blob: 4ac97b642042d0a6c8bc449d182599fb37be3734 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * Block device elevator/IO-scheduler.
3 *
4 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
5 *
6 * 30042000 Jens Axboe <axboe@suse.de> :
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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include <linux/module.h>
31#include <linux/slab.h>
32#include <linux/init.h>
33#include <linux/compiler.h>
Tejun Heocb98fc82005-10-28 08:29:39 +020034#include <linux/delay.h>
Jens Axboe2056a782006-03-23 20:00:26 +010035#include <linux/blktrace_api.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036
37#include <asm/uaccess.h>
38
39static DEFINE_SPINLOCK(elv_list_lock);
40static LIST_HEAD(elv_list);
41
42/*
43 * can we safely merge with this request?
44 */
45inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
46{
47 if (!rq_mergeable(rq))
48 return 0;
49
50 /*
51 * different data direction or already started, don't merge
52 */
53 if (bio_data_dir(bio) != rq_data_dir(rq))
54 return 0;
55
56 /*
57 * same device and no special stuff set, merge is ok
58 */
59 if (rq->rq_disk == bio->bi_bdev->bd_disk &&
60 !rq->waiting && !rq->special)
61 return 1;
62
63 return 0;
64}
65EXPORT_SYMBOL(elv_rq_merge_ok);
66
Coywolf Qi Hunt769db452005-12-28 10:55:49 +010067static inline int elv_try_merge(struct request *__rq, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -070068{
69 int ret = ELEVATOR_NO_MERGE;
70
71 /*
72 * we can merge and sequence is ok, check if it's possible
73 */
74 if (elv_rq_merge_ok(__rq, bio)) {
75 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
76 ret = ELEVATOR_BACK_MERGE;
77 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
78 ret = ELEVATOR_FRONT_MERGE;
79 }
80
81 return ret;
82}
Linus Torvalds1da177e2005-04-16 15:20:36 -070083
Linus Torvalds1da177e2005-04-16 15:20:36 -070084static struct elevator_type *elevator_find(const char *name)
85{
86 struct elevator_type *e = NULL;
87 struct list_head *entry;
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089 list_for_each(entry, &elv_list) {
90 struct elevator_type *__e;
91
92 __e = list_entry(entry, struct elevator_type, list);
93
94 if (!strcmp(__e->elevator_name, name)) {
95 e = __e;
96 break;
97 }
98 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
100 return e;
101}
102
103static void elevator_put(struct elevator_type *e)
104{
105 module_put(e->elevator_owner);
106}
107
108static struct elevator_type *elevator_get(const char *name)
109{
Tejun Heo2824bc932005-10-20 10:56:41 +0200110 struct elevator_type *e;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
Tejun Heo2824bc932005-10-20 10:56:41 +0200112 spin_lock_irq(&elv_list_lock);
113
114 e = elevator_find(name);
115 if (e && !try_module_get(e->elevator_owner))
116 e = NULL;
117
118 spin_unlock_irq(&elv_list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
120 return e;
121}
122
Jens Axboebc1c1162006-06-08 08:49:06 +0200123static void *elevator_init_queue(request_queue_t *q, struct elevator_queue *eq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124{
Jens Axboebc1c1162006-06-08 08:49:06 +0200125 return eq->ops->elevator_init_fn(q, eq);
126}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127
Jens Axboebc1c1162006-06-08 08:49:06 +0200128static void elevator_attach(request_queue_t *q, struct elevator_queue *eq,
129 void *data)
130{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 q->elevator = eq;
Jens Axboebc1c1162006-06-08 08:49:06 +0200132 eq->elevator_data = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133}
134
135static char chosen_elevator[16];
136
Nate Diller5f003972006-01-24 10:07:58 +0100137static int __init elevator_setup(char *str)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138{
Chuck Ebbert752a3b72006-01-16 09:47:37 +0100139 /*
140 * Be backwards-compatible with previous kernels, so users
141 * won't get the wrong elevator.
142 */
Nate Diller5f003972006-01-24 10:07:58 +0100143 if (!strcmp(str, "as"))
Chuck Ebbert752a3b72006-01-16 09:47:37 +0100144 strcpy(chosen_elevator, "anticipatory");
Zachary Amsdencff3ba22005-11-09 13:24:20 +0100145 else
Nate Diller5f003972006-01-24 10:07:58 +0100146 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800147 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148}
149
150__setup("elevator=", elevator_setup);
151
Al Viro3d1ab402006-03-18 18:35:43 -0500152static struct kobj_type elv_ktype;
153
154static elevator_t *elevator_alloc(struct elevator_type *e)
155{
156 elevator_t *eq = kmalloc(sizeof(elevator_t), GFP_KERNEL);
157 if (eq) {
158 memset(eq, 0, sizeof(*eq));
159 eq->ops = &e->ops;
160 eq->elevator_type = e;
161 kobject_init(&eq->kobj);
162 snprintf(eq->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
163 eq->kobj.ktype = &elv_ktype;
164 mutex_init(&eq->sysfs_lock);
165 } else {
166 elevator_put(e);
167 }
168 return eq;
169}
170
171static void elevator_release(struct kobject *kobj)
172{
173 elevator_t *e = container_of(kobj, elevator_t, kobj);
174 elevator_put(e->elevator_type);
175 kfree(e);
176}
177
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178int elevator_init(request_queue_t *q, char *name)
179{
180 struct elevator_type *e = NULL;
181 struct elevator_queue *eq;
182 int ret = 0;
Jens Axboebc1c1162006-06-08 08:49:06 +0200183 void *data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184
Tejun Heocb98fc82005-10-28 08:29:39 +0200185 INIT_LIST_HEAD(&q->queue_head);
186 q->last_merge = NULL;
187 q->end_sector = 0;
188 q->boundary_rq = NULL;
Tejun Heocb98fc82005-10-28 08:29:39 +0200189
Nate Diller5f003972006-01-24 10:07:58 +0100190 if (name && !(e = elevator_get(name)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 return -EINVAL;
192
Nate Diller248d5ca2006-01-24 10:09:14 +0100193 if (!e && *chosen_elevator && !(e = elevator_get(chosen_elevator)))
194 printk("I/O scheduler %s not found\n", chosen_elevator);
195
196 if (!e && !(e = elevator_get(CONFIG_DEFAULT_IOSCHED))) {
197 printk("Default I/O scheduler not found, using no-op\n");
198 e = elevator_get("noop");
Nate Diller5f003972006-01-24 10:07:58 +0100199 }
200
Al Viro3d1ab402006-03-18 18:35:43 -0500201 eq = elevator_alloc(e);
202 if (!eq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204
Jens Axboebc1c1162006-06-08 08:49:06 +0200205 data = elevator_init_queue(q, eq);
206 if (!data) {
Al Viro3d1ab402006-03-18 18:35:43 -0500207 kobject_put(&eq->kobj);
Jens Axboebc1c1162006-06-08 08:49:06 +0200208 return -ENOMEM;
209 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210
Jens Axboebc1c1162006-06-08 08:49:06 +0200211 elevator_attach(q, eq, data);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212 return ret;
213}
214
215void elevator_exit(elevator_t *e)
216{
Al Viro3d1ab402006-03-18 18:35:43 -0500217 mutex_lock(&e->sysfs_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218 if (e->ops->elevator_exit_fn)
219 e->ops->elevator_exit_fn(e);
Al Viro3d1ab402006-03-18 18:35:43 -0500220 e->ops = NULL;
221 mutex_unlock(&e->sysfs_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222
Al Viro3d1ab402006-03-18 18:35:43 -0500223 kobject_put(&e->kobj);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224}
225
Tejun Heo8922e162005-10-20 16:23:44 +0200226/*
227 * Insert rq into dispatch queue of q. Queue lock must be held on
228 * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
229 * appended to the dispatch queue. To be used by specific elevators.
230 */
Jens Axboe1b47f532005-10-20 16:37:00 +0200231void elv_dispatch_sort(request_queue_t *q, struct request *rq)
Tejun Heo8922e162005-10-20 16:23:44 +0200232{
233 sector_t boundary;
Tejun Heo8922e162005-10-20 16:23:44 +0200234 struct list_head *entry;
235
Tejun Heo06b86242005-10-20 16:46:23 +0200236 if (q->last_merge == rq)
237 q->last_merge = NULL;
Tejun Heo15853af2005-11-10 08:52:05 +0100238 q->nr_sorted--;
Tejun Heo06b86242005-10-20 16:46:23 +0200239
Jens Axboe1b47f532005-10-20 16:37:00 +0200240 boundary = q->end_sector;
Tejun Heocb198332005-10-24 08:35:58 +0200241
Tejun Heo8922e162005-10-20 16:23:44 +0200242 list_for_each_prev(entry, &q->queue_head) {
243 struct request *pos = list_entry_rq(entry);
244
Jens Axboe4aff5e22006-08-10 08:44:47 +0200245 if (pos->cmd_flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
Tejun Heo8922e162005-10-20 16:23:44 +0200246 break;
247 if (rq->sector >= boundary) {
248 if (pos->sector < boundary)
249 continue;
250 } else {
251 if (pos->sector >= boundary)
252 break;
253 }
254 if (rq->sector >= pos->sector)
255 break;
256 }
257
258 list_add(&rq->queuelist, entry);
259}
260
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
262{
263 elevator_t *e = q->elevator;
Tejun Heo06b86242005-10-20 16:46:23 +0200264 int ret;
265
266 if (q->last_merge) {
267 ret = elv_try_merge(q->last_merge, bio);
268 if (ret != ELEVATOR_NO_MERGE) {
269 *req = q->last_merge;
270 return ret;
271 }
272 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
274 if (e->ops->elevator_merge_fn)
275 return e->ops->elevator_merge_fn(q, req, bio);
276
277 return ELEVATOR_NO_MERGE;
278}
279
280void elv_merged_request(request_queue_t *q, struct request *rq)
281{
282 elevator_t *e = q->elevator;
283
284 if (e->ops->elevator_merged_fn)
285 e->ops->elevator_merged_fn(q, rq);
Tejun Heo06b86242005-10-20 16:46:23 +0200286
287 q->last_merge = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288}
289
290void elv_merge_requests(request_queue_t *q, struct request *rq,
291 struct request *next)
292{
293 elevator_t *e = q->elevator;
294
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295 if (e->ops->elevator_merge_req_fn)
296 e->ops->elevator_merge_req_fn(q, rq, next);
Tejun Heo15853af2005-11-10 08:52:05 +0100297 q->nr_sorted--;
Tejun Heo06b86242005-10-20 16:46:23 +0200298
299 q->last_merge = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300}
301
Tejun Heo8922e162005-10-20 16:23:44 +0200302void elv_requeue_request(request_queue_t *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303{
304 elevator_t *e = q->elevator;
305
306 /*
307 * it already went through dequeue, we need to decrement the
308 * in_flight count again
309 */
Tejun Heo8922e162005-10-20 16:23:44 +0200310 if (blk_account_rq(rq)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 q->in_flight--;
Tejun Heo8922e162005-10-20 16:23:44 +0200312 if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
313 e->ops->elevator_deactivate_req_fn(q, rq);
314 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315
Jens Axboe4aff5e22006-08-10 08:44:47 +0200316 rq->cmd_flags &= ~REQ_STARTED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317
Tejun Heo30e96562006-02-08 01:01:31 -0800318 elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319}
320
Tejun Heo15853af2005-11-10 08:52:05 +0100321static void elv_drain_elevator(request_queue_t *q)
322{
323 static int printed;
324 while (q->elevator->ops->elevator_dispatch_fn(q, 1))
325 ;
326 if (q->nr_sorted == 0)
327 return;
328 if (printed++ < 10) {
329 printk(KERN_ERR "%s: forced dispatching is broken "
330 "(nr_sorted=%u), please report this\n",
331 q->elevator->elevator_type->elevator_name, q->nr_sorted);
332 }
333}
334
Tejun Heo30e96562006-02-08 01:01:31 -0800335void elv_insert(request_queue_t *q, struct request *rq, int where)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336{
Tejun Heo797e7db2006-01-06 09:51:03 +0100337 struct list_head *pos;
338 unsigned ordseq;
Jens Axboedac07ec2006-05-11 08:20:16 +0200339 int unplug_it = 1;
Tejun Heo797e7db2006-01-06 09:51:03 +0100340
Jens Axboe2056a782006-03-23 20:00:26 +0100341 blk_add_trace_rq(q, rq, BLK_TA_INSERT);
342
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343 rq->q = q;
344
Tejun Heo8922e162005-10-20 16:23:44 +0200345 switch (where) {
346 case ELEVATOR_INSERT_FRONT:
Jens Axboe4aff5e22006-08-10 08:44:47 +0200347 rq->cmd_flags |= REQ_SOFTBARRIER;
Tejun Heo8922e162005-10-20 16:23:44 +0200348
349 list_add(&rq->queuelist, &q->queue_head);
350 break;
351
352 case ELEVATOR_INSERT_BACK:
Jens Axboe4aff5e22006-08-10 08:44:47 +0200353 rq->cmd_flags |= REQ_SOFTBARRIER;
Tejun Heo15853af2005-11-10 08:52:05 +0100354 elv_drain_elevator(q);
Tejun Heo8922e162005-10-20 16:23:44 +0200355 list_add_tail(&rq->queuelist, &q->queue_head);
356 /*
357 * We kick the queue here for the following reasons.
358 * - The elevator might have returned NULL previously
359 * to delay requests and returned them now. As the
360 * queue wasn't empty before this request, ll_rw_blk
361 * won't run the queue on return, resulting in hang.
362 * - Usually, back inserted requests won't be merged
363 * with anything. There's no point in delaying queue
364 * processing.
365 */
366 blk_remove_plug(q);
367 q->request_fn(q);
368 break;
369
370 case ELEVATOR_INSERT_SORT:
371 BUG_ON(!blk_fs_request(rq));
Jens Axboe4aff5e22006-08-10 08:44:47 +0200372 rq->cmd_flags |= REQ_SORTED;
Tejun Heo15853af2005-11-10 08:52:05 +0100373 q->nr_sorted++;
Tejun Heo06b86242005-10-20 16:46:23 +0200374 if (q->last_merge == NULL && rq_mergeable(rq))
375 q->last_merge = rq;
Tejun Heoca235092005-11-01 17:23:49 +0900376 /*
377 * Some ioscheds (cfq) run q->request_fn directly, so
378 * rq cannot be accessed after calling
379 * elevator_add_req_fn.
380 */
381 q->elevator->ops->elevator_add_req_fn(q, rq);
Tejun Heo8922e162005-10-20 16:23:44 +0200382 break;
383
Tejun Heo797e7db2006-01-06 09:51:03 +0100384 case ELEVATOR_INSERT_REQUEUE:
385 /*
386 * If ordered flush isn't in progress, we do front
387 * insertion; otherwise, requests should be requeued
388 * in ordseq order.
389 */
Jens Axboe4aff5e22006-08-10 08:44:47 +0200390 rq->cmd_flags |= REQ_SOFTBARRIER;
Tejun Heo797e7db2006-01-06 09:51:03 +0100391
392 if (q->ordseq == 0) {
393 list_add(&rq->queuelist, &q->queue_head);
394 break;
395 }
396
397 ordseq = blk_ordered_req_seq(rq);
398
399 list_for_each(pos, &q->queue_head) {
400 struct request *pos_rq = list_entry_rq(pos);
401 if (ordseq <= blk_ordered_req_seq(pos_rq))
402 break;
403 }
404
405 list_add_tail(&rq->queuelist, pos);
Jens Axboedac07ec2006-05-11 08:20:16 +0200406 /*
407 * most requeues happen because of a busy condition, don't
408 * force unplug of the queue for that case.
409 */
410 unplug_it = 0;
Tejun Heo797e7db2006-01-06 09:51:03 +0100411 break;
412
Tejun Heo8922e162005-10-20 16:23:44 +0200413 default:
414 printk(KERN_ERR "%s: bad insertion point %d\n",
415 __FUNCTION__, where);
416 BUG();
417 }
418
Jens Axboedac07ec2006-05-11 08:20:16 +0200419 if (unplug_it && blk_queue_plugged(q)) {
Tejun Heo8922e162005-10-20 16:23:44 +0200420 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
421 - q->in_flight;
422
423 if (nrq >= q->unplug_thresh)
424 __generic_unplug_device(q);
425 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426}
427
Tejun Heo30e96562006-02-08 01:01:31 -0800428void __elv_add_request(request_queue_t *q, struct request *rq, int where,
429 int plug)
430{
431 if (q->ordcolor)
Jens Axboe4aff5e22006-08-10 08:44:47 +0200432 rq->cmd_flags |= REQ_ORDERED_COLOR;
Tejun Heo30e96562006-02-08 01:01:31 -0800433
Jens Axboe4aff5e22006-08-10 08:44:47 +0200434 if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
Tejun Heo30e96562006-02-08 01:01:31 -0800435 /*
436 * toggle ordered color
437 */
438 if (blk_barrier_rq(rq))
439 q->ordcolor ^= 1;
440
441 /*
442 * barriers implicitly indicate back insertion
443 */
444 if (where == ELEVATOR_INSERT_SORT)
445 where = ELEVATOR_INSERT_BACK;
446
447 /*
448 * this request is scheduling boundary, update
449 * end_sector
450 */
451 if (blk_fs_request(rq)) {
452 q->end_sector = rq_end_sector(rq);
453 q->boundary_rq = rq;
454 }
Jens Axboe4aff5e22006-08-10 08:44:47 +0200455 } else if (!(rq->cmd_flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
Tejun Heo30e96562006-02-08 01:01:31 -0800456 where = ELEVATOR_INSERT_BACK;
457
458 if (plug)
459 blk_plug_device(q);
460
461 elv_insert(q, rq, where);
462}
463
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464void elv_add_request(request_queue_t *q, struct request *rq, int where,
465 int plug)
466{
467 unsigned long flags;
468
469 spin_lock_irqsave(q->queue_lock, flags);
470 __elv_add_request(q, rq, where, plug);
471 spin_unlock_irqrestore(q->queue_lock, flags);
472}
473
474static inline struct request *__elv_next_request(request_queue_t *q)
475{
Tejun Heo8922e162005-10-20 16:23:44 +0200476 struct request *rq;
477
Tejun Heo797e7db2006-01-06 09:51:03 +0100478 while (1) {
479 while (!list_empty(&q->queue_head)) {
480 rq = list_entry_rq(q->queue_head.next);
481 if (blk_do_ordered(q, &rq))
482 return rq;
483 }
Tejun Heo8922e162005-10-20 16:23:44 +0200484
Tejun Heo797e7db2006-01-06 09:51:03 +0100485 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
486 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488}
489
490struct request *elv_next_request(request_queue_t *q)
491{
492 struct request *rq;
493 int ret;
494
495 while ((rq = __elv_next_request(q)) != NULL) {
Jens Axboe4aff5e22006-08-10 08:44:47 +0200496 if (!(rq->cmd_flags & REQ_STARTED)) {
Tejun Heo8922e162005-10-20 16:23:44 +0200497 elevator_t *e = q->elevator;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498
Tejun Heo8922e162005-10-20 16:23:44 +0200499 /*
500 * This is the first time the device driver
501 * sees this request (possibly after
502 * requeueing). Notify IO scheduler.
503 */
504 if (blk_sorted_rq(rq) &&
505 e->ops->elevator_activate_req_fn)
506 e->ops->elevator_activate_req_fn(q, rq);
507
508 /*
509 * just mark as started even if we don't start
510 * it, a request that has been delayed should
511 * not be passed by new incoming requests
512 */
Jens Axboe4aff5e22006-08-10 08:44:47 +0200513 rq->cmd_flags |= REQ_STARTED;
Jens Axboe2056a782006-03-23 20:00:26 +0100514 blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
Tejun Heo8922e162005-10-20 16:23:44 +0200515 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516
Tejun Heo8922e162005-10-20 16:23:44 +0200517 if (!q->boundary_rq || q->boundary_rq == rq) {
Jens Axboe1b47f532005-10-20 16:37:00 +0200518 q->end_sector = rq_end_sector(rq);
Tejun Heo8922e162005-10-20 16:23:44 +0200519 q->boundary_rq = NULL;
520 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521
Jens Axboe4aff5e22006-08-10 08:44:47 +0200522 if ((rq->cmd_flags & REQ_DONTPREP) || !q->prep_rq_fn)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 break;
524
525 ret = q->prep_rq_fn(q, rq);
526 if (ret == BLKPREP_OK) {
527 break;
528 } else if (ret == BLKPREP_DEFER) {
Tejun Heo 2e759cd2005-04-24 02:04:21 -0500529 /*
530 * the request may have been (partially) prepped.
531 * we need to keep this request in the front to
Tejun Heo8922e162005-10-20 16:23:44 +0200532 * avoid resource deadlock. REQ_STARTED will
533 * prevent other fs requests from passing this one.
Tejun Heo 2e759cd2005-04-24 02:04:21 -0500534 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 rq = NULL;
536 break;
537 } else if (ret == BLKPREP_KILL) {
538 int nr_bytes = rq->hard_nr_sectors << 9;
539
540 if (!nr_bytes)
541 nr_bytes = rq->data_len;
542
543 blkdev_dequeue_request(rq);
Jens Axboe4aff5e22006-08-10 08:44:47 +0200544 rq->cmd_flags |= REQ_QUIET;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 end_that_request_chunk(rq, 0, nr_bytes);
Tejun Heo8ffdc652006-01-06 09:49:03 +0100546 end_that_request_last(rq, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 } else {
548 printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
549 ret);
550 break;
551 }
552 }
553
554 return rq;
555}
556
Tejun Heo8922e162005-10-20 16:23:44 +0200557void elv_dequeue_request(request_queue_t *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558{
Tejun Heo8922e162005-10-20 16:23:44 +0200559 BUG_ON(list_empty(&rq->queuelist));
560
561 list_del_init(&rq->queuelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562
563 /*
564 * the time frame between a request being removed from the lists
565 * and to it is freed is accounted as io that is in progress at
Tejun Heo8922e162005-10-20 16:23:44 +0200566 * the driver side.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 */
568 if (blk_account_rq(rq))
569 q->in_flight++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570}
571
572int elv_queue_empty(request_queue_t *q)
573{
574 elevator_t *e = q->elevator;
575
Tejun Heo8922e162005-10-20 16:23:44 +0200576 if (!list_empty(&q->queue_head))
577 return 0;
578
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 if (e->ops->elevator_queue_empty_fn)
580 return e->ops->elevator_queue_empty_fn(q);
581
Tejun Heo8922e162005-10-20 16:23:44 +0200582 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583}
584
585struct request *elv_latter_request(request_queue_t *q, struct request *rq)
586{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 elevator_t *e = q->elevator;
588
589 if (e->ops->elevator_latter_req_fn)
590 return e->ops->elevator_latter_req_fn(q, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 return NULL;
592}
593
594struct request *elv_former_request(request_queue_t *q, struct request *rq)
595{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596 elevator_t *e = q->elevator;
597
598 if (e->ops->elevator_former_req_fn)
599 return e->ops->elevator_former_req_fn(q, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600 return NULL;
601}
602
Jens Axboe22e2c502005-06-27 10:55:12 +0200603int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
Al Viro8267e262005-10-21 03:20:53 -0400604 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605{
606 elevator_t *e = q->elevator;
607
608 if (e->ops->elevator_set_req_fn)
Jens Axboe22e2c502005-06-27 10:55:12 +0200609 return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610
611 rq->elevator_private = NULL;
612 return 0;
613}
614
615void elv_put_request(request_queue_t *q, struct request *rq)
616{
617 elevator_t *e = q->elevator;
618
619 if (e->ops->elevator_put_req_fn)
620 e->ops->elevator_put_req_fn(q, rq);
621}
622
Jens Axboe22e2c502005-06-27 10:55:12 +0200623int elv_may_queue(request_queue_t *q, int rw, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624{
625 elevator_t *e = q->elevator;
626
627 if (e->ops->elevator_may_queue_fn)
Jens Axboe22e2c502005-06-27 10:55:12 +0200628 return e->ops->elevator_may_queue_fn(q, rw, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
630 return ELV_MQUEUE_MAY;
631}
632
633void elv_completed_request(request_queue_t *q, struct request *rq)
634{
635 elevator_t *e = q->elevator;
636
637 /*
638 * request is released from the driver, io must be done
639 */
Tejun Heo8922e162005-10-20 16:23:44 +0200640 if (blk_account_rq(rq)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 q->in_flight--;
Tejun Heo1bc691d2006-01-12 15:39:26 +0100642 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
643 e->ops->elevator_completed_req_fn(q, rq);
644 }
Tejun Heo797e7db2006-01-06 09:51:03 +0100645
Tejun Heo1bc691d2006-01-12 15:39:26 +0100646 /*
647 * Check if the queue is waiting for fs requests to be
648 * drained for flush sequence.
649 */
650 if (unlikely(q->ordseq)) {
651 struct request *first_rq = list_entry_rq(q->queue_head.next);
652 if (q->in_flight == 0 &&
Tejun Heo797e7db2006-01-06 09:51:03 +0100653 blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
654 blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
655 blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
656 q->request_fn(q);
657 }
Tejun Heo8922e162005-10-20 16:23:44 +0200658 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659}
660
Al Viro3d1ab402006-03-18 18:35:43 -0500661#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
662
663static ssize_t
664elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
665{
666 elevator_t *e = container_of(kobj, elevator_t, kobj);
667 struct elv_fs_entry *entry = to_elv(attr);
668 ssize_t error;
669
670 if (!entry->show)
671 return -EIO;
672
673 mutex_lock(&e->sysfs_lock);
674 error = e->ops ? entry->show(e, page) : -ENOENT;
675 mutex_unlock(&e->sysfs_lock);
676 return error;
677}
678
679static ssize_t
680elv_attr_store(struct kobject *kobj, struct attribute *attr,
681 const char *page, size_t length)
682{
683 elevator_t *e = container_of(kobj, elevator_t, kobj);
684 struct elv_fs_entry *entry = to_elv(attr);
685 ssize_t error;
686
687 if (!entry->store)
688 return -EIO;
689
690 mutex_lock(&e->sysfs_lock);
691 error = e->ops ? entry->store(e, page, length) : -ENOENT;
692 mutex_unlock(&e->sysfs_lock);
693 return error;
694}
695
696static struct sysfs_ops elv_sysfs_ops = {
697 .show = elv_attr_show,
698 .store = elv_attr_store,
699};
700
701static struct kobj_type elv_ktype = {
702 .sysfs_ops = &elv_sysfs_ops,
703 .release = elevator_release,
704};
705
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706int elv_register_queue(struct request_queue *q)
707{
708 elevator_t *e = q->elevator;
Al Viro3d1ab402006-03-18 18:35:43 -0500709 int error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
Al Viro3d1ab402006-03-18 18:35:43 -0500711 e->kobj.parent = &q->kobj;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712
Al Viro3d1ab402006-03-18 18:35:43 -0500713 error = kobject_add(&e->kobj);
714 if (!error) {
Al Viroe572ec72006-03-18 22:27:18 -0500715 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
Al Viro3d1ab402006-03-18 18:35:43 -0500716 if (attr) {
Al Viroe572ec72006-03-18 22:27:18 -0500717 while (attr->attr.name) {
718 if (sysfs_create_file(&e->kobj, &attr->attr))
Al Viro3d1ab402006-03-18 18:35:43 -0500719 break;
Al Viroe572ec72006-03-18 22:27:18 -0500720 attr++;
Al Viro3d1ab402006-03-18 18:35:43 -0500721 }
722 }
723 kobject_uevent(&e->kobj, KOBJ_ADD);
724 }
725 return error;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726}
727
Jens Axboebc1c1162006-06-08 08:49:06 +0200728static void __elv_unregister_queue(elevator_t *e)
729{
730 kobject_uevent(&e->kobj, KOBJ_REMOVE);
731 kobject_del(&e->kobj);
732}
733
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734void elv_unregister_queue(struct request_queue *q)
735{
Jens Axboebc1c1162006-06-08 08:49:06 +0200736 if (q)
737 __elv_unregister_queue(q->elevator);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738}
739
740int elv_register(struct elevator_type *e)
741{
Tejun Heo2824bc932005-10-20 10:56:41 +0200742 spin_lock_irq(&elv_list_lock);
Eric Sesterhennce524492006-03-24 18:43:26 +0100743 BUG_ON(elevator_find(e->elevator_name));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 list_add_tail(&e->list, &elv_list);
745 spin_unlock_irq(&elv_list_lock);
746
747 printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
Nate Diller5f003972006-01-24 10:07:58 +0100748 if (!strcmp(e->elevator_name, chosen_elevator) ||
749 (!*chosen_elevator &&
750 !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
751 printk(" (default)");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 printk("\n");
753 return 0;
754}
755EXPORT_SYMBOL_GPL(elv_register);
756
757void elv_unregister(struct elevator_type *e)
758{
Christoph Hellwig83521d32005-10-30 15:01:39 -0800759 struct task_struct *g, *p;
760
761 /*
762 * Iterate every thread in the process to remove the io contexts.
763 */
Al Viroe17a9482006-03-18 13:21:20 -0500764 if (e->ops.trim) {
765 read_lock(&tasklist_lock);
766 do_each_thread(g, p) {
767 task_lock(p);
Oleg Nesterov2d8f6132006-08-22 21:22:13 +0400768 if (p->io_context)
769 e->ops.trim(p->io_context);
Al Viroe17a9482006-03-18 13:21:20 -0500770 task_unlock(p);
771 } while_each_thread(g, p);
772 read_unlock(&tasklist_lock);
773 }
Christoph Hellwig83521d32005-10-30 15:01:39 -0800774
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 spin_lock_irq(&elv_list_lock);
776 list_del_init(&e->list);
777 spin_unlock_irq(&elv_list_lock);
778}
779EXPORT_SYMBOL_GPL(elv_unregister);
780
781/*
782 * switch to new_e io scheduler. be careful not to introduce deadlocks -
783 * we don't free the old io scheduler, before we have allocated what we
784 * need for the new one. this way we have a chance of going back to the old
Tejun Heocb98fc82005-10-28 08:29:39 +0200785 * one, if the new one fails init for some reason.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 */
Al Viro3d1ab402006-03-18 18:35:43 -0500787static int elevator_switch(request_queue_t *q, struct elevator_type *new_e)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788{
Tejun Heocb98fc82005-10-28 08:29:39 +0200789 elevator_t *old_elevator, *e;
Jens Axboebc1c1162006-06-08 08:49:06 +0200790 void *data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791
Tejun Heocb98fc82005-10-28 08:29:39 +0200792 /*
793 * Allocate new elevator
794 */
Al Viro3d1ab402006-03-18 18:35:43 -0500795 e = elevator_alloc(new_e);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 if (!e)
Al Viro3d1ab402006-03-18 18:35:43 -0500797 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798
Jens Axboebc1c1162006-06-08 08:49:06 +0200799 data = elevator_init_queue(q, e);
800 if (!data) {
801 kobject_put(&e->kobj);
802 return 0;
803 }
804
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 /*
Tejun Heocb98fc82005-10-28 08:29:39 +0200806 * Turn on BYPASS and drain all requests w/ elevator private data
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 */
Tejun Heocb98fc82005-10-28 08:29:39 +0200808 spin_lock_irq(q->queue_lock);
809
Jens Axboe64521d12005-10-28 08:30:39 +0200810 set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
Tejun Heocb98fc82005-10-28 08:29:39 +0200811
Tejun Heo15853af2005-11-10 08:52:05 +0100812 elv_drain_elevator(q);
Tejun Heocb98fc82005-10-28 08:29:39 +0200813
814 while (q->rq.elvpriv) {
Tejun Heo407df2a2005-11-10 08:48:21 +0100815 blk_remove_plug(q);
816 q->request_fn(q);
Tejun Heocb98fc82005-10-28 08:29:39 +0200817 spin_unlock_irq(q->queue_lock);
Jens Axboe64521d12005-10-28 08:30:39 +0200818 msleep(10);
Tejun Heocb98fc82005-10-28 08:29:39 +0200819 spin_lock_irq(q->queue_lock);
Tejun Heo15853af2005-11-10 08:52:05 +0100820 elv_drain_elevator(q);
Tejun Heocb98fc82005-10-28 08:29:39 +0200821 }
822
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 /*
Jens Axboebc1c1162006-06-08 08:49:06 +0200824 * Remember old elevator.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 old_elevator = q->elevator;
827
828 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 * attach and start new elevator
830 */
Jens Axboebc1c1162006-06-08 08:49:06 +0200831 elevator_attach(q, e, data);
832
833 spin_unlock_irq(q->queue_lock);
834
835 __elv_unregister_queue(old_elevator);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836
837 if (elv_register_queue(q))
838 goto fail_register;
839
840 /*
Tejun Heocb98fc82005-10-28 08:29:39 +0200841 * finally exit old elevator and turn off BYPASS.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 */
843 elevator_exit(old_elevator);
Jens Axboe64521d12005-10-28 08:30:39 +0200844 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
Al Viro3d1ab402006-03-18 18:35:43 -0500845 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846
847fail_register:
848 /*
849 * switch failed, exit the new io scheduler and reattach the old
850 * one again (along with re-adding the sysfs dir)
851 */
852 elevator_exit(e);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 q->elevator = old_elevator;
854 elv_register_queue(q);
Jens Axboe64521d12005-10-28 08:30:39 +0200855 clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
Al Viro3d1ab402006-03-18 18:35:43 -0500856 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857}
858
859ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
860{
861 char elevator_name[ELV_NAME_MAX];
Tejun Heobe561232005-11-10 08:55:01 +0100862 size_t len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 struct elevator_type *e;
864
Tejun Heobe561232005-11-10 08:55:01 +0100865 elevator_name[sizeof(elevator_name) - 1] = '\0';
866 strncpy(elevator_name, name, sizeof(elevator_name) - 1);
867 len = strlen(elevator_name);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868
Tejun Heobe561232005-11-10 08:55:01 +0100869 if (len && elevator_name[len - 1] == '\n')
870 elevator_name[len - 1] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871
872 e = elevator_get(elevator_name);
873 if (!e) {
874 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
875 return -EINVAL;
876 }
877
Nate Diller2ca7d932005-10-30 15:02:24 -0800878 if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
879 elevator_put(e);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 return count;
Nate Diller2ca7d932005-10-30 15:02:24 -0800881 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882
Al Viro3d1ab402006-03-18 18:35:43 -0500883 if (!elevator_switch(q, e))
884 printk(KERN_ERR "elevator: switch to %s failed\n",elevator_name);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 return count;
886}
887
888ssize_t elv_iosched_show(request_queue_t *q, char *name)
889{
890 elevator_t *e = q->elevator;
891 struct elevator_type *elv = e->elevator_type;
892 struct list_head *entry;
893 int len = 0;
894
895 spin_lock_irq(q->queue_lock);
896 list_for_each(entry, &elv_list) {
897 struct elevator_type *__e;
898
899 __e = list_entry(entry, struct elevator_type, list);
900 if (!strcmp(elv->elevator_name, __e->elevator_name))
901 len += sprintf(name+len, "[%s] ", elv->elevator_name);
902 else
903 len += sprintf(name+len, "%s ", __e->elevator_name);
904 }
905 spin_unlock_irq(q->queue_lock);
906
907 len += sprintf(len+name, "\n");
908 return len;
909}
910
Jens Axboe1b47f532005-10-20 16:37:00 +0200911EXPORT_SYMBOL(elv_dispatch_sort);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912EXPORT_SYMBOL(elv_add_request);
913EXPORT_SYMBOL(__elv_add_request);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914EXPORT_SYMBOL(elv_next_request);
Tejun Heo8922e162005-10-20 16:23:44 +0200915EXPORT_SYMBOL(elv_dequeue_request);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916EXPORT_SYMBOL(elv_queue_empty);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917EXPORT_SYMBOL(elevator_exit);
918EXPORT_SYMBOL(elevator_init);