Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/drivers/block/elevator.c | |
3 | * | |
4 | * Block device elevator/IO-scheduler. | |
5 | * | |
6 | * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE | |
7 | * | |
8 | * 30042000 Jens Axboe <axboe@suse.de> : | |
9 | * | |
10 | * Split the elevator a bit so that it is possible to choose a different | |
11 | * one or even write a new "plug in". There are three pieces: | |
12 | * - elevator_fn, inserts a new request in the queue list | |
13 | * - elevator_merge_fn, decides whether a new buffer can be merged with | |
14 | * an existing request | |
15 | * - elevator_dequeue_fn, called when a request is taken off the active list | |
16 | * | |
17 | * 20082000 Dave Jones <davej@suse.de> : | |
18 | * Removed tests for max-bomb-segments, which was breaking elvtune | |
19 | * when run without -bN | |
20 | * | |
21 | * Jens: | |
22 | * - Rework again to work with bio instead of buffer_heads | |
23 | * - loose bi_dev comparisons, partition handling is right now | |
24 | * - completely modularize elevator setup and teardown | |
25 | * | |
26 | */ | |
27 | #include <linux/kernel.h> | |
28 | #include <linux/fs.h> | |
29 | #include <linux/blkdev.h> | |
30 | #include <linux/elevator.h> | |
31 | #include <linux/bio.h> | |
32 | #include <linux/config.h> | |
33 | #include <linux/module.h> | |
34 | #include <linux/slab.h> | |
35 | #include <linux/init.h> | |
36 | #include <linux/compiler.h> | |
37 | ||
38 | #include <asm/uaccess.h> | |
39 | ||
40 | static DEFINE_SPINLOCK(elv_list_lock); | |
41 | static LIST_HEAD(elv_list); | |
42 | ||
43 | /* | |
44 | * can we safely merge with this request? | |
45 | */ | |
46 | inline int elv_rq_merge_ok(struct request *rq, struct bio *bio) | |
47 | { | |
48 | if (!rq_mergeable(rq)) | |
49 | return 0; | |
50 | ||
51 | /* | |
52 | * different data direction or already started, don't merge | |
53 | */ | |
54 | if (bio_data_dir(bio) != rq_data_dir(rq)) | |
55 | return 0; | |
56 | ||
57 | /* | |
58 | * same device and no special stuff set, merge is ok | |
59 | */ | |
60 | if (rq->rq_disk == bio->bi_bdev->bd_disk && | |
61 | !rq->waiting && !rq->special) | |
62 | return 1; | |
63 | ||
64 | return 0; | |
65 | } | |
66 | EXPORT_SYMBOL(elv_rq_merge_ok); | |
67 | ||
68 | inline int elv_try_merge(struct request *__rq, struct bio *bio) | |
69 | { | |
70 | int ret = ELEVATOR_NO_MERGE; | |
71 | ||
72 | /* | |
73 | * we can merge and sequence is ok, check if it's possible | |
74 | */ | |
75 | if (elv_rq_merge_ok(__rq, bio)) { | |
76 | if (__rq->sector + __rq->nr_sectors == bio->bi_sector) | |
77 | ret = ELEVATOR_BACK_MERGE; | |
78 | else if (__rq->sector - bio_sectors(bio) == bio->bi_sector) | |
79 | ret = ELEVATOR_FRONT_MERGE; | |
80 | } | |
81 | ||
82 | return ret; | |
83 | } | |
84 | EXPORT_SYMBOL(elv_try_merge); | |
85 | ||
86 | inline int elv_try_last_merge(request_queue_t *q, struct bio *bio) | |
87 | { | |
88 | if (q->last_merge) | |
89 | return elv_try_merge(q->last_merge, bio); | |
90 | ||
91 | return ELEVATOR_NO_MERGE; | |
92 | } | |
93 | EXPORT_SYMBOL(elv_try_last_merge); | |
94 | ||
95 | static struct elevator_type *elevator_find(const char *name) | |
96 | { | |
97 | struct elevator_type *e = NULL; | |
98 | struct list_head *entry; | |
99 | ||
1da177e4 LT |
100 | list_for_each(entry, &elv_list) { |
101 | struct elevator_type *__e; | |
102 | ||
103 | __e = list_entry(entry, struct elevator_type, list); | |
104 | ||
105 | if (!strcmp(__e->elevator_name, name)) { | |
106 | e = __e; | |
107 | break; | |
108 | } | |
109 | } | |
1da177e4 LT |
110 | |
111 | return e; | |
112 | } | |
113 | ||
114 | static void elevator_put(struct elevator_type *e) | |
115 | { | |
116 | module_put(e->elevator_owner); | |
117 | } | |
118 | ||
119 | static struct elevator_type *elevator_get(const char *name) | |
120 | { | |
2824bc93 | 121 | struct elevator_type *e; |
1da177e4 | 122 | |
2824bc93 TH |
123 | spin_lock_irq(&elv_list_lock); |
124 | ||
125 | e = elevator_find(name); | |
126 | if (e && !try_module_get(e->elevator_owner)) | |
127 | e = NULL; | |
128 | ||
129 | spin_unlock_irq(&elv_list_lock); | |
1da177e4 LT |
130 | |
131 | return e; | |
132 | } | |
133 | ||
134 | static int elevator_attach(request_queue_t *q, struct elevator_type *e, | |
135 | struct elevator_queue *eq) | |
136 | { | |
137 | int ret = 0; | |
138 | ||
139 | memset(eq, 0, sizeof(*eq)); | |
140 | eq->ops = &e->ops; | |
141 | eq->elevator_type = e; | |
142 | ||
143 | INIT_LIST_HEAD(&q->queue_head); | |
144 | q->last_merge = NULL; | |
145 | q->elevator = eq; | |
1b47f531 | 146 | q->end_sector = 0; |
8922e16c TH |
147 | q->boundary_rq = NULL; |
148 | q->max_back_kb = 0; | |
1da177e4 LT |
149 | |
150 | if (eq->ops->elevator_init_fn) | |
151 | ret = eq->ops->elevator_init_fn(q, eq); | |
152 | ||
153 | return ret; | |
154 | } | |
155 | ||
156 | static char chosen_elevator[16]; | |
157 | ||
158 | static void elevator_setup_default(void) | |
159 | { | |
2824bc93 TH |
160 | struct elevator_type *e; |
161 | ||
1da177e4 LT |
162 | /* |
163 | * check if default is set and exists | |
164 | */ | |
2824bc93 TH |
165 | if (chosen_elevator[0] && (e = elevator_get(chosen_elevator))) { |
166 | elevator_put(e); | |
1da177e4 | 167 | return; |
2824bc93 | 168 | } |
1da177e4 LT |
169 | |
170 | #if defined(CONFIG_IOSCHED_AS) | |
171 | strcpy(chosen_elevator, "anticipatory"); | |
172 | #elif defined(CONFIG_IOSCHED_DEADLINE) | |
173 | strcpy(chosen_elevator, "deadline"); | |
174 | #elif defined(CONFIG_IOSCHED_CFQ) | |
175 | strcpy(chosen_elevator, "cfq"); | |
176 | #elif defined(CONFIG_IOSCHED_NOOP) | |
177 | strcpy(chosen_elevator, "noop"); | |
178 | #else | |
179 | #error "You must build at least 1 IO scheduler into the kernel" | |
180 | #endif | |
181 | } | |
182 | ||
183 | static int __init elevator_setup(char *str) | |
184 | { | |
185 | strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1); | |
186 | return 0; | |
187 | } | |
188 | ||
189 | __setup("elevator=", elevator_setup); | |
190 | ||
191 | int elevator_init(request_queue_t *q, char *name) | |
192 | { | |
193 | struct elevator_type *e = NULL; | |
194 | struct elevator_queue *eq; | |
195 | int ret = 0; | |
196 | ||
197 | elevator_setup_default(); | |
198 | ||
199 | if (!name) | |
200 | name = chosen_elevator; | |
201 | ||
202 | e = elevator_get(name); | |
203 | if (!e) | |
204 | return -EINVAL; | |
205 | ||
206 | eq = kmalloc(sizeof(struct elevator_queue), GFP_KERNEL); | |
207 | if (!eq) { | |
208 | elevator_put(e->elevator_type); | |
209 | return -ENOMEM; | |
210 | } | |
211 | ||
212 | ret = elevator_attach(q, e, eq); | |
213 | if (ret) { | |
214 | kfree(eq); | |
215 | elevator_put(e->elevator_type); | |
216 | } | |
217 | ||
218 | return ret; | |
219 | } | |
220 | ||
221 | void elevator_exit(elevator_t *e) | |
222 | { | |
223 | if (e->ops->elevator_exit_fn) | |
224 | e->ops->elevator_exit_fn(e); | |
225 | ||
226 | elevator_put(e->elevator_type); | |
227 | e->elevator_type = NULL; | |
228 | kfree(e); | |
229 | } | |
230 | ||
8922e16c TH |
231 | /* |
232 | * Insert rq into dispatch queue of q. Queue lock must be held on | |
233 | * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be | |
234 | * appended to the dispatch queue. To be used by specific elevators. | |
235 | */ | |
1b47f531 | 236 | void elv_dispatch_sort(request_queue_t *q, struct request *rq) |
8922e16c TH |
237 | { |
238 | sector_t boundary; | |
239 | unsigned max_back; | |
240 | struct list_head *entry; | |
241 | ||
1b47f531 | 242 | boundary = q->end_sector; |
8922e16c TH |
243 | max_back = q->max_back_kb * 2; |
244 | boundary = boundary > max_back ? boundary - max_back : 0; | |
1b47f531 | 245 | |
8922e16c TH |
246 | list_for_each_prev(entry, &q->queue_head) { |
247 | struct request *pos = list_entry_rq(entry); | |
248 | ||
249 | if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED)) | |
250 | break; | |
251 | if (rq->sector >= boundary) { | |
252 | if (pos->sector < boundary) | |
253 | continue; | |
254 | } else { | |
255 | if (pos->sector >= boundary) | |
256 | break; | |
257 | } | |
258 | if (rq->sector >= pos->sector) | |
259 | break; | |
260 | } | |
261 | ||
262 | list_add(&rq->queuelist, entry); | |
263 | } | |
264 | ||
1da177e4 LT |
265 | int elv_merge(request_queue_t *q, struct request **req, struct bio *bio) |
266 | { | |
267 | elevator_t *e = q->elevator; | |
268 | ||
269 | if (e->ops->elevator_merge_fn) | |
270 | return e->ops->elevator_merge_fn(q, req, bio); | |
271 | ||
272 | return ELEVATOR_NO_MERGE; | |
273 | } | |
274 | ||
275 | void elv_merged_request(request_queue_t *q, struct request *rq) | |
276 | { | |
277 | elevator_t *e = q->elevator; | |
278 | ||
279 | if (e->ops->elevator_merged_fn) | |
280 | e->ops->elevator_merged_fn(q, rq); | |
281 | } | |
282 | ||
283 | void elv_merge_requests(request_queue_t *q, struct request *rq, | |
284 | struct request *next) | |
285 | { | |
286 | elevator_t *e = q->elevator; | |
287 | ||
288 | if (q->last_merge == next) | |
289 | q->last_merge = NULL; | |
290 | ||
291 | if (e->ops->elevator_merge_req_fn) | |
292 | e->ops->elevator_merge_req_fn(q, rq, next); | |
293 | } | |
294 | ||
8922e16c | 295 | void elv_requeue_request(request_queue_t *q, struct request *rq) |
1da177e4 LT |
296 | { |
297 | elevator_t *e = q->elevator; | |
298 | ||
299 | /* | |
300 | * it already went through dequeue, we need to decrement the | |
301 | * in_flight count again | |
302 | */ | |
8922e16c | 303 | if (blk_account_rq(rq)) { |
1da177e4 | 304 | q->in_flight--; |
8922e16c TH |
305 | if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn) |
306 | e->ops->elevator_deactivate_req_fn(q, rq); | |
307 | } | |
1da177e4 LT |
308 | |
309 | rq->flags &= ~REQ_STARTED; | |
310 | ||
1da177e4 LT |
311 | /* |
312 | * if this is the flush, requeue the original instead and drop the flush | |
313 | */ | |
314 | if (rq->flags & REQ_BAR_FLUSH) { | |
315 | clear_bit(QUEUE_FLAG_FLUSH, &q->queue_flags); | |
316 | rq = rq->end_io_data; | |
317 | } | |
318 | ||
8922e16c | 319 | __elv_add_request(q, rq, ELEVATOR_INSERT_FRONT, 0); |
1da177e4 LT |
320 | } |
321 | ||
322 | void __elv_add_request(request_queue_t *q, struct request *rq, int where, | |
323 | int plug) | |
324 | { | |
8922e16c TH |
325 | if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) { |
326 | /* | |
327 | * barriers implicitly indicate back insertion | |
328 | */ | |
329 | if (where == ELEVATOR_INSERT_SORT) | |
330 | where = ELEVATOR_INSERT_BACK; | |
331 | ||
332 | /* | |
1b47f531 | 333 | * this request is scheduling boundary, update end_sector |
8922e16c TH |
334 | */ |
335 | if (blk_fs_request(rq)) { | |
1b47f531 | 336 | q->end_sector = rq_end_sector(rq); |
8922e16c TH |
337 | q->boundary_rq = rq; |
338 | } | |
339 | } | |
1da177e4 LT |
340 | |
341 | if (plug) | |
342 | blk_plug_device(q); | |
343 | ||
344 | rq->q = q; | |
345 | ||
8922e16c | 346 | if (unlikely(test_bit(QUEUE_FLAG_DRAIN, &q->queue_flags))) { |
1da177e4 LT |
347 | /* |
348 | * if drain is set, store the request "locally". when the drain | |
349 | * is finished, the requests will be handed ordered to the io | |
350 | * scheduler | |
351 | */ | |
352 | list_add_tail(&rq->queuelist, &q->drain_list); | |
8922e16c TH |
353 | return; |
354 | } | |
355 | ||
356 | switch (where) { | |
357 | case ELEVATOR_INSERT_FRONT: | |
358 | rq->flags |= REQ_SOFTBARRIER; | |
359 | ||
360 | list_add(&rq->queuelist, &q->queue_head); | |
361 | break; | |
362 | ||
363 | case ELEVATOR_INSERT_BACK: | |
364 | rq->flags |= REQ_SOFTBARRIER; | |
365 | ||
366 | while (q->elevator->ops->elevator_dispatch_fn(q, 1)) | |
367 | ; | |
368 | list_add_tail(&rq->queuelist, &q->queue_head); | |
369 | /* | |
370 | * We kick the queue here for the following reasons. | |
371 | * - The elevator might have returned NULL previously | |
372 | * to delay requests and returned them now. As the | |
373 | * queue wasn't empty before this request, ll_rw_blk | |
374 | * won't run the queue on return, resulting in hang. | |
375 | * - Usually, back inserted requests won't be merged | |
376 | * with anything. There's no point in delaying queue | |
377 | * processing. | |
378 | */ | |
379 | blk_remove_plug(q); | |
380 | q->request_fn(q); | |
381 | break; | |
382 | ||
383 | case ELEVATOR_INSERT_SORT: | |
384 | BUG_ON(!blk_fs_request(rq)); | |
385 | rq->flags |= REQ_SORTED; | |
386 | q->elevator->ops->elevator_add_req_fn(q, rq); | |
387 | break; | |
388 | ||
389 | default: | |
390 | printk(KERN_ERR "%s: bad insertion point %d\n", | |
391 | __FUNCTION__, where); | |
392 | BUG(); | |
393 | } | |
394 | ||
395 | if (blk_queue_plugged(q)) { | |
396 | int nrq = q->rq.count[READ] + q->rq.count[WRITE] | |
397 | - q->in_flight; | |
398 | ||
399 | if (nrq >= q->unplug_thresh) | |
400 | __generic_unplug_device(q); | |
401 | } | |
1da177e4 LT |
402 | } |
403 | ||
404 | void elv_add_request(request_queue_t *q, struct request *rq, int where, | |
405 | int plug) | |
406 | { | |
407 | unsigned long flags; | |
408 | ||
409 | spin_lock_irqsave(q->queue_lock, flags); | |
410 | __elv_add_request(q, rq, where, plug); | |
411 | spin_unlock_irqrestore(q->queue_lock, flags); | |
412 | } | |
413 | ||
414 | static inline struct request *__elv_next_request(request_queue_t *q) | |
415 | { | |
8922e16c TH |
416 | struct request *rq; |
417 | ||
418 | if (unlikely(list_empty(&q->queue_head) && | |
419 | !q->elevator->ops->elevator_dispatch_fn(q, 0))) | |
420 | return NULL; | |
421 | ||
422 | rq = list_entry_rq(q->queue_head.next); | |
1da177e4 LT |
423 | |
424 | /* | |
425 | * if this is a barrier write and the device has to issue a | |
426 | * flush sequence to support it, check how far we are | |
427 | */ | |
8922e16c | 428 | if (blk_fs_request(rq) && blk_barrier_rq(rq)) { |
1da177e4 LT |
429 | BUG_ON(q->ordered == QUEUE_ORDERED_NONE); |
430 | ||
431 | if (q->ordered == QUEUE_ORDERED_FLUSH && | |
432 | !blk_barrier_preflush(rq)) | |
433 | rq = blk_start_pre_flush(q, rq); | |
434 | } | |
435 | ||
436 | return rq; | |
437 | } | |
438 | ||
439 | struct request *elv_next_request(request_queue_t *q) | |
440 | { | |
441 | struct request *rq; | |
442 | int ret; | |
443 | ||
444 | while ((rq = __elv_next_request(q)) != NULL) { | |
8922e16c TH |
445 | if (!(rq->flags & REQ_STARTED)) { |
446 | elevator_t *e = q->elevator; | |
447 | ||
448 | /* | |
449 | * This is the first time the device driver | |
450 | * sees this request (possibly after | |
451 | * requeueing). Notify IO scheduler. | |
452 | */ | |
453 | if (blk_sorted_rq(rq) && | |
454 | e->ops->elevator_activate_req_fn) | |
455 | e->ops->elevator_activate_req_fn(q, rq); | |
456 | ||
457 | /* | |
458 | * just mark as started even if we don't start | |
459 | * it, a request that has been delayed should | |
460 | * not be passed by new incoming requests | |
461 | */ | |
462 | rq->flags |= REQ_STARTED; | |
463 | } | |
1da177e4 LT |
464 | |
465 | if (rq == q->last_merge) | |
466 | q->last_merge = NULL; | |
467 | ||
8922e16c | 468 | if (!q->boundary_rq || q->boundary_rq == rq) { |
1b47f531 | 469 | q->end_sector = rq_end_sector(rq); |
8922e16c TH |
470 | q->boundary_rq = NULL; |
471 | } | |
472 | ||
1da177e4 LT |
473 | if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn) |
474 | break; | |
475 | ||
476 | ret = q->prep_rq_fn(q, rq); | |
477 | if (ret == BLKPREP_OK) { | |
478 | break; | |
479 | } else if (ret == BLKPREP_DEFER) { | |
2e759cd4 TH |
480 | /* |
481 | * the request may have been (partially) prepped. | |
482 | * we need to keep this request in the front to | |
8922e16c TH |
483 | * avoid resource deadlock. REQ_STARTED will |
484 | * prevent other fs requests from passing this one. | |
2e759cd4 | 485 | */ |
1da177e4 LT |
486 | rq = NULL; |
487 | break; | |
488 | } else if (ret == BLKPREP_KILL) { | |
489 | int nr_bytes = rq->hard_nr_sectors << 9; | |
490 | ||
491 | if (!nr_bytes) | |
492 | nr_bytes = rq->data_len; | |
493 | ||
494 | blkdev_dequeue_request(rq); | |
495 | rq->flags |= REQ_QUIET; | |
496 | end_that_request_chunk(rq, 0, nr_bytes); | |
497 | end_that_request_last(rq); | |
498 | } else { | |
499 | printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__, | |
500 | ret); | |
501 | break; | |
502 | } | |
503 | } | |
504 | ||
505 | return rq; | |
506 | } | |
507 | ||
8922e16c | 508 | void elv_dequeue_request(request_queue_t *q, struct request *rq) |
1da177e4 | 509 | { |
8922e16c TH |
510 | BUG_ON(list_empty(&rq->queuelist)); |
511 | ||
512 | list_del_init(&rq->queuelist); | |
1da177e4 LT |
513 | |
514 | /* | |
515 | * the time frame between a request being removed from the lists | |
516 | * and to it is freed is accounted as io that is in progress at | |
8922e16c | 517 | * the driver side. |
1da177e4 LT |
518 | */ |
519 | if (blk_account_rq(rq)) | |
520 | q->in_flight++; | |
521 | ||
522 | /* | |
523 | * the main clearing point for q->last_merge is on retrieval of | |
524 | * request by driver (it calls elv_next_request()), but it _can_ | |
525 | * also happen here if a request is added to the queue but later | |
526 | * deleted without ever being given to driver (merged with another | |
527 | * request). | |
528 | */ | |
529 | if (rq == q->last_merge) | |
530 | q->last_merge = NULL; | |
1da177e4 LT |
531 | } |
532 | ||
533 | int elv_queue_empty(request_queue_t *q) | |
534 | { | |
535 | elevator_t *e = q->elevator; | |
536 | ||
8922e16c TH |
537 | if (!list_empty(&q->queue_head)) |
538 | return 0; | |
539 | ||
1da177e4 LT |
540 | if (e->ops->elevator_queue_empty_fn) |
541 | return e->ops->elevator_queue_empty_fn(q); | |
542 | ||
8922e16c | 543 | return 1; |
1da177e4 LT |
544 | } |
545 | ||
546 | struct request *elv_latter_request(request_queue_t *q, struct request *rq) | |
547 | { | |
548 | struct list_head *next; | |
549 | ||
550 | elevator_t *e = q->elevator; | |
551 | ||
552 | if (e->ops->elevator_latter_req_fn) | |
553 | return e->ops->elevator_latter_req_fn(q, rq); | |
554 | ||
555 | next = rq->queuelist.next; | |
556 | if (next != &q->queue_head && next != &rq->queuelist) | |
557 | return list_entry_rq(next); | |
558 | ||
559 | return NULL; | |
560 | } | |
561 | ||
562 | struct request *elv_former_request(request_queue_t *q, struct request *rq) | |
563 | { | |
564 | struct list_head *prev; | |
565 | ||
566 | elevator_t *e = q->elevator; | |
567 | ||
568 | if (e->ops->elevator_former_req_fn) | |
569 | return e->ops->elevator_former_req_fn(q, rq); | |
570 | ||
571 | prev = rq->queuelist.prev; | |
572 | if (prev != &q->queue_head && prev != &rq->queuelist) | |
573 | return list_entry_rq(prev); | |
574 | ||
575 | return NULL; | |
576 | } | |
577 | ||
22e2c507 JA |
578 | int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio, |
579 | int gfp_mask) | |
1da177e4 LT |
580 | { |
581 | elevator_t *e = q->elevator; | |
582 | ||
583 | if (e->ops->elevator_set_req_fn) | |
22e2c507 | 584 | return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask); |
1da177e4 LT |
585 | |
586 | rq->elevator_private = NULL; | |
587 | return 0; | |
588 | } | |
589 | ||
590 | void elv_put_request(request_queue_t *q, struct request *rq) | |
591 | { | |
592 | elevator_t *e = q->elevator; | |
593 | ||
594 | if (e->ops->elevator_put_req_fn) | |
595 | e->ops->elevator_put_req_fn(q, rq); | |
596 | } | |
597 | ||
22e2c507 | 598 | int elv_may_queue(request_queue_t *q, int rw, struct bio *bio) |
1da177e4 LT |
599 | { |
600 | elevator_t *e = q->elevator; | |
601 | ||
602 | if (e->ops->elevator_may_queue_fn) | |
22e2c507 | 603 | return e->ops->elevator_may_queue_fn(q, rw, bio); |
1da177e4 LT |
604 | |
605 | return ELV_MQUEUE_MAY; | |
606 | } | |
607 | ||
608 | void elv_completed_request(request_queue_t *q, struct request *rq) | |
609 | { | |
610 | elevator_t *e = q->elevator; | |
611 | ||
612 | /* | |
613 | * request is released from the driver, io must be done | |
614 | */ | |
8922e16c | 615 | if (blk_account_rq(rq)) { |
1da177e4 | 616 | q->in_flight--; |
8922e16c TH |
617 | if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn) |
618 | e->ops->elevator_completed_req_fn(q, rq); | |
619 | } | |
1da177e4 LT |
620 | } |
621 | ||
622 | int elv_register_queue(struct request_queue *q) | |
623 | { | |
624 | elevator_t *e = q->elevator; | |
625 | ||
626 | e->kobj.parent = kobject_get(&q->kobj); | |
627 | if (!e->kobj.parent) | |
628 | return -EBUSY; | |
629 | ||
630 | snprintf(e->kobj.name, KOBJ_NAME_LEN, "%s", "iosched"); | |
631 | e->kobj.ktype = e->elevator_type->elevator_ktype; | |
632 | ||
633 | return kobject_register(&e->kobj); | |
634 | } | |
635 | ||
636 | void elv_unregister_queue(struct request_queue *q) | |
637 | { | |
638 | if (q) { | |
639 | elevator_t *e = q->elevator; | |
640 | kobject_unregister(&e->kobj); | |
641 | kobject_put(&q->kobj); | |
642 | } | |
643 | } | |
644 | ||
645 | int elv_register(struct elevator_type *e) | |
646 | { | |
2824bc93 | 647 | spin_lock_irq(&elv_list_lock); |
1da177e4 LT |
648 | if (elevator_find(e->elevator_name)) |
649 | BUG(); | |
1da177e4 LT |
650 | list_add_tail(&e->list, &elv_list); |
651 | spin_unlock_irq(&elv_list_lock); | |
652 | ||
653 | printk(KERN_INFO "io scheduler %s registered", e->elevator_name); | |
654 | if (!strcmp(e->elevator_name, chosen_elevator)) | |
655 | printk(" (default)"); | |
656 | printk("\n"); | |
657 | return 0; | |
658 | } | |
659 | EXPORT_SYMBOL_GPL(elv_register); | |
660 | ||
661 | void elv_unregister(struct elevator_type *e) | |
662 | { | |
663 | spin_lock_irq(&elv_list_lock); | |
664 | list_del_init(&e->list); | |
665 | spin_unlock_irq(&elv_list_lock); | |
666 | } | |
667 | EXPORT_SYMBOL_GPL(elv_unregister); | |
668 | ||
669 | /* | |
670 | * switch to new_e io scheduler. be careful not to introduce deadlocks - | |
671 | * we don't free the old io scheduler, before we have allocated what we | |
672 | * need for the new one. this way we have a chance of going back to the old | |
673 | * one, if the new one fails init for some reason. we also do an intermediate | |
674 | * switch to noop to ensure safety with stack-allocated requests, since they | |
675 | * don't originate from the block layer allocator. noop is safe here, because | |
676 | * it never needs to touch the elevator itself for completion events. DRAIN | |
677 | * flags will make sure we don't touch it for additions either. | |
678 | */ | |
679 | static void elevator_switch(request_queue_t *q, struct elevator_type *new_e) | |
680 | { | |
681 | elevator_t *e = kmalloc(sizeof(elevator_t), GFP_KERNEL); | |
682 | struct elevator_type *noop_elevator = NULL; | |
683 | elevator_t *old_elevator; | |
684 | ||
685 | if (!e) | |
686 | goto error; | |
687 | ||
688 | /* | |
689 | * first step, drain requests from the block freelist | |
690 | */ | |
691 | blk_wait_queue_drained(q, 0); | |
692 | ||
693 | /* | |
694 | * unregister old elevator data | |
695 | */ | |
696 | elv_unregister_queue(q); | |
697 | old_elevator = q->elevator; | |
698 | ||
699 | /* | |
700 | * next step, switch to noop since it uses no private rq structures | |
701 | * and doesn't allocate any memory for anything. then wait for any | |
702 | * non-fs requests in-flight | |
703 | */ | |
704 | noop_elevator = elevator_get("noop"); | |
705 | spin_lock_irq(q->queue_lock); | |
706 | elevator_attach(q, noop_elevator, e); | |
707 | spin_unlock_irq(q->queue_lock); | |
708 | ||
709 | blk_wait_queue_drained(q, 1); | |
710 | ||
711 | /* | |
712 | * attach and start new elevator | |
713 | */ | |
714 | if (elevator_attach(q, new_e, e)) | |
715 | goto fail; | |
716 | ||
717 | if (elv_register_queue(q)) | |
718 | goto fail_register; | |
719 | ||
720 | /* | |
721 | * finally exit old elevator and start queue again | |
722 | */ | |
723 | elevator_exit(old_elevator); | |
724 | blk_finish_queue_drain(q); | |
725 | elevator_put(noop_elevator); | |
726 | return; | |
727 | ||
728 | fail_register: | |
729 | /* | |
730 | * switch failed, exit the new io scheduler and reattach the old | |
731 | * one again (along with re-adding the sysfs dir) | |
732 | */ | |
733 | elevator_exit(e); | |
734 | fail: | |
735 | q->elevator = old_elevator; | |
736 | elv_register_queue(q); | |
737 | blk_finish_queue_drain(q); | |
738 | error: | |
739 | if (noop_elevator) | |
740 | elevator_put(noop_elevator); | |
741 | elevator_put(new_e); | |
742 | printk(KERN_ERR "elevator: switch to %s failed\n",new_e->elevator_name); | |
743 | } | |
744 | ||
745 | ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count) | |
746 | { | |
747 | char elevator_name[ELV_NAME_MAX]; | |
748 | struct elevator_type *e; | |
749 | ||
750 | memset(elevator_name, 0, sizeof(elevator_name)); | |
751 | strncpy(elevator_name, name, sizeof(elevator_name)); | |
752 | ||
753 | if (elevator_name[strlen(elevator_name) - 1] == '\n') | |
754 | elevator_name[strlen(elevator_name) - 1] = '\0'; | |
755 | ||
756 | e = elevator_get(elevator_name); | |
757 | if (!e) { | |
758 | printk(KERN_ERR "elevator: type %s not found\n", elevator_name); | |
759 | return -EINVAL; | |
760 | } | |
761 | ||
762 | if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) | |
763 | return count; | |
764 | ||
765 | elevator_switch(q, e); | |
766 | return count; | |
767 | } | |
768 | ||
769 | ssize_t elv_iosched_show(request_queue_t *q, char *name) | |
770 | { | |
771 | elevator_t *e = q->elevator; | |
772 | struct elevator_type *elv = e->elevator_type; | |
773 | struct list_head *entry; | |
774 | int len = 0; | |
775 | ||
776 | spin_lock_irq(q->queue_lock); | |
777 | list_for_each(entry, &elv_list) { | |
778 | struct elevator_type *__e; | |
779 | ||
780 | __e = list_entry(entry, struct elevator_type, list); | |
781 | if (!strcmp(elv->elevator_name, __e->elevator_name)) | |
782 | len += sprintf(name+len, "[%s] ", elv->elevator_name); | |
783 | else | |
784 | len += sprintf(name+len, "%s ", __e->elevator_name); | |
785 | } | |
786 | spin_unlock_irq(q->queue_lock); | |
787 | ||
788 | len += sprintf(len+name, "\n"); | |
789 | return len; | |
790 | } | |
791 | ||
1b47f531 | 792 | EXPORT_SYMBOL(elv_dispatch_sort); |
1da177e4 LT |
793 | EXPORT_SYMBOL(elv_add_request); |
794 | EXPORT_SYMBOL(__elv_add_request); | |
795 | EXPORT_SYMBOL(elv_requeue_request); | |
796 | EXPORT_SYMBOL(elv_next_request); | |
8922e16c | 797 | EXPORT_SYMBOL(elv_dequeue_request); |
1da177e4 LT |
798 | EXPORT_SYMBOL(elv_queue_empty); |
799 | EXPORT_SYMBOL(elv_completed_request); | |
800 | EXPORT_SYMBOL(elevator_exit); | |
801 | EXPORT_SYMBOL(elevator_init); |