blob: 5934c4bfd52a3d9b71331785ac6fafc26ddef992 [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 *
Jens Axboe0fe23472006-09-04 15:41:16 +02004 * Copyright (C) 2002 Jens Axboe <axboe@kernel.dk>
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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/module.h>
14#include <linux/slab.h>
15#include <linux/init.h>
16#include <linux/compiler.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/rbtree.h>
18#include <linux/interrupt.h>
19
20#define REQ_SYNC 1
21#define REQ_ASYNC 0
22
23/*
24 * See Documentation/block/as-iosched.txt
25 */
26
27/*
28 * max time before a read is submitted.
29 */
30#define default_read_expire (HZ / 8)
31
32/*
33 * ditto for writes, these limits are not hard, even
34 * if the disk is capable of satisfying them.
35 */
36#define default_write_expire (HZ / 4)
37
38/*
39 * read_batch_expire describes how long we will allow a stream of reads to
40 * persist before looking to see whether it is time to switch over to writes.
41 */
42#define default_read_batch_expire (HZ / 2)
43
44/*
45 * write_batch_expire describes how long we want a stream of writes to run for.
46 * This is not a hard limit, but a target we set for the auto-tuning thingy.
47 * See, the problem is: we can send a lot of writes to disk cache / TCQ in
48 * a short amount of time...
49 */
50#define default_write_batch_expire (HZ / 8)
51
52/*
53 * max time we may wait to anticipate a read (default around 6ms)
54 */
55#define default_antic_expire ((HZ / 150) ? HZ / 150 : 1)
56
57/*
58 * Keep track of up to 20ms thinktimes. We can go as big as we like here,
59 * however huge values tend to interfere and not decay fast enough. A program
60 * might be in a non-io phase of operation. Waiting on user input for example,
61 * or doing a lengthy computation. A small penalty can be justified there, and
62 * will still catch out those processes that constantly have large thinktimes.
63 */
64#define MAX_THINKTIME (HZ/50UL)
65
66/* Bits in as_io_context.state */
67enum as_io_states {
Nick Pigginf5b3db02005-11-07 00:59:53 -080068 AS_TASK_RUNNING=0, /* Process has not exited */
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 AS_TASK_IOSTARTED, /* Process has started some IO */
70 AS_TASK_IORUNNING, /* Process has completed some IO */
71};
72
73enum anticipation_status {
74 ANTIC_OFF=0, /* Not anticipating (normal operation) */
75 ANTIC_WAIT_REQ, /* The last read has not yet completed */
76 ANTIC_WAIT_NEXT, /* Currently anticipating a request vs
77 last read (which has completed) */
78 ANTIC_FINISHED, /* Anticipating but have found a candidate
79 * or timed out */
80};
81
82struct as_data {
83 /*
84 * run time data
85 */
86
87 struct request_queue *q; /* the "owner" queue */
88
89 /*
90 * requests (as_rq s) are present on both sort_list and fifo_list
91 */
92 struct rb_root sort_list[2];
93 struct list_head fifo_list[2];
94
Jens Axboe8a8e6742006-07-18 21:07:29 +020095 struct request *next_rq[2]; /* next in sort order */
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 sector_t last_sector[2]; /* last REQ_SYNC & REQ_ASYNC sectors */
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
98 unsigned long exit_prob; /* probability a task will exit while
99 being waited on */
Nick Pigginf5b3db02005-11-07 00:59:53 -0800100 unsigned long exit_no_coop; /* probablility an exited task will
101 not be part of a later cooperating
102 request */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 unsigned long new_ttime_total; /* mean thinktime on new proc */
104 unsigned long new_ttime_mean;
105 u64 new_seek_total; /* mean seek on new proc */
106 sector_t new_seek_mean;
107
108 unsigned long current_batch_expires;
109 unsigned long last_check_fifo[2];
110 int changed_batch; /* 1: waiting for old batch to end */
111 int new_batch; /* 1: waiting on first read complete */
112 int batch_data_dir; /* current batch REQ_SYNC / REQ_ASYNC */
113 int write_batch_count; /* max # of reqs in a write batch */
114 int current_write_count; /* how many requests left this batch */
115 int write_batch_idled; /* has the write batch gone idle? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116
117 enum anticipation_status antic_status;
118 unsigned long antic_start; /* jiffies: when it started */
119 struct timer_list antic_timer; /* anticipatory scheduling timer */
120 struct work_struct antic_work; /* Deferred unplugging */
121 struct io_context *io_context; /* Identify the expected process */
122 int ioc_finished; /* IO associated with io_context is finished */
123 int nr_dispatched;
124
125 /*
126 * settings that change how the i/o scheduler behaves
127 */
128 unsigned long fifo_expire[2];
129 unsigned long batch_expire[2];
130 unsigned long antic_expire;
131};
132
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133/*
134 * per-request data.
135 */
136enum arq_state {
137 AS_RQ_NEW=0, /* New - not referenced and not on any lists */
138 AS_RQ_QUEUED, /* In the request queue. It belongs to the
139 scheduler */
140 AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the
141 driver now */
142 AS_RQ_PRESCHED, /* Debug poisoning for requests being used */
143 AS_RQ_REMOVED,
144 AS_RQ_MERGED,
145 AS_RQ_POSTSCHED, /* when they shouldn't be */
146};
147
Jens Axboe8a8e6742006-07-18 21:07:29 +0200148#define RQ_IOC(rq) ((struct io_context *) (rq)->elevator_private)
149#define RQ_STATE(rq) ((enum arq_state)(rq)->elevator_private2)
150#define RQ_SET_STATE(rq, state) ((rq)->elevator_private2 = (void *) state)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151
Jens Axboee4313dd2006-07-19 05:10:01 +0200152static DEFINE_PER_CPU(unsigned long, ioc_count);
Al Viro334e94d2006-03-18 15:05:53 -0500153static struct completion *ioc_gone;
154
Jens Axboe8a8e6742006-07-18 21:07:29 +0200155static void as_move_to_dispatch(struct as_data *ad, struct request *rq);
Tejun Heoef9be1d2005-11-11 14:27:09 +0100156static void as_antic_stop(struct as_data *ad);
157
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158/*
159 * IO Context helper functions
160 */
161
162/* Called to deallocate the as_io_context */
163static void free_as_io_context(struct as_io_context *aic)
164{
165 kfree(aic);
Jens Axboee4313dd2006-07-19 05:10:01 +0200166 elv_ioc_count_dec(ioc_count);
167 if (ioc_gone && !elv_ioc_count_read(ioc_count))
Al Viro334e94d2006-03-18 15:05:53 -0500168 complete(ioc_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
170
Al Viroe17a9482006-03-18 13:21:20 -0500171static void as_trim(struct io_context *ioc)
172{
Al Viro334e94d2006-03-18 15:05:53 -0500173 if (ioc->aic)
174 free_as_io_context(ioc->aic);
Al Viroe17a9482006-03-18 13:21:20 -0500175 ioc->aic = NULL;
176}
177
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178/* Called when the task exits */
179static void exit_as_io_context(struct as_io_context *aic)
180{
181 WARN_ON(!test_bit(AS_TASK_RUNNING, &aic->state));
182 clear_bit(AS_TASK_RUNNING, &aic->state);
183}
184
185static struct as_io_context *alloc_as_io_context(void)
186{
187 struct as_io_context *ret;
188
189 ret = kmalloc(sizeof(*ret), GFP_ATOMIC);
190 if (ret) {
191 ret->dtor = free_as_io_context;
192 ret->exit = exit_as_io_context;
193 ret->state = 1 << AS_TASK_RUNNING;
194 atomic_set(&ret->nr_queued, 0);
195 atomic_set(&ret->nr_dispatched, 0);
196 spin_lock_init(&ret->lock);
197 ret->ttime_total = 0;
198 ret->ttime_samples = 0;
199 ret->ttime_mean = 0;
200 ret->seek_total = 0;
201 ret->seek_samples = 0;
202 ret->seek_mean = 0;
Jens Axboee4313dd2006-07-19 05:10:01 +0200203 elv_ioc_count_inc(ioc_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204 }
205
206 return ret;
207}
208
209/*
210 * If the current task has no AS IO context then create one and initialise it.
211 * Then take a ref on the task's io context and return it.
212 */
Jens Axboeb5deef92006-07-19 23:39:40 +0200213static struct io_context *as_get_io_context(int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214{
Jens Axboeb5deef92006-07-19 23:39:40 +0200215 struct io_context *ioc = get_io_context(GFP_ATOMIC, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216 if (ioc && !ioc->aic) {
217 ioc->aic = alloc_as_io_context();
218 if (!ioc->aic) {
219 put_io_context(ioc);
220 ioc = NULL;
221 }
222 }
223 return ioc;
224}
225
Jens Axboe8a8e6742006-07-18 21:07:29 +0200226static void as_put_io_context(struct request *rq)
Jens Axboeb4878f22005-10-20 16:42:29 +0200227{
228 struct as_io_context *aic;
229
Jens Axboe8a8e6742006-07-18 21:07:29 +0200230 if (unlikely(!RQ_IOC(rq)))
Jens Axboeb4878f22005-10-20 16:42:29 +0200231 return;
232
Jens Axboe8a8e6742006-07-18 21:07:29 +0200233 aic = RQ_IOC(rq)->aic;
Jens Axboeb4878f22005-10-20 16:42:29 +0200234
Jens Axboe8a8e6742006-07-18 21:07:29 +0200235 if (rq_is_sync(rq) && aic) {
Jens Axboeb4878f22005-10-20 16:42:29 +0200236 spin_lock(&aic->lock);
237 set_bit(AS_TASK_IORUNNING, &aic->state);
238 aic->last_end_request = jiffies;
239 spin_unlock(&aic->lock);
240 }
241
Jens Axboe8a8e6742006-07-18 21:07:29 +0200242 put_io_context(RQ_IOC(rq));
Jens Axboeb4878f22005-10-20 16:42:29 +0200243}
244
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246 * rb tree support functions
247 */
Jens Axboe9e2585a2006-07-28 09:26:13 +0200248#define RQ_RB_ROOT(ad, rq) (&(ad)->sort_list[rq_is_sync((rq))])
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249
Jens Axboe8a8e6742006-07-18 21:07:29 +0200250static void as_add_rq_rb(struct as_data *ad, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251{
Jens Axboee37f3462006-07-18 21:06:01 +0200252 struct request *alias;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253
Jens Axboe9e2585a2006-07-28 09:26:13 +0200254 while ((unlikely(alias = elv_rb_add(RQ_RB_ROOT(ad, rq), rq)))) {
Jens Axboe8a8e6742006-07-18 21:07:29 +0200255 as_move_to_dispatch(ad, alias);
Tejun Heoef9be1d2005-11-11 14:27:09 +0100256 as_antic_stop(ad);
257 }
258}
259
Jens Axboe8a8e6742006-07-18 21:07:29 +0200260static inline void as_del_rq_rb(struct as_data *ad, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261{
Jens Axboe9e2585a2006-07-28 09:26:13 +0200262 elv_rb_del(RQ_RB_ROOT(ad, rq), rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263}
264
265/*
266 * IO Scheduler proper
267 */
268
269#define MAXBACK (1024 * 1024) /*
270 * Maximum distance the disk will go backward
271 * for a request.
272 */
273
274#define BACK_PENALTY 2
275
276/*
277 * as_choose_req selects the preferred one of two requests of the same data_dir
278 * ignoring time - eg. timeouts, which is the job of as_dispatch_request
279 */
Jens Axboe8a8e6742006-07-18 21:07:29 +0200280static struct request *
281as_choose_req(struct as_data *ad, struct request *rq1, struct request *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282{
283 int data_dir;
284 sector_t last, s1, s2, d1, d2;
285 int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */
286 const sector_t maxback = MAXBACK;
287
Jens Axboe8a8e6742006-07-18 21:07:29 +0200288 if (rq1 == NULL || rq1 == rq2)
289 return rq2;
290 if (rq2 == NULL)
291 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
Jens Axboe8a8e6742006-07-18 21:07:29 +0200293 data_dir = rq_is_sync(rq1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294
295 last = ad->last_sector[data_dir];
Jens Axboe8a8e6742006-07-18 21:07:29 +0200296 s1 = rq1->sector;
297 s2 = rq2->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298
Jens Axboe8a8e6742006-07-18 21:07:29 +0200299 BUG_ON(data_dir != rq_is_sync(rq2));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300
301 /*
302 * Strict one way elevator _except_ in the case where we allow
303 * short backward seeks which are biased as twice the cost of a
304 * similar forward seek.
305 */
306 if (s1 >= last)
307 d1 = s1 - last;
308 else if (s1+maxback >= last)
309 d1 = (last - s1)*BACK_PENALTY;
310 else {
311 r1_wrap = 1;
312 d1 = 0; /* shut up, gcc */
313 }
314
315 if (s2 >= last)
316 d2 = s2 - last;
317 else if (s2+maxback >= last)
318 d2 = (last - s2)*BACK_PENALTY;
319 else {
320 r2_wrap = 1;
321 d2 = 0;
322 }
323
324 /* Found required data */
325 if (!r1_wrap && r2_wrap)
Jens Axboe8a8e6742006-07-18 21:07:29 +0200326 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327 else if (!r2_wrap && r1_wrap)
Jens Axboe8a8e6742006-07-18 21:07:29 +0200328 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 else if (r1_wrap && r2_wrap) {
330 /* both behind the head */
331 if (s1 <= s2)
Jens Axboe8a8e6742006-07-18 21:07:29 +0200332 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 else
Jens Axboe8a8e6742006-07-18 21:07:29 +0200334 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 }
336
337 /* Both requests in front of the head */
338 if (d1 < d2)
Jens Axboe8a8e6742006-07-18 21:07:29 +0200339 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 else if (d2 < d1)
Jens Axboe8a8e6742006-07-18 21:07:29 +0200341 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 else {
343 if (s1 >= s2)
Jens Axboe8a8e6742006-07-18 21:07:29 +0200344 return rq1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 else
Jens Axboe8a8e6742006-07-18 21:07:29 +0200346 return rq2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 }
348}
349
350/*
Jens Axboe8a8e6742006-07-18 21:07:29 +0200351 * as_find_next_rq finds the next request after @prev in elevator order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352 * this with as_choose_req form the basis for how the scheduler chooses
353 * what request to process next. Anticipation works on top of this.
354 */
Jens Axboe8a8e6742006-07-18 21:07:29 +0200355static struct request *
356as_find_next_rq(struct as_data *ad, struct request *last)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358 struct rb_node *rbnext = rb_next(&last->rb_node);
359 struct rb_node *rbprev = rb_prev(&last->rb_node);
Jens Axboe8a8e6742006-07-18 21:07:29 +0200360 struct request *next = NULL, *prev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361
Jens Axboee37f3462006-07-18 21:06:01 +0200362 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363
364 if (rbprev)
Jens Axboe8a8e6742006-07-18 21:07:29 +0200365 prev = rb_entry_rq(rbprev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366
367 if (rbnext)
Jens Axboe8a8e6742006-07-18 21:07:29 +0200368 next = rb_entry_rq(rbnext);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 else {
Jens Axboe9e2585a2006-07-28 09:26:13 +0200370 const int data_dir = rq_is_sync(last);
Jens Axboee37f3462006-07-18 21:06:01 +0200371
372 rbnext = rb_first(&ad->sort_list[data_dir]);
373 if (rbnext && rbnext != &last->rb_node)
Jens Axboe8a8e6742006-07-18 21:07:29 +0200374 next = rb_entry_rq(rbnext);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375 }
376
Jens Axboee37f3462006-07-18 21:06:01 +0200377 return as_choose_req(ad, next, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378}
379
380/*
381 * anticipatory scheduling functions follow
382 */
383
384/*
385 * as_antic_expired tells us when we have anticipated too long.
386 * The funny "absolute difference" math on the elapsed time is to handle
387 * jiffy wraps, and disks which have been idle for 0x80000000 jiffies.
388 */
389static int as_antic_expired(struct as_data *ad)
390{
391 long delta_jif;
392
393 delta_jif = jiffies - ad->antic_start;
394 if (unlikely(delta_jif < 0))
395 delta_jif = -delta_jif;
396 if (delta_jif < ad->antic_expire)
397 return 0;
398
399 return 1;
400}
401
402/*
403 * as_antic_waitnext starts anticipating that a nice request will soon be
404 * submitted. See also as_antic_waitreq
405 */
406static void as_antic_waitnext(struct as_data *ad)
407{
408 unsigned long timeout;
409
410 BUG_ON(ad->antic_status != ANTIC_OFF
411 && ad->antic_status != ANTIC_WAIT_REQ);
412
413 timeout = ad->antic_start + ad->antic_expire;
414
415 mod_timer(&ad->antic_timer, timeout);
416
417 ad->antic_status = ANTIC_WAIT_NEXT;
418}
419
420/*
421 * as_antic_waitreq starts anticipating. We don't start timing the anticipation
422 * until the request that we're anticipating on has finished. This means we
423 * are timing from when the candidate process wakes up hopefully.
424 */
425static void as_antic_waitreq(struct as_data *ad)
426{
427 BUG_ON(ad->antic_status == ANTIC_FINISHED);
428 if (ad->antic_status == ANTIC_OFF) {
429 if (!ad->io_context || ad->ioc_finished)
430 as_antic_waitnext(ad);
431 else
432 ad->antic_status = ANTIC_WAIT_REQ;
433 }
434}
435
436/*
437 * This is called directly by the functions in this file to stop anticipation.
438 * We kill the timer and schedule a call to the request_fn asap.
439 */
440static void as_antic_stop(struct as_data *ad)
441{
442 int status = ad->antic_status;
443
444 if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) {
445 if (status == ANTIC_WAIT_NEXT)
446 del_timer(&ad->antic_timer);
447 ad->antic_status = ANTIC_FINISHED;
448 /* see as_work_handler */
449 kblockd_schedule_work(&ad->antic_work);
450 }
451}
452
453/*
454 * as_antic_timeout is the timer function set by as_antic_waitnext.
455 */
456static void as_antic_timeout(unsigned long data)
457{
458 struct request_queue *q = (struct request_queue *)data;
459 struct as_data *ad = q->elevator->elevator_data;
460 unsigned long flags;
461
462 spin_lock_irqsave(q->queue_lock, flags);
463 if (ad->antic_status == ANTIC_WAIT_REQ
464 || ad->antic_status == ANTIC_WAIT_NEXT) {
465 struct as_io_context *aic = ad->io_context->aic;
466
467 ad->antic_status = ANTIC_FINISHED;
468 kblockd_schedule_work(&ad->antic_work);
469
470 if (aic->ttime_samples == 0) {
Nick Pigginf5b3db02005-11-07 00:59:53 -0800471 /* process anticipated on has exited or timed out*/
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 ad->exit_prob = (7*ad->exit_prob + 256)/8;
473 }
Nick Pigginf5b3db02005-11-07 00:59:53 -0800474 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
475 /* process not "saved" by a cooperating request */
476 ad->exit_no_coop = (7*ad->exit_no_coop + 256)/8;
477 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 }
479 spin_unlock_irqrestore(q->queue_lock, flags);
480}
481
Nick Pigginf5b3db02005-11-07 00:59:53 -0800482static void as_update_thinktime(struct as_data *ad, struct as_io_context *aic,
483 unsigned long ttime)
484{
485 /* fixed point: 1.0 == 1<<8 */
486 if (aic->ttime_samples == 0) {
487 ad->new_ttime_total = (7*ad->new_ttime_total + 256*ttime) / 8;
488 ad->new_ttime_mean = ad->new_ttime_total / 256;
489
490 ad->exit_prob = (7*ad->exit_prob)/8;
491 }
492 aic->ttime_samples = (7*aic->ttime_samples + 256) / 8;
493 aic->ttime_total = (7*aic->ttime_total + 256*ttime) / 8;
494 aic->ttime_mean = (aic->ttime_total + 128) / aic->ttime_samples;
495}
496
497static void as_update_seekdist(struct as_data *ad, struct as_io_context *aic,
498 sector_t sdist)
499{
500 u64 total;
501
502 if (aic->seek_samples == 0) {
503 ad->new_seek_total = (7*ad->new_seek_total + 256*(u64)sdist)/8;
504 ad->new_seek_mean = ad->new_seek_total / 256;
505 }
506
507 /*
508 * Don't allow the seek distance to get too large from the
509 * odd fragment, pagein, etc
510 */
511 if (aic->seek_samples <= 60) /* second&third seek */
512 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*1024);
513 else
514 sdist = min(sdist, (aic->seek_mean * 4) + 2*1024*64);
515
516 aic->seek_samples = (7*aic->seek_samples + 256) / 8;
517 aic->seek_total = (7*aic->seek_total + (u64)256*sdist) / 8;
518 total = aic->seek_total + (aic->seek_samples/2);
519 do_div(total, aic->seek_samples);
520 aic->seek_mean = (sector_t)total;
521}
522
523/*
524 * as_update_iohist keeps a decaying histogram of IO thinktimes, and
525 * updates @aic->ttime_mean based on that. It is called when a new
526 * request is queued.
527 */
528static void as_update_iohist(struct as_data *ad, struct as_io_context *aic,
529 struct request *rq)
530{
Jens Axboe9e2585a2006-07-28 09:26:13 +0200531 int data_dir = rq_is_sync(rq);
Nick Pigginf5b3db02005-11-07 00:59:53 -0800532 unsigned long thinktime = 0;
533 sector_t seek_dist;
534
535 if (aic == NULL)
536 return;
537
538 if (data_dir == REQ_SYNC) {
539 unsigned long in_flight = atomic_read(&aic->nr_queued)
540 + atomic_read(&aic->nr_dispatched);
541 spin_lock(&aic->lock);
542 if (test_bit(AS_TASK_IORUNNING, &aic->state) ||
543 test_bit(AS_TASK_IOSTARTED, &aic->state)) {
544 /* Calculate read -> read thinktime */
545 if (test_bit(AS_TASK_IORUNNING, &aic->state)
546 && in_flight == 0) {
547 thinktime = jiffies - aic->last_end_request;
548 thinktime = min(thinktime, MAX_THINKTIME-1);
549 }
550 as_update_thinktime(ad, aic, thinktime);
551
552 /* Calculate read -> read seek distance */
553 if (aic->last_request_pos < rq->sector)
554 seek_dist = rq->sector - aic->last_request_pos;
555 else
556 seek_dist = aic->last_request_pos - rq->sector;
557 as_update_seekdist(ad, aic, seek_dist);
558 }
559 aic->last_request_pos = rq->sector + rq->nr_sectors;
560 set_bit(AS_TASK_IOSTARTED, &aic->state);
561 spin_unlock(&aic->lock);
562 }
563}
564
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565/*
566 * as_close_req decides if one request is considered "close" to the
567 * previous one issued.
568 */
Nick Pigginf5b3db02005-11-07 00:59:53 -0800569static int as_close_req(struct as_data *ad, struct as_io_context *aic,
Jens Axboe8a8e6742006-07-18 21:07:29 +0200570 struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571{
572 unsigned long delay; /* milliseconds */
573 sector_t last = ad->last_sector[ad->batch_data_dir];
Jens Axboe8a8e6742006-07-18 21:07:29 +0200574 sector_t next = rq->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 sector_t delta; /* acceptable close offset (in sectors) */
Nick Pigginf5b3db02005-11-07 00:59:53 -0800576 sector_t s;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577
578 if (ad->antic_status == ANTIC_OFF || !ad->ioc_finished)
579 delay = 0;
580 else
581 delay = ((jiffies - ad->antic_start) * 1000) / HZ;
582
Nick Pigginf5b3db02005-11-07 00:59:53 -0800583 if (delay == 0)
584 delta = 8192;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 else if (delay <= 20 && delay <= ad->antic_expire)
Nick Pigginf5b3db02005-11-07 00:59:53 -0800586 delta = 8192 << delay;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 else
588 return 1;
589
Nick Pigginf5b3db02005-11-07 00:59:53 -0800590 if ((last <= next + (delta>>1)) && (next <= last + delta))
591 return 1;
592
593 if (last < next)
594 s = next - last;
595 else
596 s = last - next;
597
598 if (aic->seek_samples == 0) {
599 /*
600 * Process has just started IO. Use past statistics to
601 * gauge success possibility
602 */
603 if (ad->new_seek_mean > s) {
604 /* this request is better than what we're expecting */
605 return 1;
606 }
607
608 } else {
609 if (aic->seek_mean > s) {
610 /* this request is better than what we're expecting */
611 return 1;
612 }
613 }
614
615 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616}
617
618/*
619 * as_can_break_anticipation returns true if we have been anticipating this
620 * request.
621 *
622 * It also returns true if the process against which we are anticipating
623 * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to
624 * dispatch it ASAP, because we know that application will not be submitting
625 * any new reads.
626 *
Nick Pigginf5b3db02005-11-07 00:59:53 -0800627 * If the task which has submitted the request has exited, break anticipation.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628 *
629 * If this task has queued some other IO, do not enter enticipation.
630 */
Jens Axboe8a8e6742006-07-18 21:07:29 +0200631static int as_can_break_anticipation(struct as_data *ad, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632{
633 struct io_context *ioc;
634 struct as_io_context *aic;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635
636 ioc = ad->io_context;
637 BUG_ON(!ioc);
638
Jens Axboe8a8e6742006-07-18 21:07:29 +0200639 if (rq && ioc == RQ_IOC(rq)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640 /* request from same process */
641 return 1;
642 }
643
644 if (ad->ioc_finished && as_antic_expired(ad)) {
645 /*
646 * In this situation status should really be FINISHED,
647 * however the timer hasn't had the chance to run yet.
648 */
649 return 1;
650 }
651
652 aic = ioc->aic;
653 if (!aic)
654 return 0;
655
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 if (atomic_read(&aic->nr_queued) > 0) {
657 /* process has more requests queued */
658 return 1;
659 }
660
661 if (atomic_read(&aic->nr_dispatched) > 0) {
662 /* process has more requests dispatched */
663 return 1;
664 }
665
Jens Axboe8a8e6742006-07-18 21:07:29 +0200666 if (rq && rq_is_sync(rq) && as_close_req(ad, aic, rq)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 /*
668 * Found a close request that is not one of ours.
669 *
Nick Pigginf5b3db02005-11-07 00:59:53 -0800670 * This makes close requests from another process update
671 * our IO history. Is generally useful when there are
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 * two or more cooperating processes working in the same
673 * area.
674 */
Nick Pigginf5b3db02005-11-07 00:59:53 -0800675 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
676 if (aic->ttime_samples == 0)
677 ad->exit_prob = (7*ad->exit_prob + 256)/8;
678
679 ad->exit_no_coop = (7*ad->exit_no_coop)/8;
680 }
681
Jens Axboe8a8e6742006-07-18 21:07:29 +0200682 as_update_iohist(ad, aic, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 return 1;
684 }
685
Nick Pigginf5b3db02005-11-07 00:59:53 -0800686 if (!test_bit(AS_TASK_RUNNING, &aic->state)) {
687 /* process anticipated on has exited */
688 if (aic->ttime_samples == 0)
689 ad->exit_prob = (7*ad->exit_prob + 256)/8;
690
691 if (ad->exit_no_coop > 128)
692 return 1;
693 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694
695 if (aic->ttime_samples == 0) {
696 if (ad->new_ttime_mean > ad->antic_expire)
697 return 1;
Nick Pigginf5b3db02005-11-07 00:59:53 -0800698 if (ad->exit_prob * ad->exit_no_coop > 128*256)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 return 1;
700 } else if (aic->ttime_mean > ad->antic_expire) {
701 /* the process thinks too much between requests */
702 return 1;
703 }
704
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 return 0;
706}
707
708/*
Jens Axboe8a8e6742006-07-18 21:07:29 +0200709 * as_can_anticipate indicates whether we should either run rq
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 * or keep anticipating a better request.
711 */
Jens Axboe8a8e6742006-07-18 21:07:29 +0200712static int as_can_anticipate(struct as_data *ad, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713{
714 if (!ad->io_context)
715 /*
716 * Last request submitted was a write
717 */
718 return 0;
719
720 if (ad->antic_status == ANTIC_FINISHED)
721 /*
722 * Don't restart if we have just finished. Run the next request
723 */
724 return 0;
725
Jens Axboe8a8e6742006-07-18 21:07:29 +0200726 if (as_can_break_anticipation(ad, rq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 /*
728 * This request is a good candidate. Don't keep anticipating,
729 * run it.
730 */
731 return 0;
732
733 /*
734 * OK from here, we haven't finished, and don't have a decent request!
735 * Status is either ANTIC_OFF so start waiting,
736 * ANTIC_WAIT_REQ so continue waiting for request to finish
737 * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738 */
739
740 return 1;
741}
742
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743/*
Jens Axboe8a8e6742006-07-18 21:07:29 +0200744 * as_update_rq must be called whenever a request (rq) is added to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 * the sort_list. This function keeps caches up to date, and checks if the
746 * request might be one we are "anticipating"
747 */
Jens Axboe8a8e6742006-07-18 21:07:29 +0200748static void as_update_rq(struct as_data *ad, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749{
Jens Axboe8a8e6742006-07-18 21:07:29 +0200750 const int data_dir = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751
Jens Axboe8a8e6742006-07-18 21:07:29 +0200752 /* keep the next_rq cache up to date */
753 ad->next_rq[data_dir] = as_choose_req(ad, rq, ad->next_rq[data_dir]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754
755 /*
756 * have we been anticipating this request?
757 * or does it come from the same process as the one we are anticipating
758 * for?
759 */
760 if (ad->antic_status == ANTIC_WAIT_REQ
761 || ad->antic_status == ANTIC_WAIT_NEXT) {
Jens Axboe8a8e6742006-07-18 21:07:29 +0200762 if (as_can_break_anticipation(ad, rq))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 as_antic_stop(ad);
764 }
765}
766
767/*
768 * Gathers timings and resizes the write batch automatically
769 */
770static void update_write_batch(struct as_data *ad)
771{
772 unsigned long batch = ad->batch_expire[REQ_ASYNC];
773 long write_time;
774
775 write_time = (jiffies - ad->current_batch_expires) + batch;
776 if (write_time < 0)
777 write_time = 0;
778
779 if (write_time > batch && !ad->write_batch_idled) {
780 if (write_time > batch * 3)
781 ad->write_batch_count /= 2;
782 else
783 ad->write_batch_count--;
784 } else if (write_time < batch && ad->current_write_count == 0) {
785 if (batch > write_time * 3)
786 ad->write_batch_count *= 2;
787 else
788 ad->write_batch_count++;
789 }
790
791 if (ad->write_batch_count < 1)
792 ad->write_batch_count = 1;
793}
794
795/*
796 * as_completed_request is to be called when a request has completed and
797 * returned something to the requesting process, be it an error or data.
798 */
799static void as_completed_request(request_queue_t *q, struct request *rq)
800{
801 struct as_data *ad = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802
803 WARN_ON(!list_empty(&rq->queuelist));
804
Jens Axboe8a8e6742006-07-18 21:07:29 +0200805 if (RQ_STATE(rq) != AS_RQ_REMOVED) {
806 printk("rq->state %d\n", RQ_STATE(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 WARN_ON(1);
808 goto out;
809 }
810
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 if (ad->changed_batch && ad->nr_dispatched == 1) {
812 kblockd_schedule_work(&ad->antic_work);
813 ad->changed_batch = 0;
814
815 if (ad->batch_data_dir == REQ_SYNC)
816 ad->new_batch = 1;
817 }
818 WARN_ON(ad->nr_dispatched == 0);
819 ad->nr_dispatched--;
820
821 /*
822 * Start counting the batch from when a request of that direction is
823 * actually serviced. This should help devices with big TCQ windows
824 * and writeback caches
825 */
Jens Axboe9e2585a2006-07-28 09:26:13 +0200826 if (ad->new_batch && ad->batch_data_dir == rq_is_sync(rq)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 update_write_batch(ad);
828 ad->current_batch_expires = jiffies +
829 ad->batch_expire[REQ_SYNC];
830 ad->new_batch = 0;
831 }
832
Jens Axboe8a8e6742006-07-18 21:07:29 +0200833 if (ad->io_context == RQ_IOC(rq) && ad->io_context) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 ad->antic_start = jiffies;
835 ad->ioc_finished = 1;
836 if (ad->antic_status == ANTIC_WAIT_REQ) {
837 /*
838 * We were waiting on this request, now anticipate
839 * the next one
840 */
841 as_antic_waitnext(ad);
842 }
843 }
844
Jens Axboe8a8e6742006-07-18 21:07:29 +0200845 as_put_io_context(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846out:
Jens Axboe8a8e6742006-07-18 21:07:29 +0200847 RQ_SET_STATE(rq, AS_RQ_POSTSCHED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848}
849
850/*
851 * as_remove_queued_request removes a request from the pre dispatch queue
852 * without updating refcounts. It is expected the caller will drop the
853 * reference unless it replaces the request at somepart of the elevator
854 * (ie. the dispatch queue)
855 */
856static void as_remove_queued_request(request_queue_t *q, struct request *rq)
857{
Jens Axboe9e2585a2006-07-28 09:26:13 +0200858 const int data_dir = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 struct as_data *ad = q->elevator->elevator_data;
Jens Axboe8a8e6742006-07-18 21:07:29 +0200860 struct io_context *ioc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
Jens Axboe8a8e6742006-07-18 21:07:29 +0200862 WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863
Jens Axboe8a8e6742006-07-18 21:07:29 +0200864 ioc = RQ_IOC(rq);
865 if (ioc && ioc->aic) {
866 BUG_ON(!atomic_read(&ioc->aic->nr_queued));
867 atomic_dec(&ioc->aic->nr_queued);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868 }
869
870 /*
Jens Axboe8a8e6742006-07-18 21:07:29 +0200871 * Update the "next_rq" cache if we are about to remove its
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 * entry
873 */
Jens Axboe8a8e6742006-07-18 21:07:29 +0200874 if (ad->next_rq[data_dir] == rq)
875 ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876
Jens Axboed4f2f462006-07-13 09:12:14 +0200877 rq_fifo_clear(rq);
Jens Axboe8a8e6742006-07-18 21:07:29 +0200878 as_del_rq_rb(ad, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879}
880
881/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 * as_fifo_expired returns 0 if there are no expired reads on the fifo,
883 * 1 otherwise. It is ratelimited so that we only perform the check once per
884 * `fifo_expire' interval. Otherwise a large number of expired requests
885 * would create a hopeless seekstorm.
886 *
887 * See as_antic_expired comment.
888 */
889static int as_fifo_expired(struct as_data *ad, int adir)
890{
Jens Axboed4f2f462006-07-13 09:12:14 +0200891 struct request *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 long delta_jif;
893
894 delta_jif = jiffies - ad->last_check_fifo[adir];
895 if (unlikely(delta_jif < 0))
896 delta_jif = -delta_jif;
897 if (delta_jif < ad->fifo_expire[adir])
898 return 0;
899
900 ad->last_check_fifo[adir] = jiffies;
901
902 if (list_empty(&ad->fifo_list[adir]))
903 return 0;
904
Jens Axboed4f2f462006-07-13 09:12:14 +0200905 rq = rq_entry_fifo(ad->fifo_list[adir].next);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906
Jens Axboed4f2f462006-07-13 09:12:14 +0200907 return time_after(jiffies, rq_fifo_time(rq));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908}
909
910/*
911 * as_batch_expired returns true if the current batch has expired. A batch
912 * is a set of reads or a set of writes.
913 */
914static inline int as_batch_expired(struct as_data *ad)
915{
916 if (ad->changed_batch || ad->new_batch)
917 return 0;
918
919 if (ad->batch_data_dir == REQ_SYNC)
920 /* TODO! add a check so a complete fifo gets written? */
921 return time_after(jiffies, ad->current_batch_expires);
922
923 return time_after(jiffies, ad->current_batch_expires)
924 || ad->current_write_count == 0;
925}
926
927/*
928 * move an entry to dispatch queue
929 */
Jens Axboe8a8e6742006-07-18 21:07:29 +0200930static void as_move_to_dispatch(struct as_data *ad, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931{
Jens Axboe9e2585a2006-07-28 09:26:13 +0200932 const int data_dir = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933
Jens Axboee37f3462006-07-18 21:06:01 +0200934 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935
936 as_antic_stop(ad);
937 ad->antic_status = ANTIC_OFF;
938
939 /*
940 * This has to be set in order to be correctly updated by
Jens Axboe8a8e6742006-07-18 21:07:29 +0200941 * as_find_next_rq
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942 */
943 ad->last_sector[data_dir] = rq->sector + rq->nr_sectors;
944
945 if (data_dir == REQ_SYNC) {
Jens Axboe8a8e6742006-07-18 21:07:29 +0200946 struct io_context *ioc = RQ_IOC(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 /* In case we have to anticipate after this */
Jens Axboe8a8e6742006-07-18 21:07:29 +0200948 copy_io_context(&ad->io_context, &ioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 } else {
950 if (ad->io_context) {
951 put_io_context(ad->io_context);
952 ad->io_context = NULL;
953 }
954
955 if (ad->current_write_count != 0)
956 ad->current_write_count--;
957 }
958 ad->ioc_finished = 0;
959
Jens Axboe8a8e6742006-07-18 21:07:29 +0200960 ad->next_rq[data_dir] = as_find_next_rq(ad, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961
962 /*
963 * take it off the sort and fifo list, add to dispatch queue
964 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 as_remove_queued_request(ad->q, rq);
Jens Axboe8a8e6742006-07-18 21:07:29 +0200966 WARN_ON(RQ_STATE(rq) != AS_RQ_QUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967
Jens Axboeb4878f22005-10-20 16:42:29 +0200968 elv_dispatch_sort(ad->q, rq);
969
Jens Axboe8a8e6742006-07-18 21:07:29 +0200970 RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
971 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
972 atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 ad->nr_dispatched++;
974}
975
976/*
977 * as_dispatch_request selects the best request according to
978 * read/write expire, batch expire, etc, and moves it to the dispatch
979 * queue. Returns 1 if a request was found, 0 otherwise.
980 */
Jens Axboeb4878f22005-10-20 16:42:29 +0200981static int as_dispatch_request(request_queue_t *q, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982{
Jens Axboeb4878f22005-10-20 16:42:29 +0200983 struct as_data *ad = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]);
985 const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]);
Jens Axboe8a8e6742006-07-18 21:07:29 +0200986 struct request *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987
Jens Axboeb4878f22005-10-20 16:42:29 +0200988 if (unlikely(force)) {
989 /*
990 * Forced dispatch, accounting is useless. Reset
991 * accounting states and dump fifo_lists. Note that
992 * batch_data_dir is reset to REQ_SYNC to avoid
993 * screwing write batch accounting as write batch
994 * accounting occurs on W->R transition.
995 */
996 int dispatched = 0;
997
998 ad->batch_data_dir = REQ_SYNC;
999 ad->changed_batch = 0;
1000 ad->new_batch = 0;
1001
Jens Axboe8a8e6742006-07-18 21:07:29 +02001002 while (ad->next_rq[REQ_SYNC]) {
1003 as_move_to_dispatch(ad, ad->next_rq[REQ_SYNC]);
Jens Axboeb4878f22005-10-20 16:42:29 +02001004 dispatched++;
1005 }
1006 ad->last_check_fifo[REQ_SYNC] = jiffies;
1007
Jens Axboe8a8e6742006-07-18 21:07:29 +02001008 while (ad->next_rq[REQ_ASYNC]) {
1009 as_move_to_dispatch(ad, ad->next_rq[REQ_ASYNC]);
Jens Axboeb4878f22005-10-20 16:42:29 +02001010 dispatched++;
1011 }
1012 ad->last_check_fifo[REQ_ASYNC] = jiffies;
1013
1014 return dispatched;
1015 }
1016
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 /* Signal that the write batch was uncontended, so we can't time it */
1018 if (ad->batch_data_dir == REQ_ASYNC && !reads) {
1019 if (ad->current_write_count == 0 || !writes)
1020 ad->write_batch_idled = 1;
1021 }
1022
1023 if (!(reads || writes)
1024 || ad->antic_status == ANTIC_WAIT_REQ
1025 || ad->antic_status == ANTIC_WAIT_NEXT
1026 || ad->changed_batch)
1027 return 0;
1028
Nick Pigginf5b3db02005-11-07 00:59:53 -08001029 if (!(reads && writes && as_batch_expired(ad))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 /*
1031 * batch is still running or no reads or no writes
1032 */
Jens Axboe8a8e6742006-07-18 21:07:29 +02001033 rq = ad->next_rq[ad->batch_data_dir];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034
1035 if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) {
1036 if (as_fifo_expired(ad, REQ_SYNC))
1037 goto fifo_expired;
1038
Jens Axboe8a8e6742006-07-18 21:07:29 +02001039 if (as_can_anticipate(ad, rq)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 as_antic_waitreq(ad);
1041 return 0;
1042 }
1043 }
1044
Jens Axboe8a8e6742006-07-18 21:07:29 +02001045 if (rq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046 /* we have a "next request" */
1047 if (reads && !writes)
1048 ad->current_batch_expires =
1049 jiffies + ad->batch_expire[REQ_SYNC];
1050 goto dispatch_request;
1051 }
1052 }
1053
1054 /*
1055 * at this point we are not running a batch. select the appropriate
1056 * data direction (read / write)
1057 */
1058
1059 if (reads) {
Jens Axboedd67d052006-06-21 09:36:18 +02001060 BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_SYNC]));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061
1062 if (writes && ad->batch_data_dir == REQ_SYNC)
1063 /*
1064 * Last batch was a read, switch to writes
1065 */
1066 goto dispatch_writes;
1067
1068 if (ad->batch_data_dir == REQ_ASYNC) {
1069 WARN_ON(ad->new_batch);
1070 ad->changed_batch = 1;
1071 }
1072 ad->batch_data_dir = REQ_SYNC;
Jens Axboe8a8e6742006-07-18 21:07:29 +02001073 rq = rq_entry_fifo(ad->fifo_list[REQ_SYNC].next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 ad->last_check_fifo[ad->batch_data_dir] = jiffies;
1075 goto dispatch_request;
1076 }
1077
1078 /*
1079 * the last batch was a read
1080 */
1081
1082 if (writes) {
1083dispatch_writes:
Jens Axboedd67d052006-06-21 09:36:18 +02001084 BUG_ON(RB_EMPTY_ROOT(&ad->sort_list[REQ_ASYNC]));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085
1086 if (ad->batch_data_dir == REQ_SYNC) {
1087 ad->changed_batch = 1;
1088
1089 /*
1090 * new_batch might be 1 when the queue runs out of
1091 * reads. A subsequent submission of a write might
1092 * cause a change of batch before the read is finished.
1093 */
1094 ad->new_batch = 0;
1095 }
1096 ad->batch_data_dir = REQ_ASYNC;
1097 ad->current_write_count = ad->write_batch_count;
1098 ad->write_batch_idled = 0;
Jens Axboe8a8e6742006-07-18 21:07:29 +02001099 rq = ad->next_rq[ad->batch_data_dir];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 goto dispatch_request;
1101 }
1102
1103 BUG();
1104 return 0;
1105
1106dispatch_request:
1107 /*
1108 * If a request has expired, service it.
1109 */
1110
1111 if (as_fifo_expired(ad, ad->batch_data_dir)) {
1112fifo_expired:
Jens Axboe8a8e6742006-07-18 21:07:29 +02001113 rq = rq_entry_fifo(ad->fifo_list[ad->batch_data_dir].next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114 }
1115
1116 if (ad->changed_batch) {
1117 WARN_ON(ad->new_batch);
1118
1119 if (ad->nr_dispatched)
1120 return 0;
1121
1122 if (ad->batch_data_dir == REQ_ASYNC)
1123 ad->current_batch_expires = jiffies +
1124 ad->batch_expire[REQ_ASYNC];
1125 else
1126 ad->new_batch = 1;
1127
1128 ad->changed_batch = 0;
1129 }
1130
1131 /*
Jens Axboe8a8e6742006-07-18 21:07:29 +02001132 * rq is the selected appropriate request.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 */
Jens Axboe8a8e6742006-07-18 21:07:29 +02001134 as_move_to_dispatch(ad, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135
1136 return 1;
1137}
1138
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139/*
Jens Axboe8a8e6742006-07-18 21:07:29 +02001140 * add rq to rbtree and fifo
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 */
Jens Axboeb4878f22005-10-20 16:42:29 +02001142static void as_add_request(request_queue_t *q, struct request *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143{
Jens Axboeb4878f22005-10-20 16:42:29 +02001144 struct as_data *ad = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 int data_dir;
1146
Jens Axboe8a8e6742006-07-18 21:07:29 +02001147 RQ_SET_STATE(rq, AS_RQ_NEW);
Jens Axboeb4878f22005-10-20 16:42:29 +02001148
Jens Axboe9e2585a2006-07-28 09:26:13 +02001149 data_dir = rq_is_sync(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150
Jens Axboeb5deef92006-07-19 23:39:40 +02001151 rq->elevator_private = as_get_io_context(q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152
Jens Axboe8a8e6742006-07-18 21:07:29 +02001153 if (RQ_IOC(rq)) {
1154 as_update_iohist(ad, RQ_IOC(rq)->aic, rq);
1155 atomic_inc(&RQ_IOC(rq)->aic->nr_queued);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 }
1157
Jens Axboe8a8e6742006-07-18 21:07:29 +02001158 as_add_rq_rb(ad, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159
Tejun Heoef9be1d2005-11-11 14:27:09 +01001160 /*
1161 * set expire time (only used for reads) and add to fifo list
1162 */
Jens Axboed4f2f462006-07-13 09:12:14 +02001163 rq_set_fifo_time(rq, jiffies + ad->fifo_expire[data_dir]);
1164 list_add_tail(&rq->queuelist, &ad->fifo_list[data_dir]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165
Jens Axboe8a8e6742006-07-18 21:07:29 +02001166 as_update_rq(ad, rq); /* keep state machine up to date */
1167 RQ_SET_STATE(rq, AS_RQ_QUEUED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168}
1169
Jens Axboeb4878f22005-10-20 16:42:29 +02001170static void as_activate_request(request_queue_t *q, struct request *rq)
1171{
Jens Axboe8a8e6742006-07-18 21:07:29 +02001172 WARN_ON(RQ_STATE(rq) != AS_RQ_DISPATCHED);
1173 RQ_SET_STATE(rq, AS_RQ_REMOVED);
1174 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1175 atomic_dec(&RQ_IOC(rq)->aic->nr_dispatched);
Jens Axboeb4878f22005-10-20 16:42:29 +02001176}
1177
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178static void as_deactivate_request(request_queue_t *q, struct request *rq)
1179{
Jens Axboe8a8e6742006-07-18 21:07:29 +02001180 WARN_ON(RQ_STATE(rq) != AS_RQ_REMOVED);
1181 RQ_SET_STATE(rq, AS_RQ_DISPATCHED);
1182 if (RQ_IOC(rq) && RQ_IOC(rq)->aic)
1183 atomic_inc(&RQ_IOC(rq)->aic->nr_dispatched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184}
1185
1186/*
1187 * as_queue_empty tells us if there are requests left in the device. It may
1188 * not be the case that a driver can get the next request even if the queue
1189 * is not empty - it is used in the block layer to check for plugging and
1190 * merging opportunities
1191 */
1192static int as_queue_empty(request_queue_t *q)
1193{
1194 struct as_data *ad = q->elevator->elevator_data;
1195
Jens Axboeb4878f22005-10-20 16:42:29 +02001196 return list_empty(&ad->fifo_list[REQ_ASYNC])
1197 && list_empty(&ad->fifo_list[REQ_SYNC]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198}
1199
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200static int
1201as_merge(request_queue_t *q, struct request **req, struct bio *bio)
1202{
1203 struct as_data *ad = q->elevator->elevator_data;
1204 sector_t rb_key = bio->bi_sector + bio_sectors(bio);
1205 struct request *__rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206
1207 /*
1208 * check for front merge
1209 */
Jens Axboee37f3462006-07-18 21:06:01 +02001210 __rq = elv_rb_find(&ad->sort_list[bio_data_dir(bio)], rb_key);
Jens Axboe98170642006-07-28 09:23:08 +02001211 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1212 *req = __rq;
1213 return ELEVATOR_FRONT_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214 }
1215
1216 return ELEVATOR_NO_MERGE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217}
1218
Jens Axboee37f3462006-07-18 21:06:01 +02001219static void as_merged_request(request_queue_t *q, struct request *req, int type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220{
1221 struct as_data *ad = q->elevator->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222
1223 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 * if the merge was a front merge, we need to reposition request
1225 */
Jens Axboee37f3462006-07-18 21:06:01 +02001226 if (type == ELEVATOR_FRONT_MERGE) {
Jens Axboe8a8e6742006-07-18 21:07:29 +02001227 as_del_rq_rb(ad, req);
1228 as_add_rq_rb(ad, req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 /*
1230 * Note! At this stage of this and the next function, our next
1231 * request may not be optimal - eg the request may have "grown"
1232 * behind the disk head. We currently don't bother adjusting.
1233 */
1234 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235}
1236
Nick Pigginf5b3db02005-11-07 00:59:53 -08001237static void as_merged_requests(request_queue_t *q, struct request *req,
1238 struct request *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 /*
Jens Axboe8a8e6742006-07-18 21:07:29 +02001241 * if next expires before rq, assign its expire time to arq
1242 * and move into next position (next will be deleted) in fifo
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 */
Jens Axboed4f2f462006-07-13 09:12:14 +02001244 if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
1245 if (time_before(rq_fifo_time(next), rq_fifo_time(req))) {
Jens Axboe8a8e6742006-07-18 21:07:29 +02001246 struct io_context *rioc = RQ_IOC(req);
1247 struct io_context *nioc = RQ_IOC(next);
1248
Jens Axboed4f2f462006-07-13 09:12:14 +02001249 list_move(&req->queuelist, &next->queuelist);
1250 rq_set_fifo_time(req, rq_fifo_time(next));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 /*
1252 * Don't copy here but swap, because when anext is
1253 * removed below, it must contain the unused context
1254 */
Jens Axboe8a8e6742006-07-18 21:07:29 +02001255 swap_io_context(&rioc, &nioc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 }
1257 }
1258
1259 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260 * kill knowledge of next, this one is a goner
1261 */
1262 as_remove_queued_request(q, next);
Jens Axboe8a8e6742006-07-18 21:07:29 +02001263 as_put_io_context(next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264
Jens Axboe8a8e6742006-07-18 21:07:29 +02001265 RQ_SET_STATE(next, AS_RQ_MERGED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266}
1267
1268/*
1269 * This is executed in a "deferred" process context, by kblockd. It calls the
1270 * driver's request_fn so the driver can submit that request.
1271 *
1272 * IMPORTANT! This guy will reenter the elevator, so set up all queue global
1273 * state before calling, and don't rely on any state over calls.
1274 *
1275 * FIXME! dispatch queue is not a queue at all!
1276 */
David Howells65f27f32006-11-22 14:55:48 +00001277static void as_work_handler(struct work_struct *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278{
David Howells65f27f32006-11-22 14:55:48 +00001279 struct as_data *ad = container_of(work, struct as_data, antic_work);
1280 struct request_queue *q = ad->q;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281 unsigned long flags;
1282
1283 spin_lock_irqsave(q->queue_lock, flags);
Jens Axboedc72ef42006-07-20 14:54:05 +02001284 blk_start_queueing(q);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 spin_unlock_irqrestore(q->queue_lock, flags);
1286}
1287
Jens Axboecb78b282006-07-28 09:32:57 +02001288static int as_may_queue(request_queue_t *q, int rw)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289{
1290 int ret = ELV_MQUEUE_MAY;
1291 struct as_data *ad = q->elevator->elevator_data;
1292 struct io_context *ioc;
1293 if (ad->antic_status == ANTIC_WAIT_REQ ||
1294 ad->antic_status == ANTIC_WAIT_NEXT) {
Jens Axboeb5deef92006-07-19 23:39:40 +02001295 ioc = as_get_io_context(q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 if (ad->io_context == ioc)
1297 ret = ELV_MQUEUE_MUST;
1298 put_io_context(ioc);
1299 }
1300
1301 return ret;
1302}
1303
1304static void as_exit_queue(elevator_t *e)
1305{
1306 struct as_data *ad = e->elevator_data;
1307
1308 del_timer_sync(&ad->antic_timer);
1309 kblockd_flush();
1310
1311 BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC]));
1312 BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC]));
1313
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 put_io_context(ad->io_context);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 kfree(ad);
1316}
1317
1318/*
Jens Axboe8a8e6742006-07-18 21:07:29 +02001319 * initialize elevator private data (as_data).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 */
Jens Axboebb37b942006-12-01 10:42:33 +01001321static void *as_init_queue(request_queue_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322{
1323 struct as_data *ad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324
Christoph Lameter19460892005-06-23 00:08:19 -07001325 ad = kmalloc_node(sizeof(*ad), GFP_KERNEL, q->node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326 if (!ad)
Jens Axboebc1c1162006-06-08 08:49:06 +02001327 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 memset(ad, 0, sizeof(*ad));
1329
1330 ad->q = q; /* Identify what queue the data belongs to */
1331
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 /* anticipatory scheduling helpers */
1333 ad->antic_timer.function = as_antic_timeout;
1334 ad->antic_timer.data = (unsigned long)q;
1335 init_timer(&ad->antic_timer);
David Howells65f27f32006-11-22 14:55:48 +00001336 INIT_WORK(&ad->antic_work, as_work_handler);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]);
1339 INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]);
1340 ad->sort_list[REQ_SYNC] = RB_ROOT;
1341 ad->sort_list[REQ_ASYNC] = RB_ROOT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 ad->fifo_expire[REQ_SYNC] = default_read_expire;
1343 ad->fifo_expire[REQ_ASYNC] = default_write_expire;
1344 ad->antic_expire = default_antic_expire;
1345 ad->batch_expire[REQ_SYNC] = default_read_batch_expire;
1346 ad->batch_expire[REQ_ASYNC] = default_write_batch_expire;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347
1348 ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC];
1349 ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10;
1350 if (ad->write_batch_count < 2)
1351 ad->write_batch_count = 2;
1352
Jens Axboebc1c1162006-06-08 08:49:06 +02001353 return ad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354}
1355
1356/*
1357 * sysfs parts below
1358 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359
1360static ssize_t
1361as_var_show(unsigned int var, char *page)
1362{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363 return sprintf(page, "%d\n", var);
1364}
1365
1366static ssize_t
1367as_var_store(unsigned long *var, const char *page, size_t count)
1368{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 char *p = (char *) page;
1370
Jens Axboec9b3ad62005-07-27 11:43:37 -07001371 *var = simple_strtoul(p, &p, 10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 return count;
1373}
1374
Al Viroe572ec72006-03-18 22:27:18 -05001375static ssize_t est_time_show(elevator_t *e, char *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376{
Al Viro3d1ab402006-03-18 18:35:43 -05001377 struct as_data *ad = e->elevator_data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 int pos = 0;
1379
Nick Pigginf5b3db02005-11-07 00:59:53 -08001380 pos += sprintf(page+pos, "%lu %% exit probability\n",
1381 100*ad->exit_prob/256);
1382 pos += sprintf(page+pos, "%lu %% probability of exiting without a "
1383 "cooperating process submitting IO\n",
1384 100*ad->exit_no_coop/256);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 pos += sprintf(page+pos, "%lu ms new thinktime\n", ad->new_ttime_mean);
Nick Pigginf5b3db02005-11-07 00:59:53 -08001386 pos += sprintf(page+pos, "%llu sectors new seek distance\n",
1387 (unsigned long long)ad->new_seek_mean);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388
1389 return pos;
1390}
1391
1392#define SHOW_FUNCTION(__FUNC, __VAR) \
Al Viro3d1ab402006-03-18 18:35:43 -05001393static ssize_t __FUNC(elevator_t *e, char *page) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394{ \
Al Viro3d1ab402006-03-18 18:35:43 -05001395 struct as_data *ad = e->elevator_data; \
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 return as_var_show(jiffies_to_msecs((__VAR)), (page)); \
1397}
Al Viroe572ec72006-03-18 22:27:18 -05001398SHOW_FUNCTION(as_read_expire_show, ad->fifo_expire[REQ_SYNC]);
1399SHOW_FUNCTION(as_write_expire_show, ad->fifo_expire[REQ_ASYNC]);
1400SHOW_FUNCTION(as_antic_expire_show, ad->antic_expire);
1401SHOW_FUNCTION(as_read_batch_expire_show, ad->batch_expire[REQ_SYNC]);
1402SHOW_FUNCTION(as_write_batch_expire_show, ad->batch_expire[REQ_ASYNC]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403#undef SHOW_FUNCTION
1404
1405#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
Al Viro3d1ab402006-03-18 18:35:43 -05001406static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407{ \
Al Viro3d1ab402006-03-18 18:35:43 -05001408 struct as_data *ad = e->elevator_data; \
1409 int ret = as_var_store(__PTR, (page), count); \
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 if (*(__PTR) < (MIN)) \
1411 *(__PTR) = (MIN); \
1412 else if (*(__PTR) > (MAX)) \
1413 *(__PTR) = (MAX); \
1414 *(__PTR) = msecs_to_jiffies(*(__PTR)); \
1415 return ret; \
1416}
Al Viroe572ec72006-03-18 22:27:18 -05001417STORE_FUNCTION(as_read_expire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX);
1418STORE_FUNCTION(as_write_expire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX);
1419STORE_FUNCTION(as_antic_expire_store, &ad->antic_expire, 0, INT_MAX);
1420STORE_FUNCTION(as_read_batch_expire_store,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 &ad->batch_expire[REQ_SYNC], 0, INT_MAX);
Al Viroe572ec72006-03-18 22:27:18 -05001422STORE_FUNCTION(as_write_batch_expire_store,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423 &ad->batch_expire[REQ_ASYNC], 0, INT_MAX);
1424#undef STORE_FUNCTION
1425
Al Viroe572ec72006-03-18 22:27:18 -05001426#define AS_ATTR(name) \
1427 __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428
Al Viroe572ec72006-03-18 22:27:18 -05001429static struct elv_fs_entry as_attrs[] = {
1430 __ATTR_RO(est_time),
1431 AS_ATTR(read_expire),
1432 AS_ATTR(write_expire),
1433 AS_ATTR(antic_expire),
1434 AS_ATTR(read_batch_expire),
1435 AS_ATTR(write_batch_expire),
1436 __ATTR_NULL
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437};
1438
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439static struct elevator_type iosched_as = {
1440 .ops = {
1441 .elevator_merge_fn = as_merge,
1442 .elevator_merged_fn = as_merged_request,
1443 .elevator_merge_req_fn = as_merged_requests,
Jens Axboeb4878f22005-10-20 16:42:29 +02001444 .elevator_dispatch_fn = as_dispatch_request,
1445 .elevator_add_req_fn = as_add_request,
1446 .elevator_activate_req_fn = as_activate_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 .elevator_deactivate_req_fn = as_deactivate_request,
1448 .elevator_queue_empty_fn = as_queue_empty,
1449 .elevator_completed_req_fn = as_completed_request,
Jens Axboee37f3462006-07-18 21:06:01 +02001450 .elevator_former_req_fn = elv_rb_former_request,
1451 .elevator_latter_req_fn = elv_rb_latter_request,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452 .elevator_may_queue_fn = as_may_queue,
1453 .elevator_init_fn = as_init_queue,
1454 .elevator_exit_fn = as_exit_queue,
Al Viroe17a9482006-03-18 13:21:20 -05001455 .trim = as_trim,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 },
1457
Al Viro3d1ab402006-03-18 18:35:43 -05001458 .elevator_attrs = as_attrs,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459 .elevator_name = "anticipatory",
1460 .elevator_owner = THIS_MODULE,
1461};
1462
1463static int __init as_init(void)
1464{
1465 int ret;
1466
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 ret = elv_register(&iosched_as);
1468 if (!ret) {
1469 /*
1470 * don't allow AS to get unregistered, since we would have
1471 * to browse all tasks in the system and release their
1472 * as_io_context first
1473 */
1474 __module_get(THIS_MODULE);
1475 return 0;
1476 }
1477
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 return ret;
1479}
1480
1481static void __exit as_exit(void)
1482{
Peter Zijlstra6e9a4732006-09-30 23:28:10 -07001483 DECLARE_COMPLETION_ONSTACK(all_gone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 elv_unregister(&iosched_as);
Al Viro334e94d2006-03-18 15:05:53 -05001485 ioc_gone = &all_gone;
OGAWA Hirofumifba82272006-04-18 09:44:06 +02001486 /* ioc_gone's update must be visible before reading ioc_count */
1487 smp_wmb();
Jens Axboee4313dd2006-07-19 05:10:01 +02001488 if (elv_ioc_count_read(ioc_count))
OGAWA Hirofumifba82272006-04-18 09:44:06 +02001489 wait_for_completion(ioc_gone);
Al Viro334e94d2006-03-18 15:05:53 -05001490 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491}
1492
1493module_init(as_init);
1494module_exit(as_exit);
1495
1496MODULE_AUTHOR("Nick Piggin");
1497MODULE_LICENSE("GPL");
1498MODULE_DESCRIPTION("anticipatory IO scheduler");