]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Block device elevator/IO-scheduler. | |
3 | * | |
4 | * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE | |
5 | * | |
6 | * 30042000 Jens Axboe <axboe@kernel.dk> : | |
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> | |
30 | #include <linux/module.h> | |
31 | #include <linux/slab.h> | |
32 | #include <linux/init.h> | |
33 | #include <linux/compiler.h> | |
34 | #include <linux/delay.h> | |
35 | #include <linux/blktrace_api.h> | |
36 | #include <linux/hash.h> | |
37 | #include <linux/uaccess.h> | |
38 | ||
39 | #include <trace/events/block.h> | |
40 | ||
41 | #include "blk.h" | |
42 | ||
43 | static DEFINE_SPINLOCK(elv_list_lock); | |
44 | static LIST_HEAD(elv_list); | |
45 | ||
46 | /* | |
47 | * Merge hash stuff. | |
48 | */ | |
49 | static const int elv_hash_shift = 6; | |
50 | #define ELV_HASH_BLOCK(sec) ((sec) >> 3) | |
51 | #define ELV_HASH_FN(sec) \ | |
52 | (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift)) | |
53 | #define ELV_HASH_ENTRIES (1 << elv_hash_shift) | |
54 | #define rq_hash_key(rq) (blk_rq_pos(rq) + blk_rq_sectors(rq)) | |
55 | ||
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 | { | |
62 | struct request_queue *q = rq->q; | |
63 | struct elevator_queue *e = q->elevator; | |
64 | ||
65 | if (e->ops->elevator_allow_merge_fn) | |
66 | return e->ops->elevator_allow_merge_fn(q, rq, bio); | |
67 | ||
68 | return 1; | |
69 | } | |
70 | ||
71 | /* | |
72 | * can we safely merge with this request? | |
73 | */ | |
74 | int elv_rq_merge_ok(struct request *rq, struct bio *bio) | |
75 | { | |
76 | if (!rq_mergeable(rq)) | |
77 | return 0; | |
78 | ||
79 | /* | |
80 | * Don't merge file system requests and discard requests | |
81 | */ | |
82 | if ((bio->bi_rw & REQ_DISCARD) != (rq->bio->bi_rw & REQ_DISCARD)) | |
83 | return 0; | |
84 | ||
85 | /* | |
86 | * Don't merge discard requests and secure discard requests | |
87 | */ | |
88 | if ((bio->bi_rw & REQ_SECURE) != (rq->bio->bi_rw & REQ_SECURE)) | |
89 | return 0; | |
90 | ||
91 | /* | |
92 | * different data direction or already started, don't merge | |
93 | */ | |
94 | if (bio_data_dir(bio) != rq_data_dir(rq)) | |
95 | return 0; | |
96 | ||
97 | /* | |
98 | * must be same device and not a special request | |
99 | */ | |
100 | if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special) | |
101 | return 0; | |
102 | ||
103 | /* | |
104 | * only merge integrity protected bio into ditto rq | |
105 | */ | |
106 | if (bio_integrity(bio) != blk_integrity_rq(rq)) | |
107 | return 0; | |
108 | ||
109 | if (!elv_iosched_allow_merge(rq, bio)) | |
110 | return 0; | |
111 | ||
112 | return 1; | |
113 | } | |
114 | EXPORT_SYMBOL(elv_rq_merge_ok); | |
115 | ||
116 | int elv_try_merge(struct request *__rq, struct bio *bio) | |
117 | { | |
118 | int ret = ELEVATOR_NO_MERGE; | |
119 | ||
120 | /* | |
121 | * we can merge and sequence is ok, check if it's possible | |
122 | */ | |
123 | if (elv_rq_merge_ok(__rq, bio)) { | |
124 | if (blk_rq_pos(__rq) + blk_rq_sectors(__rq) == bio->bi_sector) | |
125 | ret = ELEVATOR_BACK_MERGE; | |
126 | else if (blk_rq_pos(__rq) - bio_sectors(bio) == bio->bi_sector) | |
127 | ret = ELEVATOR_FRONT_MERGE; | |
128 | } | |
129 | ||
130 | return ret; | |
131 | } | |
132 | ||
133 | static struct elevator_type *elevator_find(const char *name) | |
134 | { | |
135 | struct elevator_type *e; | |
136 | ||
137 | list_for_each_entry(e, &elv_list, list) { | |
138 | if (!strcmp(e->elevator_name, name)) | |
139 | return e; | |
140 | } | |
141 | ||
142 | return NULL; | |
143 | } | |
144 | ||
145 | static void elevator_put(struct elevator_type *e) | |
146 | { | |
147 | module_put(e->elevator_owner); | |
148 | } | |
149 | ||
150 | static struct elevator_type *elevator_get(const char *name) | |
151 | { | |
152 | struct elevator_type *e; | |
153 | ||
154 | spin_lock(&elv_list_lock); | |
155 | ||
156 | e = elevator_find(name); | |
157 | if (!e) { | |
158 | spin_unlock(&elv_list_lock); | |
159 | request_module("%s-iosched", name); | |
160 | spin_lock(&elv_list_lock); | |
161 | e = elevator_find(name); | |
162 | } | |
163 | ||
164 | if (e && !try_module_get(e->elevator_owner)) | |
165 | e = NULL; | |
166 | ||
167 | spin_unlock(&elv_list_lock); | |
168 | ||
169 | return e; | |
170 | } | |
171 | ||
172 | static void *elevator_init_queue(struct request_queue *q, | |
173 | struct elevator_queue *eq) | |
174 | { | |
175 | return eq->ops->elevator_init_fn(q); | |
176 | } | |
177 | ||
178 | static void elevator_attach(struct request_queue *q, struct elevator_queue *eq, | |
179 | void *data) | |
180 | { | |
181 | q->elevator = eq; | |
182 | eq->elevator_data = data; | |
183 | } | |
184 | ||
185 | static char chosen_elevator[16]; | |
186 | ||
187 | static int __init elevator_setup(char *str) | |
188 | { | |
189 | /* | |
190 | * Be backwards-compatible with previous kernels, so users | |
191 | * won't get the wrong elevator. | |
192 | */ | |
193 | strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1); | |
194 | return 1; | |
195 | } | |
196 | ||
197 | __setup("elevator=", elevator_setup); | |
198 | ||
199 | static struct kobj_type elv_ktype; | |
200 | ||
201 | static struct elevator_queue *elevator_alloc(struct request_queue *q, | |
202 | struct elevator_type *e) | |
203 | { | |
204 | struct elevator_queue *eq; | |
205 | int i; | |
206 | ||
207 | eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node); | |
208 | if (unlikely(!eq)) | |
209 | goto err; | |
210 | ||
211 | eq->ops = &e->ops; | |
212 | eq->elevator_type = e; | |
213 | kobject_init(&eq->kobj, &elv_ktype); | |
214 | mutex_init(&eq->sysfs_lock); | |
215 | ||
216 | eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES, | |
217 | GFP_KERNEL, q->node); | |
218 | if (!eq->hash) | |
219 | goto err; | |
220 | ||
221 | for (i = 0; i < ELV_HASH_ENTRIES; i++) | |
222 | INIT_HLIST_HEAD(&eq->hash[i]); | |
223 | ||
224 | return eq; | |
225 | err: | |
226 | kfree(eq); | |
227 | elevator_put(e); | |
228 | return NULL; | |
229 | } | |
230 | ||
231 | static void elevator_release(struct kobject *kobj) | |
232 | { | |
233 | struct elevator_queue *e; | |
234 | ||
235 | e = container_of(kobj, struct elevator_queue, kobj); | |
236 | elevator_put(e->elevator_type); | |
237 | kfree(e->hash); | |
238 | kfree(e); | |
239 | } | |
240 | ||
241 | int elevator_init(struct request_queue *q, char *name) | |
242 | { | |
243 | struct elevator_type *e = NULL; | |
244 | struct elevator_queue *eq; | |
245 | void *data; | |
246 | ||
247 | if (unlikely(q->elevator)) | |
248 | return 0; | |
249 | ||
250 | INIT_LIST_HEAD(&q->queue_head); | |
251 | q->last_merge = NULL; | |
252 | q->end_sector = 0; | |
253 | q->boundary_rq = NULL; | |
254 | ||
255 | if (name) { | |
256 | e = elevator_get(name); | |
257 | if (!e) | |
258 | return -EINVAL; | |
259 | } | |
260 | ||
261 | if (!e && *chosen_elevator) { | |
262 | e = elevator_get(chosen_elevator); | |
263 | if (!e) | |
264 | printk(KERN_ERR "I/O scheduler %s not found\n", | |
265 | chosen_elevator); | |
266 | } | |
267 | ||
268 | if (!e) { | |
269 | e = elevator_get(CONFIG_DEFAULT_IOSCHED); | |
270 | if (!e) { | |
271 | printk(KERN_ERR | |
272 | "Default I/O scheduler not found. " \ | |
273 | "Using noop.\n"); | |
274 | e = elevator_get("noop"); | |
275 | } | |
276 | } | |
277 | ||
278 | eq = elevator_alloc(q, e); | |
279 | if (!eq) | |
280 | return -ENOMEM; | |
281 | ||
282 | data = elevator_init_queue(q, eq); | |
283 | if (!data) { | |
284 | kobject_put(&eq->kobj); | |
285 | return -ENOMEM; | |
286 | } | |
287 | ||
288 | elevator_attach(q, eq, data); | |
289 | return 0; | |
290 | } | |
291 | EXPORT_SYMBOL(elevator_init); | |
292 | ||
293 | void elevator_exit(struct elevator_queue *e) | |
294 | { | |
295 | mutex_lock(&e->sysfs_lock); | |
296 | if (e->ops->elevator_exit_fn) | |
297 | e->ops->elevator_exit_fn(e); | |
298 | e->ops = NULL; | |
299 | mutex_unlock(&e->sysfs_lock); | |
300 | ||
301 | kobject_put(&e->kobj); | |
302 | } | |
303 | EXPORT_SYMBOL(elevator_exit); | |
304 | ||
305 | static inline void __elv_rqhash_del(struct request *rq) | |
306 | { | |
307 | hlist_del_init(&rq->hash); | |
308 | } | |
309 | ||
310 | static void elv_rqhash_del(struct request_queue *q, struct request *rq) | |
311 | { | |
312 | if (ELV_ON_HASH(rq)) | |
313 | __elv_rqhash_del(rq); | |
314 | } | |
315 | ||
316 | static void elv_rqhash_add(struct request_queue *q, struct request *rq) | |
317 | { | |
318 | struct elevator_queue *e = q->elevator; | |
319 | ||
320 | BUG_ON(ELV_ON_HASH(rq)); | |
321 | hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]); | |
322 | } | |
323 | ||
324 | static void elv_rqhash_reposition(struct request_queue *q, struct request *rq) | |
325 | { | |
326 | __elv_rqhash_del(rq); | |
327 | elv_rqhash_add(q, rq); | |
328 | } | |
329 | ||
330 | static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset) | |
331 | { | |
332 | struct elevator_queue *e = q->elevator; | |
333 | struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)]; | |
334 | struct hlist_node *entry, *next; | |
335 | struct request *rq; | |
336 | ||
337 | hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) { | |
338 | BUG_ON(!ELV_ON_HASH(rq)); | |
339 | ||
340 | if (unlikely(!rq_mergeable(rq))) { | |
341 | __elv_rqhash_del(rq); | |
342 | continue; | |
343 | } | |
344 | ||
345 | if (rq_hash_key(rq) == offset) | |
346 | return rq; | |
347 | } | |
348 | ||
349 | return NULL; | |
350 | } | |
351 | ||
352 | /* | |
353 | * RB-tree support functions for inserting/lookup/removal of requests | |
354 | * in a sorted RB tree. | |
355 | */ | |
356 | struct request *elv_rb_add(struct rb_root *root, struct request *rq) | |
357 | { | |
358 | struct rb_node **p = &root->rb_node; | |
359 | struct rb_node *parent = NULL; | |
360 | struct request *__rq; | |
361 | ||
362 | while (*p) { | |
363 | parent = *p; | |
364 | __rq = rb_entry(parent, struct request, rb_node); | |
365 | ||
366 | if (blk_rq_pos(rq) < blk_rq_pos(__rq)) | |
367 | p = &(*p)->rb_left; | |
368 | else if (blk_rq_pos(rq) > blk_rq_pos(__rq)) | |
369 | p = &(*p)->rb_right; | |
370 | else | |
371 | return __rq; | |
372 | } | |
373 | ||
374 | rb_link_node(&rq->rb_node, parent, p); | |
375 | rb_insert_color(&rq->rb_node, root); | |
376 | return NULL; | |
377 | } | |
378 | EXPORT_SYMBOL(elv_rb_add); | |
379 | ||
380 | void elv_rb_del(struct rb_root *root, struct request *rq) | |
381 | { | |
382 | BUG_ON(RB_EMPTY_NODE(&rq->rb_node)); | |
383 | rb_erase(&rq->rb_node, root); | |
384 | RB_CLEAR_NODE(&rq->rb_node); | |
385 | } | |
386 | EXPORT_SYMBOL(elv_rb_del); | |
387 | ||
388 | struct request *elv_rb_find(struct rb_root *root, sector_t sector) | |
389 | { | |
390 | struct rb_node *n = root->rb_node; | |
391 | struct request *rq; | |
392 | ||
393 | while (n) { | |
394 | rq = rb_entry(n, struct request, rb_node); | |
395 | ||
396 | if (sector < blk_rq_pos(rq)) | |
397 | n = n->rb_left; | |
398 | else if (sector > blk_rq_pos(rq)) | |
399 | n = n->rb_right; | |
400 | else | |
401 | return rq; | |
402 | } | |
403 | ||
404 | return NULL; | |
405 | } | |
406 | EXPORT_SYMBOL(elv_rb_find); | |
407 | ||
408 | /* | |
409 | * Insert rq into dispatch queue of q. Queue lock must be held on | |
410 | * entry. rq is sort instead into the dispatch queue. To be used by | |
411 | * specific elevators. | |
412 | */ | |
413 | void elv_dispatch_sort(struct request_queue *q, struct request *rq) | |
414 | { | |
415 | sector_t boundary; | |
416 | struct list_head *entry; | |
417 | int stop_flags; | |
418 | ||
419 | if (q->last_merge == rq) | |
420 | q->last_merge = NULL; | |
421 | ||
422 | elv_rqhash_del(q, rq); | |
423 | ||
424 | q->nr_sorted--; | |
425 | ||
426 | boundary = q->end_sector; | |
427 | stop_flags = REQ_SOFTBARRIER | REQ_STARTED; | |
428 | list_for_each_prev(entry, &q->queue_head) { | |
429 | struct request *pos = list_entry_rq(entry); | |
430 | ||
431 | if ((rq->cmd_flags & REQ_DISCARD) != | |
432 | (pos->cmd_flags & REQ_DISCARD)) | |
433 | break; | |
434 | if (rq_data_dir(rq) != rq_data_dir(pos)) | |
435 | break; | |
436 | if (pos->cmd_flags & stop_flags) | |
437 | break; | |
438 | if (blk_rq_pos(rq) >= boundary) { | |
439 | if (blk_rq_pos(pos) < boundary) | |
440 | continue; | |
441 | } else { | |
442 | if (blk_rq_pos(pos) >= boundary) | |
443 | break; | |
444 | } | |
445 | if (blk_rq_pos(rq) >= blk_rq_pos(pos)) | |
446 | break; | |
447 | } | |
448 | ||
449 | list_add(&rq->queuelist, entry); | |
450 | } | |
451 | EXPORT_SYMBOL(elv_dispatch_sort); | |
452 | ||
453 | /* | |
454 | * Insert rq into dispatch queue of q. Queue lock must be held on | |
455 | * entry. rq is added to the back of the dispatch queue. To be used by | |
456 | * specific elevators. | |
457 | */ | |
458 | void elv_dispatch_add_tail(struct request_queue *q, struct request *rq) | |
459 | { | |
460 | if (q->last_merge == rq) | |
461 | q->last_merge = NULL; | |
462 | ||
463 | elv_rqhash_del(q, rq); | |
464 | ||
465 | q->nr_sorted--; | |
466 | ||
467 | q->end_sector = rq_end_sector(rq); | |
468 | q->boundary_rq = rq; | |
469 | list_add_tail(&rq->queuelist, &q->queue_head); | |
470 | } | |
471 | EXPORT_SYMBOL(elv_dispatch_add_tail); | |
472 | ||
473 | int elv_merge(struct request_queue *q, struct request **req, struct bio *bio) | |
474 | { | |
475 | struct elevator_queue *e = q->elevator; | |
476 | struct request *__rq; | |
477 | int ret; | |
478 | ||
479 | /* | |
480 | * Levels of merges: | |
481 | * nomerges: No merges at all attempted | |
482 | * noxmerges: Only simple one-hit cache try | |
483 | * merges: All merge tries attempted | |
484 | */ | |
485 | if (blk_queue_nomerges(q)) | |
486 | return ELEVATOR_NO_MERGE; | |
487 | ||
488 | /* | |
489 | * First try one-hit cache. | |
490 | */ | |
491 | if (q->last_merge) { | |
492 | ret = elv_try_merge(q->last_merge, bio); | |
493 | if (ret != ELEVATOR_NO_MERGE) { | |
494 | *req = q->last_merge; | |
495 | return ret; | |
496 | } | |
497 | } | |
498 | ||
499 | if (blk_queue_noxmerges(q)) | |
500 | return ELEVATOR_NO_MERGE; | |
501 | ||
502 | /* | |
503 | * See if our hash lookup can find a potential backmerge. | |
504 | */ | |
505 | __rq = elv_rqhash_find(q, bio->bi_sector); | |
506 | if (__rq && elv_rq_merge_ok(__rq, bio)) { | |
507 | *req = __rq; | |
508 | return ELEVATOR_BACK_MERGE; | |
509 | } | |
510 | ||
511 | if (e->ops->elevator_merge_fn) | |
512 | return e->ops->elevator_merge_fn(q, req, bio); | |
513 | ||
514 | return ELEVATOR_NO_MERGE; | |
515 | } | |
516 | ||
517 | /* | |
518 | * Attempt to do an insertion back merge. Only check for the case where | |
519 | * we can append 'rq' to an existing request, so we can throw 'rq' away | |
520 | * afterwards. | |
521 | * | |
522 | * Returns true if we merged, false otherwise | |
523 | */ | |
524 | static bool elv_attempt_insert_merge(struct request_queue *q, | |
525 | struct request *rq) | |
526 | { | |
527 | struct request *__rq; | |
528 | ||
529 | if (blk_queue_nomerges(q)) | |
530 | return false; | |
531 | ||
532 | /* | |
533 | * First try one-hit cache. | |
534 | */ | |
535 | if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) | |
536 | return true; | |
537 | ||
538 | if (blk_queue_noxmerges(q)) | |
539 | return false; | |
540 | ||
541 | /* | |
542 | * See if our hash lookup can find a potential backmerge. | |
543 | */ | |
544 | __rq = elv_rqhash_find(q, blk_rq_pos(rq)); | |
545 | if (__rq && blk_attempt_req_merge(q, __rq, rq)) | |
546 | return true; | |
547 | ||
548 | return false; | |
549 | } | |
550 | ||
551 | void elv_merged_request(struct request_queue *q, struct request *rq, int type) | |
552 | { | |
553 | struct elevator_queue *e = q->elevator; | |
554 | ||
555 | if (e->ops->elevator_merged_fn) | |
556 | e->ops->elevator_merged_fn(q, rq, type); | |
557 | ||
558 | if (type == ELEVATOR_BACK_MERGE) | |
559 | elv_rqhash_reposition(q, rq); | |
560 | ||
561 | q->last_merge = rq; | |
562 | } | |
563 | ||
564 | void elv_merge_requests(struct request_queue *q, struct request *rq, | |
565 | struct request *next) | |
566 | { | |
567 | struct elevator_queue *e = q->elevator; | |
568 | const int next_sorted = next->cmd_flags & REQ_SORTED; | |
569 | ||
570 | if (next_sorted && e->ops->elevator_merge_req_fn) | |
571 | e->ops->elevator_merge_req_fn(q, rq, next); | |
572 | ||
573 | elv_rqhash_reposition(q, rq); | |
574 | ||
575 | if (next_sorted) { | |
576 | elv_rqhash_del(q, next); | |
577 | q->nr_sorted--; | |
578 | } | |
579 | ||
580 | q->last_merge = rq; | |
581 | } | |
582 | ||
583 | void elv_bio_merged(struct request_queue *q, struct request *rq, | |
584 | struct bio *bio) | |
585 | { | |
586 | struct elevator_queue *e = q->elevator; | |
587 | ||
588 | if (e->ops->elevator_bio_merged_fn) | |
589 | e->ops->elevator_bio_merged_fn(q, rq, bio); | |
590 | } | |
591 | ||
592 | void elv_requeue_request(struct request_queue *q, struct request *rq) | |
593 | { | |
594 | /* | |
595 | * it already went through dequeue, we need to decrement the | |
596 | * in_flight count again | |
597 | */ | |
598 | if (blk_account_rq(rq)) { | |
599 | q->in_flight[rq_is_sync(rq)]--; | |
600 | if (rq->cmd_flags & REQ_SORTED) | |
601 | elv_deactivate_rq(q, rq); | |
602 | } | |
603 | ||
604 | rq->cmd_flags &= ~REQ_STARTED; | |
605 | ||
606 | __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE); | |
607 | } | |
608 | ||
609 | void elv_drain_elevator(struct request_queue *q) | |
610 | { | |
611 | static int printed; | |
612 | while (q->elevator->ops->elevator_dispatch_fn(q, 1)) | |
613 | ; | |
614 | if (q->nr_sorted == 0) | |
615 | return; | |
616 | if (printed++ < 10) { | |
617 | printk(KERN_ERR "%s: forced dispatching is broken " | |
618 | "(nr_sorted=%u), please report this\n", | |
619 | q->elevator->elevator_type->elevator_name, q->nr_sorted); | |
620 | } | |
621 | } | |
622 | ||
623 | /* | |
624 | * Call with queue lock held, interrupts disabled | |
625 | */ | |
626 | void elv_quiesce_start(struct request_queue *q) | |
627 | { | |
628 | if (!q->elevator) | |
629 | return; | |
630 | ||
631 | queue_flag_set(QUEUE_FLAG_ELVSWITCH, q); | |
632 | ||
633 | /* | |
634 | * make sure we don't have any requests in flight | |
635 | */ | |
636 | elv_drain_elevator(q); | |
637 | while (q->rq.elvpriv) { | |
638 | __blk_run_queue(q); | |
639 | spin_unlock_irq(q->queue_lock); | |
640 | msleep(10); | |
641 | spin_lock_irq(q->queue_lock); | |
642 | elv_drain_elevator(q); | |
643 | } | |
644 | } | |
645 | ||
646 | void elv_quiesce_end(struct request_queue *q) | |
647 | { | |
648 | queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q); | |
649 | } | |
650 | ||
651 | void __elv_add_request(struct request_queue *q, struct request *rq, int where) | |
652 | { | |
653 | trace_block_rq_insert(q, rq); | |
654 | ||
655 | rq->q = q; | |
656 | ||
657 | if (rq->cmd_flags & REQ_SOFTBARRIER) { | |
658 | /* barriers are scheduling boundary, update end_sector */ | |
659 | if (rq->cmd_type == REQ_TYPE_FS || | |
660 | (rq->cmd_flags & REQ_DISCARD)) { | |
661 | q->end_sector = rq_end_sector(rq); | |
662 | q->boundary_rq = rq; | |
663 | } | |
664 | } else if (!(rq->cmd_flags & REQ_ELVPRIV) && | |
665 | (where == ELEVATOR_INSERT_SORT || | |
666 | where == ELEVATOR_INSERT_SORT_MERGE)) | |
667 | where = ELEVATOR_INSERT_BACK; | |
668 | ||
669 | switch (where) { | |
670 | case ELEVATOR_INSERT_REQUEUE: | |
671 | case ELEVATOR_INSERT_FRONT: | |
672 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
673 | list_add(&rq->queuelist, &q->queue_head); | |
674 | break; | |
675 | ||
676 | case ELEVATOR_INSERT_BACK: | |
677 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
678 | elv_drain_elevator(q); | |
679 | list_add_tail(&rq->queuelist, &q->queue_head); | |
680 | /* | |
681 | * We kick the queue here for the following reasons. | |
682 | * - The elevator might have returned NULL previously | |
683 | * to delay requests and returned them now. As the | |
684 | * queue wasn't empty before this request, ll_rw_blk | |
685 | * won't run the queue on return, resulting in hang. | |
686 | * - Usually, back inserted requests won't be merged | |
687 | * with anything. There's no point in delaying queue | |
688 | * processing. | |
689 | */ | |
690 | __blk_run_queue(q); | |
691 | break; | |
692 | ||
693 | case ELEVATOR_INSERT_SORT_MERGE: | |
694 | /* | |
695 | * If we succeed in merging this request with one in the | |
696 | * queue already, we are done - rq has now been freed, | |
697 | * so no need to do anything further. | |
698 | */ | |
699 | if (elv_attempt_insert_merge(q, rq)) | |
700 | break; | |
701 | case ELEVATOR_INSERT_SORT: | |
702 | BUG_ON(rq->cmd_type != REQ_TYPE_FS && | |
703 | !(rq->cmd_flags & REQ_DISCARD)); | |
704 | rq->cmd_flags |= REQ_SORTED; | |
705 | q->nr_sorted++; | |
706 | if (rq_mergeable(rq)) { | |
707 | elv_rqhash_add(q, rq); | |
708 | if (!q->last_merge) | |
709 | q->last_merge = rq; | |
710 | } | |
711 | ||
712 | /* | |
713 | * Some ioscheds (cfq) run q->request_fn directly, so | |
714 | * rq cannot be accessed after calling | |
715 | * elevator_add_req_fn. | |
716 | */ | |
717 | q->elevator->ops->elevator_add_req_fn(q, rq); | |
718 | break; | |
719 | ||
720 | case ELEVATOR_INSERT_FLUSH: | |
721 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
722 | blk_insert_flush(rq); | |
723 | break; | |
724 | default: | |
725 | printk(KERN_ERR "%s: bad insertion point %d\n", | |
726 | __func__, where); | |
727 | BUG(); | |
728 | } | |
729 | } | |
730 | EXPORT_SYMBOL(__elv_add_request); | |
731 | ||
732 | void elv_add_request(struct request_queue *q, struct request *rq, int where) | |
733 | { | |
734 | unsigned long flags; | |
735 | ||
736 | spin_lock_irqsave(q->queue_lock, flags); | |
737 | __elv_add_request(q, rq, where); | |
738 | spin_unlock_irqrestore(q->queue_lock, flags); | |
739 | } | |
740 | EXPORT_SYMBOL(elv_add_request); | |
741 | ||
742 | struct request *elv_latter_request(struct request_queue *q, struct request *rq) | |
743 | { | |
744 | struct elevator_queue *e = q->elevator; | |
745 | ||
746 | if (e->ops->elevator_latter_req_fn) | |
747 | return e->ops->elevator_latter_req_fn(q, rq); | |
748 | return NULL; | |
749 | } | |
750 | ||
751 | struct request *elv_former_request(struct request_queue *q, struct request *rq) | |
752 | { | |
753 | struct elevator_queue *e = q->elevator; | |
754 | ||
755 | if (e->ops->elevator_former_req_fn) | |
756 | return e->ops->elevator_former_req_fn(q, rq); | |
757 | return NULL; | |
758 | } | |
759 | ||
760 | int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) | |
761 | { | |
762 | struct elevator_queue *e = q->elevator; | |
763 | ||
764 | if (e->ops->elevator_set_req_fn) | |
765 | return e->ops->elevator_set_req_fn(q, rq, gfp_mask); | |
766 | ||
767 | rq->elevator_private[0] = NULL; | |
768 | return 0; | |
769 | } | |
770 | ||
771 | void elv_put_request(struct request_queue *q, struct request *rq) | |
772 | { | |
773 | struct elevator_queue *e = q->elevator; | |
774 | ||
775 | if (e->ops->elevator_put_req_fn) | |
776 | e->ops->elevator_put_req_fn(rq); | |
777 | } | |
778 | ||
779 | int elv_may_queue(struct request_queue *q, int rw) | |
780 | { | |
781 | struct elevator_queue *e = q->elevator; | |
782 | ||
783 | if (e->ops->elevator_may_queue_fn) | |
784 | return e->ops->elevator_may_queue_fn(q, rw); | |
785 | ||
786 | return ELV_MQUEUE_MAY; | |
787 | } | |
788 | ||
789 | void elv_abort_queue(struct request_queue *q) | |
790 | { | |
791 | struct request *rq; | |
792 | ||
793 | blk_abort_flushes(q); | |
794 | ||
795 | while (!list_empty(&q->queue_head)) { | |
796 | rq = list_entry_rq(q->queue_head.next); | |
797 | rq->cmd_flags |= REQ_QUIET; | |
798 | trace_block_rq_abort(q, rq); | |
799 | /* | |
800 | * Mark this request as started so we don't trigger | |
801 | * any debug logic in the end I/O path. | |
802 | */ | |
803 | blk_start_request(rq); | |
804 | __blk_end_request_all(rq, -EIO); | |
805 | } | |
806 | } | |
807 | EXPORT_SYMBOL(elv_abort_queue); | |
808 | ||
809 | void elv_completed_request(struct request_queue *q, struct request *rq) | |
810 | { | |
811 | struct elevator_queue *e = q->elevator; | |
812 | ||
813 | /* | |
814 | * request is released from the driver, io must be done | |
815 | */ | |
816 | if (blk_account_rq(rq)) { | |
817 | q->in_flight[rq_is_sync(rq)]--; | |
818 | if ((rq->cmd_flags & REQ_SORTED) && | |
819 | e->ops->elevator_completed_req_fn) | |
820 | e->ops->elevator_completed_req_fn(q, rq); | |
821 | } | |
822 | } | |
823 | ||
824 | #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr) | |
825 | ||
826 | static ssize_t | |
827 | elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page) | |
828 | { | |
829 | struct elv_fs_entry *entry = to_elv(attr); | |
830 | struct elevator_queue *e; | |
831 | ssize_t error; | |
832 | ||
833 | if (!entry->show) | |
834 | return -EIO; | |
835 | ||
836 | e = container_of(kobj, struct elevator_queue, kobj); | |
837 | mutex_lock(&e->sysfs_lock); | |
838 | error = e->ops ? entry->show(e, page) : -ENOENT; | |
839 | mutex_unlock(&e->sysfs_lock); | |
840 | return error; | |
841 | } | |
842 | ||
843 | static ssize_t | |
844 | elv_attr_store(struct kobject *kobj, struct attribute *attr, | |
845 | const char *page, size_t length) | |
846 | { | |
847 | struct elv_fs_entry *entry = to_elv(attr); | |
848 | struct elevator_queue *e; | |
849 | ssize_t error; | |
850 | ||
851 | if (!entry->store) | |
852 | return -EIO; | |
853 | ||
854 | e = container_of(kobj, struct elevator_queue, kobj); | |
855 | mutex_lock(&e->sysfs_lock); | |
856 | error = e->ops ? entry->store(e, page, length) : -ENOENT; | |
857 | mutex_unlock(&e->sysfs_lock); | |
858 | return error; | |
859 | } | |
860 | ||
861 | static const struct sysfs_ops elv_sysfs_ops = { | |
862 | .show = elv_attr_show, | |
863 | .store = elv_attr_store, | |
864 | }; | |
865 | ||
866 | static struct kobj_type elv_ktype = { | |
867 | .sysfs_ops = &elv_sysfs_ops, | |
868 | .release = elevator_release, | |
869 | }; | |
870 | ||
871 | int elv_register_queue(struct request_queue *q) | |
872 | { | |
873 | struct elevator_queue *e = q->elevator; | |
874 | int error; | |
875 | ||
876 | error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched"); | |
877 | if (!error) { | |
878 | struct elv_fs_entry *attr = e->elevator_type->elevator_attrs; | |
879 | if (attr) { | |
880 | while (attr->attr.name) { | |
881 | if (sysfs_create_file(&e->kobj, &attr->attr)) | |
882 | break; | |
883 | attr++; | |
884 | } | |
885 | } | |
886 | kobject_uevent(&e->kobj, KOBJ_ADD); | |
887 | e->registered = 1; | |
888 | } | |
889 | return error; | |
890 | } | |
891 | EXPORT_SYMBOL(elv_register_queue); | |
892 | ||
893 | static void __elv_unregister_queue(struct elevator_queue *e) | |
894 | { | |
895 | kobject_uevent(&e->kobj, KOBJ_REMOVE); | |
896 | kobject_del(&e->kobj); | |
897 | e->registered = 0; | |
898 | } | |
899 | ||
900 | void elv_unregister_queue(struct request_queue *q) | |
901 | { | |
902 | if (q) | |
903 | __elv_unregister_queue(q->elevator); | |
904 | } | |
905 | EXPORT_SYMBOL(elv_unregister_queue); | |
906 | ||
907 | void elv_register(struct elevator_type *e) | |
908 | { | |
909 | char *def = ""; | |
910 | ||
911 | spin_lock(&elv_list_lock); | |
912 | BUG_ON(elevator_find(e->elevator_name)); | |
913 | list_add_tail(&e->list, &elv_list); | |
914 | spin_unlock(&elv_list_lock); | |
915 | ||
916 | if (!strcmp(e->elevator_name, chosen_elevator) || | |
917 | (!*chosen_elevator && | |
918 | !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED))) | |
919 | def = " (default)"; | |
920 | ||
921 | printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name, | |
922 | def); | |
923 | } | |
924 | EXPORT_SYMBOL_GPL(elv_register); | |
925 | ||
926 | void elv_unregister(struct elevator_type *e) | |
927 | { | |
928 | struct task_struct *g, *p; | |
929 | ||
930 | /* | |
931 | * Iterate every thread in the process to remove the io contexts. | |
932 | */ | |
933 | if (e->ops.trim) { | |
934 | read_lock(&tasklist_lock); | |
935 | do_each_thread(g, p) { | |
936 | task_lock(p); | |
937 | if (p->io_context) | |
938 | e->ops.trim(p->io_context); | |
939 | task_unlock(p); | |
940 | } while_each_thread(g, p); | |
941 | read_unlock(&tasklist_lock); | |
942 | } | |
943 | ||
944 | spin_lock(&elv_list_lock); | |
945 | list_del_init(&e->list); | |
946 | spin_unlock(&elv_list_lock); | |
947 | } | |
948 | EXPORT_SYMBOL_GPL(elv_unregister); | |
949 | ||
950 | /* | |
951 | * switch to new_e io scheduler. be careful not to introduce deadlocks - | |
952 | * we don't free the old io scheduler, before we have allocated what we | |
953 | * need for the new one. this way we have a chance of going back to the old | |
954 | * one, if the new one fails init for some reason. | |
955 | */ | |
956 | static int elevator_switch(struct request_queue *q, struct elevator_type *new_e) | |
957 | { | |
958 | struct elevator_queue *old_elevator, *e; | |
959 | void *data; | |
960 | int err; | |
961 | ||
962 | /* | |
963 | * Allocate new elevator | |
964 | */ | |
965 | e = elevator_alloc(q, new_e); | |
966 | if (!e) | |
967 | return -ENOMEM; | |
968 | ||
969 | data = elevator_init_queue(q, e); | |
970 | if (!data) { | |
971 | kobject_put(&e->kobj); | |
972 | return -ENOMEM; | |
973 | } | |
974 | ||
975 | /* | |
976 | * Turn on BYPASS and drain all requests w/ elevator private data | |
977 | */ | |
978 | spin_lock_irq(q->queue_lock); | |
979 | elv_quiesce_start(q); | |
980 | ||
981 | /* | |
982 | * Remember old elevator. | |
983 | */ | |
984 | old_elevator = q->elevator; | |
985 | ||
986 | /* | |
987 | * attach and start new elevator | |
988 | */ | |
989 | elevator_attach(q, e, data); | |
990 | ||
991 | spin_unlock_irq(q->queue_lock); | |
992 | ||
993 | if (old_elevator->registered) { | |
994 | __elv_unregister_queue(old_elevator); | |
995 | ||
996 | err = elv_register_queue(q); | |
997 | if (err) | |
998 | goto fail_register; | |
999 | } | |
1000 | ||
1001 | /* | |
1002 | * finally exit old elevator and turn off BYPASS. | |
1003 | */ | |
1004 | elevator_exit(old_elevator); | |
1005 | spin_lock_irq(q->queue_lock); | |
1006 | elv_quiesce_end(q); | |
1007 | spin_unlock_irq(q->queue_lock); | |
1008 | ||
1009 | blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name); | |
1010 | ||
1011 | return 0; | |
1012 | ||
1013 | fail_register: | |
1014 | /* | |
1015 | * switch failed, exit the new io scheduler and reattach the old | |
1016 | * one again (along with re-adding the sysfs dir) | |
1017 | */ | |
1018 | elevator_exit(e); | |
1019 | q->elevator = old_elevator; | |
1020 | elv_register_queue(q); | |
1021 | ||
1022 | spin_lock_irq(q->queue_lock); | |
1023 | queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q); | |
1024 | spin_unlock_irq(q->queue_lock); | |
1025 | ||
1026 | return err; | |
1027 | } | |
1028 | ||
1029 | /* | |
1030 | * Switch this queue to the given IO scheduler. | |
1031 | */ | |
1032 | int elevator_change(struct request_queue *q, const char *name) | |
1033 | { | |
1034 | char elevator_name[ELV_NAME_MAX]; | |
1035 | struct elevator_type *e; | |
1036 | ||
1037 | if (!q->elevator) | |
1038 | return -ENXIO; | |
1039 | ||
1040 | strlcpy(elevator_name, name, sizeof(elevator_name)); | |
1041 | e = elevator_get(strstrip(elevator_name)); | |
1042 | if (!e) { | |
1043 | printk(KERN_ERR "elevator: type %s not found\n", elevator_name); | |
1044 | return -EINVAL; | |
1045 | } | |
1046 | ||
1047 | if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) { | |
1048 | elevator_put(e); | |
1049 | return 0; | |
1050 | } | |
1051 | ||
1052 | return elevator_switch(q, e); | |
1053 | } | |
1054 | EXPORT_SYMBOL(elevator_change); | |
1055 | ||
1056 | ssize_t elv_iosched_store(struct request_queue *q, const char *name, | |
1057 | size_t count) | |
1058 | { | |
1059 | int ret; | |
1060 | ||
1061 | if (!q->elevator) | |
1062 | return count; | |
1063 | ||
1064 | ret = elevator_change(q, name); | |
1065 | if (!ret) | |
1066 | return count; | |
1067 | ||
1068 | printk(KERN_ERR "elevator: switch to %s failed\n", name); | |
1069 | return ret; | |
1070 | } | |
1071 | ||
1072 | ssize_t elv_iosched_show(struct request_queue *q, char *name) | |
1073 | { | |
1074 | struct elevator_queue *e = q->elevator; | |
1075 | struct elevator_type *elv; | |
1076 | struct elevator_type *__e; | |
1077 | int len = 0; | |
1078 | ||
1079 | if (!q->elevator || !blk_queue_stackable(q)) | |
1080 | return sprintf(name, "none\n"); | |
1081 | ||
1082 | elv = e->elevator_type; | |
1083 | ||
1084 | spin_lock(&elv_list_lock); | |
1085 | list_for_each_entry(__e, &elv_list, list) { | |
1086 | if (!strcmp(elv->elevator_name, __e->elevator_name)) | |
1087 | len += sprintf(name+len, "[%s] ", elv->elevator_name); | |
1088 | else | |
1089 | len += sprintf(name+len, "%s ", __e->elevator_name); | |
1090 | } | |
1091 | spin_unlock(&elv_list_lock); | |
1092 | ||
1093 | len += sprintf(len+name, "\n"); | |
1094 | return len; | |
1095 | } | |
1096 | ||
1097 | struct request *elv_rb_former_request(struct request_queue *q, | |
1098 | struct request *rq) | |
1099 | { | |
1100 | struct rb_node *rbprev = rb_prev(&rq->rb_node); | |
1101 | ||
1102 | if (rbprev) | |
1103 | return rb_entry_rq(rbprev); | |
1104 | ||
1105 | return NULL; | |
1106 | } | |
1107 | EXPORT_SYMBOL(elv_rb_former_request); | |
1108 | ||
1109 | struct request *elv_rb_latter_request(struct request_queue *q, | |
1110 | struct request *rq) | |
1111 | { | |
1112 | struct rb_node *rbnext = rb_next(&rq->rb_node); | |
1113 | ||
1114 | if (rbnext) | |
1115 | return rb_entry_rq(rbnext); | |
1116 | ||
1117 | return NULL; | |
1118 | } | |
1119 | EXPORT_SYMBOL(elv_rb_latter_request); |