2 * Anticipatory & deadline i/o scheduler.
4 * Copyright (C) 2002 Jens Axboe <axboe@suse.de>
5 * Nick Piggin <nickpiggin@yahoo.com.au>
8 #include <linux/kernel.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>
26 * See Documentation/block/as-iosched.txt
30 * max time before a read is submitted.
32 #define default_read_expire (HZ / 8)
35 * ditto for writes, these limits are not hard, even
36 * if the disk is capable of satisfying them.
38 #define default_write_expire (HZ / 4)
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.
44 #define default_read_batch_expire (HZ / 2)
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...
52 #define default_write_batch_expire (HZ / 8)
55 * max time we may wait to anticipate a read (default around 6ms)
57 #define default_antic_expire ((HZ / 150) ? HZ / 150 : 1)
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.
66 #define MAX_THINKTIME (HZ/50UL)
68 /* Bits in as_io_context.state */
70 AS_TASK_RUNNING
=0, /* Process has not exited */
71 AS_TASK_IOSTARTED
, /* Process has started some IO */
72 AS_TASK_IORUNNING
, /* Process has completed some IO */
75 enum 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
89 struct request_queue
*q
; /* the "owner" queue */
92 * requests (as_rq s) are present on both sort_list and fifo_list
94 struct rb_root sort_list
[2];
95 struct list_head fifo_list
[2];
97 struct as_rq
*next_arq
[2]; /* next in sort order */
98 sector_t last_sector
[2]; /* last REQ_SYNC & REQ_ASYNC sectors */
99 struct list_head
*hash
; /* request hash */
101 unsigned long exit_prob
; /* probability a task will exit while
103 unsigned long exit_no_coop
; /* probablility an exited task will
104 not be part of a later cooperating
106 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
;
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? */
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 */
130 * settings that change how the i/o scheduler behaves
132 unsigned long fifo_expire
[2];
133 unsigned long batch_expire
[2];
134 unsigned long antic_expire
;
137 #define list_entry_fifo(ptr) list_entry((ptr), struct as_rq, fifo)
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
146 AS_RQ_DISPATCHED
, /* On the dispatch list. It belongs to the
148 AS_RQ_PRESCHED
, /* Debug poisoning for requests being used */
151 AS_RQ_POSTSCHED
, /* when they shouldn't be */
156 * rbtree index, key is the starting offset
158 struct rb_node rb_node
;
161 struct request
*request
;
163 struct io_context
*io_context
; /* The submitting task */
166 * request hash, key is the ending offset (for back merge lookup)
168 struct list_head hash
;
169 unsigned int on_hash
;
174 struct list_head fifo
;
175 unsigned long expires
;
177 unsigned int is_sync
;
178 enum arq_state state
;
181 #define RQ_DATA(rq) ((struct as_rq *) (rq)->elevator_private)
183 static kmem_cache_t
*arq_pool
;
185 static atomic_t ioc_count
= ATOMIC_INIT(0);
186 static struct completion
*ioc_gone
;
188 static void as_move_to_dispatch(struct as_data
*ad
, struct as_rq
*arq
);
189 static void as_antic_stop(struct as_data
*ad
);
192 * IO Context helper functions
195 /* Called to deallocate the as_io_context */
196 static void free_as_io_context(struct as_io_context
*aic
)
199 if (atomic_dec_and_test(&ioc_count
) && ioc_gone
)
203 static void as_trim(struct io_context
*ioc
)
206 free_as_io_context(ioc
->aic
);
210 /* Called when the task exits */
211 static void exit_as_io_context(struct as_io_context
*aic
)
213 WARN_ON(!test_bit(AS_TASK_RUNNING
, &aic
->state
));
214 clear_bit(AS_TASK_RUNNING
, &aic
->state
);
217 static struct as_io_context
*alloc_as_io_context(void)
219 struct as_io_context
*ret
;
221 ret
= kmalloc(sizeof(*ret
), GFP_ATOMIC
);
223 ret
->dtor
= free_as_io_context
;
224 ret
->exit
= exit_as_io_context
;
225 ret
->state
= 1 << AS_TASK_RUNNING
;
226 atomic_set(&ret
->nr_queued
, 0);
227 atomic_set(&ret
->nr_dispatched
, 0);
228 spin_lock_init(&ret
->lock
);
229 ret
->ttime_total
= 0;
230 ret
->ttime_samples
= 0;
233 ret
->seek_samples
= 0;
235 atomic_inc(&ioc_count
);
242 * If the current task has no AS IO context then create one and initialise it.
243 * Then take a ref on the task's io context and return it.
245 static struct io_context
*as_get_io_context(void)
247 struct io_context
*ioc
= get_io_context(GFP_ATOMIC
);
248 if (ioc
&& !ioc
->aic
) {
249 ioc
->aic
= alloc_as_io_context();
258 static void as_put_io_context(struct as_rq
*arq
)
260 struct as_io_context
*aic
;
262 if (unlikely(!arq
->io_context
))
265 aic
= arq
->io_context
->aic
;
267 if (arq
->is_sync
== REQ_SYNC
&& aic
) {
268 spin_lock(&aic
->lock
);
269 set_bit(AS_TASK_IORUNNING
, &aic
->state
);
270 aic
->last_end_request
= jiffies
;
271 spin_unlock(&aic
->lock
);
274 put_io_context(arq
->io_context
);
278 * the back merge hash support functions
280 static const int as_hash_shift
= 6;
281 #define AS_HASH_BLOCK(sec) ((sec) >> 3)
282 #define AS_HASH_FN(sec) (hash_long(AS_HASH_BLOCK((sec)), as_hash_shift))
283 #define AS_HASH_ENTRIES (1 << as_hash_shift)
284 #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
285 #define list_entry_hash(ptr) list_entry((ptr), struct as_rq, hash)
287 static inline void __as_del_arq_hash(struct as_rq
*arq
)
290 list_del_init(&arq
->hash
);
293 static inline void as_del_arq_hash(struct as_rq
*arq
)
296 __as_del_arq_hash(arq
);
299 static void as_add_arq_hash(struct as_data
*ad
, struct as_rq
*arq
)
301 struct request
*rq
= arq
->request
;
303 BUG_ON(arq
->on_hash
);
306 list_add(&arq
->hash
, &ad
->hash
[AS_HASH_FN(rq_hash_key(rq
))]);
310 * move hot entry to front of chain
312 static inline void as_hot_arq_hash(struct as_data
*ad
, struct as_rq
*arq
)
314 struct request
*rq
= arq
->request
;
315 struct list_head
*head
= &ad
->hash
[AS_HASH_FN(rq_hash_key(rq
))];
322 if (arq
->hash
.prev
!= head
) {
323 list_del(&arq
->hash
);
324 list_add(&arq
->hash
, head
);
328 static struct request
*as_find_arq_hash(struct as_data
*ad
, sector_t offset
)
330 struct list_head
*hash_list
= &ad
->hash
[AS_HASH_FN(offset
)];
331 struct list_head
*entry
, *next
= hash_list
->next
;
333 while ((entry
= next
) != hash_list
) {
334 struct as_rq
*arq
= list_entry_hash(entry
);
335 struct request
*__rq
= arq
->request
;
339 BUG_ON(!arq
->on_hash
);
341 if (!rq_mergeable(__rq
)) {
342 as_del_arq_hash(arq
);
346 if (rq_hash_key(__rq
) == offset
)
354 * rb tree support functions
356 #define RB_EMPTY(root) ((root)->rb_node == NULL)
357 #define ON_RB(node) (rb_parent(node) != node)
358 #define RB_CLEAR(node) (rb_set_parent(node, node))
359 #define rb_entry_arq(node) rb_entry((node), struct as_rq, rb_node)
360 #define ARQ_RB_ROOT(ad, arq) (&(ad)->sort_list[(arq)->is_sync])
361 #define rq_rb_key(rq) (rq)->sector
364 * as_find_first_arq finds the first (lowest sector numbered) request
365 * for the specified data_dir. Used to sweep back to the start of the disk
366 * (1-way elevator) after we process the last (highest sector) request.
368 static struct as_rq
*as_find_first_arq(struct as_data
*ad
, int data_dir
)
370 struct rb_node
*n
= ad
->sort_list
[data_dir
].rb_node
;
376 if (n
->rb_left
== NULL
)
377 return rb_entry_arq(n
);
384 * Add the request to the rb tree if it is unique. If there is an alias (an
385 * existing request against the same sector), which can happen when using
386 * direct IO, then return the alias.
388 static struct as_rq
*__as_add_arq_rb(struct as_data
*ad
, struct as_rq
*arq
)
390 struct rb_node
**p
= &ARQ_RB_ROOT(ad
, arq
)->rb_node
;
391 struct rb_node
*parent
= NULL
;
393 struct request
*rq
= arq
->request
;
395 arq
->rb_key
= rq_rb_key(rq
);
399 __arq
= rb_entry_arq(parent
);
401 if (arq
->rb_key
< __arq
->rb_key
)
403 else if (arq
->rb_key
> __arq
->rb_key
)
409 rb_link_node(&arq
->rb_node
, parent
, p
);
410 rb_insert_color(&arq
->rb_node
, ARQ_RB_ROOT(ad
, arq
));
415 static void as_add_arq_rb(struct as_data
*ad
, struct as_rq
*arq
)
419 while ((unlikely(alias
= __as_add_arq_rb(ad
, arq
)))) {
420 as_move_to_dispatch(ad
, alias
);
425 static inline void as_del_arq_rb(struct as_data
*ad
, struct as_rq
*arq
)
427 if (!ON_RB(&arq
->rb_node
)) {
432 rb_erase(&arq
->rb_node
, ARQ_RB_ROOT(ad
, arq
));
433 RB_CLEAR(&arq
->rb_node
);
436 static struct request
*
437 as_find_arq_rb(struct as_data
*ad
, sector_t sector
, int data_dir
)
439 struct rb_node
*n
= ad
->sort_list
[data_dir
].rb_node
;
443 arq
= rb_entry_arq(n
);
445 if (sector
< arq
->rb_key
)
447 else if (sector
> arq
->rb_key
)
457 * IO Scheduler proper
460 #define MAXBACK (1024 * 1024) /*
461 * Maximum distance the disk will go backward
465 #define BACK_PENALTY 2
468 * as_choose_req selects the preferred one of two requests of the same data_dir
469 * ignoring time - eg. timeouts, which is the job of as_dispatch_request
471 static struct as_rq
*
472 as_choose_req(struct as_data
*ad
, struct as_rq
*arq1
, struct as_rq
*arq2
)
475 sector_t last
, s1
, s2
, d1
, d2
;
476 int r1_wrap
=0, r2_wrap
=0; /* requests are behind the disk head */
477 const sector_t maxback
= MAXBACK
;
479 if (arq1
== NULL
|| arq1
== arq2
)
484 data_dir
= arq1
->is_sync
;
486 last
= ad
->last_sector
[data_dir
];
487 s1
= arq1
->request
->sector
;
488 s2
= arq2
->request
->sector
;
490 BUG_ON(data_dir
!= arq2
->is_sync
);
493 * Strict one way elevator _except_ in the case where we allow
494 * short backward seeks which are biased as twice the cost of a
495 * similar forward seek.
499 else if (s1
+maxback
>= last
)
500 d1
= (last
- s1
)*BACK_PENALTY
;
503 d1
= 0; /* shut up, gcc */
508 else if (s2
+maxback
>= last
)
509 d2
= (last
- s2
)*BACK_PENALTY
;
515 /* Found required data */
516 if (!r1_wrap
&& r2_wrap
)
518 else if (!r2_wrap
&& r1_wrap
)
520 else if (r1_wrap
&& r2_wrap
) {
521 /* both behind the head */
528 /* Both requests in front of the head */
542 * as_find_next_arq finds the next request after @prev in elevator order.
543 * this with as_choose_req form the basis for how the scheduler chooses
544 * what request to process next. Anticipation works on top of this.
546 static struct as_rq
*as_find_next_arq(struct as_data
*ad
, struct as_rq
*last
)
548 const int data_dir
= last
->is_sync
;
550 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
551 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
552 struct as_rq
*arq_next
, *arq_prev
;
554 BUG_ON(!ON_RB(&last
->rb_node
));
557 arq_prev
= rb_entry_arq(rbprev
);
562 arq_next
= rb_entry_arq(rbnext
);
564 arq_next
= as_find_first_arq(ad
, data_dir
);
565 if (arq_next
== last
)
569 ret
= as_choose_req(ad
, arq_next
, arq_prev
);
575 * anticipatory scheduling functions follow
579 * as_antic_expired tells us when we have anticipated too long.
580 * The funny "absolute difference" math on the elapsed time is to handle
581 * jiffy wraps, and disks which have been idle for 0x80000000 jiffies.
583 static int as_antic_expired(struct as_data
*ad
)
587 delta_jif
= jiffies
- ad
->antic_start
;
588 if (unlikely(delta_jif
< 0))
589 delta_jif
= -delta_jif
;
590 if (delta_jif
< ad
->antic_expire
)
597 * as_antic_waitnext starts anticipating that a nice request will soon be
598 * submitted. See also as_antic_waitreq
600 static void as_antic_waitnext(struct as_data
*ad
)
602 unsigned long timeout
;
604 BUG_ON(ad
->antic_status
!= ANTIC_OFF
605 && ad
->antic_status
!= ANTIC_WAIT_REQ
);
607 timeout
= ad
->antic_start
+ ad
->antic_expire
;
609 mod_timer(&ad
->antic_timer
, timeout
);
611 ad
->antic_status
= ANTIC_WAIT_NEXT
;
615 * as_antic_waitreq starts anticipating. We don't start timing the anticipation
616 * until the request that we're anticipating on has finished. This means we
617 * are timing from when the candidate process wakes up hopefully.
619 static void as_antic_waitreq(struct as_data
*ad
)
621 BUG_ON(ad
->antic_status
== ANTIC_FINISHED
);
622 if (ad
->antic_status
== ANTIC_OFF
) {
623 if (!ad
->io_context
|| ad
->ioc_finished
)
624 as_antic_waitnext(ad
);
626 ad
->antic_status
= ANTIC_WAIT_REQ
;
631 * This is called directly by the functions in this file to stop anticipation.
632 * We kill the timer and schedule a call to the request_fn asap.
634 static void as_antic_stop(struct as_data
*ad
)
636 int status
= ad
->antic_status
;
638 if (status
== ANTIC_WAIT_REQ
|| status
== ANTIC_WAIT_NEXT
) {
639 if (status
== ANTIC_WAIT_NEXT
)
640 del_timer(&ad
->antic_timer
);
641 ad
->antic_status
= ANTIC_FINISHED
;
642 /* see as_work_handler */
643 kblockd_schedule_work(&ad
->antic_work
);
648 * as_antic_timeout is the timer function set by as_antic_waitnext.
650 static void as_antic_timeout(unsigned long data
)
652 struct request_queue
*q
= (struct request_queue
*)data
;
653 struct as_data
*ad
= q
->elevator
->elevator_data
;
656 spin_lock_irqsave(q
->queue_lock
, flags
);
657 if (ad
->antic_status
== ANTIC_WAIT_REQ
658 || ad
->antic_status
== ANTIC_WAIT_NEXT
) {
659 struct as_io_context
*aic
= ad
->io_context
->aic
;
661 ad
->antic_status
= ANTIC_FINISHED
;
662 kblockd_schedule_work(&ad
->antic_work
);
664 if (aic
->ttime_samples
== 0) {
665 /* process anticipated on has exited or timed out*/
666 ad
->exit_prob
= (7*ad
->exit_prob
+ 256)/8;
668 if (!test_bit(AS_TASK_RUNNING
, &aic
->state
)) {
669 /* process not "saved" by a cooperating request */
670 ad
->exit_no_coop
= (7*ad
->exit_no_coop
+ 256)/8;
673 spin_unlock_irqrestore(q
->queue_lock
, flags
);
676 static void as_update_thinktime(struct as_data
*ad
, struct as_io_context
*aic
,
679 /* fixed point: 1.0 == 1<<8 */
680 if (aic
->ttime_samples
== 0) {
681 ad
->new_ttime_total
= (7*ad
->new_ttime_total
+ 256*ttime
) / 8;
682 ad
->new_ttime_mean
= ad
->new_ttime_total
/ 256;
684 ad
->exit_prob
= (7*ad
->exit_prob
)/8;
686 aic
->ttime_samples
= (7*aic
->ttime_samples
+ 256) / 8;
687 aic
->ttime_total
= (7*aic
->ttime_total
+ 256*ttime
) / 8;
688 aic
->ttime_mean
= (aic
->ttime_total
+ 128) / aic
->ttime_samples
;
691 static void as_update_seekdist(struct as_data
*ad
, struct as_io_context
*aic
,
696 if (aic
->seek_samples
== 0) {
697 ad
->new_seek_total
= (7*ad
->new_seek_total
+ 256*(u64
)sdist
)/8;
698 ad
->new_seek_mean
= ad
->new_seek_total
/ 256;
702 * Don't allow the seek distance to get too large from the
703 * odd fragment, pagein, etc
705 if (aic
->seek_samples
<= 60) /* second&third seek */
706 sdist
= min(sdist
, (aic
->seek_mean
* 4) + 2*1024*1024);
708 sdist
= min(sdist
, (aic
->seek_mean
* 4) + 2*1024*64);
710 aic
->seek_samples
= (7*aic
->seek_samples
+ 256) / 8;
711 aic
->seek_total
= (7*aic
->seek_total
+ (u64
)256*sdist
) / 8;
712 total
= aic
->seek_total
+ (aic
->seek_samples
/2);
713 do_div(total
, aic
->seek_samples
);
714 aic
->seek_mean
= (sector_t
)total
;
718 * as_update_iohist keeps a decaying histogram of IO thinktimes, and
719 * updates @aic->ttime_mean based on that. It is called when a new
722 static void as_update_iohist(struct as_data
*ad
, struct as_io_context
*aic
,
725 struct as_rq
*arq
= RQ_DATA(rq
);
726 int data_dir
= arq
->is_sync
;
727 unsigned long thinktime
= 0;
733 if (data_dir
== REQ_SYNC
) {
734 unsigned long in_flight
= atomic_read(&aic
->nr_queued
)
735 + atomic_read(&aic
->nr_dispatched
);
736 spin_lock(&aic
->lock
);
737 if (test_bit(AS_TASK_IORUNNING
, &aic
->state
) ||
738 test_bit(AS_TASK_IOSTARTED
, &aic
->state
)) {
739 /* Calculate read -> read thinktime */
740 if (test_bit(AS_TASK_IORUNNING
, &aic
->state
)
742 thinktime
= jiffies
- aic
->last_end_request
;
743 thinktime
= min(thinktime
, MAX_THINKTIME
-1);
745 as_update_thinktime(ad
, aic
, thinktime
);
747 /* Calculate read -> read seek distance */
748 if (aic
->last_request_pos
< rq
->sector
)
749 seek_dist
= rq
->sector
- aic
->last_request_pos
;
751 seek_dist
= aic
->last_request_pos
- rq
->sector
;
752 as_update_seekdist(ad
, aic
, seek_dist
);
754 aic
->last_request_pos
= rq
->sector
+ rq
->nr_sectors
;
755 set_bit(AS_TASK_IOSTARTED
, &aic
->state
);
756 spin_unlock(&aic
->lock
);
761 * as_close_req decides if one request is considered "close" to the
762 * previous one issued.
764 static int as_close_req(struct as_data
*ad
, struct as_io_context
*aic
,
767 unsigned long delay
; /* milliseconds */
768 sector_t last
= ad
->last_sector
[ad
->batch_data_dir
];
769 sector_t next
= arq
->request
->sector
;
770 sector_t delta
; /* acceptable close offset (in sectors) */
773 if (ad
->antic_status
== ANTIC_OFF
|| !ad
->ioc_finished
)
776 delay
= ((jiffies
- ad
->antic_start
) * 1000) / HZ
;
780 else if (delay
<= 20 && delay
<= ad
->antic_expire
)
781 delta
= 8192 << delay
;
785 if ((last
<= next
+ (delta
>>1)) && (next
<= last
+ delta
))
793 if (aic
->seek_samples
== 0) {
795 * Process has just started IO. Use past statistics to
796 * gauge success possibility
798 if (ad
->new_seek_mean
> s
) {
799 /* this request is better than what we're expecting */
804 if (aic
->seek_mean
> s
) {
805 /* this request is better than what we're expecting */
814 * as_can_break_anticipation returns true if we have been anticipating this
817 * It also returns true if the process against which we are anticipating
818 * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to
819 * dispatch it ASAP, because we know that application will not be submitting
822 * If the task which has submitted the request has exited, break anticipation.
824 * If this task has queued some other IO, do not enter enticipation.
826 static int as_can_break_anticipation(struct as_data
*ad
, struct as_rq
*arq
)
828 struct io_context
*ioc
;
829 struct as_io_context
*aic
;
831 ioc
= ad
->io_context
;
834 if (arq
&& ioc
== arq
->io_context
) {
835 /* request from same process */
839 if (ad
->ioc_finished
&& as_antic_expired(ad
)) {
841 * In this situation status should really be FINISHED,
842 * however the timer hasn't had the chance to run yet.
851 if (atomic_read(&aic
->nr_queued
) > 0) {
852 /* process has more requests queued */
856 if (atomic_read(&aic
->nr_dispatched
) > 0) {
857 /* process has more requests dispatched */
861 if (arq
&& arq
->is_sync
== REQ_SYNC
&& as_close_req(ad
, aic
, arq
)) {
863 * Found a close request that is not one of ours.
865 * This makes close requests from another process update
866 * our IO history. Is generally useful when there are
867 * two or more cooperating processes working in the same
870 if (!test_bit(AS_TASK_RUNNING
, &aic
->state
)) {
871 if (aic
->ttime_samples
== 0)
872 ad
->exit_prob
= (7*ad
->exit_prob
+ 256)/8;
874 ad
->exit_no_coop
= (7*ad
->exit_no_coop
)/8;
877 as_update_iohist(ad
, aic
, arq
->request
);
881 if (!test_bit(AS_TASK_RUNNING
, &aic
->state
)) {
882 /* process anticipated on has exited */
883 if (aic
->ttime_samples
== 0)
884 ad
->exit_prob
= (7*ad
->exit_prob
+ 256)/8;
886 if (ad
->exit_no_coop
> 128)
890 if (aic
->ttime_samples
== 0) {
891 if (ad
->new_ttime_mean
> ad
->antic_expire
)
893 if (ad
->exit_prob
* ad
->exit_no_coop
> 128*256)
895 } else if (aic
->ttime_mean
> ad
->antic_expire
) {
896 /* the process thinks too much between requests */
904 * as_can_anticipate indicates weather we should either run arq
905 * or keep anticipating a better request.
907 static int as_can_anticipate(struct as_data
*ad
, struct as_rq
*arq
)
911 * Last request submitted was a write
915 if (ad
->antic_status
== ANTIC_FINISHED
)
917 * Don't restart if we have just finished. Run the next request
921 if (as_can_break_anticipation(ad
, arq
))
923 * This request is a good candidate. Don't keep anticipating,
929 * OK from here, we haven't finished, and don't have a decent request!
930 * Status is either ANTIC_OFF so start waiting,
931 * ANTIC_WAIT_REQ so continue waiting for request to finish
932 * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request.
939 * as_update_arq must be called whenever a request (arq) is added to
940 * the sort_list. This function keeps caches up to date, and checks if the
941 * request might be one we are "anticipating"
943 static void as_update_arq(struct as_data
*ad
, struct as_rq
*arq
)
945 const int data_dir
= arq
->is_sync
;
947 /* keep the next_arq cache up to date */
948 ad
->next_arq
[data_dir
] = as_choose_req(ad
, arq
, ad
->next_arq
[data_dir
]);
951 * have we been anticipating this request?
952 * or does it come from the same process as the one we are anticipating
955 if (ad
->antic_status
== ANTIC_WAIT_REQ
956 || ad
->antic_status
== ANTIC_WAIT_NEXT
) {
957 if (as_can_break_anticipation(ad
, arq
))
963 * Gathers timings and resizes the write batch automatically
965 static void update_write_batch(struct as_data
*ad
)
967 unsigned long batch
= ad
->batch_expire
[REQ_ASYNC
];
970 write_time
= (jiffies
- ad
->current_batch_expires
) + batch
;
974 if (write_time
> batch
&& !ad
->write_batch_idled
) {
975 if (write_time
> batch
* 3)
976 ad
->write_batch_count
/= 2;
978 ad
->write_batch_count
--;
979 } else if (write_time
< batch
&& ad
->current_write_count
== 0) {
980 if (batch
> write_time
* 3)
981 ad
->write_batch_count
*= 2;
983 ad
->write_batch_count
++;
986 if (ad
->write_batch_count
< 1)
987 ad
->write_batch_count
= 1;
991 * as_completed_request is to be called when a request has completed and
992 * returned something to the requesting process, be it an error or data.
994 static void as_completed_request(request_queue_t
*q
, struct request
*rq
)
996 struct as_data
*ad
= q
->elevator
->elevator_data
;
997 struct as_rq
*arq
= RQ_DATA(rq
);
999 WARN_ON(!list_empty(&rq
->queuelist
));
1001 if (arq
->state
!= AS_RQ_REMOVED
) {
1002 printk("arq->state %d\n", arq
->state
);
1007 if (ad
->changed_batch
&& ad
->nr_dispatched
== 1) {
1008 kblockd_schedule_work(&ad
->antic_work
);
1009 ad
->changed_batch
= 0;
1011 if (ad
->batch_data_dir
== REQ_SYNC
)
1014 WARN_ON(ad
->nr_dispatched
== 0);
1015 ad
->nr_dispatched
--;
1018 * Start counting the batch from when a request of that direction is
1019 * actually serviced. This should help devices with big TCQ windows
1020 * and writeback caches
1022 if (ad
->new_batch
&& ad
->batch_data_dir
== arq
->is_sync
) {
1023 update_write_batch(ad
);
1024 ad
->current_batch_expires
= jiffies
+
1025 ad
->batch_expire
[REQ_SYNC
];
1029 if (ad
->io_context
== arq
->io_context
&& ad
->io_context
) {
1030 ad
->antic_start
= jiffies
;
1031 ad
->ioc_finished
= 1;
1032 if (ad
->antic_status
== ANTIC_WAIT_REQ
) {
1034 * We were waiting on this request, now anticipate
1037 as_antic_waitnext(ad
);
1041 as_put_io_context(arq
);
1043 arq
->state
= AS_RQ_POSTSCHED
;
1047 * as_remove_queued_request removes a request from the pre dispatch queue
1048 * without updating refcounts. It is expected the caller will drop the
1049 * reference unless it replaces the request at somepart of the elevator
1050 * (ie. the dispatch queue)
1052 static void as_remove_queued_request(request_queue_t
*q
, struct request
*rq
)
1054 struct as_rq
*arq
= RQ_DATA(rq
);
1055 const int data_dir
= arq
->is_sync
;
1056 struct as_data
*ad
= q
->elevator
->elevator_data
;
1058 WARN_ON(arq
->state
!= AS_RQ_QUEUED
);
1060 if (arq
->io_context
&& arq
->io_context
->aic
) {
1061 BUG_ON(!atomic_read(&arq
->io_context
->aic
->nr_queued
));
1062 atomic_dec(&arq
->io_context
->aic
->nr_queued
);
1066 * Update the "next_arq" cache if we are about to remove its
1069 if (ad
->next_arq
[data_dir
] == arq
)
1070 ad
->next_arq
[data_dir
] = as_find_next_arq(ad
, arq
);
1072 list_del_init(&arq
->fifo
);
1073 as_del_arq_hash(arq
);
1074 as_del_arq_rb(ad
, arq
);
1078 * as_fifo_expired returns 0 if there are no expired reads on the fifo,
1079 * 1 otherwise. It is ratelimited so that we only perform the check once per
1080 * `fifo_expire' interval. Otherwise a large number of expired requests
1081 * would create a hopeless seekstorm.
1083 * See as_antic_expired comment.
1085 static int as_fifo_expired(struct as_data
*ad
, int adir
)
1090 delta_jif
= jiffies
- ad
->last_check_fifo
[adir
];
1091 if (unlikely(delta_jif
< 0))
1092 delta_jif
= -delta_jif
;
1093 if (delta_jif
< ad
->fifo_expire
[adir
])
1096 ad
->last_check_fifo
[adir
] = jiffies
;
1098 if (list_empty(&ad
->fifo_list
[adir
]))
1101 arq
= list_entry_fifo(ad
->fifo_list
[adir
].next
);
1103 return time_after(jiffies
, arq
->expires
);
1107 * as_batch_expired returns true if the current batch has expired. A batch
1108 * is a set of reads or a set of writes.
1110 static inline int as_batch_expired(struct as_data
*ad
)
1112 if (ad
->changed_batch
|| ad
->new_batch
)
1115 if (ad
->batch_data_dir
== REQ_SYNC
)
1116 /* TODO! add a check so a complete fifo gets written? */
1117 return time_after(jiffies
, ad
->current_batch_expires
);
1119 return time_after(jiffies
, ad
->current_batch_expires
)
1120 || ad
->current_write_count
== 0;
1124 * move an entry to dispatch queue
1126 static void as_move_to_dispatch(struct as_data
*ad
, struct as_rq
*arq
)
1128 struct request
*rq
= arq
->request
;
1129 const int data_dir
= arq
->is_sync
;
1131 BUG_ON(!ON_RB(&arq
->rb_node
));
1134 ad
->antic_status
= ANTIC_OFF
;
1137 * This has to be set in order to be correctly updated by
1140 ad
->last_sector
[data_dir
] = rq
->sector
+ rq
->nr_sectors
;
1142 if (data_dir
== REQ_SYNC
) {
1143 /* In case we have to anticipate after this */
1144 copy_io_context(&ad
->io_context
, &arq
->io_context
);
1146 if (ad
->io_context
) {
1147 put_io_context(ad
->io_context
);
1148 ad
->io_context
= NULL
;
1151 if (ad
->current_write_count
!= 0)
1152 ad
->current_write_count
--;
1154 ad
->ioc_finished
= 0;
1156 ad
->next_arq
[data_dir
] = as_find_next_arq(ad
, arq
);
1159 * take it off the sort and fifo list, add to dispatch queue
1161 as_remove_queued_request(ad
->q
, rq
);
1162 WARN_ON(arq
->state
!= AS_RQ_QUEUED
);
1164 elv_dispatch_sort(ad
->q
, rq
);
1166 arq
->state
= AS_RQ_DISPATCHED
;
1167 if (arq
->io_context
&& arq
->io_context
->aic
)
1168 atomic_inc(&arq
->io_context
->aic
->nr_dispatched
);
1169 ad
->nr_dispatched
++;
1173 * as_dispatch_request selects the best request according to
1174 * read/write expire, batch expire, etc, and moves it to the dispatch
1175 * queue. Returns 1 if a request was found, 0 otherwise.
1177 static int as_dispatch_request(request_queue_t
*q
, int force
)
1179 struct as_data
*ad
= q
->elevator
->elevator_data
;
1181 const int reads
= !list_empty(&ad
->fifo_list
[REQ_SYNC
]);
1182 const int writes
= !list_empty(&ad
->fifo_list
[REQ_ASYNC
]);
1184 if (unlikely(force
)) {
1186 * Forced dispatch, accounting is useless. Reset
1187 * accounting states and dump fifo_lists. Note that
1188 * batch_data_dir is reset to REQ_SYNC to avoid
1189 * screwing write batch accounting as write batch
1190 * accounting occurs on W->R transition.
1194 ad
->batch_data_dir
= REQ_SYNC
;
1195 ad
->changed_batch
= 0;
1198 while (ad
->next_arq
[REQ_SYNC
]) {
1199 as_move_to_dispatch(ad
, ad
->next_arq
[REQ_SYNC
]);
1202 ad
->last_check_fifo
[REQ_SYNC
] = jiffies
;
1204 while (ad
->next_arq
[REQ_ASYNC
]) {
1205 as_move_to_dispatch(ad
, ad
->next_arq
[REQ_ASYNC
]);
1208 ad
->last_check_fifo
[REQ_ASYNC
] = jiffies
;
1213 /* Signal that the write batch was uncontended, so we can't time it */
1214 if (ad
->batch_data_dir
== REQ_ASYNC
&& !reads
) {
1215 if (ad
->current_write_count
== 0 || !writes
)
1216 ad
->write_batch_idled
= 1;
1219 if (!(reads
|| writes
)
1220 || ad
->antic_status
== ANTIC_WAIT_REQ
1221 || ad
->antic_status
== ANTIC_WAIT_NEXT
1222 || ad
->changed_batch
)
1225 if (!(reads
&& writes
&& as_batch_expired(ad
))) {
1227 * batch is still running or no reads or no writes
1229 arq
= ad
->next_arq
[ad
->batch_data_dir
];
1231 if (ad
->batch_data_dir
== REQ_SYNC
&& ad
->antic_expire
) {
1232 if (as_fifo_expired(ad
, REQ_SYNC
))
1235 if (as_can_anticipate(ad
, arq
)) {
1236 as_antic_waitreq(ad
);
1242 /* we have a "next request" */
1243 if (reads
&& !writes
)
1244 ad
->current_batch_expires
=
1245 jiffies
+ ad
->batch_expire
[REQ_SYNC
];
1246 goto dispatch_request
;
1251 * at this point we are not running a batch. select the appropriate
1252 * data direction (read / write)
1256 BUG_ON(RB_EMPTY(&ad
->sort_list
[REQ_SYNC
]));
1258 if (writes
&& ad
->batch_data_dir
== REQ_SYNC
)
1260 * Last batch was a read, switch to writes
1262 goto dispatch_writes
;
1264 if (ad
->batch_data_dir
== REQ_ASYNC
) {
1265 WARN_ON(ad
->new_batch
);
1266 ad
->changed_batch
= 1;
1268 ad
->batch_data_dir
= REQ_SYNC
;
1269 arq
= list_entry_fifo(ad
->fifo_list
[ad
->batch_data_dir
].next
);
1270 ad
->last_check_fifo
[ad
->batch_data_dir
] = jiffies
;
1271 goto dispatch_request
;
1275 * the last batch was a read
1280 BUG_ON(RB_EMPTY(&ad
->sort_list
[REQ_ASYNC
]));
1282 if (ad
->batch_data_dir
== REQ_SYNC
) {
1283 ad
->changed_batch
= 1;
1286 * new_batch might be 1 when the queue runs out of
1287 * reads. A subsequent submission of a write might
1288 * cause a change of batch before the read is finished.
1292 ad
->batch_data_dir
= REQ_ASYNC
;
1293 ad
->current_write_count
= ad
->write_batch_count
;
1294 ad
->write_batch_idled
= 0;
1295 arq
= ad
->next_arq
[ad
->batch_data_dir
];
1296 goto dispatch_request
;
1304 * If a request has expired, service it.
1307 if (as_fifo_expired(ad
, ad
->batch_data_dir
)) {
1309 arq
= list_entry_fifo(ad
->fifo_list
[ad
->batch_data_dir
].next
);
1310 BUG_ON(arq
== NULL
);
1313 if (ad
->changed_batch
) {
1314 WARN_ON(ad
->new_batch
);
1316 if (ad
->nr_dispatched
)
1319 if (ad
->batch_data_dir
== REQ_ASYNC
)
1320 ad
->current_batch_expires
= jiffies
+
1321 ad
->batch_expire
[REQ_ASYNC
];
1325 ad
->changed_batch
= 0;
1329 * arq is the selected appropriate request.
1331 as_move_to_dispatch(ad
, arq
);
1337 * add arq to rbtree and fifo
1339 static void as_add_request(request_queue_t
*q
, struct request
*rq
)
1341 struct as_data
*ad
= q
->elevator
->elevator_data
;
1342 struct as_rq
*arq
= RQ_DATA(rq
);
1345 arq
->state
= AS_RQ_NEW
;
1347 if (rq_data_dir(arq
->request
) == READ
1348 || current
->flags
&PF_SYNCWRITE
)
1352 data_dir
= arq
->is_sync
;
1354 arq
->io_context
= as_get_io_context();
1356 if (arq
->io_context
) {
1357 as_update_iohist(ad
, arq
->io_context
->aic
, arq
->request
);
1358 atomic_inc(&arq
->io_context
->aic
->nr_queued
);
1361 as_add_arq_rb(ad
, arq
);
1362 if (rq_mergeable(arq
->request
))
1363 as_add_arq_hash(ad
, arq
);
1366 * set expire time (only used for reads) and add to fifo list
1368 arq
->expires
= jiffies
+ ad
->fifo_expire
[data_dir
];
1369 list_add_tail(&arq
->fifo
, &ad
->fifo_list
[data_dir
]);
1371 as_update_arq(ad
, arq
); /* keep state machine up to date */
1372 arq
->state
= AS_RQ_QUEUED
;
1375 static void as_activate_request(request_queue_t
*q
, struct request
*rq
)
1377 struct as_rq
*arq
= RQ_DATA(rq
);
1379 WARN_ON(arq
->state
!= AS_RQ_DISPATCHED
);
1380 arq
->state
= AS_RQ_REMOVED
;
1381 if (arq
->io_context
&& arq
->io_context
->aic
)
1382 atomic_dec(&arq
->io_context
->aic
->nr_dispatched
);
1385 static void as_deactivate_request(request_queue_t
*q
, struct request
*rq
)
1387 struct as_rq
*arq
= RQ_DATA(rq
);
1389 WARN_ON(arq
->state
!= AS_RQ_REMOVED
);
1390 arq
->state
= AS_RQ_DISPATCHED
;
1391 if (arq
->io_context
&& arq
->io_context
->aic
)
1392 atomic_inc(&arq
->io_context
->aic
->nr_dispatched
);
1396 * as_queue_empty tells us if there are requests left in the device. It may
1397 * not be the case that a driver can get the next request even if the queue
1398 * is not empty - it is used in the block layer to check for plugging and
1399 * merging opportunities
1401 static int as_queue_empty(request_queue_t
*q
)
1403 struct as_data
*ad
= q
->elevator
->elevator_data
;
1405 return list_empty(&ad
->fifo_list
[REQ_ASYNC
])
1406 && list_empty(&ad
->fifo_list
[REQ_SYNC
]);
1409 static struct request
*as_former_request(request_queue_t
*q
,
1412 struct as_rq
*arq
= RQ_DATA(rq
);
1413 struct rb_node
*rbprev
= rb_prev(&arq
->rb_node
);
1414 struct request
*ret
= NULL
;
1417 ret
= rb_entry_arq(rbprev
)->request
;
1422 static struct request
*as_latter_request(request_queue_t
*q
,
1425 struct as_rq
*arq
= RQ_DATA(rq
);
1426 struct rb_node
*rbnext
= rb_next(&arq
->rb_node
);
1427 struct request
*ret
= NULL
;
1430 ret
= rb_entry_arq(rbnext
)->request
;
1436 as_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
1438 struct as_data
*ad
= q
->elevator
->elevator_data
;
1439 sector_t rb_key
= bio
->bi_sector
+ bio_sectors(bio
);
1440 struct request
*__rq
;
1444 * see if the merge hash can satisfy a back merge
1446 __rq
= as_find_arq_hash(ad
, bio
->bi_sector
);
1448 BUG_ON(__rq
->sector
+ __rq
->nr_sectors
!= bio
->bi_sector
);
1450 if (elv_rq_merge_ok(__rq
, bio
)) {
1451 ret
= ELEVATOR_BACK_MERGE
;
1457 * check for front merge
1459 __rq
= as_find_arq_rb(ad
, rb_key
, bio_data_dir(bio
));
1461 BUG_ON(rb_key
!= rq_rb_key(__rq
));
1463 if (elv_rq_merge_ok(__rq
, bio
)) {
1464 ret
= ELEVATOR_FRONT_MERGE
;
1469 return ELEVATOR_NO_MERGE
;
1472 if (rq_mergeable(__rq
))
1473 as_hot_arq_hash(ad
, RQ_DATA(__rq
));
1479 static void as_merged_request(request_queue_t
*q
, struct request
*req
)
1481 struct as_data
*ad
= q
->elevator
->elevator_data
;
1482 struct as_rq
*arq
= RQ_DATA(req
);
1485 * hash always needs to be repositioned, key is end sector
1487 as_del_arq_hash(arq
);
1488 as_add_arq_hash(ad
, arq
);
1491 * if the merge was a front merge, we need to reposition request
1493 if (rq_rb_key(req
) != arq
->rb_key
) {
1494 as_del_arq_rb(ad
, arq
);
1495 as_add_arq_rb(ad
, arq
);
1497 * Note! At this stage of this and the next function, our next
1498 * request may not be optimal - eg the request may have "grown"
1499 * behind the disk head. We currently don't bother adjusting.
1504 static void as_merged_requests(request_queue_t
*q
, struct request
*req
,
1505 struct request
*next
)
1507 struct as_data
*ad
= q
->elevator
->elevator_data
;
1508 struct as_rq
*arq
= RQ_DATA(req
);
1509 struct as_rq
*anext
= RQ_DATA(next
);
1515 * reposition arq (this is the merged request) in hash, and in rbtree
1516 * in case of a front merge
1518 as_del_arq_hash(arq
);
1519 as_add_arq_hash(ad
, arq
);
1521 if (rq_rb_key(req
) != arq
->rb_key
) {
1522 as_del_arq_rb(ad
, arq
);
1523 as_add_arq_rb(ad
, arq
);
1527 * if anext expires before arq, assign its expire time to arq
1528 * and move into anext position (anext will be deleted) in fifo
1530 if (!list_empty(&arq
->fifo
) && !list_empty(&anext
->fifo
)) {
1531 if (time_before(anext
->expires
, arq
->expires
)) {
1532 list_move(&arq
->fifo
, &anext
->fifo
);
1533 arq
->expires
= anext
->expires
;
1535 * Don't copy here but swap, because when anext is
1536 * removed below, it must contain the unused context
1538 swap_io_context(&arq
->io_context
, &anext
->io_context
);
1543 * kill knowledge of next, this one is a goner
1545 as_remove_queued_request(q
, next
);
1546 as_put_io_context(anext
);
1548 anext
->state
= AS_RQ_MERGED
;
1552 * This is executed in a "deferred" process context, by kblockd. It calls the
1553 * driver's request_fn so the driver can submit that request.
1555 * IMPORTANT! This guy will reenter the elevator, so set up all queue global
1556 * state before calling, and don't rely on any state over calls.
1558 * FIXME! dispatch queue is not a queue at all!
1560 static void as_work_handler(void *data
)
1562 struct request_queue
*q
= data
;
1563 unsigned long flags
;
1565 spin_lock_irqsave(q
->queue_lock
, flags
);
1566 if (!as_queue_empty(q
))
1568 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1571 static void as_put_request(request_queue_t
*q
, struct request
*rq
)
1573 struct as_data
*ad
= q
->elevator
->elevator_data
;
1574 struct as_rq
*arq
= RQ_DATA(rq
);
1581 if (unlikely(arq
->state
!= AS_RQ_POSTSCHED
&&
1582 arq
->state
!= AS_RQ_PRESCHED
&&
1583 arq
->state
!= AS_RQ_MERGED
)) {
1584 printk("arq->state %d\n", arq
->state
);
1588 mempool_free(arq
, ad
->arq_pool
);
1589 rq
->elevator_private
= NULL
;
1592 static int as_set_request(request_queue_t
*q
, struct request
*rq
,
1593 struct bio
*bio
, gfp_t gfp_mask
)
1595 struct as_data
*ad
= q
->elevator
->elevator_data
;
1596 struct as_rq
*arq
= mempool_alloc(ad
->arq_pool
, gfp_mask
);
1599 memset(arq
, 0, sizeof(*arq
));
1600 RB_CLEAR(&arq
->rb_node
);
1602 arq
->state
= AS_RQ_PRESCHED
;
1603 arq
->io_context
= NULL
;
1604 INIT_LIST_HEAD(&arq
->hash
);
1606 INIT_LIST_HEAD(&arq
->fifo
);
1607 rq
->elevator_private
= arq
;
1614 static int as_may_queue(request_queue_t
*q
, int rw
, struct bio
*bio
)
1616 int ret
= ELV_MQUEUE_MAY
;
1617 struct as_data
*ad
= q
->elevator
->elevator_data
;
1618 struct io_context
*ioc
;
1619 if (ad
->antic_status
== ANTIC_WAIT_REQ
||
1620 ad
->antic_status
== ANTIC_WAIT_NEXT
) {
1621 ioc
= as_get_io_context();
1622 if (ad
->io_context
== ioc
)
1623 ret
= ELV_MQUEUE_MUST
;
1624 put_io_context(ioc
);
1630 static void as_exit_queue(elevator_t
*e
)
1632 struct as_data
*ad
= e
->elevator_data
;
1634 del_timer_sync(&ad
->antic_timer
);
1637 BUG_ON(!list_empty(&ad
->fifo_list
[REQ_SYNC
]));
1638 BUG_ON(!list_empty(&ad
->fifo_list
[REQ_ASYNC
]));
1640 mempool_destroy(ad
->arq_pool
);
1641 put_io_context(ad
->io_context
);
1647 * initialize elevator private data (as_data), and alloc a arq for
1648 * each request on the free lists
1650 static void *as_init_queue(request_queue_t
*q
, elevator_t
*e
)
1658 ad
= kmalloc_node(sizeof(*ad
), GFP_KERNEL
, q
->node
);
1661 memset(ad
, 0, sizeof(*ad
));
1663 ad
->q
= q
; /* Identify what queue the data belongs to */
1665 ad
->hash
= kmalloc_node(sizeof(struct list_head
)*AS_HASH_ENTRIES
,
1666 GFP_KERNEL
, q
->node
);
1672 ad
->arq_pool
= mempool_create_node(BLKDEV_MIN_RQ
, mempool_alloc_slab
,
1673 mempool_free_slab
, arq_pool
, q
->node
);
1674 if (!ad
->arq_pool
) {
1680 /* anticipatory scheduling helpers */
1681 ad
->antic_timer
.function
= as_antic_timeout
;
1682 ad
->antic_timer
.data
= (unsigned long)q
;
1683 init_timer(&ad
->antic_timer
);
1684 INIT_WORK(&ad
->antic_work
, as_work_handler
, q
);
1686 for (i
= 0; i
< AS_HASH_ENTRIES
; i
++)
1687 INIT_LIST_HEAD(&ad
->hash
[i
]);
1689 INIT_LIST_HEAD(&ad
->fifo_list
[REQ_SYNC
]);
1690 INIT_LIST_HEAD(&ad
->fifo_list
[REQ_ASYNC
]);
1691 ad
->sort_list
[REQ_SYNC
] = RB_ROOT
;
1692 ad
->sort_list
[REQ_ASYNC
] = RB_ROOT
;
1693 ad
->fifo_expire
[REQ_SYNC
] = default_read_expire
;
1694 ad
->fifo_expire
[REQ_ASYNC
] = default_write_expire
;
1695 ad
->antic_expire
= default_antic_expire
;
1696 ad
->batch_expire
[REQ_SYNC
] = default_read_batch_expire
;
1697 ad
->batch_expire
[REQ_ASYNC
] = default_write_batch_expire
;
1699 ad
->current_batch_expires
= jiffies
+ ad
->batch_expire
[REQ_SYNC
];
1700 ad
->write_batch_count
= ad
->batch_expire
[REQ_ASYNC
] / 10;
1701 if (ad
->write_batch_count
< 2)
1702 ad
->write_batch_count
= 2;
1712 as_var_show(unsigned int var
, char *page
)
1714 return sprintf(page
, "%d\n", var
);
1718 as_var_store(unsigned long *var
, const char *page
, size_t count
)
1720 char *p
= (char *) page
;
1722 *var
= simple_strtoul(p
, &p
, 10);
1726 static ssize_t
est_time_show(elevator_t
*e
, char *page
)
1728 struct as_data
*ad
= e
->elevator_data
;
1731 pos
+= sprintf(page
+pos
, "%lu %% exit probability\n",
1732 100*ad
->exit_prob
/256);
1733 pos
+= sprintf(page
+pos
, "%lu %% probability of exiting without a "
1734 "cooperating process submitting IO\n",
1735 100*ad
->exit_no_coop
/256);
1736 pos
+= sprintf(page
+pos
, "%lu ms new thinktime\n", ad
->new_ttime_mean
);
1737 pos
+= sprintf(page
+pos
, "%llu sectors new seek distance\n",
1738 (unsigned long long)ad
->new_seek_mean
);
1743 #define SHOW_FUNCTION(__FUNC, __VAR) \
1744 static ssize_t __FUNC(elevator_t *e, char *page) \
1746 struct as_data *ad = e->elevator_data; \
1747 return as_var_show(jiffies_to_msecs((__VAR)), (page)); \
1749 SHOW_FUNCTION(as_read_expire_show
, ad
->fifo_expire
[REQ_SYNC
]);
1750 SHOW_FUNCTION(as_write_expire_show
, ad
->fifo_expire
[REQ_ASYNC
]);
1751 SHOW_FUNCTION(as_antic_expire_show
, ad
->antic_expire
);
1752 SHOW_FUNCTION(as_read_batch_expire_show
, ad
->batch_expire
[REQ_SYNC
]);
1753 SHOW_FUNCTION(as_write_batch_expire_show
, ad
->batch_expire
[REQ_ASYNC
]);
1754 #undef SHOW_FUNCTION
1756 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
1757 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
1759 struct as_data *ad = e->elevator_data; \
1760 int ret = as_var_store(__PTR, (page), count); \
1761 if (*(__PTR) < (MIN)) \
1763 else if (*(__PTR) > (MAX)) \
1765 *(__PTR) = msecs_to_jiffies(*(__PTR)); \
1768 STORE_FUNCTION(as_read_expire_store
, &ad
->fifo_expire
[REQ_SYNC
], 0, INT_MAX
);
1769 STORE_FUNCTION(as_write_expire_store
, &ad
->fifo_expire
[REQ_ASYNC
], 0, INT_MAX
);
1770 STORE_FUNCTION(as_antic_expire_store
, &ad
->antic_expire
, 0, INT_MAX
);
1771 STORE_FUNCTION(as_read_batch_expire_store
,
1772 &ad
->batch_expire
[REQ_SYNC
], 0, INT_MAX
);
1773 STORE_FUNCTION(as_write_batch_expire_store
,
1774 &ad
->batch_expire
[REQ_ASYNC
], 0, INT_MAX
);
1775 #undef STORE_FUNCTION
1777 #define AS_ATTR(name) \
1778 __ATTR(name, S_IRUGO|S_IWUSR, as_##name##_show, as_##name##_store)
1780 static struct elv_fs_entry as_attrs
[] = {
1781 __ATTR_RO(est_time
),
1782 AS_ATTR(read_expire
),
1783 AS_ATTR(write_expire
),
1784 AS_ATTR(antic_expire
),
1785 AS_ATTR(read_batch_expire
),
1786 AS_ATTR(write_batch_expire
),
1790 static struct elevator_type iosched_as
= {
1792 .elevator_merge_fn
= as_merge
,
1793 .elevator_merged_fn
= as_merged_request
,
1794 .elevator_merge_req_fn
= as_merged_requests
,
1795 .elevator_dispatch_fn
= as_dispatch_request
,
1796 .elevator_add_req_fn
= as_add_request
,
1797 .elevator_activate_req_fn
= as_activate_request
,
1798 .elevator_deactivate_req_fn
= as_deactivate_request
,
1799 .elevator_queue_empty_fn
= as_queue_empty
,
1800 .elevator_completed_req_fn
= as_completed_request
,
1801 .elevator_former_req_fn
= as_former_request
,
1802 .elevator_latter_req_fn
= as_latter_request
,
1803 .elevator_set_req_fn
= as_set_request
,
1804 .elevator_put_req_fn
= as_put_request
,
1805 .elevator_may_queue_fn
= as_may_queue
,
1806 .elevator_init_fn
= as_init_queue
,
1807 .elevator_exit_fn
= as_exit_queue
,
1811 .elevator_attrs
= as_attrs
,
1812 .elevator_name
= "anticipatory",
1813 .elevator_owner
= THIS_MODULE
,
1816 static int __init
as_init(void)
1820 arq_pool
= kmem_cache_create("as_arq", sizeof(struct as_rq
),
1825 ret
= elv_register(&iosched_as
);
1828 * don't allow AS to get unregistered, since we would have
1829 * to browse all tasks in the system and release their
1830 * as_io_context first
1832 __module_get(THIS_MODULE
);
1836 kmem_cache_destroy(arq_pool
);
1840 static void __exit
as_exit(void)
1842 DECLARE_COMPLETION(all_gone
);
1843 elv_unregister(&iosched_as
);
1844 ioc_gone
= &all_gone
;
1845 /* ioc_gone's update must be visible before reading ioc_count */
1847 if (atomic_read(&ioc_count
))
1848 wait_for_completion(ioc_gone
);
1850 kmem_cache_destroy(arq_pool
);
1853 module_init(as_init
);
1854 module_exit(as_exit
);
1856 MODULE_AUTHOR("Nick Piggin");
1857 MODULE_LICENSE("GPL");
1858 MODULE_DESCRIPTION("anticipatory IO scheduler");