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@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/hash.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
19 static const int cfq_quantum
= 4; /* max queue in one round of service */
20 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
21 static const int cfq_back_max
= 16 * 1024; /* maximum backwards seek, in KiB */
22 static const int cfq_back_penalty
= 2; /* penalty of a backwards seek */
24 static const int cfq_slice_sync
= HZ
/ 10;
25 static int cfq_slice_async
= HZ
/ 25;
26 static const int cfq_slice_async_rq
= 2;
27 static int cfq_slice_idle
= HZ
/ 125;
30 * grace period before allowing idle class to get disk access
32 #define CFQ_IDLE_GRACE (HZ / 10)
35 * below this threshold, we consider thinktime immediate
37 #define CFQ_MIN_TT (2)
39 #define CFQ_SLICE_SCALE (5)
41 #define CFQ_KEY_ASYNC (0)
44 * for the hash of cfqq inside the cfqd
46 #define CFQ_QHASH_SHIFT 6
47 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
48 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
50 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
52 #define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private)
53 #define RQ_CFQQ(rq) ((rq)->elevator_private2)
55 static struct kmem_cache
*cfq_pool
;
56 static struct kmem_cache
*cfq_ioc_pool
;
58 static DEFINE_PER_CPU(unsigned long, ioc_count
);
59 static struct completion
*ioc_gone
;
61 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
62 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
63 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68 #define cfq_cfqq_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
70 #define sample_valid(samples) ((samples) > 80)
73 * Per block device queue structure
76 request_queue_t
*queue
;
79 * rr list of queues with requests and the count of them
81 struct rb_root service_tree
;
82 struct list_head cur_rr
;
83 struct list_head idle_rr
;
84 unsigned int busy_queues
;
89 struct hlist_head
*cfq_hash
;
95 * idle window management
97 struct timer_list idle_slice_timer
;
98 struct work_struct unplug_work
;
100 struct cfq_queue
*active_queue
;
101 struct cfq_io_context
*active_cic
;
102 unsigned int dispatch_slice
;
104 struct timer_list idle_class_timer
;
106 sector_t last_position
;
107 unsigned long last_end_request
;
110 * tunables, see top of file
112 unsigned int cfq_quantum
;
113 unsigned int cfq_fifo_expire
[2];
114 unsigned int cfq_back_penalty
;
115 unsigned int cfq_back_max
;
116 unsigned int cfq_slice
[2];
117 unsigned int cfq_slice_async_rq
;
118 unsigned int cfq_slice_idle
;
120 struct list_head cic_list
;
122 sector_t new_seek_mean
;
127 * Per process-grouping structure
130 /* reference count */
132 /* parent cfq_data */
133 struct cfq_data
*cfqd
;
134 /* cfqq lookup hash */
135 struct hlist_node cfq_hash
;
138 /* member of the rr/busy/cur/idle cfqd list */
139 struct list_head cfq_list
;
140 /* service_tree member */
141 struct rb_node rb_node
;
142 /* service_tree key */
143 unsigned long rb_key
;
144 /* sorted list of pending requests */
145 struct rb_root sort_list
;
146 /* if fifo isn't expired, next request to serve */
147 struct request
*next_rq
;
148 /* requests queued in sort_list */
150 /* currently allocated requests */
152 /* pending metadata requests */
154 /* fifo list of requests in sort_list */
155 struct list_head fifo
;
157 unsigned long slice_end
;
160 /* number of requests that are on the dispatch list or inside driver */
163 /* io prio of this group */
164 unsigned short ioprio
, org_ioprio
;
165 unsigned short ioprio_class
, org_ioprio_class
;
167 /* various state flags, see below */
170 sector_t last_request_pos
;
173 enum cfqq_state_flags
{
174 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
175 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
176 CFQ_CFQQ_FLAG_must_alloc
, /* must be allowed rq alloc */
177 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
178 CFQ_CFQQ_FLAG_must_dispatch
, /* must dispatch, even if expired */
179 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
180 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
181 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
182 CFQ_CFQQ_FLAG_queue_new
, /* queue never been serviced */
183 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
186 #define CFQ_CFQQ_FNS(name) \
187 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
189 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
191 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
193 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
195 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
197 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
201 CFQ_CFQQ_FNS(wait_request
);
202 CFQ_CFQQ_FNS(must_alloc
);
203 CFQ_CFQQ_FNS(must_alloc_slice
);
204 CFQ_CFQQ_FNS(must_dispatch
);
205 CFQ_CFQQ_FNS(fifo_expire
);
206 CFQ_CFQQ_FNS(idle_window
);
207 CFQ_CFQQ_FNS(prio_changed
);
208 CFQ_CFQQ_FNS(queue_new
);
209 CFQ_CFQQ_FNS(slice_new
);
212 static struct cfq_queue
*cfq_find_cfq_hash(struct cfq_data
*, unsigned int, unsigned short);
213 static void cfq_dispatch_insert(request_queue_t
*, struct request
*);
214 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
, gfp_t gfp_mask
);
217 * scheduler run of queue, if there are requests pending and no one in the
218 * driver that will restart queueing
220 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
222 if (cfqd
->busy_queues
)
223 kblockd_schedule_work(&cfqd
->unplug_work
);
226 static int cfq_queue_empty(request_queue_t
*q
)
228 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
230 return !cfqd
->busy_queues
;
233 static inline pid_t
cfq_queue_pid(struct task_struct
*task
, int rw
, int is_sync
)
236 * Use the per-process queue, for read requests and syncronous writes
238 if (!(rw
& REQ_RW
) || is_sync
)
241 return CFQ_KEY_ASYNC
;
245 * Scale schedule slice based on io priority. Use the sync time slice only
246 * if a queue is marked sync and has sync io queued. A sync queue with async
247 * io only, should not get full sync slice length.
249 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, int sync
,
252 const int base_slice
= cfqd
->cfq_slice
[sync
];
254 WARN_ON(prio
>= IOPRIO_BE_NR
);
256 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
260 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
262 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
266 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
268 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
272 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
273 * isn't valid until the first request from the dispatch is activated
274 * and the slice time set.
276 static inline int cfq_slice_used(struct cfq_queue
*cfqq
)
278 if (cfq_cfqq_slice_new(cfqq
))
280 if (time_before(jiffies
, cfqq
->slice_end
))
287 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
288 * We choose the request that is closest to the head right now. Distance
289 * behind the head is penalized and only allowed to a certain extent.
291 static struct request
*
292 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
)
294 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
295 unsigned long back_max
;
296 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
297 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
298 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
300 if (rq1
== NULL
|| rq1
== rq2
)
305 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
307 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
309 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
311 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
317 last
= cfqd
->last_position
;
320 * by definition, 1KiB is 2 sectors
322 back_max
= cfqd
->cfq_back_max
* 2;
325 * Strict one way elevator _except_ in the case where we allow
326 * short backward seeks which are biased as twice the cost of a
327 * similar forward seek.
331 else if (s1
+ back_max
>= last
)
332 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
334 wrap
|= CFQ_RQ1_WRAP
;
338 else if (s2
+ back_max
>= last
)
339 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
341 wrap
|= CFQ_RQ2_WRAP
;
343 /* Found required data */
346 * By doing switch() on the bit mask "wrap" we avoid having to
347 * check two variables for all permutations: --> faster!
350 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
366 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
369 * Since both rqs are wrapped,
370 * start with the one that's further behind head
371 * (--> only *one* back seek required),
372 * since back seek takes more time than forward.
382 * would be nice to take fifo expire time into account as well
384 static struct request
*
385 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
386 struct request
*last
)
388 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
389 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
390 struct request
*next
= NULL
, *prev
= NULL
;
392 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
395 prev
= rb_entry_rq(rbprev
);
398 next
= rb_entry_rq(rbnext
);
400 rbnext
= rb_first(&cfqq
->sort_list
);
401 if (rbnext
&& rbnext
!= &last
->rb_node
)
402 next
= rb_entry_rq(rbnext
);
405 return cfq_choose_req(cfqd
, next
, prev
);
408 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
409 struct cfq_queue
*cfqq
)
412 * just an approximation, should be ok.
414 return ((cfqd
->busy_queues
- 1) * cfq_prio_slice(cfqd
, 1, 0));
417 static void cfq_service_tree_add(struct cfq_data
*cfqd
,
418 struct cfq_queue
*cfqq
)
420 struct rb_node
**p
= &cfqd
->service_tree
.rb_node
;
421 struct rb_node
*parent
= NULL
;
422 struct cfq_queue
*__cfqq
;
423 unsigned long rb_key
;
425 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
426 rb_key
+= cfqq
->slice_resid
;
427 cfqq
->slice_resid
= 0;
429 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
431 * same position, nothing more to do
433 if (rb_key
== cfqq
->rb_key
)
436 rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
441 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
443 if (rb_key
< __cfqq
->rb_key
)
449 cfqq
->rb_key
= rb_key
;
450 rb_link_node(&cfqq
->rb_node
, parent
, p
);
451 rb_insert_color(&cfqq
->rb_node
, &cfqd
->service_tree
);
454 static void cfq_resort_rr_list(struct cfq_queue
*cfqq
, int preempted
)
456 struct cfq_data
*cfqd
= cfqq
->cfqd
;
460 * Resorting requires the cfqq to be on the RR list already.
462 if (!cfq_cfqq_on_rr(cfqq
))
465 list_del_init(&cfqq
->cfq_list
);
467 if (cfq_class_rt(cfqq
)) {
469 * At to the front of the current list, but behind other
473 while (n
->next
!= &cfqd
->cur_rr
)
474 if (!cfq_class_rt(cfqq
))
477 list_add(&cfqq
->cfq_list
, n
);
478 } else if (cfq_class_idle(cfqq
)) {
480 * IDLE goes to the tail of the idle list
482 list_add_tail(&cfqq
->cfq_list
, &cfqd
->idle_rr
);
485 * So we get here, ergo the queue is a regular best-effort queue
487 cfq_service_tree_add(cfqd
, cfqq
);
492 * add to busy list of queues for service, trying to be fair in ordering
493 * the pending list according to last request service
496 cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
498 BUG_ON(cfq_cfqq_on_rr(cfqq
));
499 cfq_mark_cfqq_on_rr(cfqq
);
502 cfq_resort_rr_list(cfqq
, 0);
506 cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
508 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
509 cfq_clear_cfqq_on_rr(cfqq
);
510 list_del_init(&cfqq
->cfq_list
);
512 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
513 rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
514 RB_CLEAR_NODE(&cfqq
->rb_node
);
517 BUG_ON(!cfqd
->busy_queues
);
522 * rb tree support functions
524 static inline void cfq_del_rq_rb(struct request
*rq
)
526 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
527 struct cfq_data
*cfqd
= cfqq
->cfqd
;
528 const int sync
= rq_is_sync(rq
);
530 BUG_ON(!cfqq
->queued
[sync
]);
531 cfqq
->queued
[sync
]--;
533 elv_rb_del(&cfqq
->sort_list
, rq
);
535 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
536 cfq_del_cfqq_rr(cfqd
, cfqq
);
539 static void cfq_add_rq_rb(struct request
*rq
)
541 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
542 struct cfq_data
*cfqd
= cfqq
->cfqd
;
543 struct request
*__alias
;
545 cfqq
->queued
[rq_is_sync(rq
)]++;
548 * looks a little odd, but the first insert might return an alias.
549 * if that happens, put the alias on the dispatch list
551 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
552 cfq_dispatch_insert(cfqd
->queue
, __alias
);
554 if (!cfq_cfqq_on_rr(cfqq
))
555 cfq_add_cfqq_rr(cfqd
, cfqq
);
558 * check if this request is a better next-serve candidate
560 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
561 BUG_ON(!cfqq
->next_rq
);
565 cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
567 elv_rb_del(&cfqq
->sort_list
, rq
);
568 cfqq
->queued
[rq_is_sync(rq
)]--;
572 static struct request
*
573 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
575 struct task_struct
*tsk
= current
;
576 pid_t key
= cfq_queue_pid(tsk
, bio_data_dir(bio
), bio_sync(bio
));
577 struct cfq_queue
*cfqq
;
579 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
581 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
583 return elv_rb_find(&cfqq
->sort_list
, sector
);
589 static void cfq_activate_request(request_queue_t
*q
, struct request
*rq
)
591 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
593 cfqd
->rq_in_driver
++;
596 * If the depth is larger 1, it really could be queueing. But lets
597 * make the mark a little higher - idling could still be good for
598 * low queueing, and a low queueing number could also just indicate
599 * a SCSI mid layer like behaviour where limit+1 is often seen.
601 if (!cfqd
->hw_tag
&& cfqd
->rq_in_driver
> 4)
604 cfqd
->last_position
= rq
->hard_sector
+ rq
->hard_nr_sectors
;
607 static void cfq_deactivate_request(request_queue_t
*q
, struct request
*rq
)
609 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
611 WARN_ON(!cfqd
->rq_in_driver
);
612 cfqd
->rq_in_driver
--;
615 static void cfq_remove_request(struct request
*rq
)
617 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
619 if (cfqq
->next_rq
== rq
)
620 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
622 list_del_init(&rq
->queuelist
);
625 if (rq_is_meta(rq
)) {
626 WARN_ON(!cfqq
->meta_pending
);
627 cfqq
->meta_pending
--;
632 cfq_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
634 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
635 struct request
*__rq
;
637 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
638 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
640 return ELEVATOR_FRONT_MERGE
;
643 return ELEVATOR_NO_MERGE
;
646 static void cfq_merged_request(request_queue_t
*q
, struct request
*req
,
649 if (type
== ELEVATOR_FRONT_MERGE
) {
650 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
652 cfq_reposition_rq_rb(cfqq
, req
);
657 cfq_merged_requests(request_queue_t
*q
, struct request
*rq
,
658 struct request
*next
)
661 * reposition in fifo if next is older than rq
663 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
664 time_before(next
->start_time
, rq
->start_time
))
665 list_move(&rq
->queuelist
, &next
->queuelist
);
667 cfq_remove_request(next
);
670 static int cfq_allow_merge(request_queue_t
*q
, struct request
*rq
,
673 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
674 const int rw
= bio_data_dir(bio
);
675 struct cfq_queue
*cfqq
;
679 * Disallow merge of a sync bio into an async request.
681 if ((bio_data_dir(bio
) == READ
|| bio_sync(bio
)) && !rq_is_sync(rq
))
685 * Lookup the cfqq that this bio will be queued with. Allow
686 * merge only if rq is queued there.
688 key
= cfq_queue_pid(current
, rw
, bio_sync(bio
));
689 cfqq
= cfq_find_cfq_hash(cfqd
, key
, current
->ioprio
);
691 if (cfqq
== RQ_CFQQ(rq
))
698 __cfq_set_active_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
702 * stop potential idle class queues waiting service
704 del_timer(&cfqd
->idle_class_timer
);
707 cfq_clear_cfqq_must_alloc_slice(cfqq
);
708 cfq_clear_cfqq_fifo_expire(cfqq
);
709 cfq_mark_cfqq_slice_new(cfqq
);
710 cfq_clear_cfqq_queue_new(cfqq
);
713 cfqd
->active_queue
= cfqq
;
717 * current cfqq expired its slice (or was too idle), select new one
720 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
721 int preempted
, int timed_out
)
723 if (cfq_cfqq_wait_request(cfqq
))
724 del_timer(&cfqd
->idle_slice_timer
);
726 cfq_clear_cfqq_must_dispatch(cfqq
);
727 cfq_clear_cfqq_wait_request(cfqq
);
730 * store what was left of this slice, if the queue idled out
733 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
))
734 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
736 cfq_resort_rr_list(cfqq
, preempted
);
738 if (cfqq
== cfqd
->active_queue
)
739 cfqd
->active_queue
= NULL
;
741 if (cfqd
->active_cic
) {
742 put_io_context(cfqd
->active_cic
->ioc
);
743 cfqd
->active_cic
= NULL
;
746 cfqd
->dispatch_slice
= 0;
749 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, int preempted
,
752 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
755 __cfq_slice_expired(cfqd
, cfqq
, preempted
, timed_out
);
758 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
760 struct cfq_queue
*cfqq
= NULL
;
762 if (!list_empty(&cfqd
->cur_rr
)) {
764 * if current list is non-empty, grab first entry.
766 cfqq
= list_entry_cfqq(cfqd
->cur_rr
.next
);
767 } else if (!RB_EMPTY_ROOT(&cfqd
->service_tree
)) {
768 struct rb_node
*n
= rb_first(&cfqd
->service_tree
);
770 cfqq
= rb_entry(n
, struct cfq_queue
, rb_node
);
771 } else if (!list_empty(&cfqd
->idle_rr
)) {
773 * if we have idle queues and no rt or be queues had pending
774 * requests, either allow immediate service if the grace period
775 * has passed or arm the idle grace timer
777 unsigned long end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
779 if (time_after_eq(jiffies
, end
))
780 cfqq
= list_entry_cfqq(cfqd
->idle_rr
.next
);
782 mod_timer(&cfqd
->idle_class_timer
, end
);
788 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
)
790 struct cfq_queue
*cfqq
;
792 cfqq
= cfq_get_next_queue(cfqd
);
793 __cfq_set_active_queue(cfqd
, cfqq
);
797 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
800 if (rq
->sector
>= cfqd
->last_position
)
801 return rq
->sector
- cfqd
->last_position
;
803 return cfqd
->last_position
- rq
->sector
;
806 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct request
*rq
)
808 struct cfq_io_context
*cic
= cfqd
->active_cic
;
810 if (!sample_valid(cic
->seek_samples
))
813 return cfq_dist_from_last(cfqd
, rq
) <= cic
->seek_mean
;
816 static int cfq_close_cooperator(struct cfq_data
*cfq_data
,
817 struct cfq_queue
*cfqq
)
820 * We should notice if some of the queues are cooperating, eg
821 * working closely on the same area of the disk. In that case,
822 * we can group them together and don't waste time idling.
827 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
829 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
831 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
832 struct cfq_io_context
*cic
;
835 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
836 WARN_ON(cfq_cfqq_slice_new(cfqq
));
839 * idle is disabled, either manually or by past process history
841 if (!cfqd
->cfq_slice_idle
|| !cfq_cfqq_idle_window(cfqq
))
845 * task has exited, don't wait
847 cic
= cfqd
->active_cic
;
848 if (!cic
|| !cic
->ioc
->task
)
852 * See if this prio level has a good candidate
854 if (cfq_close_cooperator(cfqd
, cfqq
) &&
855 (sample_valid(cic
->ttime_samples
) && cic
->ttime_mean
> 2))
858 cfq_mark_cfqq_must_dispatch(cfqq
);
859 cfq_mark_cfqq_wait_request(cfqq
);
862 * we don't want to idle for seeks, but we do want to allow
863 * fair distribution of slice time for a process doing back-to-back
864 * seeks. so allow a little bit of time for him to submit a new rq
866 sl
= cfqd
->cfq_slice_idle
;
867 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
868 sl
= min(sl
, msecs_to_jiffies(CFQ_MIN_TT
));
870 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
873 static void cfq_dispatch_insert(request_queue_t
*q
, struct request
*rq
)
875 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
877 cfq_remove_request(rq
);
879 elv_dispatch_sort(q
, rq
);
883 * return expired entry, or NULL to just start from scratch in rbtree
885 static inline struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
887 struct cfq_data
*cfqd
= cfqq
->cfqd
;
891 if (cfq_cfqq_fifo_expire(cfqq
))
894 cfq_mark_cfqq_fifo_expire(cfqq
);
896 if (list_empty(&cfqq
->fifo
))
899 fifo
= cfq_cfqq_sync(cfqq
);
900 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
902 if (time_before(jiffies
, rq
->start_time
+ cfqd
->cfq_fifo_expire
[fifo
]))
909 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
911 const int base_rq
= cfqd
->cfq_slice_async_rq
;
913 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
915 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
919 * get next queue for service
921 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
923 struct cfq_queue
*cfqq
;
925 cfqq
= cfqd
->active_queue
;
930 * The active queue has run out of time, expire it and select new.
932 if (cfq_slice_used(cfqq
))
936 * The active queue has requests and isn't expired, allow it to
939 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
943 * No requests pending. If the active queue still has requests in
944 * flight or is idling for a new request, allow either of these
945 * conditions to happen (or time out) before selecting a new queue.
947 if (cfqq
->dispatched
|| timer_pending(&cfqd
->idle_slice_timer
)) {
953 cfq_slice_expired(cfqd
, 0, 0);
955 cfqq
= cfq_set_active_queue(cfqd
);
961 __cfq_dispatch_requests(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
966 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
972 * follow expired path, else get first next available
974 if ((rq
= cfq_check_fifo(cfqq
)) == NULL
)
978 * finally, insert request into driver dispatch list
980 cfq_dispatch_insert(cfqd
->queue
, rq
);
982 cfqd
->dispatch_slice
++;
985 if (!cfqd
->active_cic
) {
986 atomic_inc(&RQ_CIC(rq
)->ioc
->refcount
);
987 cfqd
->active_cic
= RQ_CIC(rq
);
990 if (RB_EMPTY_ROOT(&cfqq
->sort_list
))
993 } while (dispatched
< max_dispatch
);
996 * expire an async queue immediately if it has used up its slice. idle
997 * queue always expire after 1 dispatch round.
999 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
1000 cfqd
->dispatch_slice
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
1001 cfq_class_idle(cfqq
))) {
1002 cfqq
->slice_end
= jiffies
+ 1;
1003 cfq_slice_expired(cfqd
, 0, 0);
1009 static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
1013 while (cfqq
->next_rq
) {
1014 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
1018 BUG_ON(!list_empty(&cfqq
->fifo
));
1022 static int cfq_forced_dispatch_cfqqs(struct list_head
*list
)
1024 struct cfq_queue
*cfqq
, *next
;
1028 list_for_each_entry_safe(cfqq
, next
, list
, cfq_list
)
1029 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1034 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
1039 while ((n
= rb_first(&cfqd
->service_tree
)) != NULL
) {
1040 struct cfq_queue
*cfqq
= rb_entry(n
, struct cfq_queue
, rb_node
);
1042 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1045 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->cur_rr
);
1046 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->idle_rr
);
1048 cfq_slice_expired(cfqd
, 0, 0);
1050 BUG_ON(cfqd
->busy_queues
);
1055 static int cfq_dispatch_requests(request_queue_t
*q
, int force
)
1057 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1058 struct cfq_queue
*cfqq
;
1061 if (!cfqd
->busy_queues
)
1064 if (unlikely(force
))
1065 return cfq_forced_dispatch(cfqd
);
1068 while ((cfqq
= cfq_select_queue(cfqd
)) != NULL
) {
1071 if (cfqd
->busy_queues
> 1) {
1073 * So we have dispatched before in this round, if the
1074 * next queue has idling enabled (must be sync), don't
1075 * allow it service until the previous have completed.
1077 if (cfqd
->rq_in_driver
&& cfq_cfqq_idle_window(cfqq
) &&
1080 if (cfqq
->dispatched
>= cfqd
->cfq_quantum
)
1084 cfq_clear_cfqq_must_dispatch(cfqq
);
1085 cfq_clear_cfqq_wait_request(cfqq
);
1086 del_timer(&cfqd
->idle_slice_timer
);
1088 max_dispatch
= cfqd
->cfq_quantum
;
1089 if (cfq_class_idle(cfqq
))
1092 dispatched
+= __cfq_dispatch_requests(cfqd
, cfqq
, max_dispatch
);
1099 * task holds one reference to the queue, dropped when task exits. each rq
1100 * in-flight on this queue also holds a reference, dropped when rq is freed.
1102 * queue lock must be held here.
1104 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1106 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1108 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1110 if (!atomic_dec_and_test(&cfqq
->ref
))
1113 BUG_ON(rb_first(&cfqq
->sort_list
));
1114 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1115 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1117 if (unlikely(cfqd
->active_queue
== cfqq
)) {
1118 __cfq_slice_expired(cfqd
, cfqq
, 0, 0);
1119 cfq_schedule_dispatch(cfqd
);
1123 * it's on the empty list and still hashed
1125 hlist_del(&cfqq
->cfq_hash
);
1126 kmem_cache_free(cfq_pool
, cfqq
);
1129 static struct cfq_queue
*
1130 __cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned int prio
,
1133 struct hlist_head
*hash_list
= &cfqd
->cfq_hash
[hashval
];
1134 struct hlist_node
*entry
;
1135 struct cfq_queue
*__cfqq
;
1137 hlist_for_each_entry(__cfqq
, entry
, hash_list
, cfq_hash
) {
1138 const unsigned short __p
= IOPRIO_PRIO_VALUE(__cfqq
->org_ioprio_class
, __cfqq
->org_ioprio
);
1140 if (__cfqq
->key
== key
&& (__p
== prio
|| !prio
))
1147 static struct cfq_queue
*
1148 cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned short prio
)
1150 return __cfq_find_cfq_hash(cfqd
, key
, prio
, hash_long(key
, CFQ_QHASH_SHIFT
));
1153 static void cfq_free_io_context(struct io_context
*ioc
)
1155 struct cfq_io_context
*__cic
;
1159 while ((n
= rb_first(&ioc
->cic_root
)) != NULL
) {
1160 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1161 rb_erase(&__cic
->rb_node
, &ioc
->cic_root
);
1162 kmem_cache_free(cfq_ioc_pool
, __cic
);
1166 elv_ioc_count_mod(ioc_count
, -freed
);
1168 if (ioc_gone
&& !elv_ioc_count_read(ioc_count
))
1172 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1174 if (unlikely(cfqq
== cfqd
->active_queue
)) {
1175 __cfq_slice_expired(cfqd
, cfqq
, 0, 0);
1176 cfq_schedule_dispatch(cfqd
);
1179 cfq_put_queue(cfqq
);
1182 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1183 struct cfq_io_context
*cic
)
1185 list_del_init(&cic
->queue_list
);
1189 if (cic
->cfqq
[ASYNC
]) {
1190 cfq_exit_cfqq(cfqd
, cic
->cfqq
[ASYNC
]);
1191 cic
->cfqq
[ASYNC
] = NULL
;
1194 if (cic
->cfqq
[SYNC
]) {
1195 cfq_exit_cfqq(cfqd
, cic
->cfqq
[SYNC
]);
1196 cic
->cfqq
[SYNC
] = NULL
;
1202 * Called with interrupts disabled
1204 static void cfq_exit_single_io_context(struct cfq_io_context
*cic
)
1206 struct cfq_data
*cfqd
= cic
->key
;
1209 request_queue_t
*q
= cfqd
->queue
;
1211 spin_lock_irq(q
->queue_lock
);
1212 __cfq_exit_single_io_context(cfqd
, cic
);
1213 spin_unlock_irq(q
->queue_lock
);
1217 static void cfq_exit_io_context(struct io_context
*ioc
)
1219 struct cfq_io_context
*__cic
;
1223 * put the reference this task is holding to the various queues
1226 n
= rb_first(&ioc
->cic_root
);
1228 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1230 cfq_exit_single_io_context(__cic
);
1235 static struct cfq_io_context
*
1236 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1238 struct cfq_io_context
*cic
;
1240 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
, cfqd
->queue
->node
);
1242 memset(cic
, 0, sizeof(*cic
));
1243 cic
->last_end_request
= jiffies
;
1244 INIT_LIST_HEAD(&cic
->queue_list
);
1245 cic
->dtor
= cfq_free_io_context
;
1246 cic
->exit
= cfq_exit_io_context
;
1247 elv_ioc_count_inc(ioc_count
);
1253 static void cfq_init_prio_data(struct cfq_queue
*cfqq
)
1255 struct task_struct
*tsk
= current
;
1258 if (!cfq_cfqq_prio_changed(cfqq
))
1261 ioprio_class
= IOPRIO_PRIO_CLASS(tsk
->ioprio
);
1262 switch (ioprio_class
) {
1264 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1265 case IOPRIO_CLASS_NONE
:
1267 * no prio set, place us in the middle of the BE classes
1269 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1270 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1272 case IOPRIO_CLASS_RT
:
1273 cfqq
->ioprio
= task_ioprio(tsk
);
1274 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1276 case IOPRIO_CLASS_BE
:
1277 cfqq
->ioprio
= task_ioprio(tsk
);
1278 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1280 case IOPRIO_CLASS_IDLE
:
1281 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1283 cfq_clear_cfqq_idle_window(cfqq
);
1288 * keep track of original prio settings in case we have to temporarily
1289 * elevate the priority of this queue
1291 cfqq
->org_ioprio
= cfqq
->ioprio
;
1292 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1293 cfq_clear_cfqq_prio_changed(cfqq
);
1296 static inline void changed_ioprio(struct cfq_io_context
*cic
)
1298 struct cfq_data
*cfqd
= cic
->key
;
1299 struct cfq_queue
*cfqq
;
1300 unsigned long flags
;
1302 if (unlikely(!cfqd
))
1305 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1307 cfqq
= cic
->cfqq
[ASYNC
];
1309 struct cfq_queue
*new_cfqq
;
1310 new_cfqq
= cfq_get_queue(cfqd
, CFQ_KEY_ASYNC
, cic
->ioc
->task
,
1313 cic
->cfqq
[ASYNC
] = new_cfqq
;
1314 cfq_put_queue(cfqq
);
1318 cfqq
= cic
->cfqq
[SYNC
];
1320 cfq_mark_cfqq_prio_changed(cfqq
);
1322 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1325 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1327 struct cfq_io_context
*cic
;
1330 ioc
->ioprio_changed
= 0;
1332 n
= rb_first(&ioc
->cic_root
);
1334 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1336 changed_ioprio(cic
);
1341 static struct cfq_queue
*
1342 cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
,
1345 const int hashval
= hash_long(key
, CFQ_QHASH_SHIFT
);
1346 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1347 unsigned short ioprio
;
1350 ioprio
= tsk
->ioprio
;
1351 cfqq
= __cfq_find_cfq_hash(cfqd
, key
, ioprio
, hashval
);
1357 } else if (gfp_mask
& __GFP_WAIT
) {
1359 * Inform the allocator of the fact that we will
1360 * just repeat this allocation if it fails, to allow
1361 * the allocator to do whatever it needs to attempt to
1364 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1365 new_cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
|__GFP_NOFAIL
, cfqd
->queue
->node
);
1366 spin_lock_irq(cfqd
->queue
->queue_lock
);
1369 cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
, cfqd
->queue
->node
);
1374 memset(cfqq
, 0, sizeof(*cfqq
));
1376 INIT_HLIST_NODE(&cfqq
->cfq_hash
);
1377 INIT_LIST_HEAD(&cfqq
->cfq_list
);
1378 RB_CLEAR_NODE(&cfqq
->rb_node
);
1379 INIT_LIST_HEAD(&cfqq
->fifo
);
1382 hlist_add_head(&cfqq
->cfq_hash
, &cfqd
->cfq_hash
[hashval
]);
1383 atomic_set(&cfqq
->ref
, 0);
1386 if (key
!= CFQ_KEY_ASYNC
)
1387 cfq_mark_cfqq_idle_window(cfqq
);
1389 cfq_mark_cfqq_prio_changed(cfqq
);
1390 cfq_mark_cfqq_queue_new(cfqq
);
1391 cfq_init_prio_data(cfqq
);
1395 kmem_cache_free(cfq_pool
, new_cfqq
);
1397 atomic_inc(&cfqq
->ref
);
1399 WARN_ON((gfp_mask
& __GFP_WAIT
) && !cfqq
);
1404 cfq_drop_dead_cic(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1406 WARN_ON(!list_empty(&cic
->queue_list
));
1407 rb_erase(&cic
->rb_node
, &ioc
->cic_root
);
1408 kmem_cache_free(cfq_ioc_pool
, cic
);
1409 elv_ioc_count_dec(ioc_count
);
1412 static struct cfq_io_context
*
1413 cfq_cic_rb_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1416 struct cfq_io_context
*cic
;
1417 void *k
, *key
= cfqd
;
1420 n
= ioc
->cic_root
.rb_node
;
1422 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1423 /* ->key must be copied to avoid race with cfq_exit_queue() */
1426 cfq_drop_dead_cic(ioc
, cic
);
1442 cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1443 struct cfq_io_context
*cic
)
1446 struct rb_node
*parent
;
1447 struct cfq_io_context
*__cic
;
1448 unsigned long flags
;
1456 p
= &ioc
->cic_root
.rb_node
;
1459 __cic
= rb_entry(parent
, struct cfq_io_context
, rb_node
);
1460 /* ->key must be copied to avoid race with cfq_exit_queue() */
1463 cfq_drop_dead_cic(ioc
, __cic
);
1469 else if (cic
->key
> k
)
1470 p
= &(*p
)->rb_right
;
1475 rb_link_node(&cic
->rb_node
, parent
, p
);
1476 rb_insert_color(&cic
->rb_node
, &ioc
->cic_root
);
1478 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1479 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1480 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1484 * Setup general io context and cfq io context. There can be several cfq
1485 * io contexts per general io context, if this process is doing io to more
1486 * than one device managed by cfq.
1488 static struct cfq_io_context
*
1489 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1491 struct io_context
*ioc
= NULL
;
1492 struct cfq_io_context
*cic
;
1494 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1496 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
1500 cic
= cfq_cic_rb_lookup(cfqd
, ioc
);
1504 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1508 cfq_cic_link(cfqd
, ioc
, cic
);
1510 smp_read_barrier_depends();
1511 if (unlikely(ioc
->ioprio_changed
))
1512 cfq_ioc_set_ioprio(ioc
);
1516 put_io_context(ioc
);
1521 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1523 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
1524 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1526 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1527 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1528 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1532 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
1538 if (cic
->last_request_pos
< rq
->sector
)
1539 sdist
= rq
->sector
- cic
->last_request_pos
;
1541 sdist
= cic
->last_request_pos
- rq
->sector
;
1543 if (!cic
->seek_samples
) {
1544 cfqd
->new_seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1545 cfqd
->new_seek_mean
= cfqd
->new_seek_total
/ 256;
1549 * Don't allow the seek distance to get too large from the
1550 * odd fragment, pagein, etc
1552 if (cic
->seek_samples
<= 60) /* second&third seek */
1553 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1555 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1557 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1558 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1559 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1560 do_div(total
, cic
->seek_samples
);
1561 cic
->seek_mean
= (sector_t
)total
;
1565 * Disable idle window if the process thinks too long or seeks so much that
1569 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1570 struct cfq_io_context
*cic
)
1572 int enable_idle
= cfq_cfqq_idle_window(cfqq
);
1574 if (!cic
->ioc
->task
|| !cfqd
->cfq_slice_idle
||
1575 (cfqd
->hw_tag
&& CIC_SEEKY(cic
)))
1577 else if (sample_valid(cic
->ttime_samples
)) {
1578 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
1585 cfq_mark_cfqq_idle_window(cfqq
);
1587 cfq_clear_cfqq_idle_window(cfqq
);
1591 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1592 * no or if we aren't sure, a 1 will cause a preempt.
1595 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
1598 struct cfq_queue
*cfqq
;
1600 cfqq
= cfqd
->active_queue
;
1604 if (cfq_slice_used(cfqq
))
1607 if (cfq_class_idle(new_cfqq
))
1610 if (cfq_class_idle(cfqq
))
1614 * if the new request is sync, but the currently running queue is
1615 * not, let the sync request have priority.
1617 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
1621 * So both queues are sync. Let the new request get disk time if
1622 * it's a metadata request and the current queue is doing regular IO.
1624 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
1627 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
1631 * if this request is as-good as one we would expect from the
1632 * current cfqq, let it preempt
1634 if (cfq_rq_close(cfqd
, rq
))
1641 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1642 * let it have half of its nominal slice.
1644 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1646 cfq_slice_expired(cfqd
, 1, 1);
1649 * Put the new queue at the front of the of the current list,
1650 * so we know that it will be selected next.
1652 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1653 list_del_init(&cfqq
->cfq_list
);
1654 list_add(&cfqq
->cfq_list
, &cfqd
->cur_rr
);
1656 cfqq
->slice_end
= 0;
1657 cfq_mark_cfqq_slice_new(cfqq
);
1661 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1662 * something we should do about it
1665 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1668 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1671 cfqq
->meta_pending
++;
1673 cfq_update_io_thinktime(cfqd
, cic
);
1674 cfq_update_io_seektime(cfqd
, cic
, rq
);
1675 cfq_update_idle_window(cfqd
, cfqq
, cic
);
1677 cic
->last_request_pos
= rq
->sector
+ rq
->nr_sectors
;
1678 cfqq
->last_request_pos
= cic
->last_request_pos
;
1680 if (cfqq
== cfqd
->active_queue
) {
1682 * if we are waiting for a request for this queue, let it rip
1683 * immediately and flag that we must not expire this queue
1686 if (cfq_cfqq_wait_request(cfqq
)) {
1687 cfq_mark_cfqq_must_dispatch(cfqq
);
1688 del_timer(&cfqd
->idle_slice_timer
);
1689 blk_start_queueing(cfqd
->queue
);
1691 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
1693 * not the active queue - expire current slice if it is
1694 * idle and has expired it's mean thinktime or this new queue
1695 * has some old slice time left and is of higher priority
1697 cfq_preempt_queue(cfqd
, cfqq
);
1698 cfq_mark_cfqq_must_dispatch(cfqq
);
1699 blk_start_queueing(cfqd
->queue
);
1703 static void cfq_insert_request(request_queue_t
*q
, struct request
*rq
)
1705 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1706 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1708 cfq_init_prio_data(cfqq
);
1712 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
1714 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
1717 static void cfq_completed_request(request_queue_t
*q
, struct request
*rq
)
1719 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1720 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1721 const int sync
= rq_is_sync(rq
);
1726 WARN_ON(!cfqd
->rq_in_driver
);
1727 WARN_ON(!cfqq
->dispatched
);
1728 cfqd
->rq_in_driver
--;
1731 if (!cfq_class_idle(cfqq
))
1732 cfqd
->last_end_request
= now
;
1735 RQ_CIC(rq
)->last_end_request
= now
;
1738 * If this is the active queue, check if it needs to be expired,
1739 * or if we want to idle in case it has no pending requests.
1741 if (cfqd
->active_queue
== cfqq
) {
1742 if (cfq_cfqq_slice_new(cfqq
)) {
1743 cfq_set_prio_slice(cfqd
, cfqq
);
1744 cfq_clear_cfqq_slice_new(cfqq
);
1746 if (cfq_slice_used(cfqq
))
1747 cfq_slice_expired(cfqd
, 0, 1);
1748 else if (sync
&& RB_EMPTY_ROOT(&cfqq
->sort_list
))
1749 cfq_arm_slice_timer(cfqd
);
1752 if (!cfqd
->rq_in_driver
)
1753 cfq_schedule_dispatch(cfqd
);
1757 * we temporarily boost lower priority queues if they are holding fs exclusive
1758 * resources. they are boosted to normal prio (CLASS_BE/4)
1760 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
1762 if (has_fs_excl()) {
1764 * boost idle prio on transactions that would lock out other
1765 * users of the filesystem
1767 if (cfq_class_idle(cfqq
))
1768 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1769 if (cfqq
->ioprio
> IOPRIO_NORM
)
1770 cfqq
->ioprio
= IOPRIO_NORM
;
1773 * check if we need to unboost the queue
1775 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
1776 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
1777 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
1778 cfqq
->ioprio
= cfqq
->org_ioprio
;
1782 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
1784 if ((cfq_cfqq_wait_request(cfqq
) || cfq_cfqq_must_alloc(cfqq
)) &&
1785 !cfq_cfqq_must_alloc_slice(cfqq
)) {
1786 cfq_mark_cfqq_must_alloc_slice(cfqq
);
1787 return ELV_MQUEUE_MUST
;
1790 return ELV_MQUEUE_MAY
;
1793 static int cfq_may_queue(request_queue_t
*q
, int rw
)
1795 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1796 struct task_struct
*tsk
= current
;
1797 struct cfq_queue
*cfqq
;
1800 key
= cfq_queue_pid(tsk
, rw
, rw
& REQ_RW_SYNC
);
1803 * don't force setup of a queue from here, as a call to may_queue
1804 * does not necessarily imply that a request actually will be queued.
1805 * so just lookup a possibly existing queue, or return 'may queue'
1808 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
1810 cfq_init_prio_data(cfqq
);
1811 cfq_prio_boost(cfqq
);
1813 return __cfq_may_queue(cfqq
);
1816 return ELV_MQUEUE_MAY
;
1820 * queue lock held here
1822 static void cfq_put_request(struct request
*rq
)
1824 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1827 const int rw
= rq_data_dir(rq
);
1829 BUG_ON(!cfqq
->allocated
[rw
]);
1830 cfqq
->allocated
[rw
]--;
1832 put_io_context(RQ_CIC(rq
)->ioc
);
1834 rq
->elevator_private
= NULL
;
1835 rq
->elevator_private2
= NULL
;
1837 cfq_put_queue(cfqq
);
1842 * Allocate cfq data structures associated with this request.
1845 cfq_set_request(request_queue_t
*q
, struct request
*rq
, gfp_t gfp_mask
)
1847 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1848 struct task_struct
*tsk
= current
;
1849 struct cfq_io_context
*cic
;
1850 const int rw
= rq_data_dir(rq
);
1851 const int is_sync
= rq_is_sync(rq
);
1852 pid_t key
= cfq_queue_pid(tsk
, rw
, is_sync
);
1853 struct cfq_queue
*cfqq
;
1854 unsigned long flags
;
1856 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1858 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
1860 spin_lock_irqsave(q
->queue_lock
, flags
);
1865 if (!cic
->cfqq
[is_sync
]) {
1866 cfqq
= cfq_get_queue(cfqd
, key
, tsk
, gfp_mask
);
1870 cic
->cfqq
[is_sync
] = cfqq
;
1872 cfqq
= cic
->cfqq
[is_sync
];
1874 cfqq
->allocated
[rw
]++;
1875 cfq_clear_cfqq_must_alloc(cfqq
);
1876 atomic_inc(&cfqq
->ref
);
1878 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1880 rq
->elevator_private
= cic
;
1881 rq
->elevator_private2
= cfqq
;
1886 put_io_context(cic
->ioc
);
1888 cfq_schedule_dispatch(cfqd
);
1889 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1893 static void cfq_kick_queue(struct work_struct
*work
)
1895 struct cfq_data
*cfqd
=
1896 container_of(work
, struct cfq_data
, unplug_work
);
1897 request_queue_t
*q
= cfqd
->queue
;
1898 unsigned long flags
;
1900 spin_lock_irqsave(q
->queue_lock
, flags
);
1901 blk_start_queueing(q
);
1902 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1906 * Timer running if the active_queue is currently idling inside its time slice
1908 static void cfq_idle_slice_timer(unsigned long data
)
1910 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1911 struct cfq_queue
*cfqq
;
1912 unsigned long flags
;
1915 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1917 if ((cfqq
= cfqd
->active_queue
) != NULL
) {
1923 if (cfq_slice_used(cfqq
))
1927 * only expire and reinvoke request handler, if there are
1928 * other queues with pending requests
1930 if (!cfqd
->busy_queues
)
1934 * not expired and it has a request pending, let it dispatch
1936 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1937 cfq_mark_cfqq_must_dispatch(cfqq
);
1942 cfq_slice_expired(cfqd
, 0, timed_out
);
1944 cfq_schedule_dispatch(cfqd
);
1946 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1950 * Timer running if an idle class queue is waiting for service
1952 static void cfq_idle_class_timer(unsigned long data
)
1954 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1955 unsigned long flags
, end
;
1957 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1960 * race with a non-idle queue, reset timer
1962 end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
1963 if (!time_after_eq(jiffies
, end
))
1964 mod_timer(&cfqd
->idle_class_timer
, end
);
1966 cfq_schedule_dispatch(cfqd
);
1968 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1971 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
1973 del_timer_sync(&cfqd
->idle_slice_timer
);
1974 del_timer_sync(&cfqd
->idle_class_timer
);
1975 blk_sync_queue(cfqd
->queue
);
1978 static void cfq_exit_queue(elevator_t
*e
)
1980 struct cfq_data
*cfqd
= e
->elevator_data
;
1981 request_queue_t
*q
= cfqd
->queue
;
1983 cfq_shutdown_timer_wq(cfqd
);
1985 spin_lock_irq(q
->queue_lock
);
1987 if (cfqd
->active_queue
)
1988 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0, 0);
1990 while (!list_empty(&cfqd
->cic_list
)) {
1991 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
1992 struct cfq_io_context
,
1995 __cfq_exit_single_io_context(cfqd
, cic
);
1998 spin_unlock_irq(q
->queue_lock
);
2000 cfq_shutdown_timer_wq(cfqd
);
2002 kfree(cfqd
->cfq_hash
);
2006 static void *cfq_init_queue(request_queue_t
*q
)
2008 struct cfq_data
*cfqd
;
2011 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
, q
->node
);
2015 memset(cfqd
, 0, sizeof(*cfqd
));
2017 cfqd
->service_tree
= RB_ROOT
;
2018 INIT_LIST_HEAD(&cfqd
->cur_rr
);
2019 INIT_LIST_HEAD(&cfqd
->idle_rr
);
2020 INIT_LIST_HEAD(&cfqd
->cic_list
);
2022 cfqd
->cfq_hash
= kmalloc_node(sizeof(struct hlist_head
) * CFQ_QHASH_ENTRIES
, GFP_KERNEL
, q
->node
);
2023 if (!cfqd
->cfq_hash
)
2026 for (i
= 0; i
< CFQ_QHASH_ENTRIES
; i
++)
2027 INIT_HLIST_HEAD(&cfqd
->cfq_hash
[i
]);
2031 init_timer(&cfqd
->idle_slice_timer
);
2032 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2033 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2035 init_timer(&cfqd
->idle_class_timer
);
2036 cfqd
->idle_class_timer
.function
= cfq_idle_class_timer
;
2037 cfqd
->idle_class_timer
.data
= (unsigned long) cfqd
;
2039 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
2041 cfqd
->cfq_quantum
= cfq_quantum
;
2042 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2043 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2044 cfqd
->cfq_back_max
= cfq_back_max
;
2045 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2046 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2047 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2048 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2049 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2057 static void cfq_slab_kill(void)
2060 kmem_cache_destroy(cfq_pool
);
2062 kmem_cache_destroy(cfq_ioc_pool
);
2065 static int __init
cfq_slab_setup(void)
2067 cfq_pool
= kmem_cache_create("cfq_pool", sizeof(struct cfq_queue
), 0, 0,
2072 cfq_ioc_pool
= kmem_cache_create("cfq_ioc_pool",
2073 sizeof(struct cfq_io_context
), 0, 0, NULL
, NULL
);
2084 * sysfs parts below -->
2087 cfq_var_show(unsigned int var
, char *page
)
2089 return sprintf(page
, "%d\n", var
);
2093 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2095 char *p
= (char *) page
;
2097 *var
= simple_strtoul(p
, &p
, 10);
2101 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2102 static ssize_t __FUNC(elevator_t *e, char *page) \
2104 struct cfq_data *cfqd = e->elevator_data; \
2105 unsigned int __data = __VAR; \
2107 __data = jiffies_to_msecs(__data); \
2108 return cfq_var_show(__data, (page)); \
2110 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2111 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2112 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2113 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2114 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2115 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2116 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2117 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2118 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2119 #undef SHOW_FUNCTION
2121 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2122 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2124 struct cfq_data *cfqd = e->elevator_data; \
2125 unsigned int __data; \
2126 int ret = cfq_var_store(&__data, (page), count); \
2127 if (__data < (MIN)) \
2129 else if (__data > (MAX)) \
2132 *(__PTR) = msecs_to_jiffies(__data); \
2134 *(__PTR) = __data; \
2137 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2138 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1, UINT_MAX
, 1);
2139 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1, UINT_MAX
, 1);
2140 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2141 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1, UINT_MAX
, 0);
2142 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2143 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2144 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2145 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1, UINT_MAX
, 0);
2146 #undef STORE_FUNCTION
2148 #define CFQ_ATTR(name) \
2149 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2151 static struct elv_fs_entry cfq_attrs
[] = {
2153 CFQ_ATTR(fifo_expire_sync
),
2154 CFQ_ATTR(fifo_expire_async
),
2155 CFQ_ATTR(back_seek_max
),
2156 CFQ_ATTR(back_seek_penalty
),
2157 CFQ_ATTR(slice_sync
),
2158 CFQ_ATTR(slice_async
),
2159 CFQ_ATTR(slice_async_rq
),
2160 CFQ_ATTR(slice_idle
),
2164 static struct elevator_type iosched_cfq
= {
2166 .elevator_merge_fn
= cfq_merge
,
2167 .elevator_merged_fn
= cfq_merged_request
,
2168 .elevator_merge_req_fn
= cfq_merged_requests
,
2169 .elevator_allow_merge_fn
= cfq_allow_merge
,
2170 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2171 .elevator_add_req_fn
= cfq_insert_request
,
2172 .elevator_activate_req_fn
= cfq_activate_request
,
2173 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2174 .elevator_queue_empty_fn
= cfq_queue_empty
,
2175 .elevator_completed_req_fn
= cfq_completed_request
,
2176 .elevator_former_req_fn
= elv_rb_former_request
,
2177 .elevator_latter_req_fn
= elv_rb_latter_request
,
2178 .elevator_set_req_fn
= cfq_set_request
,
2179 .elevator_put_req_fn
= cfq_put_request
,
2180 .elevator_may_queue_fn
= cfq_may_queue
,
2181 .elevator_init_fn
= cfq_init_queue
,
2182 .elevator_exit_fn
= cfq_exit_queue
,
2183 .trim
= cfq_free_io_context
,
2185 .elevator_attrs
= cfq_attrs
,
2186 .elevator_name
= "cfq",
2187 .elevator_owner
= THIS_MODULE
,
2190 static int __init
cfq_init(void)
2195 * could be 0 on HZ < 1000 setups
2197 if (!cfq_slice_async
)
2198 cfq_slice_async
= 1;
2199 if (!cfq_slice_idle
)
2202 if (cfq_slab_setup())
2205 ret
= elv_register(&iosched_cfq
);
2212 static void __exit
cfq_exit(void)
2214 DECLARE_COMPLETION_ONSTACK(all_gone
);
2215 elv_unregister(&iosched_cfq
);
2216 ioc_gone
= &all_gone
;
2217 /* ioc_gone's update must be visible before reading ioc_count */
2219 if (elv_ioc_count_read(ioc_count
))
2220 wait_for_completion(ioc_gone
);
2225 module_init(cfq_init
);
2226 module_exit(cfq_exit
);
2228 MODULE_AUTHOR("Jens Axboe");
2229 MODULE_LICENSE("GPL");
2230 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");