f746e2 TH |
694 | if (e->type->ops.elevator_may_queue_fn) |
695 | return e->type->ops.elevator_may_queue_fn(q, rw); | |
1da177e4 LT |
696 | |
697 | return ELV_MQUEUE_MAY; | |
698 | } | |
699 | ||
11914a53 MA |
700 | void elv_abort_queue(struct request_queue *q) |
701 | { | |
702 | struct request *rq; | |
703 | ||
ae1b1539 TH |
704 | blk_abort_flushes(q); |
705 | ||
11914a53 MA |
706 | while (!list_empty(&q->queue_head)) { |
707 | rq = list_entry_rq(q->queue_head.next); | |
708 | rq->cmd_flags |= REQ_QUIET; | |
5f3ea37c | 709 | trace_block_rq_abort(q, rq); |
53c663ce KU |
710 | /* |
711 | * Mark this request as started so we don't trigger | |
712 | * any debug logic in the end I/O path. | |
713 | */ | |
714 | blk_start_request(rq); | |
40cbbb78 | 715 | __blk_end_request_all(rq, -EIO); |
11914a53 MA |
716 | } |
717 | } | |
718 | EXPORT_SYMBOL(elv_abort_queue); | |
719 | ||
165125e1 | 720 | void elv_completed_request(struct request_queue *q, struct request *rq) |
1da177e4 | 721 | { |
b374d18a | 722 | struct elevator_queue *e = q->elevator; |
1da177e4 LT |
723 | |
724 | /* | |
725 | * request is released from the driver, io must be done | |
726 | */ | |
8922e16c | 727 | if (blk_account_rq(rq)) { |
0a7ae2ff | 728 | q->in_flight[rq_is_sync(rq)]--; |
33659ebb | 729 | if ((rq->cmd_flags & REQ_SORTED) && |
22f746e2 TH |
730 | e->type->ops.elevator_completed_req_fn) |
731 | e->type->ops.elevator_completed_req_fn(q, rq); | |
1bc691d3 | 732 | } |
1da177e4 LT |
733 | } |
734 | ||
3d1ab40f AV |
735 | #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr) |
736 | ||
737 | static ssize_t | |
738 | elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page) | |
1da177e4 | 739 | { |
3d1ab40f | 740 | struct elv_fs_entry *entry = to_elv(attr); |
b374d18a | 741 | struct elevator_queue *e; |
3d1ab40f AV |
742 | ssize_t error; |
743 | ||
744 | if (!entry->show) | |
745 | return -EIO; | |
746 | ||
b374d18a | 747 | e = container_of(kobj, struct elevator_queue, kobj); |
3d1ab40f | 748 | mutex_lock(&e->sysfs_lock); |
22f746e2 | 749 | error = e->type ? entry->show(e, page) : -ENOENT; |
3d1ab40f AV |
750 | mutex_unlock(&e->sysfs_lock); |
751 | return error; | |
752 | } | |
1da177e4 | 753 | |
3d1ab40f AV |
754 | static ssize_t |
755 | elv_attr_store(struct kobject *kobj, struct attribute *attr, | |
756 | const char *page, size_t length) | |
757 | { | |
3d1ab40f | 758 | struct elv_fs_entry *entry = to_elv(attr); |
b374d18a | 759 | struct elevator_queue *e; |
3d1ab40f | 760 | ssize_t error; |
1da177e4 | 761 | |
3d1ab40f AV |
762 | if (!entry->store) |
763 | return -EIO; | |
1da177e4 | 764 | |
b374d18a | 765 | e = container_of(kobj, struct elevator_queue, kobj); |
3d1ab40f | 766 | mutex_lock(&e->sysfs_lock); |
22f746e2 | 767 | error = e->type ? entry->store(e, page, length) : -ENOENT; |
3d1ab40f AV |
768 | mutex_unlock(&e->sysfs_lock); |
769 | return error; | |
770 | } | |
771 | ||
52cf25d0 | 772 | static const struct sysfs_ops elv_sysfs_ops = { |
3d1ab40f AV |
773 | .show = elv_attr_show, |
774 | .store = elv_attr_store, | |
775 | }; | |
776 | ||
777 | static struct kobj_type elv_ktype = { | |
778 | .sysfs_ops = &elv_sysfs_ops, | |
779 | .release = elevator_release, | |
780 | }; | |
781 | ||
5a5bafdc | 782 | int elv_register_queue(struct request_queue *q) |
3d1ab40f | 783 | { |
5a5bafdc | 784 | struct elevator_queue *e = q->elevator; |
3d1ab40f AV |
785 | int error; |
786 | ||
b2d6db58 | 787 | error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched"); |
3d1ab40f | 788 | if (!error) { |
22f746e2 | 789 | struct elv_fs_entry *attr = e->type->elevator_attrs; |
3d1ab40f | 790 | if (attr) { |
e572ec7e AV |
791 | while (attr->attr.name) { |
792 | if (sysfs_create_file(&e->kobj, &attr->attr)) | |
3d1ab40f | 793 | break; |
e572ec7e | 794 | attr++; |
3d1ab40f AV |
795 | } |
796 | } | |
797 | kobject_uevent(&e->kobj, KOBJ_ADD); | |
430c62fb | 798 | e->registered = 1; |
3d1ab40f AV |
799 | } |
800 | return error; | |
1da177e4 | 801 | } |
f8fc877d | 802 | EXPORT_SYMBOL(elv_register_queue); |
bc1c1169 | 803 | |
1da177e4 LT |
804 | void elv_unregister_queue(struct request_queue *q) |
805 | { | |
f8fc877d TH |
806 | if (q) { |
807 | struct elevator_queue *e = q->elevator; | |
808 | ||
809 | kobject_uevent(&e->kobj, KOBJ_REMOVE); | |
810 | kobject_del(&e->kobj); | |
811 | e->registered = 0; | |
812 | } | |
1da177e4 | 813 | } |
01effb0d | 814 | EXPORT_SYMBOL(elv_unregister_queue); |
1da177e4 | 815 | |
3d3c2379 | 816 | int elv_register(struct elevator_type *e) |
1da177e4 | 817 | { |
1ffb96c5 | 818 | char *def = ""; |
2a12dcd7 | 819 | |
3d3c2379 TH |
820 | /* create icq_cache if requested */ |
821 | if (e->icq_size) { | |
822 | if (WARN_ON(e->icq_size < sizeof(struct io_cq)) || | |
823 | WARN_ON(e->icq_align < __alignof__(struct io_cq))) | |
824 | return -EINVAL; | |
825 | ||
826 | snprintf(e->icq_cache_name, sizeof(e->icq_cache_name), | |
827 | "%s_io_cq", e->elevator_name); | |
828 | e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size, | |
829 | e->icq_align, 0, NULL); | |
830 | if (!e->icq_cache) | |
831 | return -ENOMEM; | |
832 | } | |
833 | ||
834 | /* register, don't allow duplicate names */ | |
2a12dcd7 | 835 | spin_lock(&elv_list_lock); |
3d3c2379 TH |
836 | if (elevator_find(e->elevator_name)) { |
837 | spin_unlock(&elv_list_lock); | |
838 | if (e->icq_cache) | |
839 | kmem_cache_destroy(e->icq_cache); | |
840 | return -EBUSY; | |
841 | } | |
1da177e4 | 842 | list_add_tail(&e->list, &elv_list); |
2a12dcd7 | 843 | spin_unlock(&elv_list_lock); |
1da177e4 | 844 | |
3d3c2379 | 845 | /* print pretty message */ |
5f003976 ND |
846 | if (!strcmp(e->elevator_name, chosen_elevator) || |
847 | (!*chosen_elevator && | |
848 | !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED))) | |
1ffb96c5 TV |
849 | def = " (default)"; |
850 | ||
4eb166d9 JA |
851 | printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name, |
852 | def); | |
3d3c2379 | 853 | return 0; |
1da177e4 LT |
854 | } |
855 | EXPORT_SYMBOL_GPL(elv_register); | |
856 | ||
857 | void elv_unregister(struct elevator_type *e) | |
858 | { | |
3d3c2379 | 859 | /* unregister */ |
2a12dcd7 | 860 | spin_lock(&elv_list_lock); |
1da177e4 | 861 | list_del_init(&e->list); |
2a12dcd7 | 862 | spin_unlock(&elv_list_lock); |
3d3c2379 TH |
863 | |
864 | /* | |
865 | * Destroy icq_cache if it exists. icq's are RCU managed. Make | |
866 | * sure all RCU operations are complete before proceeding. | |
867 | */ | |
868 | if (e->icq_cache) { | |
869 | rcu_barrier(); | |
870 | kmem_cache_destroy(e->icq_cache); | |
871 | e->icq_cache = NULL; | |
872 | } | |
1da177e4 LT |
873 | } |
874 | EXPORT_SYMBOL_GPL(elv_unregister); | |
875 | ||
876 | /* | |
877 | * switch to new_e io scheduler. be careful not to introduce deadlocks - | |
878 | * we don't free the old io scheduler, before we have allocated what we | |
879 | * need for the new one. this way we have a chance of going back to the old | |
cb98fc8b | 880 | * one, if the new one fails init for some reason. |
1da177e4 | 881 | */ |
165125e1 | 882 | static int elevator_switch(struct request_queue *q, struct elevator_type *new_e) |
1da177e4 | 883 | { |
5a5bafdc TH |
884 | struct elevator_queue *old = q->elevator; |
885 | bool registered = old->registered; | |
e8989fae | 886 | int err; |
1da177e4 | 887 | |
5a5bafdc TH |
888 | /* |
889 | * Turn on BYPASS and drain all requests w/ elevator private data. | |
890 | * Block layer doesn't call into a quiesced elevator - all requests | |
891 | * are directly put on the dispatch list without elevator data | |
892 | * using INSERT_BACK. All requests have SOFTBARRIER set and no | |
893 | * merge happens either. | |
894 | */ | |
d732580b | 895 | blk_queue_bypass_start(q); |
cb98fc8b | 896 | |
5a5bafdc TH |
897 | /* unregister and clear all auxiliary data of the old elevator */ |
898 | if (registered) | |
f8fc877d | 899 | elv_unregister_queue(q); |
1da177e4 | 900 | |
f8fc877d | 901 | spin_lock_irq(q->queue_lock); |
7e5a8794 | 902 | ioc_clear_queue(q); |
f8fc877d TH |
903 | spin_unlock_irq(q->queue_lock); |
904 | ||
5a5bafdc TH |
905 | /* allocate, init and register new elevator */ |
906 | err = -ENOMEM; | |
907 | q->elevator = elevator_alloc(q, new_e); | |
908 | if (!q->elevator) | |
909 | goto fail_init; | |
910 | ||
b2fab5ac | 911 | err = new_e->ops.elevator_init_fn(q); |
5a5bafdc TH |
912 | if (err) { |
913 | kobject_put(&q->elevator->kobj); | |
914 | goto fail_init; | |
915 | } | |
916 | ||
917 | if (registered) { | |
918 | err = elv_register_queue(q); | |
919 | if (err) | |
920 | goto fail_register; | |
921 | } | |
922 | ||
923 | /* done, kill the old one and finish */ | |
924 | elevator_exit(old); | |
d732580b | 925 | blk_queue_bypass_end(q); |
75ad23bc | 926 | |
5a5bafdc | 927 | blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name); |
4722dc52 | 928 | |
5dd531a0 | 929 | return 0; |
1da177e4 LT |
930 | |
931 | fail_register: | |
5a5bafdc TH |
932 | elevator_exit(q->elevator); |
933 | fail_init: | |
934 | /* switch failed, restore and re-register old elevator */ | |
935 | q->elevator = old; | |
1da177e4 | 936 | elv_register_queue(q); |
d732580b | 937 | blk_queue_bypass_end(q); |
75ad23bc | 938 | |
5dd531a0 | 939 | return err; |
1da177e4 LT |
940 | } |
941 | ||
5dd531a0 JA |
942 | /* |
943 | * Switch this queue to the given IO scheduler. | |
944 | */ | |
945 | int elevator_change(struct request_queue *q, const char *name) | |
1da177e4 LT |
946 | { |
947 | char elevator_name[ELV_NAME_MAX]; | |
948 | struct elevator_type *e; | |
949 | ||
cd43e26f | 950 | if (!q->elevator) |
5dd531a0 | 951 | return -ENXIO; |
cd43e26f | 952 | |
ee2e992c | 953 | strlcpy(elevator_name, name, sizeof(elevator_name)); |
8c279598 | 954 | e = elevator_get(strstrip(elevator_name)); |
1da177e4 LT |
955 | if (!e) { |
956 | printk(KERN_ERR "elevator: type %s not found\n", elevator_name); | |
957 | return -EINVAL; | |
958 | } | |
959 | ||
22f746e2 | 960 | if (!strcmp(elevator_name, q->elevator->type->elevator_name)) { |
2ca7d93b | 961 | elevator_put(e); |
5dd531a0 | 962 | return 0; |
2ca7d93b | 963 | } |
1da177e4 | 964 | |
5dd531a0 JA |
965 | return elevator_switch(q, e); |
966 | } | |
967 | EXPORT_SYMBOL(elevator_change); | |
968 | ||
969 | ssize_t elv_iosched_store(struct request_queue *q, const char *name, | |
970 | size_t count) | |
971 | { | |
972 | int ret; | |
973 | ||
974 | if (!q->elevator) | |
975 | return count; | |
976 | ||
977 | ret = elevator_change(q, name); | |
978 | if (!ret) | |
979 | return count; | |
980 | ||
981 | printk(KERN_ERR "elevator: switch to %s failed\n", name); | |
982 | return ret; | |
1da177e4 LT |
983 | } |
984 | ||
165125e1 | 985 | ssize_t elv_iosched_show(struct request_queue *q, char *name) |
1da177e4 | 986 | { |
b374d18a | 987 | struct elevator_queue *e = q->elevator; |
cd43e26f | 988 | struct elevator_type *elv; |
70cee26e | 989 | struct elevator_type *__e; |
1da177e4 LT |
990 | int len = 0; |
991 | ||
e36f724b | 992 | if (!q->elevator || !blk_queue_stackable(q)) |
cd43e26f MP |
993 | return sprintf(name, "none\n"); |
994 | ||
22f746e2 | 995 | elv = e->type; |
cd43e26f | 996 | |
2a12dcd7 | 997 | spin_lock(&elv_list_lock); |
70cee26e | 998 | list_for_each_entry(__e, &elv_list, list) { |
1da177e4 LT |
999 | if (!strcmp(elv->elevator_name, __e->elevator_name)) |
1000 | len += sprintf(name+len, "[%s] ", elv->elevator_name); | |
1001 | else | |
1002 | len += sprintf(name+len, "%s ", __e->elevator_name); | |
1003 | } | |
2a12dcd7 | 1004 | spin_unlock(&elv_list_lock); |
1da177e4 LT |
1005 | |
1006 | len += sprintf(len+name, "\n"); | |
1007 | return len; | |
1008 | } | |
1009 | ||
165125e1 JA |
1010 | struct request *elv_rb_former_request(struct request_queue *q, |
1011 | struct request *rq) | |
2e662b65 JA |
1012 | { |
1013 | struct rb_node *rbprev = rb_prev(&rq->rb_node); | |
1014 | ||
1015 | if (rbprev) | |
1016 | return rb_entry_rq(rbprev); | |
1017 | ||
1018 | return NULL; | |
1019 | } | |
2e662b65 JA |
1020 | EXPORT_SYMBOL(elv_rb_former_request); |
1021 | ||
165125e1 JA |
1022 | struct request *elv_rb_latter_request(struct request_queue *q, |
1023 | struct request *rq) | |
2e662b65 JA |
1024 | { |
1025 | struct rb_node *rbnext = rb_next(&rq->rb_node); | |
1026 | ||
1027 | if (rbnext) | |
1028 | return rb_entry_rq(rbnext); | |
1029 | ||
1030 | return NULL; | |
1031 | } | |
2e662b65 | 1032 | EXPORT_SYMBOL(elv_rb_latter_request); |