2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@suse.de>
9 #include <linux/config.h>
10 #include <linux/module.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/hash.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
20 static const int cfq_quantum
= 4; /* max queue in one round of service */
21 static const int cfq_queued
= 8; /* minimum rq allocate limit per-queue*/
22 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
23 static const int cfq_back_max
= 16 * 1024; /* maximum backwards seek, in KiB */
24 static const int cfq_back_penalty
= 2; /* penalty of a backwards seek */
26 static const int cfq_slice_sync
= HZ
/ 10;
27 static int cfq_slice_async
= HZ
/ 25;
28 static const int cfq_slice_async_rq
= 2;
29 static int cfq_slice_idle
= HZ
/ 70;
31 #define CFQ_IDLE_GRACE (HZ / 10)
32 #define CFQ_SLICE_SCALE (5)
34 #define CFQ_KEY_ASYNC (0)
36 static DEFINE_SPINLOCK(cfq_exit_lock
);
39 * for the hash of cfqq inside the cfqd
41 #define CFQ_QHASH_SHIFT 6
42 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
43 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
46 * for the hash of crq inside the cfqq
48 #define CFQ_MHASH_SHIFT 6
49 #define CFQ_MHASH_BLOCK(sec) ((sec) >> 3)
50 #define CFQ_MHASH_ENTRIES (1 << CFQ_MHASH_SHIFT)
51 #define CFQ_MHASH_FN(sec) hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
52 #define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors)
53 #define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash)
55 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
56 #define list_entry_fifo(ptr) list_entry((ptr), struct request, queuelist)
58 #define RQ_DATA(rq) (rq)->elevator_private
63 #define RB_EMPTY(node) ((node)->rb_node == NULL)
64 #define RB_CLEAR(node) do { \
65 memset(node, 0, sizeof(*node)); \
67 #define RB_CLEAR_ROOT(root) ((root)->rb_node = NULL)
68 #define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node)
69 #define rq_rb_key(rq) (rq)->sector
71 static kmem_cache_t
*crq_pool
;
72 static kmem_cache_t
*cfq_pool
;
73 static kmem_cache_t
*cfq_ioc_pool
;
75 static atomic_t ioc_count
= ATOMIC_INIT(0);
76 static struct completion
*ioc_gone
;
78 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
79 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
80 #define cfq_class_be(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
81 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
86 #define cfq_cfqq_dispatched(cfqq) \
87 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
89 #define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
91 #define cfq_cfqq_sync(cfqq) \
92 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
94 #define sample_valid(samples) ((samples) > 80)
97 * Per block device queue structure
100 request_queue_t
*queue
;
103 * rr list of queues with requests and the count of them
105 struct list_head rr_list
[CFQ_PRIO_LISTS
];
106 struct list_head busy_rr
;
107 struct list_head cur_rr
;
108 struct list_head idle_rr
;
109 unsigned int busy_queues
;
112 * non-ordered list of empty cfqq's
114 struct list_head empty_list
;
119 struct hlist_head
*cfq_hash
;
122 * global crq hash for all queues
124 struct hlist_head
*crq_hash
;
126 unsigned int max_queued
;
134 * schedule slice state info
137 * idle window management
139 struct timer_list idle_slice_timer
;
140 struct work_struct unplug_work
;
142 struct cfq_queue
*active_queue
;
143 struct cfq_io_context
*active_cic
;
144 int cur_prio
, cur_end_prio
;
145 unsigned int dispatch_slice
;
147 struct timer_list idle_class_timer
;
149 sector_t last_sector
;
150 unsigned long last_end_request
;
152 unsigned int rq_starved
;
155 * tunables, see top of file
157 unsigned int cfq_quantum
;
158 unsigned int cfq_queued
;
159 unsigned int cfq_fifo_expire
[2];
160 unsigned int cfq_back_penalty
;
161 unsigned int cfq_back_max
;
162 unsigned int cfq_slice
[2];
163 unsigned int cfq_slice_async_rq
;
164 unsigned int cfq_slice_idle
;
166 struct list_head cic_list
;
170 * Per process-grouping structure
173 /* reference count */
175 /* parent cfq_data */
176 struct cfq_data
*cfqd
;
177 /* cfqq lookup hash */
178 struct hlist_node cfq_hash
;
181 /* on either rr or empty list of cfqd */
182 struct list_head cfq_list
;
183 /* sorted list of pending requests */
184 struct rb_root sort_list
;
185 /* if fifo isn't expired, next request to serve */
186 struct cfq_rq
*next_crq
;
187 /* requests queued in sort_list */
189 /* currently allocated requests */
191 /* fifo list of requests in sort_list */
192 struct list_head fifo
;
194 unsigned long slice_start
;
195 unsigned long slice_end
;
196 unsigned long slice_left
;
197 unsigned long service_last
;
199 /* number of requests that are on the dispatch list */
202 /* io prio of this group */
203 unsigned short ioprio
, org_ioprio
;
204 unsigned short ioprio_class
, org_ioprio_class
;
206 /* various state flags, see below */
211 struct rb_node rb_node
;
213 struct request
*request
;
214 struct hlist_node hash
;
216 struct cfq_queue
*cfq_queue
;
217 struct cfq_io_context
*io_context
;
219 unsigned int crq_flags
;
222 enum cfqq_state_flags
{
223 CFQ_CFQQ_FLAG_on_rr
= 0,
224 CFQ_CFQQ_FLAG_wait_request
,
225 CFQ_CFQQ_FLAG_must_alloc
,
226 CFQ_CFQQ_FLAG_must_alloc_slice
,
227 CFQ_CFQQ_FLAG_must_dispatch
,
228 CFQ_CFQQ_FLAG_fifo_expire
,
229 CFQ_CFQQ_FLAG_idle_window
,
230 CFQ_CFQQ_FLAG_prio_changed
,
233 #define CFQ_CFQQ_FNS(name) \
234 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
236 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
238 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
240 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
242 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
244 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
248 CFQ_CFQQ_FNS(wait_request
);
249 CFQ_CFQQ_FNS(must_alloc
);
250 CFQ_CFQQ_FNS(must_alloc_slice
);
251 CFQ_CFQQ_FNS(must_dispatch
);
252 CFQ_CFQQ_FNS(fifo_expire
);
253 CFQ_CFQQ_FNS(idle_window
);
254 CFQ_CFQQ_FNS(prio_changed
);
257 enum cfq_rq_state_flags
{
258 CFQ_CRQ_FLAG_is_sync
= 0,
261 #define CFQ_CRQ_FNS(name) \
262 static inline void cfq_mark_crq_##name(struct cfq_rq *crq) \
264 crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name); \
266 static inline void cfq_clear_crq_##name(struct cfq_rq *crq) \
268 crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name); \
270 static inline int cfq_crq_##name(const struct cfq_rq *crq) \
272 return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0; \
275 CFQ_CRQ_FNS(is_sync
);
278 static struct cfq_queue
*cfq_find_cfq_hash(struct cfq_data
*, unsigned int, unsigned short);
279 static void cfq_dispatch_insert(request_queue_t
*, struct cfq_rq
*);
280 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
, gfp_t gfp_mask
);
282 #define process_sync(tsk) ((tsk)->flags & PF_SYNCWRITE)
285 * lots of deadline iosched dupes, can be abstracted later...
287 static inline void cfq_del_crq_hash(struct cfq_rq
*crq
)
289 hlist_del_init(&crq
->hash
);
292 static inline void cfq_add_crq_hash(struct cfq_data
*cfqd
, struct cfq_rq
*crq
)
294 const int hash_idx
= CFQ_MHASH_FN(rq_hash_key(crq
->request
));
296 hlist_add_head(&crq
->hash
, &cfqd
->crq_hash
[hash_idx
]);
299 static struct request
*cfq_find_rq_hash(struct cfq_data
*cfqd
, sector_t offset
)
301 struct hlist_head
*hash_list
= &cfqd
->crq_hash
[CFQ_MHASH_FN(offset
)];
302 struct hlist_node
*entry
, *next
;
304 hlist_for_each_safe(entry
, next
, hash_list
) {
305 struct cfq_rq
*crq
= list_entry_hash(entry
);
306 struct request
*__rq
= crq
->request
;
308 if (!rq_mergeable(__rq
)) {
309 cfq_del_crq_hash(crq
);
313 if (rq_hash_key(__rq
) == offset
)
321 * scheduler run of queue, if there are requests pending and no one in the
322 * driver that will restart queueing
324 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
326 if (cfqd
->busy_queues
)
327 kblockd_schedule_work(&cfqd
->unplug_work
);
330 static int cfq_queue_empty(request_queue_t
*q
)
332 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
334 return !cfqd
->busy_queues
;
337 static inline pid_t
cfq_queue_pid(struct task_struct
*task
, int rw
)
339 if (rw
== READ
|| process_sync(task
))
342 return CFQ_KEY_ASYNC
;
346 * Lifted from AS - choose which of crq1 and crq2 that is best served now.
347 * We choose the request that is closest to the head right now. Distance
348 * behind the head is penalized and only allowed to a certain extent.
350 static struct cfq_rq
*
351 cfq_choose_req(struct cfq_data
*cfqd
, struct cfq_rq
*crq1
, struct cfq_rq
*crq2
)
353 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
354 unsigned long back_max
;
355 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
356 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
357 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
359 if (crq1
== NULL
|| crq1
== crq2
)
364 if (cfq_crq_is_sync(crq1
) && !cfq_crq_is_sync(crq2
))
366 else if (cfq_crq_is_sync(crq2
) && !cfq_crq_is_sync(crq1
))
369 s1
= crq1
->request
->sector
;
370 s2
= crq2
->request
->sector
;
372 last
= cfqd
->last_sector
;
375 * by definition, 1KiB is 2 sectors
377 back_max
= cfqd
->cfq_back_max
* 2;
380 * Strict one way elevator _except_ in the case where we allow
381 * short backward seeks which are biased as twice the cost of a
382 * similar forward seek.
386 else if (s1
+ back_max
>= last
)
387 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
389 wrap
|= CFQ_RQ1_WRAP
;
393 else if (s2
+ back_max
>= last
)
394 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
396 wrap
|= CFQ_RQ2_WRAP
;
398 /* Found required data */
401 * By doing switch() on the bit mask "wrap" we avoid having to
402 * check two variables for all permutations: --> faster!
405 case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
421 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both crqs wrapped */
424 * Since both rqs are wrapped,
425 * start with the one that's further behind head
426 * (--> only *one* back seek required),
427 * since back seek takes more time than forward.
437 * would be nice to take fifo expire time into account as well
439 static struct cfq_rq
*
440 cfq_find_next_crq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
443 struct cfq_rq
*crq_next
= NULL
, *crq_prev
= NULL
;
444 struct rb_node
*rbnext
, *rbprev
;
446 if (!(rbnext
= rb_next(&last
->rb_node
))) {
447 rbnext
= rb_first(&cfqq
->sort_list
);
448 if (rbnext
== &last
->rb_node
)
452 rbprev
= rb_prev(&last
->rb_node
);
455 crq_prev
= rb_entry_crq(rbprev
);
457 crq_next
= rb_entry_crq(rbnext
);
459 return cfq_choose_req(cfqd
, crq_next
, crq_prev
);
462 static void cfq_update_next_crq(struct cfq_rq
*crq
)
464 struct cfq_queue
*cfqq
= crq
->cfq_queue
;
466 if (cfqq
->next_crq
== crq
)
467 cfqq
->next_crq
= cfq_find_next_crq(cfqq
->cfqd
, cfqq
, crq
);
470 static void cfq_resort_rr_list(struct cfq_queue
*cfqq
, int preempted
)
472 struct cfq_data
*cfqd
= cfqq
->cfqd
;
473 struct list_head
*list
, *entry
;
475 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
477 list_del(&cfqq
->cfq_list
);
479 if (cfq_class_rt(cfqq
))
480 list
= &cfqd
->cur_rr
;
481 else if (cfq_class_idle(cfqq
))
482 list
= &cfqd
->idle_rr
;
485 * if cfqq has requests in flight, don't allow it to be
486 * found in cfq_set_active_queue before it has finished them.
487 * this is done to increase fairness between a process that
488 * has lots of io pending vs one that only generates one
489 * sporadically or synchronously
491 if (cfq_cfqq_dispatched(cfqq
))
492 list
= &cfqd
->busy_rr
;
494 list
= &cfqd
->rr_list
[cfqq
->ioprio
];
498 * if queue was preempted, just add to front to be fair. busy_rr
499 * isn't sorted, but insert at the back for fairness.
501 if (preempted
|| list
== &cfqd
->busy_rr
) {
505 list_add_tail(&cfqq
->cfq_list
, list
);
510 * sort by when queue was last serviced
513 while ((entry
= entry
->prev
) != list
) {
514 struct cfq_queue
*__cfqq
= list_entry_cfqq(entry
);
516 if (!__cfqq
->service_last
)
518 if (time_before(__cfqq
->service_last
, cfqq
->service_last
))
522 list_add(&cfqq
->cfq_list
, entry
);
526 * add to busy list of queues for service, trying to be fair in ordering
527 * the pending list according to last request service
530 cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
532 BUG_ON(cfq_cfqq_on_rr(cfqq
));
533 cfq_mark_cfqq_on_rr(cfqq
);
536 cfq_resort_rr_list(cfqq
, 0);
540 cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
542 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
543 cfq_clear_cfqq_on_rr(cfqq
);
544 list_move(&cfqq
->cfq_list
, &cfqd
->empty_list
);
546 BUG_ON(!cfqd
->busy_queues
);
551 * rb tree support functions
553 static inline void cfq_del_crq_rb(struct cfq_rq
*crq
)
555 struct cfq_queue
*cfqq
= crq
->cfq_queue
;
556 struct cfq_data
*cfqd
= cfqq
->cfqd
;
557 const int sync
= cfq_crq_is_sync(crq
);
559 BUG_ON(!cfqq
->queued
[sync
]);
560 cfqq
->queued
[sync
]--;
562 cfq_update_next_crq(crq
);
564 rb_erase(&crq
->rb_node
, &cfqq
->sort_list
);
566 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY(&cfqq
->sort_list
))
567 cfq_del_cfqq_rr(cfqd
, cfqq
);
570 static struct cfq_rq
*
571 __cfq_add_crq_rb(struct cfq_rq
*crq
)
573 struct rb_node
**p
= &crq
->cfq_queue
->sort_list
.rb_node
;
574 struct rb_node
*parent
= NULL
;
575 struct cfq_rq
*__crq
;
579 __crq
= rb_entry_crq(parent
);
581 if (crq
->rb_key
< __crq
->rb_key
)
583 else if (crq
->rb_key
> __crq
->rb_key
)
589 rb_link_node(&crq
->rb_node
, parent
, p
);
593 static void cfq_add_crq_rb(struct cfq_rq
*crq
)
595 struct cfq_queue
*cfqq
= crq
->cfq_queue
;
596 struct cfq_data
*cfqd
= cfqq
->cfqd
;
597 struct request
*rq
= crq
->request
;
598 struct cfq_rq
*__alias
;
600 crq
->rb_key
= rq_rb_key(rq
);
601 cfqq
->queued
[cfq_crq_is_sync(crq
)]++;
604 * looks a little odd, but the first insert might return an alias.
605 * if that happens, put the alias on the dispatch list
607 while ((__alias
= __cfq_add_crq_rb(crq
)) != NULL
)
608 cfq_dispatch_insert(cfqd
->queue
, __alias
);
610 rb_insert_color(&crq
->rb_node
, &cfqq
->sort_list
);
612 if (!cfq_cfqq_on_rr(cfqq
))
613 cfq_add_cfqq_rr(cfqd
, cfqq
);
616 * check if this request is a better next-serve candidate
618 cfqq
->next_crq
= cfq_choose_req(cfqd
, cfqq
->next_crq
, crq
);
622 cfq_reposition_crq_rb(struct cfq_queue
*cfqq
, struct cfq_rq
*crq
)
624 rb_erase(&crq
->rb_node
, &cfqq
->sort_list
);
625 cfqq
->queued
[cfq_crq_is_sync(crq
)]--;
630 static struct request
*
631 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
633 struct task_struct
*tsk
= current
;
634 pid_t key
= cfq_queue_pid(tsk
, bio_data_dir(bio
));
635 struct cfq_queue
*cfqq
;
639 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
643 sector
= bio
->bi_sector
+ bio_sectors(bio
);
644 n
= cfqq
->sort_list
.rb_node
;
646 struct cfq_rq
*crq
= rb_entry_crq(n
);
648 if (sector
< crq
->rb_key
)
650 else if (sector
> crq
->rb_key
)
660 static void cfq_activate_request(request_queue_t
*q
, struct request
*rq
)
662 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
664 cfqd
->rq_in_driver
++;
667 * If the depth is larger 1, it really could be queueing. But lets
668 * make the mark a little higher - idling could still be good for
669 * low queueing, and a low queueing number could also just indicate
670 * a SCSI mid layer like behaviour where limit+1 is often seen.
672 if (!cfqd
->hw_tag
&& cfqd
->rq_in_driver
> 4)
676 static void cfq_deactivate_request(request_queue_t
*q
, struct request
*rq
)
678 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
680 WARN_ON(!cfqd
->rq_in_driver
);
681 cfqd
->rq_in_driver
--;
684 static void cfq_remove_request(struct request
*rq
)
686 struct cfq_rq
*crq
= RQ_DATA(rq
);
688 list_del_init(&rq
->queuelist
);
690 cfq_del_crq_hash(crq
);
694 cfq_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
696 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
697 struct request
*__rq
;
700 __rq
= cfq_find_rq_hash(cfqd
, bio
->bi_sector
);
701 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
702 ret
= ELEVATOR_BACK_MERGE
;
706 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
707 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
708 ret
= ELEVATOR_FRONT_MERGE
;
712 return ELEVATOR_NO_MERGE
;
718 static void cfq_merged_request(request_queue_t
*q
, struct request
*req
)
720 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
721 struct cfq_rq
*crq
= RQ_DATA(req
);
723 cfq_del_crq_hash(crq
);
724 cfq_add_crq_hash(cfqd
, crq
);
726 if (rq_rb_key(req
) != crq
->rb_key
) {
727 struct cfq_queue
*cfqq
= crq
->cfq_queue
;
729 cfq_update_next_crq(crq
);
730 cfq_reposition_crq_rb(cfqq
, crq
);
735 cfq_merged_requests(request_queue_t
*q
, struct request
*rq
,
736 struct request
*next
)
738 cfq_merged_request(q
, rq
);
741 * reposition in fifo if next is older than rq
743 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
744 time_before(next
->start_time
, rq
->start_time
))
745 list_move(&rq
->queuelist
, &next
->queuelist
);
747 cfq_remove_request(next
);
751 __cfq_set_active_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
755 * stop potential idle class queues waiting service
757 del_timer(&cfqd
->idle_class_timer
);
759 cfqq
->slice_start
= jiffies
;
761 cfqq
->slice_left
= 0;
762 cfq_clear_cfqq_must_alloc_slice(cfqq
);
763 cfq_clear_cfqq_fifo_expire(cfqq
);
766 cfqd
->active_queue
= cfqq
;
770 * current cfqq expired its slice (or was too idle), select new one
773 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
776 unsigned long now
= jiffies
;
778 if (cfq_cfqq_wait_request(cfqq
))
779 del_timer(&cfqd
->idle_slice_timer
);
781 if (!preempted
&& !cfq_cfqq_dispatched(cfqq
)) {
782 cfqq
->service_last
= now
;
783 cfq_schedule_dispatch(cfqd
);
786 cfq_clear_cfqq_must_dispatch(cfqq
);
787 cfq_clear_cfqq_wait_request(cfqq
);
790 * store what was left of this slice, if the queue idled out
793 if (time_after(cfqq
->slice_end
, now
))
794 cfqq
->slice_left
= cfqq
->slice_end
- now
;
796 cfqq
->slice_left
= 0;
798 if (cfq_cfqq_on_rr(cfqq
))
799 cfq_resort_rr_list(cfqq
, preempted
);
801 if (cfqq
== cfqd
->active_queue
)
802 cfqd
->active_queue
= NULL
;
804 if (cfqd
->active_cic
) {
805 put_io_context(cfqd
->active_cic
->ioc
);
806 cfqd
->active_cic
= NULL
;
809 cfqd
->dispatch_slice
= 0;
812 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, int preempted
)
814 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
817 __cfq_slice_expired(cfqd
, cfqq
, preempted
);
830 static int cfq_get_next_prio_level(struct cfq_data
*cfqd
)
839 for (p
= cfqd
->cur_prio
; p
<= cfqd
->cur_end_prio
; p
++) {
840 if (!list_empty(&cfqd
->rr_list
[p
])) {
849 if (++cfqd
->cur_end_prio
== CFQ_PRIO_LISTS
) {
850 cfqd
->cur_end_prio
= 0;
857 if (unlikely(prio
== -1))
860 BUG_ON(prio
>= CFQ_PRIO_LISTS
);
862 list_splice_init(&cfqd
->rr_list
[prio
], &cfqd
->cur_rr
);
864 cfqd
->cur_prio
= prio
+ 1;
865 if (cfqd
->cur_prio
> cfqd
->cur_end_prio
) {
866 cfqd
->cur_end_prio
= cfqd
->cur_prio
;
869 if (cfqd
->cur_end_prio
== CFQ_PRIO_LISTS
) {
871 cfqd
->cur_end_prio
= 0;
877 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
)
879 struct cfq_queue
*cfqq
= NULL
;
882 * if current list is non-empty, grab first entry. if it is empty,
883 * get next prio level and grab first entry then if any are spliced
885 if (!list_empty(&cfqd
->cur_rr
) || cfq_get_next_prio_level(cfqd
) != -1)
886 cfqq
= list_entry_cfqq(cfqd
->cur_rr
.next
);
889 * If no new queues are available, check if the busy list has some
890 * before falling back to idle io.
892 if (!cfqq
&& !list_empty(&cfqd
->busy_rr
))
893 cfqq
= list_entry_cfqq(cfqd
->busy_rr
.next
);
896 * if we have idle queues and no rt or be queues had pending
897 * requests, either allow immediate service if the grace period
898 * has passed or arm the idle grace timer
900 if (!cfqq
&& !list_empty(&cfqd
->idle_rr
)) {
901 unsigned long end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
903 if (time_after_eq(jiffies
, end
))
904 cfqq
= list_entry_cfqq(cfqd
->idle_rr
.next
);
906 mod_timer(&cfqd
->idle_class_timer
, end
);
909 __cfq_set_active_queue(cfqd
, cfqq
);
913 static int cfq_arm_slice_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
916 struct cfq_io_context
*cic
;
919 WARN_ON(!RB_EMPTY(&cfqq
->sort_list
));
920 WARN_ON(cfqq
!= cfqd
->active_queue
);
923 * idle is disabled, either manually or by past process history
925 if (!cfqd
->cfq_slice_idle
)
927 if (!cfq_cfqq_idle_window(cfqq
))
930 * task has exited, don't wait
932 cic
= cfqd
->active_cic
;
933 if (!cic
|| !cic
->ioc
->task
)
936 cfq_mark_cfqq_must_dispatch(cfqq
);
937 cfq_mark_cfqq_wait_request(cfqq
);
939 sl
= min(cfqq
->slice_end
- 1, (unsigned long) cfqd
->cfq_slice_idle
);
942 * we don't want to idle for seeks, but we do want to allow
943 * fair distribution of slice time for a process doing back-to-back
944 * seeks. so allow a little bit of time for him to submit a new rq
946 if (sample_valid(cic
->seek_samples
) && cic
->seek_mean
> 131072)
949 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
953 static void cfq_dispatch_insert(request_queue_t
*q
, struct cfq_rq
*crq
)
955 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
956 struct cfq_queue
*cfqq
= crq
->cfq_queue
;
958 cfqq
->next_crq
= cfq_find_next_crq(cfqd
, cfqq
, crq
);
959 cfq_remove_request(crq
->request
);
960 cfqq
->on_dispatch
[cfq_crq_is_sync(crq
)]++;
961 elv_dispatch_sort(q
, crq
->request
);
965 * return expired entry, or NULL to just start from scratch in rbtree
967 static inline struct cfq_rq
*cfq_check_fifo(struct cfq_queue
*cfqq
)
969 struct cfq_data
*cfqd
= cfqq
->cfqd
;
973 if (cfq_cfqq_fifo_expire(cfqq
))
976 if (!list_empty(&cfqq
->fifo
)) {
977 int fifo
= cfq_cfqq_class_sync(cfqq
);
979 crq
= RQ_DATA(list_entry_fifo(cfqq
->fifo
.next
));
981 if (time_after(jiffies
, rq
->start_time
+ cfqd
->cfq_fifo_expire
[fifo
])) {
982 cfq_mark_cfqq_fifo_expire(cfqq
);
991 * Scale schedule slice based on io priority. Use the sync time slice only
992 * if a queue is marked sync and has sync io queued. A sync queue with async
993 * io only, should not get full sync slice length.
996 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
998 const int base_slice
= cfqd
->cfq_slice
[cfq_cfqq_sync(cfqq
)];
1000 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1002 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - cfqq
->ioprio
));
1006 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1008 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
1012 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1014 const int base_rq
= cfqd
->cfq_slice_async_rq
;
1016 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1018 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
1022 * get next queue for service
1024 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
1026 unsigned long now
= jiffies
;
1027 struct cfq_queue
*cfqq
;
1029 cfqq
= cfqd
->active_queue
;
1036 if (!cfq_cfqq_must_dispatch(cfqq
) && time_after(now
, cfqq
->slice_end
))
1040 * if queue has requests, dispatch one. if not, check if
1041 * enough slice is left to wait for one
1043 if (!RB_EMPTY(&cfqq
->sort_list
))
1045 else if (cfq_cfqq_class_sync(cfqq
) &&
1046 time_before(now
, cfqq
->slice_end
)) {
1047 if (cfq_arm_slice_timer(cfqd
, cfqq
))
1052 cfq_slice_expired(cfqd
, 0);
1054 cfqq
= cfq_set_active_queue(cfqd
);
1060 __cfq_dispatch_requests(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1065 BUG_ON(RB_EMPTY(&cfqq
->sort_list
));
1071 * follow expired path, else get first next available
1073 if ((crq
= cfq_check_fifo(cfqq
)) == NULL
)
1074 crq
= cfqq
->next_crq
;
1077 * finally, insert request into driver dispatch list
1079 cfq_dispatch_insert(cfqd
->queue
, crq
);
1081 cfqd
->dispatch_slice
++;
1084 if (!cfqd
->active_cic
) {
1085 atomic_inc(&crq
->io_context
->ioc
->refcount
);
1086 cfqd
->active_cic
= crq
->io_context
;
1089 if (RB_EMPTY(&cfqq
->sort_list
))
1092 } while (dispatched
< max_dispatch
);
1095 * if slice end isn't set yet, set it. if at least one request was
1096 * sync, use the sync time slice value
1098 if (!cfqq
->slice_end
)
1099 cfq_set_prio_slice(cfqd
, cfqq
);
1102 * expire an async queue immediately if it has used up its slice. idle
1103 * queue always expire after 1 dispatch round.
1105 if ((!cfq_cfqq_sync(cfqq
) &&
1106 cfqd
->dispatch_slice
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
1107 cfq_class_idle(cfqq
))
1108 cfq_slice_expired(cfqd
, 0);
1114 cfq_forced_dispatch_cfqqs(struct list_head
*list
)
1117 struct cfq_queue
*cfqq
, *next
;
1120 list_for_each_entry_safe(cfqq
, next
, list
, cfq_list
) {
1121 while ((crq
= cfqq
->next_crq
)) {
1122 cfq_dispatch_insert(cfqq
->cfqd
->queue
, crq
);
1125 BUG_ON(!list_empty(&cfqq
->fifo
));
1131 cfq_forced_dispatch(struct cfq_data
*cfqd
)
1133 int i
, dispatched
= 0;
1135 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
1136 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->rr_list
[i
]);
1138 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->busy_rr
);
1139 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->cur_rr
);
1140 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->idle_rr
);
1142 cfq_slice_expired(cfqd
, 0);
1144 BUG_ON(cfqd
->busy_queues
);
1150 cfq_dispatch_requests(request_queue_t
*q
, int force
)
1152 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1153 struct cfq_queue
*cfqq
;
1155 if (!cfqd
->busy_queues
)
1158 if (unlikely(force
))
1159 return cfq_forced_dispatch(cfqd
);
1161 cfqq
= cfq_select_queue(cfqd
);
1165 cfq_clear_cfqq_must_dispatch(cfqq
);
1166 cfq_clear_cfqq_wait_request(cfqq
);
1167 del_timer(&cfqd
->idle_slice_timer
);
1169 max_dispatch
= cfqd
->cfq_quantum
;
1170 if (cfq_class_idle(cfqq
))
1173 return __cfq_dispatch_requests(cfqd
, cfqq
, max_dispatch
);
1180 * task holds one reference to the queue, dropped when task exits. each crq
1181 * in-flight on this queue also holds a reference, dropped when crq is freed.
1183 * queue lock must be held here.
1185 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1187 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1189 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1191 if (!atomic_dec_and_test(&cfqq
->ref
))
1194 BUG_ON(rb_first(&cfqq
->sort_list
));
1195 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1196 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1198 if (unlikely(cfqd
->active_queue
== cfqq
))
1199 __cfq_slice_expired(cfqd
, cfqq
, 0);
1202 * it's on the empty list and still hashed
1204 list_del(&cfqq
->cfq_list
);
1205 hlist_del(&cfqq
->cfq_hash
);
1206 kmem_cache_free(cfq_pool
, cfqq
);
1209 static inline struct cfq_queue
*
1210 __cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned int prio
,
1213 struct hlist_head
*hash_list
= &cfqd
->cfq_hash
[hashval
];
1214 struct hlist_node
*entry
;
1215 struct cfq_queue
*__cfqq
;
1217 hlist_for_each_entry(__cfqq
, entry
, hash_list
, cfq_hash
) {
1218 const unsigned short __p
= IOPRIO_PRIO_VALUE(__cfqq
->org_ioprio_class
, __cfqq
->org_ioprio
);
1220 if (__cfqq
->key
== key
&& (__p
== prio
|| !prio
))
1227 static struct cfq_queue
*
1228 cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned short prio
)
1230 return __cfq_find_cfq_hash(cfqd
, key
, prio
, hash_long(key
, CFQ_QHASH_SHIFT
));
1233 static void cfq_free_io_context(struct io_context
*ioc
)
1235 struct cfq_io_context
*__cic
;
1239 while ((n
= rb_first(&ioc
->cic_root
)) != NULL
) {
1240 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1241 rb_erase(&__cic
->rb_node
, &ioc
->cic_root
);
1242 kmem_cache_free(cfq_ioc_pool
, __cic
);
1246 if (atomic_sub_and_test(freed
, &ioc_count
) && ioc_gone
)
1250 static void cfq_trim(struct io_context
*ioc
)
1252 ioc
->set_ioprio
= NULL
;
1253 cfq_free_io_context(ioc
);
1257 * Called with interrupts disabled
1259 static void cfq_exit_single_io_context(struct cfq_io_context
*cic
)
1261 struct cfq_data
*cfqd
= cic
->key
;
1269 WARN_ON(!irqs_disabled());
1271 spin_lock(q
->queue_lock
);
1273 if (cic
->cfqq
[ASYNC
]) {
1274 if (unlikely(cic
->cfqq
[ASYNC
] == cfqd
->active_queue
))
1275 __cfq_slice_expired(cfqd
, cic
->cfqq
[ASYNC
], 0);
1276 cfq_put_queue(cic
->cfqq
[ASYNC
]);
1277 cic
->cfqq
[ASYNC
] = NULL
;
1280 if (cic
->cfqq
[SYNC
]) {
1281 if (unlikely(cic
->cfqq
[SYNC
] == cfqd
->active_queue
))
1282 __cfq_slice_expired(cfqd
, cic
->cfqq
[SYNC
], 0);
1283 cfq_put_queue(cic
->cfqq
[SYNC
]);
1284 cic
->cfqq
[SYNC
] = NULL
;
1288 list_del_init(&cic
->queue_list
);
1289 spin_unlock(q
->queue_lock
);
1292 static void cfq_exit_io_context(struct io_context
*ioc
)
1294 struct cfq_io_context
*__cic
;
1295 unsigned long flags
;
1299 * put the reference this task is holding to the various queues
1301 spin_lock_irqsave(&cfq_exit_lock
, flags
);
1303 n
= rb_first(&ioc
->cic_root
);
1305 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1307 cfq_exit_single_io_context(__cic
);
1311 spin_unlock_irqrestore(&cfq_exit_lock
, flags
);
1314 static struct cfq_io_context
*
1315 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1317 struct cfq_io_context
*cic
= kmem_cache_alloc(cfq_ioc_pool
, gfp_mask
);
1320 memset(cic
, 0, sizeof(*cic
));
1321 cic
->last_end_request
= jiffies
;
1322 INIT_LIST_HEAD(&cic
->queue_list
);
1323 cic
->dtor
= cfq_free_io_context
;
1324 cic
->exit
= cfq_exit_io_context
;
1325 atomic_inc(&ioc_count
);
1331 static void cfq_init_prio_data(struct cfq_queue
*cfqq
)
1333 struct task_struct
*tsk
= current
;
1336 if (!cfq_cfqq_prio_changed(cfqq
))
1339 ioprio_class
= IOPRIO_PRIO_CLASS(tsk
->ioprio
);
1340 switch (ioprio_class
) {
1342 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1343 case IOPRIO_CLASS_NONE
:
1345 * no prio set, place us in the middle of the BE classes
1347 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1348 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1350 case IOPRIO_CLASS_RT
:
1351 cfqq
->ioprio
= task_ioprio(tsk
);
1352 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1354 case IOPRIO_CLASS_BE
:
1355 cfqq
->ioprio
= task_ioprio(tsk
);
1356 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1358 case IOPRIO_CLASS_IDLE
:
1359 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1361 cfq_clear_cfqq_idle_window(cfqq
);
1366 * keep track of original prio settings in case we have to temporarily
1367 * elevate the priority of this queue
1369 cfqq
->org_ioprio
= cfqq
->ioprio
;
1370 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1372 if (cfq_cfqq_on_rr(cfqq
))
1373 cfq_resort_rr_list(cfqq
, 0);
1375 cfq_clear_cfqq_prio_changed(cfqq
);
1378 static inline void changed_ioprio(struct cfq_io_context
*cic
)
1380 struct cfq_data
*cfqd
= cic
->key
;
1381 struct cfq_queue
*cfqq
;
1383 spin_lock(cfqd
->queue
->queue_lock
);
1384 cfqq
= cic
->cfqq
[ASYNC
];
1386 struct cfq_queue
*new_cfqq
;
1387 new_cfqq
= cfq_get_queue(cfqd
, CFQ_KEY_ASYNC
,
1388 cic
->ioc
->task
, GFP_ATOMIC
);
1390 cic
->cfqq
[ASYNC
] = new_cfqq
;
1391 cfq_put_queue(cfqq
);
1394 cfqq
= cic
->cfqq
[SYNC
];
1396 cfq_mark_cfqq_prio_changed(cfqq
);
1397 cfq_init_prio_data(cfqq
);
1399 spin_unlock(cfqd
->queue
->queue_lock
);
1404 * callback from sys_ioprio_set, irqs are disabled
1406 static int cfq_ioc_set_ioprio(struct io_context
*ioc
, unsigned int ioprio
)
1408 struct cfq_io_context
*cic
;
1411 spin_lock(&cfq_exit_lock
);
1413 n
= rb_first(&ioc
->cic_root
);
1415 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1417 changed_ioprio(cic
);
1421 spin_unlock(&cfq_exit_lock
);
1426 static struct cfq_queue
*
1427 cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
,
1430 const int hashval
= hash_long(key
, CFQ_QHASH_SHIFT
);
1431 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1432 unsigned short ioprio
;
1435 ioprio
= tsk
->ioprio
;
1436 cfqq
= __cfq_find_cfq_hash(cfqd
, key
, ioprio
, hashval
);
1442 } else if (gfp_mask
& __GFP_WAIT
) {
1443 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1444 new_cfqq
= kmem_cache_alloc(cfq_pool
, gfp_mask
);
1445 spin_lock_irq(cfqd
->queue
->queue_lock
);
1448 cfqq
= kmem_cache_alloc(cfq_pool
, gfp_mask
);
1453 memset(cfqq
, 0, sizeof(*cfqq
));
1455 INIT_HLIST_NODE(&cfqq
->cfq_hash
);
1456 INIT_LIST_HEAD(&cfqq
->cfq_list
);
1457 RB_CLEAR_ROOT(&cfqq
->sort_list
);
1458 INIT_LIST_HEAD(&cfqq
->fifo
);
1461 hlist_add_head(&cfqq
->cfq_hash
, &cfqd
->cfq_hash
[hashval
]);
1462 atomic_set(&cfqq
->ref
, 0);
1464 cfqq
->service_last
= 0;
1466 * set ->slice_left to allow preemption for a new process
1468 cfqq
->slice_left
= 2 * cfqd
->cfq_slice_idle
;
1470 cfq_mark_cfqq_idle_window(cfqq
);
1471 cfq_mark_cfqq_prio_changed(cfqq
);
1472 cfq_init_prio_data(cfqq
);
1476 kmem_cache_free(cfq_pool
, new_cfqq
);
1478 atomic_inc(&cfqq
->ref
);
1480 WARN_ON((gfp_mask
& __GFP_WAIT
) && !cfqq
);
1485 cfq_drop_dead_cic(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1487 spin_lock(&cfq_exit_lock
);
1488 rb_erase(&cic
->rb_node
, &ioc
->cic_root
);
1489 list_del_init(&cic
->queue_list
);
1490 spin_unlock(&cfq_exit_lock
);
1491 kmem_cache_free(cfq_ioc_pool
, cic
);
1492 atomic_dec(&ioc_count
);
1495 static struct cfq_io_context
*
1496 cfq_cic_rb_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1499 struct cfq_io_context
*cic
;
1500 void *k
, *key
= cfqd
;
1503 n
= ioc
->cic_root
.rb_node
;
1505 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1506 /* ->key must be copied to avoid race with cfq_exit_queue() */
1509 cfq_drop_dead_cic(ioc
, cic
);
1525 cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1526 struct cfq_io_context
*cic
)
1529 struct rb_node
*parent
;
1530 struct cfq_io_context
*__cic
;
1536 ioc
->set_ioprio
= cfq_ioc_set_ioprio
;
1539 p
= &ioc
->cic_root
.rb_node
;
1542 __cic
= rb_entry(parent
, struct cfq_io_context
, rb_node
);
1543 /* ->key must be copied to avoid race with cfq_exit_queue() */
1546 cfq_drop_dead_cic(ioc
, cic
);
1552 else if (cic
->key
> k
)
1553 p
= &(*p
)->rb_right
;
1558 spin_lock(&cfq_exit_lock
);
1559 rb_link_node(&cic
->rb_node
, parent
, p
);
1560 rb_insert_color(&cic
->rb_node
, &ioc
->cic_root
);
1561 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1562 spin_unlock(&cfq_exit_lock
);
1566 * Setup general io context and cfq io context. There can be several cfq
1567 * io contexts per general io context, if this process is doing io to more
1568 * than one device managed by cfq.
1570 static struct cfq_io_context
*
1571 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1573 struct io_context
*ioc
= NULL
;
1574 struct cfq_io_context
*cic
;
1576 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1578 ioc
= get_io_context(gfp_mask
);
1582 cic
= cfq_cic_rb_lookup(cfqd
, ioc
);
1586 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1590 cfq_cic_link(cfqd
, ioc
, cic
);
1594 put_io_context(ioc
);
1599 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1601 unsigned long elapsed
, ttime
;
1604 * if this context already has stuff queued, thinktime is from
1605 * last queue not last end
1608 if (time_after(cic
->last_end_request
, cic
->last_queue
))
1609 elapsed
= jiffies
- cic
->last_end_request
;
1611 elapsed
= jiffies
- cic
->last_queue
;
1613 elapsed
= jiffies
- cic
->last_end_request
;
1616 ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1618 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1619 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1620 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1624 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
1630 if (cic
->last_request_pos
< crq
->request
->sector
)
1631 sdist
= crq
->request
->sector
- cic
->last_request_pos
;
1633 sdist
= cic
->last_request_pos
- crq
->request
->sector
;
1636 * Don't allow the seek distance to get too large from the
1637 * odd fragment, pagein, etc
1639 if (cic
->seek_samples
<= 60) /* second&third seek */
1640 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1642 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1644 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1645 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1646 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1647 do_div(total
, cic
->seek_samples
);
1648 cic
->seek_mean
= (sector_t
)total
;
1652 * Disable idle window if the process thinks too long or seeks so much that
1656 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1657 struct cfq_io_context
*cic
)
1659 int enable_idle
= cfq_cfqq_idle_window(cfqq
);
1661 if (!cic
->ioc
->task
|| !cfqd
->cfq_slice_idle
|| cfqd
->hw_tag
)
1663 else if (sample_valid(cic
->ttime_samples
)) {
1664 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
1671 cfq_mark_cfqq_idle_window(cfqq
);
1673 cfq_clear_cfqq_idle_window(cfqq
);
1678 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1679 * no or if we aren't sure, a 1 will cause a preempt.
1682 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
1685 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1687 if (cfq_class_idle(new_cfqq
))
1693 if (cfq_class_idle(cfqq
))
1695 if (!cfq_cfqq_wait_request(new_cfqq
))
1698 * if it doesn't have slice left, forget it
1700 if (new_cfqq
->slice_left
< cfqd
->cfq_slice_idle
)
1702 if (cfq_crq_is_sync(crq
) && !cfq_cfqq_sync(cfqq
))
1709 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1710 * let it have half of its nominal slice.
1712 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1714 struct cfq_queue
*__cfqq
, *next
;
1716 list_for_each_entry_safe(__cfqq
, next
, &cfqd
->cur_rr
, cfq_list
)
1717 cfq_resort_rr_list(__cfqq
, 1);
1719 if (!cfqq
->slice_left
)
1720 cfqq
->slice_left
= cfq_prio_to_slice(cfqd
, cfqq
) / 2;
1722 cfqq
->slice_end
= cfqq
->slice_left
+ jiffies
;
1723 __cfq_slice_expired(cfqd
, cfqq
, 1);
1724 __cfq_set_active_queue(cfqd
, cfqq
);
1728 * should really be a ll_rw_blk.c helper
1730 static void cfq_start_queueing(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1732 request_queue_t
*q
= cfqd
->queue
;
1734 if (!blk_queue_plugged(q
))
1737 __generic_unplug_device(q
);
1741 * Called when a new fs request (crq) is added (to cfqq). Check if there's
1742 * something we should do about it
1745 cfq_crq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1748 struct cfq_io_context
*cic
;
1750 cfqq
->next_crq
= cfq_choose_req(cfqd
, cfqq
->next_crq
, crq
);
1752 cic
= crq
->io_context
;
1755 * we never wait for an async request and we don't allow preemption
1756 * of an async request. so just return early
1758 if (!cfq_crq_is_sync(crq
)) {
1760 * sync process issued an async request, if it's waiting
1761 * then expire it and kick rq handling.
1763 if (cic
== cfqd
->active_cic
&&
1764 del_timer(&cfqd
->idle_slice_timer
)) {
1765 cfq_slice_expired(cfqd
, 0);
1766 cfq_start_queueing(cfqd
, cfqq
);
1771 cfq_update_io_thinktime(cfqd
, cic
);
1772 cfq_update_io_seektime(cfqd
, cic
, crq
);
1773 cfq_update_idle_window(cfqd
, cfqq
, cic
);
1775 cic
->last_queue
= jiffies
;
1776 cic
->last_request_pos
= crq
->request
->sector
+ crq
->request
->nr_sectors
;
1778 if (cfqq
== cfqd
->active_queue
) {
1780 * if we are waiting for a request for this queue, let it rip
1781 * immediately and flag that we must not expire this queue
1784 if (cfq_cfqq_wait_request(cfqq
)) {
1785 cfq_mark_cfqq_must_dispatch(cfqq
);
1786 del_timer(&cfqd
->idle_slice_timer
);
1787 cfq_start_queueing(cfqd
, cfqq
);
1789 } else if (cfq_should_preempt(cfqd
, cfqq
, crq
)) {
1791 * not the active queue - expire current slice if it is
1792 * idle and has expired it's mean thinktime or this new queue
1793 * has some old slice time left and is of higher priority
1795 cfq_preempt_queue(cfqd
, cfqq
);
1796 cfq_mark_cfqq_must_dispatch(cfqq
);
1797 cfq_start_queueing(cfqd
, cfqq
);
1801 static void cfq_insert_request(request_queue_t
*q
, struct request
*rq
)
1803 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1804 struct cfq_rq
*crq
= RQ_DATA(rq
);
1805 struct cfq_queue
*cfqq
= crq
->cfq_queue
;
1807 cfq_init_prio_data(cfqq
);
1809 cfq_add_crq_rb(crq
);
1811 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
1813 if (rq_mergeable(rq
))
1814 cfq_add_crq_hash(cfqd
, crq
);
1816 cfq_crq_enqueued(cfqd
, cfqq
, crq
);
1819 static void cfq_completed_request(request_queue_t
*q
, struct request
*rq
)
1821 struct cfq_rq
*crq
= RQ_DATA(rq
);
1822 struct cfq_queue
*cfqq
= crq
->cfq_queue
;
1823 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1824 const int sync
= cfq_crq_is_sync(crq
);
1829 WARN_ON(!cfqd
->rq_in_driver
);
1830 WARN_ON(!cfqq
->on_dispatch
[sync
]);
1831 cfqd
->rq_in_driver
--;
1832 cfqq
->on_dispatch
[sync
]--;
1834 if (!cfq_class_idle(cfqq
))
1835 cfqd
->last_end_request
= now
;
1837 if (!cfq_cfqq_dispatched(cfqq
)) {
1838 if (cfq_cfqq_on_rr(cfqq
)) {
1839 cfqq
->service_last
= now
;
1840 cfq_resort_rr_list(cfqq
, 0);
1842 cfq_schedule_dispatch(cfqd
);
1845 if (cfq_crq_is_sync(crq
))
1846 crq
->io_context
->last_end_request
= now
;
1849 static struct request
*
1850 cfq_former_request(request_queue_t
*q
, struct request
*rq
)
1852 struct cfq_rq
*crq
= RQ_DATA(rq
);
1853 struct rb_node
*rbprev
= rb_prev(&crq
->rb_node
);
1856 return rb_entry_crq(rbprev
)->request
;
1861 static struct request
*
1862 cfq_latter_request(request_queue_t
*q
, struct request
*rq
)
1864 struct cfq_rq
*crq
= RQ_DATA(rq
);
1865 struct rb_node
*rbnext
= rb_next(&crq
->rb_node
);
1868 return rb_entry_crq(rbnext
)->request
;
1874 * we temporarily boost lower priority queues if they are holding fs exclusive
1875 * resources. they are boosted to normal prio (CLASS_BE/4)
1877 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
1879 const int ioprio_class
= cfqq
->ioprio_class
;
1880 const int ioprio
= cfqq
->ioprio
;
1882 if (has_fs_excl()) {
1884 * boost idle prio on transactions that would lock out other
1885 * users of the filesystem
1887 if (cfq_class_idle(cfqq
))
1888 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1889 if (cfqq
->ioprio
> IOPRIO_NORM
)
1890 cfqq
->ioprio
= IOPRIO_NORM
;
1893 * check if we need to unboost the queue
1895 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
1896 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
1897 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
1898 cfqq
->ioprio
= cfqq
->org_ioprio
;
1902 * refile between round-robin lists if we moved the priority class
1904 if ((ioprio_class
!= cfqq
->ioprio_class
|| ioprio
!= cfqq
->ioprio
) &&
1905 cfq_cfqq_on_rr(cfqq
))
1906 cfq_resort_rr_list(cfqq
, 0);
1910 __cfq_may_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1911 struct task_struct
*task
, int rw
)
1914 if ((cfq_cfqq_wait_request(cfqq
) || cfq_cfqq_must_alloc(cfqq
)) &&
1915 !cfq_cfqq_must_alloc_slice(cfqq
)) {
1916 cfq_mark_cfqq_must_alloc_slice(cfqq
);
1917 return ELV_MQUEUE_MUST
;
1920 return ELV_MQUEUE_MAY
;
1922 if (!cfqq
|| task
->flags
& PF_MEMALLOC
)
1923 return ELV_MQUEUE_MAY
;
1924 if (!cfqq
->allocated
[rw
] || cfq_cfqq_must_alloc(cfqq
)) {
1925 if (cfq_cfqq_wait_request(cfqq
))
1926 return ELV_MQUEUE_MUST
;
1929 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1930 * can quickly flood the queue with writes from a single task
1932 if (rw
== READ
|| !cfq_cfqq_must_alloc_slice(cfqq
)) {
1933 cfq_mark_cfqq_must_alloc_slice(cfqq
);
1934 return ELV_MQUEUE_MUST
;
1937 return ELV_MQUEUE_MAY
;
1939 if (cfq_class_idle(cfqq
))
1940 return ELV_MQUEUE_NO
;
1941 if (cfqq
->allocated
[rw
] >= cfqd
->max_queued
) {
1942 struct io_context
*ioc
= get_io_context(GFP_ATOMIC
);
1943 int ret
= ELV_MQUEUE_NO
;
1945 if (ioc
&& ioc
->nr_batch_requests
)
1946 ret
= ELV_MQUEUE_MAY
;
1948 put_io_context(ioc
);
1952 return ELV_MQUEUE_MAY
;
1956 static int cfq_may_queue(request_queue_t
*q
, int rw
, struct bio
*bio
)
1958 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1959 struct task_struct
*tsk
= current
;
1960 struct cfq_queue
*cfqq
;
1963 * don't force setup of a queue from here, as a call to may_queue
1964 * does not necessarily imply that a request actually will be queued.
1965 * so just lookup a possibly existing queue, or return 'may queue'
1968 cfqq
= cfq_find_cfq_hash(cfqd
, cfq_queue_pid(tsk
, rw
), tsk
->ioprio
);
1970 cfq_init_prio_data(cfqq
);
1971 cfq_prio_boost(cfqq
);
1973 return __cfq_may_queue(cfqd
, cfqq
, tsk
, rw
);
1976 return ELV_MQUEUE_MAY
;
1979 static void cfq_check_waiters(request_queue_t
*q
, struct cfq_queue
*cfqq
)
1981 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1982 struct request_list
*rl
= &q
->rq
;
1984 if (cfqq
->allocated
[READ
] <= cfqd
->max_queued
|| cfqd
->rq_starved
) {
1986 if (waitqueue_active(&rl
->wait
[READ
]))
1987 wake_up(&rl
->wait
[READ
]);
1990 if (cfqq
->allocated
[WRITE
] <= cfqd
->max_queued
|| cfqd
->rq_starved
) {
1992 if (waitqueue_active(&rl
->wait
[WRITE
]))
1993 wake_up(&rl
->wait
[WRITE
]);
1998 * queue lock held here
2000 static void cfq_put_request(request_queue_t
*q
, struct request
*rq
)
2002 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2003 struct cfq_rq
*crq
= RQ_DATA(rq
);
2006 struct cfq_queue
*cfqq
= crq
->cfq_queue
;
2007 const int rw
= rq_data_dir(rq
);
2009 BUG_ON(!cfqq
->allocated
[rw
]);
2010 cfqq
->allocated
[rw
]--;
2012 put_io_context(crq
->io_context
->ioc
);
2014 mempool_free(crq
, cfqd
->crq_pool
);
2015 rq
->elevator_private
= NULL
;
2017 cfq_check_waiters(q
, cfqq
);
2018 cfq_put_queue(cfqq
);
2023 * Allocate cfq data structures associated with this request.
2026 cfq_set_request(request_queue_t
*q
, struct request
*rq
, struct bio
*bio
,
2029 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2030 struct task_struct
*tsk
= current
;
2031 struct cfq_io_context
*cic
;
2032 const int rw
= rq_data_dir(rq
);
2033 pid_t key
= cfq_queue_pid(tsk
, rw
);
2034 struct cfq_queue
*cfqq
;
2036 unsigned long flags
;
2037 int is_sync
= key
!= CFQ_KEY_ASYNC
;
2039 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2041 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
2043 spin_lock_irqsave(q
->queue_lock
, flags
);
2048 if (!cic
->cfqq
[is_sync
]) {
2049 cfqq
= cfq_get_queue(cfqd
, key
, tsk
, gfp_mask
);
2053 cic
->cfqq
[is_sync
] = cfqq
;
2055 cfqq
= cic
->cfqq
[is_sync
];
2057 cfqq
->allocated
[rw
]++;
2058 cfq_clear_cfqq_must_alloc(cfqq
);
2059 cfqd
->rq_starved
= 0;
2060 atomic_inc(&cfqq
->ref
);
2061 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2063 crq
= mempool_alloc(cfqd
->crq_pool
, gfp_mask
);
2065 RB_CLEAR(&crq
->rb_node
);
2068 INIT_HLIST_NODE(&crq
->hash
);
2069 crq
->cfq_queue
= cfqq
;
2070 crq
->io_context
= cic
;
2073 cfq_mark_crq_is_sync(crq
);
2075 cfq_clear_crq_is_sync(crq
);
2077 rq
->elevator_private
= crq
;
2081 spin_lock_irqsave(q
->queue_lock
, flags
);
2082 cfqq
->allocated
[rw
]--;
2083 if (!(cfqq
->allocated
[0] + cfqq
->allocated
[1]))
2084 cfq_mark_cfqq_must_alloc(cfqq
);
2085 cfq_put_queue(cfqq
);
2088 put_io_context(cic
->ioc
);
2090 * mark us rq allocation starved. we need to kickstart the process
2091 * ourselves if there are no pending requests that can do it for us.
2092 * that would be an extremely rare OOM situation
2094 cfqd
->rq_starved
= 1;
2095 cfq_schedule_dispatch(cfqd
);
2096 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2100 static void cfq_kick_queue(void *data
)
2102 request_queue_t
*q
= data
;
2103 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2104 unsigned long flags
;
2106 spin_lock_irqsave(q
->queue_lock
, flags
);
2108 if (cfqd
->rq_starved
) {
2109 struct request_list
*rl
= &q
->rq
;
2112 * we aren't guaranteed to get a request after this, but we
2113 * have to be opportunistic
2116 if (waitqueue_active(&rl
->wait
[READ
]))
2117 wake_up(&rl
->wait
[READ
]);
2118 if (waitqueue_active(&rl
->wait
[WRITE
]))
2119 wake_up(&rl
->wait
[WRITE
]);
2124 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2128 * Timer running if the active_queue is currently idling inside its time slice
2130 static void cfq_idle_slice_timer(unsigned long data
)
2132 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
2133 struct cfq_queue
*cfqq
;
2134 unsigned long flags
;
2136 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2138 if ((cfqq
= cfqd
->active_queue
) != NULL
) {
2139 unsigned long now
= jiffies
;
2144 if (time_after(now
, cfqq
->slice_end
))
2148 * only expire and reinvoke request handler, if there are
2149 * other queues with pending requests
2151 if (!cfqd
->busy_queues
) {
2152 cfqd
->idle_slice_timer
.expires
= min(now
+ cfqd
->cfq_slice_idle
, cfqq
->slice_end
);
2153 add_timer(&cfqd
->idle_slice_timer
);
2158 * not expired and it has a request pending, let it dispatch
2160 if (!RB_EMPTY(&cfqq
->sort_list
)) {
2161 cfq_mark_cfqq_must_dispatch(cfqq
);
2166 cfq_slice_expired(cfqd
, 0);
2168 cfq_schedule_dispatch(cfqd
);
2170 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2174 * Timer running if an idle class queue is waiting for service
2176 static void cfq_idle_class_timer(unsigned long data
)
2178 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
2179 unsigned long flags
, end
;
2181 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2184 * race with a non-idle queue, reset timer
2186 end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
2187 if (!time_after_eq(jiffies
, end
))
2188 mod_timer(&cfqd
->idle_class_timer
, end
);
2190 cfq_schedule_dispatch(cfqd
);
2192 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2195 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
2197 del_timer_sync(&cfqd
->idle_slice_timer
);
2198 del_timer_sync(&cfqd
->idle_class_timer
);
2199 blk_sync_queue(cfqd
->queue
);
2202 static void cfq_exit_queue(elevator_t
*e
)
2204 struct cfq_data
*cfqd
= e
->elevator_data
;
2205 request_queue_t
*q
= cfqd
->queue
;
2207 cfq_shutdown_timer_wq(cfqd
);
2209 spin_lock(&cfq_exit_lock
);
2210 spin_lock_irq(q
->queue_lock
);
2212 if (cfqd
->active_queue
)
2213 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
2215 while (!list_empty(&cfqd
->cic_list
)) {
2216 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
2217 struct cfq_io_context
,
2219 if (cic
->cfqq
[ASYNC
]) {
2220 cfq_put_queue(cic
->cfqq
[ASYNC
]);
2221 cic
->cfqq
[ASYNC
] = NULL
;
2223 if (cic
->cfqq
[SYNC
]) {
2224 cfq_put_queue(cic
->cfqq
[SYNC
]);
2225 cic
->cfqq
[SYNC
] = NULL
;
2228 list_del_init(&cic
->queue_list
);
2231 spin_unlock_irq(q
->queue_lock
);
2232 spin_unlock(&cfq_exit_lock
);
2234 cfq_shutdown_timer_wq(cfqd
);
2236 mempool_destroy(cfqd
->crq_pool
);
2237 kfree(cfqd
->crq_hash
);
2238 kfree(cfqd
->cfq_hash
);
2242 static void *cfq_init_queue(request_queue_t
*q
, elevator_t
*e
)
2244 struct cfq_data
*cfqd
;
2247 cfqd
= kmalloc(sizeof(*cfqd
), GFP_KERNEL
);
2251 memset(cfqd
, 0, sizeof(*cfqd
));
2253 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
2254 INIT_LIST_HEAD(&cfqd
->rr_list
[i
]);
2256 INIT_LIST_HEAD(&cfqd
->busy_rr
);
2257 INIT_LIST_HEAD(&cfqd
->cur_rr
);
2258 INIT_LIST_HEAD(&cfqd
->idle_rr
);
2259 INIT_LIST_HEAD(&cfqd
->empty_list
);
2260 INIT_LIST_HEAD(&cfqd
->cic_list
);
2262 cfqd
->crq_hash
= kmalloc(sizeof(struct hlist_head
) * CFQ_MHASH_ENTRIES
, GFP_KERNEL
);
2263 if (!cfqd
->crq_hash
)
2266 cfqd
->cfq_hash
= kmalloc(sizeof(struct hlist_head
) * CFQ_QHASH_ENTRIES
, GFP_KERNEL
);
2267 if (!cfqd
->cfq_hash
)
2270 cfqd
->crq_pool
= mempool_create_slab_pool(BLKDEV_MIN_RQ
, crq_pool
);
2271 if (!cfqd
->crq_pool
)
2274 for (i
= 0; i
< CFQ_MHASH_ENTRIES
; i
++)
2275 INIT_HLIST_HEAD(&cfqd
->crq_hash
[i
]);
2276 for (i
= 0; i
< CFQ_QHASH_ENTRIES
; i
++)
2277 INIT_HLIST_HEAD(&cfqd
->cfq_hash
[i
]);
2281 cfqd
->max_queued
= q
->nr_requests
/ 4;
2282 q
->nr_batching
= cfq_queued
;
2284 init_timer(&cfqd
->idle_slice_timer
);
2285 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2286 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2288 init_timer(&cfqd
->idle_class_timer
);
2289 cfqd
->idle_class_timer
.function
= cfq_idle_class_timer
;
2290 cfqd
->idle_class_timer
.data
= (unsigned long) cfqd
;
2292 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
, q
);
2294 cfqd
->cfq_queued
= cfq_queued
;
2295 cfqd
->cfq_quantum
= cfq_quantum
;
2296 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2297 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2298 cfqd
->cfq_back_max
= cfq_back_max
;
2299 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2300 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2301 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2302 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2303 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2307 kfree(cfqd
->cfq_hash
);
2309 kfree(cfqd
->crq_hash
);
2315 static void cfq_slab_kill(void)
2318 kmem_cache_destroy(crq_pool
);
2320 kmem_cache_destroy(cfq_pool
);
2322 kmem_cache_destroy(cfq_ioc_pool
);
2325 static int __init
cfq_slab_setup(void)
2327 crq_pool
= kmem_cache_create("crq_pool", sizeof(struct cfq_rq
), 0, 0,
2332 cfq_pool
= kmem_cache_create("cfq_pool", sizeof(struct cfq_queue
), 0, 0,
2337 cfq_ioc_pool
= kmem_cache_create("cfq_ioc_pool",
2338 sizeof(struct cfq_io_context
), 0, 0, NULL
, NULL
);
2349 * sysfs parts below -->
2353 cfq_var_show(unsigned int var
, char *page
)
2355 return sprintf(page
, "%d\n", var
);
2359 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2361 char *p
= (char *) page
;
2363 *var
= simple_strtoul(p
, &p
, 10);
2367 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2368 static ssize_t __FUNC(elevator_t *e, char *page) \
2370 struct cfq_data *cfqd = e->elevator_data; \
2371 unsigned int __data = __VAR; \
2373 __data = jiffies_to_msecs(__data); \
2374 return cfq_var_show(__data, (page)); \
2376 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2377 SHOW_FUNCTION(cfq_queued_show
, cfqd
->cfq_queued
, 0);
2378 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2379 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2380 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2381 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2382 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2383 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2384 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2385 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2386 #undef SHOW_FUNCTION
2388 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2389 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2391 struct cfq_data *cfqd = e->elevator_data; \
2392 unsigned int __data; \
2393 int ret = cfq_var_store(&__data, (page), count); \
2394 if (__data < (MIN)) \
2396 else if (__data > (MAX)) \
2399 *(__PTR) = msecs_to_jiffies(__data); \
2401 *(__PTR) = __data; \
2404 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2405 STORE_FUNCTION(cfq_queued_store
, &cfqd
->cfq_queued
, 1, UINT_MAX
, 0);
2406 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1, UINT_MAX
, 1);
2407 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1, UINT_MAX
, 1);
2408 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2409 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1, UINT_MAX
, 0);
2410 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2411 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2412 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2413 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1, UINT_MAX
, 0);
2414 #undef STORE_FUNCTION
2416 #define CFQ_ATTR(name) \
2417 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2419 static struct elv_fs_entry cfq_attrs
[] = {
2422 CFQ_ATTR(fifo_expire_sync
),
2423 CFQ_ATTR(fifo_expire_async
),
2424 CFQ_ATTR(back_seek_max
),
2425 CFQ_ATTR(back_seek_penalty
),
2426 CFQ_ATTR(slice_sync
),
2427 CFQ_ATTR(slice_async
),
2428 CFQ_ATTR(slice_async_rq
),
2429 CFQ_ATTR(slice_idle
),
2433 static struct elevator_type iosched_cfq
= {
2435 .elevator_merge_fn
= cfq_merge
,
2436 .elevator_merged_fn
= cfq_merged_request
,
2437 .elevator_merge_req_fn
= cfq_merged_requests
,
2438 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2439 .elevator_add_req_fn
= cfq_insert_request
,
2440 .elevator_activate_req_fn
= cfq_activate_request
,
2441 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2442 .elevator_queue_empty_fn
= cfq_queue_empty
,
2443 .elevator_completed_req_fn
= cfq_completed_request
,
2444 .elevator_former_req_fn
= cfq_former_request
,
2445 .elevator_latter_req_fn
= cfq_latter_request
,
2446 .elevator_set_req_fn
= cfq_set_request
,
2447 .elevator_put_req_fn
= cfq_put_request
,
2448 .elevator_may_queue_fn
= cfq_may_queue
,
2449 .elevator_init_fn
= cfq_init_queue
,
2450 .elevator_exit_fn
= cfq_exit_queue
,
2453 .elevator_attrs
= cfq_attrs
,
2454 .elevator_name
= "cfq",
2455 .elevator_owner
= THIS_MODULE
,
2458 static int __init
cfq_init(void)
2463 * could be 0 on HZ < 1000 setups
2465 if (!cfq_slice_async
)
2466 cfq_slice_async
= 1;
2467 if (!cfq_slice_idle
)
2470 if (cfq_slab_setup())
2473 ret
= elv_register(&iosched_cfq
);
2480 static void __exit
cfq_exit(void)
2482 DECLARE_COMPLETION(all_gone
);
2483 elv_unregister(&iosched_cfq
);
2484 ioc_gone
= &all_gone
;
2485 /* ioc_gone's update must be visible before reading ioc_count */
2487 if (atomic_read(&ioc_count
))
2488 wait_for_completion(ioc_gone
);
2493 module_init(cfq_init
);
2494 module_exit(cfq_exit
);
2496 MODULE_AUTHOR("Jens Axboe");
2497 MODULE_LICENSE("GPL");
2498 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");