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