blob: 8da3cf66894c379b1b4908d94602c49f77cb473e [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * Anticipatory & deadline i/o scheduler.
3 *
4 * Copyright (C) 2002 Jens Axboe <axboe@suse.de>
Nick Pigginf5b3db02005-11-07 00:59:53 -08005 * Nick Piggin <nickpiggin@yahoo.com.au>
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 */
8#include <linux/kernel.h>
9#include <linux/fs.h>
10#include <linux/blkdev.h>
11#include <linux/elevator.h>
12#include <linux/bio.h>
13#include <linux/config.h>
14#include <linux/module.h>
15#include <linux/slab.h>
16#include <linux/init.h>
17#include <linux/compiler.h>
18#include <linux/hash.h>
19#include <linux/rbtree.h>
20#include <linux/interrupt.h>
21
22#define REQ_SYNC 1
23#define REQ_ASYNC 0
24
25/*
26 * See Documentation/block/as-iosched.txt
27 */
28
29/*
30 * max time before a read is submitted.
31 */
32#define default_read_expire (HZ / 8)
33
34/*
35 * ditto for writes, these limits are not hard, even
36 * if the disk is capable of satisfying them.
37 */
38#define default_write_expire (HZ / 4)
39
40/*
41 * read_batch_expire describes how long we will allow a stream of reads to
42 * persist before looking to see whether it is time to switch over to writes.
43 */
44#define default_read_batch_expire (HZ / 2)
45
46/*
47 * write_batch_expire describes how long we want a stream of writes to run for.
48 * This is not a hard limit, but a target we set for the auto-tuning thingy.
49 * See, the problem is: we can send a lot of writes to disk cache / TCQ in
50 * a short amount of time...
51 */
52#define default_write_batch_expire (HZ / 8)
53
54/*
55 * max time we may wait to anticipate a read (default around 6ms)
56 */
57#define default_antic_expire ((HZ / 150) ? HZ / 150 : 1)
58
59/*
60 * Keep track of up to 20ms thinktimes. We can go as big as we like here,
61 * however huge values tend to interfere and not decay fast enough. A program
62 * might be in a non-io phase of operation. Waiting on user input for example,
63 * or doing a lengthy computation. A small penalty can be justified there, and
64 * will still catch out those processes that constantly have large thinktimes.
65 */
66#define MAX_THINKTIME (HZ/50UL)
67
68/* Bits in as_io_context.state */
69enum as_io_states {
Nick Pigginf5b3db02005-11-07 00:59:53 -080070 AS_TASK_RUNNING=0, /* Process has not exited */
Linus Torvalds1da177e2005-04-16 15:20:36 -070071 AS_TASK_IOSTARTED, /* Process has started some IO */
72 AS_TASK_IORUNNING, /* Process has completed some IO */
73};
74
75enum anticipation_status {
76 ANTIC_OFF=0, /* Not anticipating (normal operation) */
77 ANTIC_WAIT_REQ, /* The last read has not yet completed */
78 ANTIC_WAIT_NEXT, /* Currently anticipating a request vs
79 last read (which has completed) */
80 ANTIC_FINISHED, /* Anticipating but have found a candidate
81 * or timed out */
82};
83
84struct as_data {
85 /*
86 * run time data
87 */
88
89 struct request_queue *q; /* the "owner" queue */
90
91 /*
92 * requests (as_rq s) are present on both sort_list and fifo_list
93 */
94 struct rb_root sort_list[2];
95 struct list_head fifo_list[2];
96
97 struct as_rq *next_arq[2]; /* next in sort order */
98 sector_t last_sector[2]; /* last REQ_SYNC & REQ_ASYNC sectors */
Linus Torvalds1da177e2005-04-16 15:20:36 -070099 struct list_head *hash; /* request hash */
100
101 unsigned long exit_prob; /* probability a task will exit while
102 being waited on */
Nick Pigginf5b3db02005-11-07 00:59:53 -0800103 unsigned long exit_no_coop; /* probablility an exited task will
104 not be part of a later cooperating
105 request */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106 unsigned long new_ttime_total; /* mean thinktime on new proc */
107 unsigned long new_ttime_mean;
108 u64 new_seek_total; /* mean seek on new proc */
109 sector_t new_seek_mean;
110
111 unsigned long current_batch_expires;
112 unsigned long last_check_fifo[2];
113 int changed_batch; /* 1: waiting for old batch to end */
114 int new_batch; /* 1: waiting on first read complete */
115 int batch_data_dir; /* current batch REQ_SYNC / REQ_ASYNC */
116 int write_batch_count; /* max # of reqs in a write batch */
117 int current_write_count; /* how many requests left this batch */
118 int write_batch_idled; /* has the write batch gone idle? */
119 mempool_t *arq_pool;
120
121 enum anticipation_status antic_status;
122 unsigned long antic_start; /* jiffies: when it started */
123 struct timer_list antic_timer; /* anticipatory scheduling timer */
124 struct work_struct antic_work; /* Deferred unplugging */
125 struct io_context *io_context; /* Identify the expected process */
126 int ioc_finished; /* IO associated with io_context is finished */
127 int nr_dispatched;
128
129 /*
130 * settings that change how the i/o scheduler behaves
131 */
132 unsigned long fifo_expire[2];
133 unsigned long batch_expire[2];
134 unsigned long antic_expire;
135};
136
137#define list_entry_fifo(ptr) list_entry((ptr), struct as_rq, fifo)
138
139/*
140 * per-request data.
141 */
142enum arq_state {
143 AS_RQ_NEW=0, /* New - not referenced and not on any lists */
144 AS_RQ_QUEUED, /* In the request queue. It belongs to the
145 scheduler */
146 AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the
147 driver now */
148 AS_RQ_PRESCHED, /* Debug poisoning for requests being used */
149 AS_RQ_REMOVED,
150 AS_RQ_MERGED,
151 AS_RQ_POSTSCHED, /* when they shouldn't be */
152};
153
154struct as_rq {
155 /*
156 * rbtree index, key is the starting offset
157 */
158 struct rb_node rb_node;
159 sector_t rb_key;
160
161 struct request *request;
162
163 struct io_context *io_context; /* The submitting task */
164
165 /*
166 * request hash, key is the ending offset (for back merge lookup)
167 */
168 struct list_head hash;
169 unsigned int on_hash;
170
171 /*
172 * expire fifo
173 */
174 struct list_head fifo;
175 unsigned long expires;
176
177 unsigned int is_sync;
178 enum arq_state state;
179};
180
181#define RQ_DATA(rq) ((struct as_rq *) (rq)->elevator_private)
182
183static kmem_cache_t *arq_pool;
184
Tejun Heoef9be1d2005-11-11 14:27:09 +0100185static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq);
186static void as_antic_stop(struct as_data *ad);
187
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188/*
189 * IO Context helper functions
190 */
191
192/* Called to deallocate the as_io_context */
193static void free_as_io_context(struct as_io_context *aic)
194{
195 kfree(aic);
196}
197
198/* Called when the task exits */
199static void exit_as_io_context(struct as_io_context *aic)
200{
201 WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state));
202 clear_bit(AS_TASK_RUNNING, &aic->state);
203}
204
205static struct as_io_context *alloc_as_io_context(void)
206{
207 struct as_io_context *ret;
208
209 ret = kmalloc(sizeof(*ret), GFP_ATOMIC);
210 if (ret) {
211 ret->dtor = free_as_io_context;
212 ret->exit = exit_as_io_context;
213 ret->state = 1 << AS_TASK_RUNNING;
214 atomic_set(&ret->nr_queued, 0);
215 atomic_set(&ret->nr_dispatched, 0);
216 spin_lock_init(&ret->lock);
217 ret->ttime_total = 0;
218 ret->ttime_samples = 0;
219 ret->ttime_mean = 0;
220 ret->seek_total = 0;
221 ret->seek_samples = 0;
222 ret->seek_mean = 0;
223 }
224
225 return ret;
226}
227
228/*
229 * If the current task has no AS IO context then create one and initialise it.
230 * Then take a ref on the task's io context and return it.
231 */
232static struct io_context *as_get_io_context(void)
233{
234 struct io_context *ioc = get_io_context(GFP_ATOMIC);
235 if (ioc && !ioc->aic) {
236 ioc->aic = alloc_as_io_context();
237 if (!ioc->aic) {
238 put_io_context(ioc);
239 ioc = NULL;
240 }
241 }
242 return ioc;
243}
244
Jens Axboeb4878f22005-10-20 16:42:29 +0200245static void as_put_io_context(struct as_rq *arq)
246{
247 struct as_io_context *aic;
248
249 if (unlikely(!arq->io_context))
250 return;
251
252 aic = arq->io_context->aic;
253
254 if (arq->is_sync == REQ_SYNC && aic) {
255 spin_lock(&aic->lock);
256 set_bit(AS_TASK_IORUNNING, &aic->state);
257 aic->last_end_request = jiffies;
258 spin_unlock(&aic->lock);
259 }
260
261 put_io_context(arq->io_context);
262}
263
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264/*
265 * the back merge hash support functions
266 */
267static const int as_hash_shift = 6;
268#define AS_HASH_BLOCK(sec) ((sec) >> 3)
269#define AS_HASH_FN(sec) (hash_long(AS_HASH_BLOCK((sec)), as_hash_shift))
270#define AS_HASH_ENTRIES (1 << as_hash_shift)
271#define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
272#define list_entry_hash(ptr) list_entry((ptr), struct as_rq, hash)
273
274static inline void __as_del_arq_hash(struct as_rq *arq)
275{
276 arq->on_hash = 0;
277 list_del_init(&arq->hash);
278}
279
280static inline void as_del_arq_hash(struct as_rq *arq)
281{
282 if (arq->on_hash)
283 __as_del_arq_hash(arq);
284}
285
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286static void as_add_arq_hash(struct as_data *ad, struct as_rq *arq)
287{
288 struct request *rq = arq->request;
289
290 BUG_ON(arq->on_hash);
291
292 arq->on_hash = 1;
293 list_add(&arq->hash, &ad->hash[AS_HASH_FN(rq_hash_key(rq))]);
294}
295
296/*
297 * move hot entry to front of chain
298 */
299static inline void as_hot_arq_hash(struct as_data *ad, struct as_rq *arq)
300{
301 struct request *rq = arq->request;
302 struct list_head *head = &ad->hash[AS_HASH_FN(rq_hash_key(rq))];
303
304 if (!arq->on_hash) {
305 WARN_ON(1);
306 return;
307 }
308
309 if (arq->hash.prev != head) {
310 list_del(&arq->hash);
311 list_add(&arq->hash, head);
312 }
313}
314
315static struct request *as_find_arq_hash(struct as_data *ad, sector_t offset)
316{
317 struct list_head *hash_list = &ad->hash[AS_HASH_FN(offset)];
318 struct list_head *entry, *next = hash_list->next;
319
320 while ((entry = next) != hash_list) {
321 struct as_rq *arq = list_entry_hash(entry);
322 struct request *__rq = arq->request;
323
324 next = entry->next;
325
326 BUG_ON(!arq->on_hash);
327
328 if (!rq_mergeable(__rq)) {
Tejun Heo98b11472005-10-20 16:46:54 +0200329 as_del_arq_hash(arq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330 continue;
331 }
332
333 if (rq_hash_key(__rq) == offset)
334 return __rq;
335 }
336
337 return NULL;
338}
339
340/*
341 * rb tree support functions
342 */
343#define RB_NONE (2)
344#define RB_EMPTY(root) ((root)->rb_node == NULL)
345#define ON_RB(node) ((node)->rb_color != RB_NONE)
346#define RB_CLEAR(node) ((node)->rb_color = RB_NONE)
347#define rb_entry_arq(node) rb_entry((node), struct as_rq, rb_node)
348#define ARQ_RB_ROOT(ad, arq) (&(ad)->sort_list[(arq)->is_sync])
349#define rq_rb_key(rq) (rq)->sector
350
351/*
352 * as_find_first_arq finds the first (lowest sector numbered) request
353 * for the specified data_dir. Used to sweep back to the start of the disk
354 * (1-way elevator) after we process the last (highest sector) request.
355 */
356static struct as_rq *as_find_first_arq(struct as_data *ad, int data_dir)
357{
358 struct rb_node *n = ad->sort_list[data_dir].rb_node;
359
360 if (n == NULL)
361 return NULL;
362
363 for (;;) {
364 if (n->rb_left == NULL)
365 return rb_entry_arq(n);
366
367 n = n->rb_left;
368 }
369}
370
371/*
372 * Add the request to the rb tree if it is unique. If there is an alias (an
373 * existing request against the same sector), which can happen when using
374 * direct IO, then return the alias.
375 */
Tejun Heoef9be1d2005-11-11 14:27:09 +0100376static struct as_rq *__as_add_arq_rb(struct as_data *ad, struct as_rq *arq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377{
378 struct rb_node **p = &ARQ_RB_ROOT(ad, arq)->rb_node;
379 struct rb_node *parent = NULL;
380 struct as_rq *__arq;
381 struct request *rq = arq->request;
382
383 arq->rb_key = rq_rb_key(rq);
384
385 while (*p) {
386 parent = *p;
387 __arq = rb_entry_arq(parent);
388
389 if (arq->rb_key < __arq->rb_key)
390 p = &(*p)->rb_left;
391 else if (arq->rb_key > __arq->rb_key)
392 p = &(*p)->rb_right;
393 else
394 return __arq;
395 }
396
397 rb_link_node(&arq->rb_node, parent, p);
398 rb_insert_color(&arq->rb_node, ARQ_RB_ROOT(ad, arq));
399
400 return NULL;
401}
402
Tejun Heoef9be1d2005-11-11 14:27:09 +0100403static void as_add_arq_rb(struct as_data *ad, struct as_rq *arq)
404{
405 struct as_rq *alias;
406
407 while ((unlikely(alias = __as_add_arq_rb(ad, arq)))) {
408 as_move_to_dispatch(ad, alias);
409 as_antic_stop(ad);
410 }
411}
412
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413static inline void as_del_arq_rb(struct as_data *ad, struct as_rq *arq)
414{
415 if (!ON_RB(&arq->rb_node)) {
416 WARN_ON(1);
417 return;
418 }
419
420 rb_erase(&arq->rb_node, ARQ_RB_ROOT(ad, arq));
421 RB_CLEAR(&arq->rb_node);
422}
423
424static struct request *
425as_find_arq_rb(struct as_data *ad, sector_t sector, int data_dir)
426{
427 struct rb_node *n = ad->sort_list[data_dir].rb_node;
428 struct as_rq *arq;
429
430 while (n) {
431 arq = rb_entry_arq(n);
432
433 if (sector < arq->rb_key)
434 n = n->rb_left;
435 else if (sector > arq->rb_key)
436 n = n->rb_right;
437 else
438 return arq->request;
439 }
440
441 return NULL;
442}
443
444/*
445 * IO Scheduler proper
446 */
447
448#define MAXBACK (1024 * 1024) /*
449 * Maximum distance the disk will go backward
450 * for a request.
451 */
452
453#define BACK_PENALTY 2
454
455/*
456 * as_choose_req selects the preferred one of two requests of the same data_dir
457 * ignoring time - eg. timeouts, which is the job of as_dispatch_request
458 */
459static struct as_rq *
460as_choose_req(struct as_data *ad, struct as_rq *arq1, struct as_rq *arq2)
461{
462 int data_dir;
463 sector_t last, s1, s2, d1, d2;
464 int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */
465 const sector_t maxback = MAXBACK;
466
467 if (arq1 == NULL || arq1 == arq2)
468 return arq2;
469 if (arq2 == NULL)
470 return arq1;
471
472 data_dir = arq1->is_sync;
473
474 last = ad->last_sector[data_dir];
475 s1 = arq1->request->sector;
476 s2 = arq2->request->sector;
477
478 BUG_ON(data_dir != arq2->is_sync);
479
480 /*
481 * Strict one way elevator _except_ in the case where we allow
482 * short backward seeks which are biased as twice the cost of a
483 * similar forward seek.
484 */
485 if (s1 >= last)
486 d1 = s1 - last;
487 else if (s1+maxback >= last)
488 d1 = (last - s1)*BACK_PENALTY;
489 else {
490 r1_wrap = 1;
491 d1 = 0; /* shut up, gcc */
492 }
493
494 if (s2 >= last)
495 d2 = s2 - last;
496 else if (s2+maxback >= last)
497 d2 = (last - s2)*BACK_PENALTY;
498 else {
499 r2_wrap = 1;
500 d2 = 0;
501 }
502
503 /* Found required data */
504 if (!r1_wrap && r2_wrap)
505 return arq1;
506 else if (!r2_wrap && r1_wrap)
507 return arq2;
508 else if (r1_wrap && r2_wrap) {
509 /* both behind the head */
510 if (s1 <= s2)
511 return arq1;
512 else
513 return arq2;
514 }
515
516 /* Both requests in front of the head */
517 if (d1 < d2)
518 return arq1;
519 else if (d2 < d1)
520 return arq2;
521 else {
522 if (s1 >= s2)
523 return arq1;
524 else
525 return arq2;
526 }
527}
528
529/*
530 * as_find_next_arq finds the next request after @prev in elevator order.
531 * this with as_choose_req form the basis for how the scheduler chooses
532 * what request to process next. Anticipation works on top of this.
533 */
534static struct as_rq *as_find_next_arq(struct as_data *ad, struct as_rq *last)
535{
536 const int data_dir = last->is_sync;
537 struct as_rq *ret;
538 struct rb_node *rbnext = rb_next(&last->rb_node);
539 struct rb_node *rbprev = rb_prev(&last->rb_node);
540 struct as_rq *arq_next, *arq_prev;
541
542 BUG_ON(!ON_RB(&last->rb_node));
543
544 if (rbprev)
545 arq_prev = rb_entry_arq(rbprev);
546 else
547 arq_prev = NULL;
548
549 if (rbnext)
550 arq_next = rb_entry_arq(rbnext);
551 else {
552 arq_next = as_find_first_arq(ad, data_dir);
553 if (arq_next == last)
554 arq_next = NULL;
555 }
556
557 ret = as_choose_req(ad, arq_next, arq_prev);
558
559 return ret;
560}
561
562/*
563 * anticipatory scheduling functions follow
564 */
565
566/*
567 * as_antic_expired tells us when we have anticipated too long.
568 * The funny "absolute difference" math on the elapsed time is to handle
569 * jiffy wraps, and disks which have been idle for 0x80000000 jiffies.
570 */
571static int as_antic_expired(struct as_data *ad)
572{
573 long delta_jif;
574
575 delta_jif = jiffies - ad->antic_start;
576 if (unlikely(delta_jif < 0))
577 delta_jif = -delta_jif;
578 if (delta_jif < ad->antic_expire)
579 return 0;
580
581 return 1;
582}
583
584/*
585 * as_antic_waitnext starts anticipating that a nice request will soon be
586 * submitted. See also as_antic_waitreq
587 */
588static void as_antic_waitnext(struct as_data *ad)
589{
590 unsigned long timeout;
591
592 BUG_ON(ad->antic_status != ANTIC_OFF
593 && ad->antic_status != ANTIC_WAIT_REQ);
594
595 timeout = ad->antic_start + ad->antic_expire;
596
597 mod_timer(&ad->antic_timer, timeout);
598
599 ad->antic_status = ANTIC_WAIT_NEXT;
600}
601
602/*
603 * as_antic_waitreq starts anticipating. We don't start timing the anticipation
604 * until the request that we're anticipating on has finished. This means we
605 * are timing from when the candidate process wakes up hopefully.
606 */
607static void as_antic_waitreq(struct as_data *ad)
608{
609 BUG_ON(ad->antic_status == ANTIC_FINISHED);
610 if (ad->antic_status == ANTIC_OFF) {
611 if (!ad->io_context || ad->ioc_finished)
612 as_antic_waitnext(ad);
613 else
614 ad->antic_status = ANTIC_WAIT_REQ;
615 }
616}
617
618/*
619 * This is called directly by the functions in this file to stop anticipation.
620 * We kill the timer and schedule a call to the request_fn asap.
621 */
622static void as_antic_stop(struct as_data *ad)
623{
624 int status = ad->antic_status;
625
626 if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) {
627 if (status == ANTIC_WAIT_NEXT)
628 del_timer(&ad->antic_timer);
629 ad->antic_status = ANTIC_FINISHED;
630 /* see as_work_handler */
631 kblockd_schedule_work(&ad->antic_work);
632 }
633}
634
635/*
636 * as_antic_timeout is the timer function set by as_antic_waitnext.
637 */
638static void as_antic_timeout(unsigned long data)
639{
640 struct request_queue *q = (struct request_queue *)data;
641 struct as_data *ad = q->elevator->elevator_data;
642 unsigned long flags;
643
644 spin_lock_irqsave(q->queue_lock, flags);
645 if (ad->antic_status == ANTIC_WAIT_REQ
646 || ad->antic_status == ANTIC_WAIT_NEXT) {
647 struct as_io_context *aic = ad->io_context->aic;
648
649 ad->antic_status = ANTIC_FINISHED;
650 kblockd_schedule_work(&ad->antic_work);
651
652 if (aic->ttime_samples == 0) {
Nick Pigginf5b3db02005-11-07 00:59:53 -0800653 /* process anticipated on has exited or timed out*/
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 ad->exit_prob = (7*ad->exit_prob + 256)/8;
655 }
Nick Pigginf5b3db02005-11-07 00:59:53 -0800656 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
657 /* process not "saved" by a cooperating request */
658 ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8;
659 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 }
661 spin_unlock_irqrestore(q->queue_lock, flags);
662}
663
Nick Pigginf5b3db02005-11-07 00:59:53 -0800664static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic,
665 unsigned long ttime)
666{
667 /* fixed point: 1.0 == 1<<8 */
668 if (aic->ttime_samples == 0) {
669 ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8;
670 ad->new_ttime_mean = ad->new_ttime_total / 256;
671
672 ad->exit_prob = (7*ad->exit_prob)/8;
673 }
674 aic->ttime_samples = (7*aic->ttime_samples + 256) / 8;
675 aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8;
676 aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples;
677}
678
679static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic,
680 sector_t sdist)
681{
682 u64 total;
683
684 if (aic->seek_samples == 0) {
685 ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8;
686 ad->new_seek_mean = ad->new_seek_total / 256;
687 }
688
689 /*
690 * Don't allow the seek distance to get too large from the
691 * odd fragment, pagein, etc
692 */
693 if (aic->seek_samples <= 60) /* second&third seek */
694 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024);
695 else
696 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64);
697
698 aic->seek_samples = (7*aic->seek_samples + 256) / 8;
699 aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8;
700 total = aic->seek_total + (aic->seek_samples/2);
701 do_div(total, aic->seek_samples);
702 aic->seek_mean = (sector_t)total;
703}
704
705/*
706 * as_update_iohist keeps a decaying histogram of IO thinktimes, and
707 * updates @aic->ttime_mean based on that. It is called when a new
708 * request is queued.
709 */
710static void as_update_iohist(struct as_data *ad, struct as_io_context *aic,
711 struct request *rq)
712{
713 struct as_rq *arq = RQ_DATA(rq);
714 int data_dir = arq->is_sync;
715 unsigned long thinktime = 0;
716 sector_t seek_dist;
717
718 if (aic == NULL)
719 return;
720
721 if (data_dir == REQ_SYNC) {
722 unsigned long in_flight = atomic_read(&aic->nr_queued)
723 + atomic_read(&aic->nr_dispatched);
724 spin_lock(&aic->lock);
725 if (test_bit(AS_TASK_IORUNNING, &aic->state) ||
726 test_bit(AS_TASK_IOSTARTED, &aic->state)) {
727 /* Calculate read -> read thinktime */
728 if (test_bit(AS_TASK_IORUNNING, &aic->state)
729 && in_flight == 0) {
730 thinktime = jiffies - aic->last_end_request;
731 thinktime = min(thinktime, MAX_THINKTIME-1);
732 }
733 as_update_thinktime(ad, aic, thinktime);
734
735 /* Calculate read -> read seek distance */
736 if (aic->last_request_pos < rq->sector)
737 seek_dist = rq->sector - aic->last_request_pos;
738 else
739 seek_dist = aic->last_request_pos - rq->sector;
740 as_update_seekdist(ad, aic, seek_dist);
741 }
742 aic->last_request_pos = rq->sector + rq->nr_sectors;
743 set_bit(AS_TASK_IOSTARTED, &aic->state);
744 spin_unlock(&aic->lock);
745 }
746}
747
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748/*
749 * as_close_req decides if one request is considered "close" to the
750 * previous one issued.
751 */
Nick Pigginf5b3db02005-11-07 00:59:53 -0800752static int as_close_req(struct as_data *ad, struct as_io_context *aic,
753 struct as_rq *arq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754{
755 unsigned long delay; /* milliseconds */
756 sector_t last = ad->last_sector[ad->batch_data_dir];
757 sector_t next = arq->request->sector;
758 sector_t delta; /* acceptable close offset (in sectors) */
Nick Pigginf5b3db02005-11-07 00:59:53 -0800759 sector_t s;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760
761 if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished)
762 delay = 0;
763 else
764 delay = ((jiffies - ad->antic_start) * 1000) / HZ;
765
Nick Pigginf5b3db02005-11-07 00:59:53 -0800766 if (delay == 0)
767 delta = 8192;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 else if (delay <= 20 && delay <= ad->antic_expire)
Nick Pigginf5b3db02005-11-07 00:59:53 -0800769 delta = 8192 << delay;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 else
771 return 1;
772
Nick Pigginf5b3db02005-11-07 00:59:53 -0800773 if ((last <= next + (delta>>1)) && (next <= last + delta))
774 return 1;
775
776 if (last < next)
777 s = next - last;
778 else
779 s = last - next;
780
781 if (aic->seek_samples == 0) {
782 /*
783 * Process has just started IO. Use past statistics to
784 * gauge success possibility
785 */
786 if (ad->new_seek_mean > s) {
787 /* this request is better than what we're expecting */
788 return 1;
789 }
790
791 } else {
792 if (aic->seek_mean > s) {
793 /* this request is better than what we're expecting */
794 return 1;
795 }
796 }
797
798 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799}
800
801/*
802 * as_can_break_anticipation returns true if we have been anticipating this
803 * request.
804 *
805 * It also returns true if the process against which we are anticipating
806 * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to
807 * dispatch it ASAP, because we know that application will not be submitting
808 * any new reads.
809 *
Nick Pigginf5b3db02005-11-07 00:59:53 -0800810 * If the task which has submitted the request has exited, break anticipation.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 *
812 * If this task has queued some other IO, do not enter enticipation.
813 */
814static int as_can_break_anticipation(struct as_data *ad, struct as_rq *arq)
815{
816 struct io_context *ioc;
817 struct as_io_context *aic;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818
819 ioc = ad->io_context;
820 BUG_ON(!ioc);
821
822 if (arq && ioc == arq->io_context) {
823 /* request from same process */
824 return 1;
825 }
826
827 if (ad->ioc_finished && as_antic_expired(ad)) {
828 /*
829 * In this situation status should really be FINISHED,
830 * however the timer hasn't had the chance to run yet.
831 */
832 return 1;
833 }
834
835 aic = ioc->aic;
836 if (!aic)
837 return 0;
838
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 if (atomic_read(&aic->nr_queued) > 0) {
840 /* process has more requests queued */
841 return 1;
842 }
843
844 if (atomic_read(&aic->nr_dispatched) > 0) {
845 /* process has more requests dispatched */
846 return 1;
847 }
848
Nick Pigginf5b3db02005-11-07 00:59:53 -0800849 if (arq && arq->is_sync == REQ_SYNC && as_close_req(ad, aic, arq)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 /*
851 * Found a close request that is not one of ours.
852 *
Nick Pigginf5b3db02005-11-07 00:59:53 -0800853 * This makes close requests from another process update
854 * our IO history. Is generally useful when there are
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 * two or more cooperating processes working in the same
856 * area.
857 */
Nick Pigginf5b3db02005-11-07 00:59:53 -0800858 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
859 if (aic->ttime_samples == 0)
860 ad->exit_prob = (7*ad->exit_prob + 256)/8;
861
862 ad->exit_no_coop = (7*ad->exit_no_coop)/8;
863 }
864
865 as_update_iohist(ad, aic, arq->request);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 return 1;
867 }
868
Nick Pigginf5b3db02005-11-07 00:59:53 -0800869 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
870 /* process anticipated on has exited */
871 if (aic->ttime_samples == 0)
872 ad->exit_prob = (7*ad->exit_prob + 256)/8;
873
874 if (ad->exit_no_coop > 128)
875 return 1;
876 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877
878 if (aic->ttime_samples == 0) {
879 if (ad->new_ttime_mean > ad->antic_expire)
880 return 1;
Nick Pigginf5b3db02005-11-07 00:59:53 -0800881 if (ad->exit_prob * ad->exit_no_coop > 128*256)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 return 1;
883 } else if (aic->ttime_mean > ad->antic_expire) {
884 /* the process thinks too much between requests */
885 return 1;
886 }
887
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 return 0;
889}
890
891/*
892 * as_can_anticipate indicates weather we should either run arq
893 * or keep anticipating a better request.
894 */
895static int as_can_anticipate(struct as_data *ad, struct as_rq *arq)
896{
897 if (!ad->io_context)
898 /*
899 * Last request submitted was a write
900 */
901 return 0;
902
903 if (ad->antic_status == ANTIC_FINISHED)
904 /*
905 * Don't restart if we have just finished. Run the next request
906 */
907 return 0;
908
909 if (as_can_break_anticipation(ad, arq))
910 /*
911 * This request is a good candidate. Don't keep anticipating,
912 * run it.
913 */
914 return 0;
915
916 /*
917 * OK from here, we haven't finished, and don't have a decent request!
918 * Status is either ANTIC_OFF so start waiting,
919 * ANTIC_WAIT_REQ so continue waiting for request to finish
920 * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 */
922
923 return 1;
924}
925
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926/*
927 * as_update_arq must be called whenever a request (arq) is added to
928 * the sort_list. This function keeps caches up to date, and checks if the
929 * request might be one we are "anticipating"
930 */
931static void as_update_arq(struct as_data *ad, struct as_rq *arq)
932{
933 const int data_dir = arq->is_sync;
934
935 /* keep the next_arq cache up to date */
936 ad->next_arq[data_dir] = as_choose_req(ad, arq, ad->next_arq[data_dir]);
937
938 /*
939 * have we been anticipating this request?
940 * or does it come from the same process as the one we are anticipating
941 * for?
942 */
943 if (ad->antic_status == ANTIC_WAIT_REQ
944 || ad->antic_status == ANTIC_WAIT_NEXT) {
945 if (as_can_break_anticipation(ad, arq))
946 as_antic_stop(ad);
947 }
948}
949
950/*
951 * Gathers timings and resizes the write batch automatically
952 */
953static void update_write_batch(struct as_data *ad)
954{
955 unsigned long batch = ad->batch_expire[REQ_ASYNC];
956 long write_time;
957
958 write_time = (jiffies - ad->current_batch_expires) + batch;
959 if (write_time < 0)
960 write_time = 0;
961
962 if (write_time > batch && !ad->write_batch_idled) {
963 if (write_time > batch * 3)
964 ad->write_batch_count /= 2;
965 else
966 ad->write_batch_count--;
967 } else if (write_time < batch && ad->current_write_count == 0) {
968 if (batch > write_time * 3)
969 ad->write_batch_count *= 2;
970 else
971 ad->write_batch_count++;
972 }
973
974 if (ad->write_batch_count < 1)
975 ad->write_batch_count = 1;
976}
977
978/*
979 * as_completed_request is to be called when a request has completed and
980 * returned something to the requesting process, be it an error or data.
981 */
982static void as_completed_request(request_queue_t *q, struct request *rq)
983{
984 struct as_data *ad = q->elevator->elevator_data;
985 struct as_rq *arq = RQ_DATA(rq);
986
987 WARN_ON(!list_empty(&rq->queuelist));
988
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 if (arq->state != AS_RQ_REMOVED) {
990 printk("arq->state %d\n", arq->state);
991 WARN_ON(1);
992 goto out;
993 }
994
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 if (ad->changed_batch && ad->nr_dispatched == 1) {
996 kblockd_schedule_work(&ad->antic_work);
997 ad->changed_batch = 0;
998
999 if (ad->batch_data_dir == REQ_SYNC)
1000 ad->new_batch = 1;
1001 }
1002 WARN_ON(ad->nr_dispatched == 0);
1003 ad->nr_dispatched--;
1004
1005 /*
1006 * Start counting the batch from when a request of that direction is
1007 * actually serviced. This should help devices with big TCQ windows
1008 * and writeback caches
1009 */
1010 if (ad->new_batch && ad->batch_data_dir == arq->is_sync) {
1011 update_write_batch(ad);
1012 ad->current_batch_expires = jiffies +
1013 ad->batch_expire[REQ_SYNC];
1014 ad->new_batch = 0;
1015 }
1016
1017 if (ad->io_context == arq->io_context && ad->io_context) {
1018 ad->antic_start = jiffies;
1019 ad->ioc_finished = 1;
1020 if (ad->antic_status == ANTIC_WAIT_REQ) {
1021 /*
1022 * We were waiting on this request, now anticipate
1023 * the next one
1024 */
1025 as_antic_waitnext(ad);
1026 }
1027 }
1028
Jens Axboeb4878f22005-10-20 16:42:29 +02001029 as_put_io_context(arq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030out:
1031 arq->state = AS_RQ_POSTSCHED;
1032}
1033
1034/*
1035 * as_remove_queued_request removes a request from the pre dispatch queue
1036 * without updating refcounts. It is expected the caller will drop the
1037 * reference unless it replaces the request at somepart of the elevator
1038 * (ie. the dispatch queue)
1039 */
1040static void as_remove_queued_request(request_queue_t *q, struct request *rq)
1041{
1042 struct as_rq *arq = RQ_DATA(rq);
1043 const int data_dir = arq->is_sync;
1044 struct as_data *ad = q->elevator->elevator_data;
1045
1046 WARN_ON(arq->state != AS_RQ_QUEUED);
1047
1048 if (arq->io_context && arq->io_context->aic) {
1049 BUG_ON(!atomic_read(&arq->io_context->aic->nr_queued));
1050 atomic_dec(&arq->io_context->aic->nr_queued);
1051 }
1052
1053 /*
1054 * Update the "next_arq" cache if we are about to remove its
1055 * entry
1056 */
1057 if (ad->next_arq[data_dir] == arq)
1058 ad->next_arq[data_dir] = as_find_next_arq(ad, arq);
1059
1060 list_del_init(&arq->fifo);
Tejun Heo98b11472005-10-20 16:46:54 +02001061 as_del_arq_hash(arq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 as_del_arq_rb(ad, arq);
1063}
1064
1065/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066 * as_fifo_expired returns 0 if there are no expired reads on the fifo,
1067 * 1 otherwise. It is ratelimited so that we only perform the check once per
1068 * `fifo_expire' interval. Otherwise a large number of expired requests
1069 * would create a hopeless seekstorm.
1070 *
1071 * See as_antic_expired comment.
1072 */
1073static int as_fifo_expired(struct as_data *ad, int adir)
1074{
1075 struct as_rq *arq;
1076 long delta_jif;
1077
1078 delta_jif = jiffies - ad->last_check_fifo[adir];
1079 if (unlikely(delta_jif < 0))
1080 delta_jif = -delta_jif;
1081 if (delta_jif < ad->fifo_expire[adir])
1082 return 0;
1083
1084 ad->last_check_fifo[adir] = jiffies;
1085
1086 if (list_empty(&ad->fifo_list[adir]))
1087 return 0;
1088
1089 arq = list_entry_fifo(ad->fifo_list[adir].next);
1090
1091 return time_after(jiffies, arq->expires);
1092}
1093
1094/*
1095 * as_batch_expired returns true if the current batch has expired. A batch
1096 * is a set of reads or a set of writes.
1097 */
1098static inline int as_batch_expired(struct as_data *ad)
1099{
1100 if (ad->changed_batch || ad->new_batch)
1101 return 0;
1102
1103 if (ad->batch_data_dir == REQ_SYNC)
1104 /* TODO! add a check so a complete fifo gets written? */
1105 return time_after(jiffies, ad->current_batch_expires);
1106
1107 return time_after(jiffies, ad->current_batch_expires)
1108 || ad->current_write_count == 0;
1109}
1110
1111/*
1112 * move an entry to dispatch queue
1113 */
1114static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq)
1115{
1116 struct request *rq = arq->request;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 const int data_dir = arq->is_sync;
1118
1119 BUG_ON(!ON_RB(&arq->rb_node));
1120
1121 as_antic_stop(ad);
1122 ad->antic_status = ANTIC_OFF;
1123
1124 /*
1125 * This has to be set in order to be correctly updated by
1126 * as_find_next_arq
1127 */
1128 ad->last_sector[data_dir] = rq->sector + rq->nr_sectors;
1129
1130 if (data_dir == REQ_SYNC) {
1131 /* In case we have to anticipate after this */
1132 copy_io_context(&ad->io_context, &arq->io_context);
1133 } else {
1134 if (ad->io_context) {
1135 put_io_context(ad->io_context);
1136 ad->io_context = NULL;
1137 }
1138
1139 if (ad->current_write_count != 0)
1140 ad->current_write_count--;
1141 }
1142 ad->ioc_finished = 0;
1143
1144 ad->next_arq[data_dir] = as_find_next_arq(ad, arq);
1145
1146 /*
1147 * take it off the sort and fifo list, add to dispatch queue
1148 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 as_remove_queued_request(ad->q, rq);
1150 WARN_ON(arq->state != AS_RQ_QUEUED);
1151
Jens Axboeb4878f22005-10-20 16:42:29 +02001152 elv_dispatch_sort(ad->q, rq);
1153
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 arq->state = AS_RQ_DISPATCHED;
1155 if (arq->io_context && arq->io_context->aic)
1156 atomic_inc(&arq->io_context->aic->nr_dispatched);
1157 ad->nr_dispatched++;
1158}
1159
1160/*
1161 * as_dispatch_request selects the best request according to
1162 * read/write expire, batch expire, etc, and moves it to the dispatch
1163 * queue. Returns 1 if a request was found, 0 otherwise.
1164 */
Jens Axboeb4878f22005-10-20 16:42:29 +02001165static int as_dispatch_request(request_queue_t *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166{
Jens Axboeb4878f22005-10-20 16:42:29 +02001167 struct as_data *ad = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 struct as_rq *arq;
1169 const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]);
1170 const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]);
1171
Jens Axboeb4878f22005-10-20 16:42:29 +02001172 if (unlikely(force)) {
1173 /*
1174 * Forced dispatch, accounting is useless. Reset
1175 * accounting states and dump fifo_lists. Note that
1176 * batch_data_dir is reset to REQ_SYNC to avoid
1177 * screwing write batch accounting as write batch
1178 * accounting occurs on W->R transition.
1179 */
1180 int dispatched = 0;
1181
1182 ad->batch_data_dir = REQ_SYNC;
1183 ad->changed_batch = 0;
1184 ad->new_batch = 0;
1185
1186 while (ad->next_arq[REQ_SYNC]) {
1187 as_move_to_dispatch(ad, ad->next_arq[REQ_SYNC]);
1188 dispatched++;
1189 }
1190 ad->last_check_fifo[REQ_SYNC] = jiffies;
1191
1192 while (ad->next_arq[REQ_ASYNC]) {
1193 as_move_to_dispatch(ad, ad->next_arq[REQ_ASYNC]);
1194 dispatched++;
1195 }
1196 ad->last_check_fifo[REQ_ASYNC] = jiffies;
1197
1198 return dispatched;
1199 }
1200
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201 /* Signal that the write batch was uncontended, so we can't time it */
1202 if (ad->batch_data_dir == REQ_ASYNC && !reads) {
1203 if (ad->current_write_count == 0 || !writes)
1204 ad->write_batch_idled = 1;
1205 }
1206
1207 if (!(reads || writes)
1208 || ad->antic_status == ANTIC_WAIT_REQ
1209 || ad->antic_status == ANTIC_WAIT_NEXT
1210 || ad->changed_batch)
1211 return 0;
1212
Nick Pigginf5b3db02005-11-07 00:59:53 -08001213 if (!(reads && writes && as_batch_expired(ad))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214 /*
1215 * batch is still running or no reads or no writes
1216 */
1217 arq = ad->next_arq[ad->batch_data_dir];
1218
1219 if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) {
1220 if (as_fifo_expired(ad, REQ_SYNC))
1221 goto fifo_expired;
1222
1223 if (as_can_anticipate(ad, arq)) {
1224 as_antic_waitreq(ad);
1225 return 0;
1226 }
1227 }
1228
1229 if (arq) {
1230 /* we have a "next request" */
1231 if (reads && !writes)
1232 ad->current_batch_expires =
1233 jiffies + ad->batch_expire[REQ_SYNC];
1234 goto dispatch_request;
1235 }
1236 }
1237
1238 /*
1239 * at this point we are not running a batch. select the appropriate
1240 * data direction (read / write)
1241 */
1242
1243 if (reads) {
1244 BUG_ON(RB_EMPTY(&ad->sort_list[REQ_SYNC]));
1245
1246 if (writes && ad->batch_data_dir == REQ_SYNC)
1247 /*
1248 * Last batch was a read, switch to writes
1249 */
1250 goto dispatch_writes;
1251
1252 if (ad->batch_data_dir == REQ_ASYNC) {
1253 WARN_ON(ad->new_batch);
1254 ad->changed_batch = 1;
1255 }
1256 ad->batch_data_dir = REQ_SYNC;
1257 arq = list_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
1258 ad->last_check_fifo[ad->batch_data_dir] = jiffies;
1259 goto dispatch_request;
1260 }
1261
1262 /*
1263 * the last batch was a read
1264 */
1265
1266 if (writes) {
1267dispatch_writes:
1268 BUG_ON(RB_EMPTY(&ad->sort_list[REQ_ASYNC]));
1269
1270 if (ad->batch_data_dir == REQ_SYNC) {
1271 ad->changed_batch = 1;
1272
1273 /*
1274 * new_batch might be 1 when the queue runs out of
1275 * reads. A subsequent submission of a write might
1276 * cause a change of batch before the read is finished.
1277 */
1278 ad->new_batch = 0;
1279 }
1280 ad->batch_data_dir = REQ_ASYNC;
1281 ad->current_write_count = ad->write_batch_count;
1282 ad->write_batch_idled = 0;
1283 arq = ad->next_arq[ad->batch_data_dir];
1284 goto dispatch_request;
1285 }
1286
1287 BUG();
1288 return 0;
1289
1290dispatch_request:
1291 /*
1292 * If a request has expired, service it.
1293 */
1294
1295 if (as_fifo_expired(ad, ad->batch_data_dir)) {
1296fifo_expired:
1297 arq = list_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
1298 BUG_ON(arq == NULL);
1299 }
1300
1301 if (ad->changed_batch) {
1302 WARN_ON(ad->new_batch);
1303
1304 if (ad->nr_dispatched)
1305 return 0;
1306
1307 if (ad->batch_data_dir == REQ_ASYNC)
1308 ad->current_batch_expires = jiffies +
1309 ad->batch_expire[REQ_ASYNC];
1310 else
1311 ad->new_batch = 1;
1312
1313 ad->changed_batch = 0;
1314 }
1315
1316 /*
1317 * arq is the selected appropriate request.
1318 */
1319 as_move_to_dispatch(ad, arq);
1320
1321 return 1;
1322}
1323
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 * add arq to rbtree and fifo
1326 */
Jens Axboeb4878f22005-10-20 16:42:29 +02001327static void as_add_request(request_queue_t *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328{
Jens Axboeb4878f22005-10-20 16:42:29 +02001329 struct as_data *ad = q->elevator->elevator_data;
1330 struct as_rq *arq = RQ_DATA(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 int data_dir;
1332
Jens Axboeb4878f22005-10-20 16:42:29 +02001333 arq->state = AS_RQ_NEW;
1334
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 if (rq_data_dir(arq->request) == READ
1336 || current->flags&PF_SYNCWRITE)
1337 arq->is_sync = 1;
1338 else
1339 arq->is_sync = 0;
1340 data_dir = arq->is_sync;
1341
1342 arq->io_context = as_get_io_context();
1343
1344 if (arq->io_context) {
1345 as_update_iohist(ad, arq->io_context->aic, arq->request);
1346 atomic_inc(&arq->io_context->aic->nr_queued);
1347 }
1348
Tejun Heoef9be1d2005-11-11 14:27:09 +01001349 as_add_arq_rb(ad, arq);
1350 if (rq_mergeable(arq->request))
1351 as_add_arq_hash(ad, arq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352
Tejun Heoef9be1d2005-11-11 14:27:09 +01001353 /*
1354 * set expire time (only used for reads) and add to fifo list
1355 */
1356 arq->expires = jiffies + ad->fifo_expire[data_dir];
1357 list_add_tail(&arq->fifo, &ad->fifo_list[data_dir]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358
Tejun Heoef9be1d2005-11-11 14:27:09 +01001359 as_update_arq(ad, arq); /* keep state machine up to date */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 arq->state = AS_RQ_QUEUED;
1361}
1362
Jens Axboeb4878f22005-10-20 16:42:29 +02001363static void as_activate_request(request_queue_t *q, struct request *rq)
1364{
1365 struct as_rq *arq = RQ_DATA(rq);
1366
1367 WARN_ON(arq->state != AS_RQ_DISPATCHED);
1368 arq->state = AS_RQ_REMOVED;
1369 if (arq->io_context && arq->io_context->aic)
1370 atomic_dec(&arq->io_context->aic->nr_dispatched);
1371}
1372
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373static void as_deactivate_request(request_queue_t *q, struct request *rq)
1374{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375 struct as_rq *arq = RQ_DATA(rq);
1376
Jens Axboeb4878f22005-10-20 16:42:29 +02001377 WARN_ON(arq->state != AS_RQ_REMOVED);
1378 arq->state = AS_RQ_DISPATCHED;
1379 if (arq->io_context && arq->io_context->aic)
1380 atomic_inc(&arq->io_context->aic->nr_dispatched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381}
1382
1383/*
1384 * as_queue_empty tells us if there are requests left in the device. It may
1385 * not be the case that a driver can get the next request even if the queue
1386 * is not empty - it is used in the block layer to check for plugging and
1387 * merging opportunities
1388 */
1389static int as_queue_empty(request_queue_t *q)
1390{
1391 struct as_data *ad = q->elevator->elevator_data;
1392
Jens Axboeb4878f22005-10-20 16:42:29 +02001393 return list_empty(&ad->fifo_list[REQ_ASYNC])
1394 && list_empty(&ad->fifo_list[REQ_SYNC]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395}
1396
Nick Pigginf5b3db02005-11-07 00:59:53 -08001397static struct request *as_former_request(request_queue_t *q,
1398 struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399{
1400 struct as_rq *arq = RQ_DATA(rq);
1401 struct rb_node *rbprev = rb_prev(&arq->rb_node);
1402 struct request *ret = NULL;
1403
1404 if (rbprev)
1405 ret = rb_entry_arq(rbprev)->request;
1406
1407 return ret;
1408}
1409
Nick Pigginf5b3db02005-11-07 00:59:53 -08001410static struct request *as_latter_request(request_queue_t *q,
1411 struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412{
1413 struct as_rq *arq = RQ_DATA(rq);
1414 struct rb_node *rbnext = rb_next(&arq->rb_node);
1415 struct request *ret = NULL;
1416
1417 if (rbnext)
1418 ret = rb_entry_arq(rbnext)->request;
1419
1420 return ret;
1421}
1422
1423static int
1424as_merge(request_queue_t *q, struct request **req, struct bio *bio)
1425{
1426 struct as_data *ad = q->elevator->elevator_data;
1427 sector_t rb_key = bio->bi_sector + bio_sectors(bio);
1428 struct request *__rq;
1429 int ret;
1430
1431 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 * see if the merge hash can satisfy a back merge
1433 */
1434 __rq = as_find_arq_hash(ad, bio->bi_sector);
1435 if (__rq) {
1436 BUG_ON(__rq->sector + __rq->nr_sectors != bio->bi_sector);
1437
1438 if (elv_rq_merge_ok(__rq, bio)) {
1439 ret = ELEVATOR_BACK_MERGE;
1440 goto out;
1441 }
1442 }
1443
1444 /*
1445 * check for front merge
1446 */
1447 __rq = as_find_arq_rb(ad, rb_key, bio_data_dir(bio));
1448 if (__rq) {
1449 BUG_ON(rb_key != rq_rb_key(__rq));
1450
1451 if (elv_rq_merge_ok(__rq, bio)) {
1452 ret = ELEVATOR_FRONT_MERGE;
1453 goto out;
1454 }
1455 }
1456
1457 return ELEVATOR_NO_MERGE;
1458out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459 if (ret) {
1460 if (rq_mergeable(__rq))
1461 as_hot_arq_hash(ad, RQ_DATA(__rq));
1462 }
1463 *req = __rq;
1464 return ret;
1465}
1466
1467static void as_merged_request(request_queue_t *q, struct request *req)
1468{
1469 struct as_data *ad = q->elevator->elevator_data;
1470 struct as_rq *arq = RQ_DATA(req);
1471
1472 /*
1473 * hash always needs to be repositioned, key is end sector
1474 */
1475 as_del_arq_hash(arq);
1476 as_add_arq_hash(ad, arq);
1477
1478 /*
1479 * if the merge was a front merge, we need to reposition request
1480 */
1481 if (rq_rb_key(req) != arq->rb_key) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 as_del_arq_rb(ad, arq);
Tejun Heoef9be1d2005-11-11 14:27:09 +01001483 as_add_arq_rb(ad, arq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 /*
1485 * Note! At this stage of this and the next function, our next
1486 * request may not be optimal - eg the request may have "grown"
1487 * behind the disk head. We currently don't bother adjusting.
1488 */
1489 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490}
1491
Nick Pigginf5b3db02005-11-07 00:59:53 -08001492static void as_merged_requests(request_queue_t *q, struct request *req,
1493 struct request *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494{
1495 struct as_data *ad = q->elevator->elevator_data;
1496 struct as_rq *arq = RQ_DATA(req);
1497 struct as_rq *anext = RQ_DATA(next);
1498
1499 BUG_ON(!arq);
1500 BUG_ON(!anext);
1501
1502 /*
1503 * reposition arq (this is the merged request) in hash, and in rbtree
1504 * in case of a front merge
1505 */
1506 as_del_arq_hash(arq);
1507 as_add_arq_hash(ad, arq);
1508
1509 if (rq_rb_key(req) != arq->rb_key) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510 as_del_arq_rb(ad, arq);
Tejun Heoef9be1d2005-11-11 14:27:09 +01001511 as_add_arq_rb(ad, arq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 }
1513
1514 /*
1515 * if anext expires before arq, assign its expire time to arq
1516 * and move into anext position (anext will be deleted) in fifo
1517 */
1518 if (!list_empty(&arq->fifo) && !list_empty(&anext->fifo)) {
1519 if (time_before(anext->expires, arq->expires)) {
1520 list_move(&arq->fifo, &anext->fifo);
1521 arq->expires = anext->expires;
1522 /*
1523 * Don't copy here but swap, because when anext is
1524 * removed below, it must contain the unused context
1525 */
1526 swap_io_context(&arq->io_context, &anext->io_context);
1527 }
1528 }
1529
1530 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 * kill knowledge of next, this one is a goner
1532 */
1533 as_remove_queued_request(q, next);
Jens Axboeb4878f22005-10-20 16:42:29 +02001534 as_put_io_context(anext);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535
1536 anext->state = AS_RQ_MERGED;
1537}
1538
1539/*
1540 * This is executed in a "deferred" process context, by kblockd. It calls the
1541 * driver's request_fn so the driver can submit that request.
1542 *
1543 * IMPORTANT! This guy will reenter the elevator, so set up all queue global
1544 * state before calling, and don't rely on any state over calls.
1545 *
1546 * FIXME! dispatch queue is not a queue at all!
1547 */
1548static void as_work_handler(void *data)
1549{
1550 struct request_queue *q = data;
1551 unsigned long flags;
1552
1553 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboeb4878f22005-10-20 16:42:29 +02001554 if (!as_queue_empty(q))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555 q->request_fn(q);
1556 spin_unlock_irqrestore(q->queue_lock, flags);
1557}
1558
1559static void as_put_request(request_queue_t *q, struct request *rq)
1560{
1561 struct as_data *ad = q->elevator->elevator_data;
1562 struct as_rq *arq = RQ_DATA(rq);
1563
1564 if (!arq) {
1565 WARN_ON(1);
1566 return;
1567 }
1568
Jens Axboeb4878f22005-10-20 16:42:29 +02001569 if (unlikely(arq->state != AS_RQ_POSTSCHED &&
1570 arq->state != AS_RQ_PRESCHED &&
1571 arq->state != AS_RQ_MERGED)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 printk("arq->state %d\n", arq->state);
1573 WARN_ON(1);
1574 }
1575
1576 mempool_free(arq, ad->arq_pool);
1577 rq->elevator_private = NULL;
1578}
1579
Jens Axboe22e2c502005-06-27 10:55:12 +02001580static int as_set_request(request_queue_t *q, struct request *rq,
Al Viro8267e262005-10-21 03:20:53 -04001581 struct bio *bio, gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582{
1583 struct as_data *ad = q->elevator->elevator_data;
1584 struct as_rq *arq = mempool_alloc(ad->arq_pool, gfp_mask);
1585
1586 if (arq) {
1587 memset(arq, 0, sizeof(*arq));
1588 RB_CLEAR(&arq->rb_node);
1589 arq->request = rq;
1590 arq->state = AS_RQ_PRESCHED;
1591 arq->io_context = NULL;
1592 INIT_LIST_HEAD(&arq->hash);
1593 arq->on_hash = 0;
1594 INIT_LIST_HEAD(&arq->fifo);
1595 rq->elevator_private = arq;
1596 return 0;
1597 }
1598
1599 return 1;
1600}
1601
Jens Axboe22e2c502005-06-27 10:55:12 +02001602static int as_may_queue(request_queue_t *q, int rw, struct bio *bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603{
1604 int ret = ELV_MQUEUE_MAY;
1605 struct as_data *ad = q->elevator->elevator_data;
1606 struct io_context *ioc;
1607 if (ad->antic_status == ANTIC_WAIT_REQ ||
1608 ad->antic_status == ANTIC_WAIT_NEXT) {
1609 ioc = as_get_io_context();
1610 if (ad->io_context == ioc)
1611 ret = ELV_MQUEUE_MUST;
1612 put_io_context(ioc);
1613 }
1614
1615 return ret;
1616}
1617
1618static void as_exit_queue(elevator_t *e)
1619{
1620 struct as_data *ad = e->elevator_data;
1621
1622 del_timer_sync(&ad->antic_timer);
1623 kblockd_flush();
1624
1625 BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC]));
1626 BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC]));
1627
1628 mempool_destroy(ad->arq_pool);
1629 put_io_context(ad->io_context);
1630 kfree(ad->hash);
1631 kfree(ad);
1632}
1633
1634/*
1635 * initialize elevator private data (as_data), and alloc a arq for
1636 * each request on the free lists
1637 */
1638static int as_init_queue(request_queue_t *q, elevator_t *e)
1639{
1640 struct as_data *ad;
1641 int i;
1642
1643 if (!arq_pool)
1644 return -ENOMEM;
1645
Christoph Lameter19460892005-06-23 00:08:19 -07001646 ad = kmalloc_node(sizeof(*ad), GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647 if (!ad)
1648 return -ENOMEM;
1649 memset(ad, 0, sizeof(*ad));
1650
1651 ad->q = q; /* Identify what queue the data belongs to */
1652
Christoph Lameter19460892005-06-23 00:08:19 -07001653 ad->hash = kmalloc_node(sizeof(struct list_head)*AS_HASH_ENTRIES,
1654 GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 if (!ad->hash) {
1656 kfree(ad);
1657 return -ENOMEM;
1658 }
1659
Christoph Lameter19460892005-06-23 00:08:19 -07001660 ad->arq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
1661 mempool_free_slab, arq_pool, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662 if (!ad->arq_pool) {
1663 kfree(ad->hash);
1664 kfree(ad);
1665 return -ENOMEM;
1666 }
1667
1668 /* anticipatory scheduling helpers */
1669 ad->antic_timer.function = as_antic_timeout;
1670 ad->antic_timer.data = (unsigned long)q;
1671 init_timer(&ad->antic_timer);
1672 INIT_WORK(&ad->antic_work, as_work_handler, q);
1673
1674 for (i = 0; i < AS_HASH_ENTRIES; i++)
1675 INIT_LIST_HEAD(&ad->hash[i]);
1676
1677 INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]);
1678 INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]);
1679 ad->sort_list[REQ_SYNC] = RB_ROOT;
1680 ad->sort_list[REQ_ASYNC] = RB_ROOT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 ad->fifo_expire[REQ_SYNC] = default_read_expire;
1682 ad->fifo_expire[REQ_ASYNC] = default_write_expire;
1683 ad->antic_expire = default_antic_expire;
1684 ad->batch_expire[REQ_SYNC] = default_read_batch_expire;
1685 ad->batch_expire[REQ_ASYNC] = default_write_batch_expire;
1686 e->elevator_data = ad;
1687
1688 ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC];
1689 ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10;
1690 if (ad->write_batch_count < 2)
1691 ad->write_batch_count = 2;
1692
1693 return 0;
1694}
1695
1696/*
1697 * sysfs parts below
1698 */
1699struct as_fs_entry {
1700 struct attribute attr;
1701 ssize_t (*show)(struct as_data *, char *);
1702 ssize_t (*store)(struct as_data *, const char *, size_t);
1703};
1704
1705static ssize_t
1706as_var_show(unsigned int var, char *page)
1707{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 return sprintf(page, "%d\n", var);
1709}
1710
1711static ssize_t
1712as_var_store(unsigned long *var, const char *page, size_t count)
1713{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 char *p = (char *) page;
1715
Jens Axboec9b3ad62005-07-27 11:43:37 -07001716 *var = simple_strtoul(p, &p, 10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 return count;
1718}
1719
1720static ssize_t as_est_show(struct as_data *ad, char *page)
1721{
1722 int pos = 0;
1723
Nick Pigginf5b3db02005-11-07 00:59:53 -08001724 pos += sprintf(page+pos, "%lu %% exit probability\n",
1725 100*ad->exit_prob/256);
1726 pos += sprintf(page+pos, "%lu %% probability of exiting without a "
1727 "cooperating process submitting IO\n",
1728 100*ad->exit_no_coop/256);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean);
Nick Pigginf5b3db02005-11-07 00:59:53 -08001730 pos += sprintf(page+pos, "%llu sectors new seek distance\n",
1731 (unsigned long long)ad->new_seek_mean);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732
1733 return pos;
1734}
1735
1736#define SHOW_FUNCTION(__FUNC, __VAR) \
1737static ssize_t __FUNC(struct as_data *ad, char *page) \
1738{ \
1739 return as_var_show(jiffies_to_msecs((__VAR)), (page)); \
1740}
1741SHOW_FUNCTION(as_readexpire_show, ad->fifo_expire[REQ_SYNC]);
1742SHOW_FUNCTION(as_writeexpire_show, ad->fifo_expire[REQ_ASYNC]);
1743SHOW_FUNCTION(as_anticexpire_show, ad->antic_expire);
1744SHOW_FUNCTION(as_read_batchexpire_show, ad->batch_expire[REQ_SYNC]);
1745SHOW_FUNCTION(as_write_batchexpire_show, ad->batch_expire[REQ_ASYNC]);
1746#undef SHOW_FUNCTION
1747
1748#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
1749static ssize_t __FUNC(struct as_data *ad, const char *page, size_t count) \
1750{ \
1751 int ret = as_var_store(__PTR, (page), count); \
1752 if (*(__PTR) < (MIN)) \
1753 *(__PTR) = (MIN); \
1754 else if (*(__PTR) > (MAX)) \
1755 *(__PTR) = (MAX); \
1756 *(__PTR) = msecs_to_jiffies(*(__PTR)); \
1757 return ret; \
1758}
1759STORE_FUNCTION(as_readexpire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX);
1760STORE_FUNCTION(as_writeexpire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX);
1761STORE_FUNCTION(as_anticexpire_store, &ad->antic_expire, 0, INT_MAX);
1762STORE_FUNCTION(as_read_batchexpire_store,
1763 &ad->batch_expire[REQ_SYNC], 0, INT_MAX);
1764STORE_FUNCTION(as_write_batchexpire_store,
1765 &ad->batch_expire[REQ_ASYNC], 0, INT_MAX);
1766#undef STORE_FUNCTION
1767
1768static struct as_fs_entry as_est_entry = {
1769 .attr = {.name = "est_time", .mode = S_IRUGO },
1770 .show = as_est_show,
1771};
1772static struct as_fs_entry as_readexpire_entry = {
1773 .attr = {.name = "read_expire", .mode = S_IRUGO | S_IWUSR },
1774 .show = as_readexpire_show,
1775 .store = as_readexpire_store,
1776};
1777static struct as_fs_entry as_writeexpire_entry = {
1778 .attr = {.name = "write_expire", .mode = S_IRUGO | S_IWUSR },
1779 .show = as_writeexpire_show,
1780 .store = as_writeexpire_store,
1781};
1782static struct as_fs_entry as_anticexpire_entry = {
1783 .attr = {.name = "antic_expire", .mode = S_IRUGO | S_IWUSR },
1784 .show = as_anticexpire_show,
1785 .store = as_anticexpire_store,
1786};
1787static struct as_fs_entry as_read_batchexpire_entry = {
1788 .attr = {.name = "read_batch_expire", .mode = S_IRUGO | S_IWUSR },
1789 .show = as_read_batchexpire_show,
1790 .store = as_read_batchexpire_store,
1791};
1792static struct as_fs_entry as_write_batchexpire_entry = {
1793 .attr = {.name = "write_batch_expire", .mode = S_IRUGO | S_IWUSR },
1794 .show = as_write_batchexpire_show,
1795 .store = as_write_batchexpire_store,
1796};
1797
1798static struct attribute *default_attrs[] = {
1799 &as_est_entry.attr,
1800 &as_readexpire_entry.attr,
1801 &as_writeexpire_entry.attr,
1802 &as_anticexpire_entry.attr,
1803 &as_read_batchexpire_entry.attr,
1804 &as_write_batchexpire_entry.attr,
1805 NULL,
1806};
1807
1808#define to_as(atr) container_of((atr), struct as_fs_entry, attr)
1809
1810static ssize_t
1811as_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
1812{
1813 elevator_t *e = container_of(kobj, elevator_t, kobj);
1814 struct as_fs_entry *entry = to_as(attr);
1815
1816 if (!entry->show)
Dmitry Torokhov6c1852a2005-04-29 01:26:06 -05001817 return -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818
1819 return entry->show(e->elevator_data, page);
1820}
1821
1822static ssize_t
1823as_attr_store(struct kobject *kobj, struct attribute *attr,
1824 const char *page, size_t length)
1825{
1826 elevator_t *e = container_of(kobj, elevator_t, kobj);
1827 struct as_fs_entry *entry = to_as(attr);
1828
1829 if (!entry->store)
Dmitry Torokhov6c1852a2005-04-29 01:26:06 -05001830 return -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831
1832 return entry->store(e->elevator_data, page, length);
1833}
1834
1835static struct sysfs_ops as_sysfs_ops = {
1836 .show = as_attr_show,
1837 .store = as_attr_store,
1838};
1839
1840static struct kobj_type as_ktype = {
1841 .sysfs_ops = &as_sysfs_ops,
1842 .default_attrs = default_attrs,
1843};
1844
1845static struct elevator_type iosched_as = {
1846 .ops = {
1847 .elevator_merge_fn = as_merge,
1848 .elevator_merged_fn = as_merged_request,
1849 .elevator_merge_req_fn = as_merged_requests,
Jens Axboeb4878f22005-10-20 16:42:29 +02001850 .elevator_dispatch_fn = as_dispatch_request,
1851 .elevator_add_req_fn = as_add_request,
1852 .elevator_activate_req_fn = as_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 .elevator_deactivate_req_fn = as_deactivate_request,
1854 .elevator_queue_empty_fn = as_queue_empty,
1855 .elevator_completed_req_fn = as_completed_request,
1856 .elevator_former_req_fn = as_former_request,
1857 .elevator_latter_req_fn = as_latter_request,
1858 .elevator_set_req_fn = as_set_request,
1859 .elevator_put_req_fn = as_put_request,
1860 .elevator_may_queue_fn = as_may_queue,
1861 .elevator_init_fn = as_init_queue,
1862 .elevator_exit_fn = as_exit_queue,
1863 },
1864
1865 .elevator_ktype = &as_ktype,
1866 .elevator_name = "anticipatory",
1867 .elevator_owner = THIS_MODULE,
1868};
1869
1870static int __init as_init(void)
1871{
1872 int ret;
1873
1874 arq_pool = kmem_cache_create("as_arq", sizeof(struct as_rq),
1875 0, 0, NULL, NULL);
1876 if (!arq_pool)
1877 return -ENOMEM;
1878
1879 ret = elv_register(&iosched_as);
1880 if (!ret) {
1881 /*
1882 * don't allow AS to get unregistered, since we would have
1883 * to browse all tasks in the system and release their
1884 * as_io_context first
1885 */
1886 __module_get(THIS_MODULE);
1887 return 0;
1888 }
1889
1890 kmem_cache_destroy(arq_pool);
1891 return ret;
1892}
1893
1894static void __exit as_exit(void)
1895{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 elv_unregister(&iosched_as);
Christoph Hellwig83521d32005-10-30 15:01:39 -08001897 kmem_cache_destroy(arq_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898}
1899
1900module_init(as_init);
1901module_exit(as_exit);
1902
1903MODULE_AUTHOR("Nick Piggin");
1904MODULE_LICENSE("GPL");
1905MODULE_DESCRIPTION("anticipatory IO scheduler");