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 * Most of our rbtree usage is for sorting with min extraction, so
74 * if we cache the leftmost node we don't have to walk down the tree
75 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
76 * move this into the elevator for the rq sorting as well.
82 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
85 * Per block device queue structure
88 request_queue_t
*queue
;
91 * rr list of queues with requests and the count of them
93 struct cfq_rb_root service_tree
;
94 struct list_head cur_rr
;
95 unsigned int busy_queues
;
100 struct hlist_head
*cfq_hash
;
106 * idle window management
108 struct timer_list idle_slice_timer
;
109 struct work_struct unplug_work
;
111 struct cfq_queue
*active_queue
;
112 struct cfq_io_context
*active_cic
;
113 unsigned int dispatch_slice
;
115 struct timer_list idle_class_timer
;
117 sector_t last_position
;
118 unsigned long last_end_request
;
121 * tunables, see top of file
123 unsigned int cfq_quantum
;
124 unsigned int cfq_fifo_expire
[2];
125 unsigned int cfq_back_penalty
;
126 unsigned int cfq_back_max
;
127 unsigned int cfq_slice
[2];
128 unsigned int cfq_slice_async_rq
;
129 unsigned int cfq_slice_idle
;
131 struct list_head cic_list
;
133 sector_t new_seek_mean
;
138 * Per process-grouping structure
141 /* reference count */
143 /* parent cfq_data */
144 struct cfq_data
*cfqd
;
145 /* cfqq lookup hash */
146 struct hlist_node cfq_hash
;
149 /* member of the rr/busy/cur/idle cfqd list */
150 struct list_head cfq_list
;
151 /* service_tree member */
152 struct rb_node rb_node
;
153 /* service_tree key */
154 unsigned long rb_key
;
155 /* sorted list of pending requests */
156 struct rb_root sort_list
;
157 /* if fifo isn't expired, next request to serve */
158 struct request
*next_rq
;
159 /* requests queued in sort_list */
161 /* currently allocated requests */
163 /* pending metadata requests */
165 /* fifo list of requests in sort_list */
166 struct list_head fifo
;
168 unsigned long slice_end
;
171 /* number of requests that are on the dispatch list or inside driver */
174 /* io prio of this group */
175 unsigned short ioprio
, org_ioprio
;
176 unsigned short ioprio_class
, org_ioprio_class
;
178 /* various state flags, see below */
181 sector_t last_request_pos
;
184 enum cfqq_state_flags
{
185 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
186 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
187 CFQ_CFQQ_FLAG_must_alloc
, /* must be allowed rq alloc */
188 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
189 CFQ_CFQQ_FLAG_must_dispatch
, /* must dispatch, even if expired */
190 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
191 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
192 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
193 CFQ_CFQQ_FLAG_queue_new
, /* queue never been serviced */
194 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
197 #define CFQ_CFQQ_FNS(name) \
198 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
200 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
202 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
204 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
206 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
208 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
212 CFQ_CFQQ_FNS(wait_request
);
213 CFQ_CFQQ_FNS(must_alloc
);
214 CFQ_CFQQ_FNS(must_alloc_slice
);
215 CFQ_CFQQ_FNS(must_dispatch
);
216 CFQ_CFQQ_FNS(fifo_expire
);
217 CFQ_CFQQ_FNS(idle_window
);
218 CFQ_CFQQ_FNS(prio_changed
);
219 CFQ_CFQQ_FNS(queue_new
);
220 CFQ_CFQQ_FNS(slice_new
);
223 static struct cfq_queue
*cfq_find_cfq_hash(struct cfq_data
*, unsigned int, unsigned short);
224 static void cfq_dispatch_insert(request_queue_t
*, struct request
*);
225 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, unsigned int, struct task_struct
*, gfp_t
);
228 * scheduler run of queue, if there are requests pending and no one in the
229 * driver that will restart queueing
231 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
233 if (cfqd
->busy_queues
)
234 kblockd_schedule_work(&cfqd
->unplug_work
);
237 static int cfq_queue_empty(request_queue_t
*q
)
239 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
241 return !cfqd
->busy_queues
;
244 static inline pid_t
cfq_queue_pid(struct task_struct
*task
, int rw
, int is_sync
)
247 * Use the per-process queue, for read requests and syncronous writes
249 if (!(rw
& REQ_RW
) || is_sync
)
252 return CFQ_KEY_ASYNC
;
256 * Scale schedule slice based on io priority. Use the sync time slice only
257 * if a queue is marked sync and has sync io queued. A sync queue with async
258 * io only, should not get full sync slice length.
260 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, int sync
,
263 const int base_slice
= cfqd
->cfq_slice
[sync
];
265 WARN_ON(prio
>= IOPRIO_BE_NR
);
267 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
271 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
273 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
277 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
279 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
283 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
284 * isn't valid until the first request from the dispatch is activated
285 * and the slice time set.
287 static inline int cfq_slice_used(struct cfq_queue
*cfqq
)
289 if (cfq_cfqq_slice_new(cfqq
))
291 if (time_before(jiffies
, cfqq
->slice_end
))
298 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
299 * We choose the request that is closest to the head right now. Distance
300 * behind the head is penalized and only allowed to a certain extent.
302 static struct request
*
303 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
)
305 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
306 unsigned long back_max
;
307 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
308 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
309 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
311 if (rq1
== NULL
|| rq1
== rq2
)
316 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
318 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
320 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
322 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
328 last
= cfqd
->last_position
;
331 * by definition, 1KiB is 2 sectors
333 back_max
= cfqd
->cfq_back_max
* 2;
336 * Strict one way elevator _except_ in the case where we allow
337 * short backward seeks which are biased as twice the cost of a
338 * similar forward seek.
342 else if (s1
+ back_max
>= last
)
343 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
345 wrap
|= CFQ_RQ1_WRAP
;
349 else if (s2
+ back_max
>= last
)
350 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
352 wrap
|= CFQ_RQ2_WRAP
;
354 /* Found required data */
357 * By doing switch() on the bit mask "wrap" we avoid having to
358 * check two variables for all permutations: --> faster!
361 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
377 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
380 * Since both rqs are wrapped,
381 * start with the one that's further behind head
382 * (--> only *one* back seek required),
383 * since back seek takes more time than forward.
393 * The below is leftmost cache rbtree addon
395 static struct rb_node
*cfq_rb_first(struct cfq_rb_root
*root
)
398 root
->left
= rb_first(&root
->rb
);
403 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
408 rb_erase(n
, &root
->rb
);
413 * would be nice to take fifo expire time into account as well
415 static struct request
*
416 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
417 struct request
*last
)
419 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
420 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
421 struct request
*next
= NULL
, *prev
= NULL
;
423 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
426 prev
= rb_entry_rq(rbprev
);
429 next
= rb_entry_rq(rbnext
);
431 rbnext
= rb_first(&cfqq
->sort_list
);
432 if (rbnext
&& rbnext
!= &last
->rb_node
)
433 next
= rb_entry_rq(rbnext
);
436 return cfq_choose_req(cfqd
, next
, prev
);
439 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
440 struct cfq_queue
*cfqq
)
443 * just an approximation, should be ok.
445 return ((cfqd
->busy_queues
- 1) * cfq_prio_slice(cfqd
, 1, 0));
449 * The cfqd->service_tree holds all pending cfq_queue's that have
450 * requests waiting to be processed. It is sorted in the order that
451 * we will service the queues.
453 static void cfq_service_tree_add(struct cfq_data
*cfqd
,
454 struct cfq_queue
*cfqq
)
456 struct rb_node
**p
= &cfqd
->service_tree
.rb
.rb_node
;
457 struct rb_node
*parent
= NULL
;
458 unsigned long rb_key
;
461 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
462 rb_key
+= cfqq
->slice_resid
;
463 cfqq
->slice_resid
= 0;
465 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
467 * same position, nothing more to do
469 if (rb_key
== cfqq
->rb_key
)
472 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
477 struct cfq_queue
*__cfqq
;
481 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
484 * sort RT queues first, we always want to give
485 * preference to them. IDLE queues goes to the back.
486 * after that, sort on the next service time.
488 if (cfq_class_rt(cfqq
) > cfq_class_rt(__cfqq
))
490 else if (cfq_class_rt(cfqq
) < cfq_class_rt(__cfqq
))
492 else if (cfq_class_idle(cfqq
) < cfq_class_idle(__cfqq
))
494 else if (cfq_class_idle(cfqq
) > cfq_class_idle(__cfqq
))
496 else if (rb_key
< __cfqq
->rb_key
)
501 if (n
== &(*p
)->rb_right
)
508 cfqd
->service_tree
.left
= &cfqq
->rb_node
;
510 cfqq
->rb_key
= rb_key
;
511 rb_link_node(&cfqq
->rb_node
, parent
, p
);
512 rb_insert_color(&cfqq
->rb_node
, &cfqd
->service_tree
.rb
);
516 * Update cfqq's position in the service tree.
518 static void cfq_resort_rr_list(struct cfq_queue
*cfqq
, int preempted
)
521 * Resorting requires the cfqq to be on the RR list already.
523 if (cfq_cfqq_on_rr(cfqq
))
524 cfq_service_tree_add(cfqq
->cfqd
, cfqq
);
528 * add to busy list of queues for service, trying to be fair in ordering
529 * the pending list according to last request service
532 cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
534 BUG_ON(cfq_cfqq_on_rr(cfqq
));
535 cfq_mark_cfqq_on_rr(cfqq
);
538 cfq_resort_rr_list(cfqq
, 0);
542 * Called when the cfqq no longer has requests pending, remove it from
546 cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
548 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
549 cfq_clear_cfqq_on_rr(cfqq
);
550 list_del_init(&cfqq
->cfq_list
);
552 if (!RB_EMPTY_NODE(&cfqq
->rb_node
))
553 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
555 BUG_ON(!cfqd
->busy_queues
);
560 * rb tree support functions
562 static inline void cfq_del_rq_rb(struct request
*rq
)
564 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
565 struct cfq_data
*cfqd
= cfqq
->cfqd
;
566 const int sync
= rq_is_sync(rq
);
568 BUG_ON(!cfqq
->queued
[sync
]);
569 cfqq
->queued
[sync
]--;
571 elv_rb_del(&cfqq
->sort_list
, rq
);
573 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
574 cfq_del_cfqq_rr(cfqd
, cfqq
);
577 static void cfq_add_rq_rb(struct request
*rq
)
579 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
580 struct cfq_data
*cfqd
= cfqq
->cfqd
;
581 struct request
*__alias
;
583 cfqq
->queued
[rq_is_sync(rq
)]++;
586 * looks a little odd, but the first insert might return an alias.
587 * if that happens, put the alias on the dispatch list
589 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
590 cfq_dispatch_insert(cfqd
->queue
, __alias
);
592 if (!cfq_cfqq_on_rr(cfqq
))
593 cfq_add_cfqq_rr(cfqd
, cfqq
);
596 * check if this request is a better next-serve candidate
598 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
599 BUG_ON(!cfqq
->next_rq
);
603 cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
605 elv_rb_del(&cfqq
->sort_list
, rq
);
606 cfqq
->queued
[rq_is_sync(rq
)]--;
610 static struct request
*
611 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
613 struct task_struct
*tsk
= current
;
614 pid_t key
= cfq_queue_pid(tsk
, bio_data_dir(bio
), bio_sync(bio
));
615 struct cfq_queue
*cfqq
;
617 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
619 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
621 return elv_rb_find(&cfqq
->sort_list
, sector
);
627 static void cfq_activate_request(request_queue_t
*q
, struct request
*rq
)
629 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
631 cfqd
->rq_in_driver
++;
634 * If the depth is larger 1, it really could be queueing. But lets
635 * make the mark a little higher - idling could still be good for
636 * low queueing, and a low queueing number could also just indicate
637 * a SCSI mid layer like behaviour where limit+1 is often seen.
639 if (!cfqd
->hw_tag
&& cfqd
->rq_in_driver
> 4)
642 cfqd
->last_position
= rq
->hard_sector
+ rq
->hard_nr_sectors
;
645 static void cfq_deactivate_request(request_queue_t
*q
, struct request
*rq
)
647 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
649 WARN_ON(!cfqd
->rq_in_driver
);
650 cfqd
->rq_in_driver
--;
653 static void cfq_remove_request(struct request
*rq
)
655 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
657 if (cfqq
->next_rq
== rq
)
658 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
660 list_del_init(&rq
->queuelist
);
663 if (rq_is_meta(rq
)) {
664 WARN_ON(!cfqq
->meta_pending
);
665 cfqq
->meta_pending
--;
669 static int cfq_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
671 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
672 struct request
*__rq
;
674 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
675 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
677 return ELEVATOR_FRONT_MERGE
;
680 return ELEVATOR_NO_MERGE
;
683 static void cfq_merged_request(request_queue_t
*q
, struct request
*req
,
686 if (type
== ELEVATOR_FRONT_MERGE
) {
687 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
689 cfq_reposition_rq_rb(cfqq
, req
);
694 cfq_merged_requests(request_queue_t
*q
, struct request
*rq
,
695 struct request
*next
)
698 * reposition in fifo if next is older than rq
700 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
701 time_before(next
->start_time
, rq
->start_time
))
702 list_move(&rq
->queuelist
, &next
->queuelist
);
704 cfq_remove_request(next
);
707 static int cfq_allow_merge(request_queue_t
*q
, struct request
*rq
,
710 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
711 const int rw
= bio_data_dir(bio
);
712 struct cfq_queue
*cfqq
;
716 * Disallow merge of a sync bio into an async request.
718 if ((bio_data_dir(bio
) == READ
|| bio_sync(bio
)) && !rq_is_sync(rq
))
722 * Lookup the cfqq that this bio will be queued with. Allow
723 * merge only if rq is queued there.
725 key
= cfq_queue_pid(current
, rw
, bio_sync(bio
));
726 cfqq
= cfq_find_cfq_hash(cfqd
, key
, current
->ioprio
);
728 if (cfqq
== RQ_CFQQ(rq
))
735 __cfq_set_active_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
739 * stop potential idle class queues waiting service
741 del_timer(&cfqd
->idle_class_timer
);
744 cfq_clear_cfqq_must_alloc_slice(cfqq
);
745 cfq_clear_cfqq_fifo_expire(cfqq
);
746 cfq_mark_cfqq_slice_new(cfqq
);
747 cfq_clear_cfqq_queue_new(cfqq
);
750 cfqd
->active_queue
= cfqq
;
754 * current cfqq expired its slice (or was too idle), select new one
757 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
758 int preempted
, int timed_out
)
760 if (cfq_cfqq_wait_request(cfqq
))
761 del_timer(&cfqd
->idle_slice_timer
);
763 cfq_clear_cfqq_must_dispatch(cfqq
);
764 cfq_clear_cfqq_wait_request(cfqq
);
767 * store what was left of this slice, if the queue idled out
770 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
))
771 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
773 cfq_resort_rr_list(cfqq
, preempted
);
775 if (cfqq
== cfqd
->active_queue
)
776 cfqd
->active_queue
= NULL
;
778 if (cfqd
->active_cic
) {
779 put_io_context(cfqd
->active_cic
->ioc
);
780 cfqd
->active_cic
= NULL
;
783 cfqd
->dispatch_slice
= 0;
786 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, int preempted
,
789 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
792 __cfq_slice_expired(cfqd
, cfqq
, preempted
, timed_out
);
796 * Get next queue for service. Unless we have a queue preemption,
797 * we'll simply select the first cfqq in the service tree.
799 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
801 struct cfq_queue
*cfqq
= NULL
;
803 if (!list_empty(&cfqd
->cur_rr
)) {
805 * if current list is non-empty, grab first entry.
807 cfqq
= list_entry_cfqq(cfqd
->cur_rr
.next
);
808 } else if (!RB_EMPTY_ROOT(&cfqd
->service_tree
.rb
)) {
809 struct rb_node
*n
= cfq_rb_first(&cfqd
->service_tree
);
811 cfqq
= rb_entry(n
, struct cfq_queue
, rb_node
);
812 if (cfq_class_idle(cfqq
)) {
816 * if we have idle queues and no rt or be queues had
817 * pending requests, either allow immediate service if
818 * the grace period has passed or arm the idle grace
821 end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
822 if (time_before(jiffies
, end
)) {
823 mod_timer(&cfqd
->idle_class_timer
, end
);
833 * Get and set a new active queue for service.
835 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
)
837 struct cfq_queue
*cfqq
;
839 cfqq
= cfq_get_next_queue(cfqd
);
840 __cfq_set_active_queue(cfqd
, cfqq
);
844 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
847 if (rq
->sector
>= cfqd
->last_position
)
848 return rq
->sector
- cfqd
->last_position
;
850 return cfqd
->last_position
- rq
->sector
;
853 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct request
*rq
)
855 struct cfq_io_context
*cic
= cfqd
->active_cic
;
857 if (!sample_valid(cic
->seek_samples
))
860 return cfq_dist_from_last(cfqd
, rq
) <= cic
->seek_mean
;
863 static int cfq_close_cooperator(struct cfq_data
*cfq_data
,
864 struct cfq_queue
*cfqq
)
867 * We should notice if some of the queues are cooperating, eg
868 * working closely on the same area of the disk. In that case,
869 * we can group them together and don't waste time idling.
874 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
876 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
878 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
879 struct cfq_io_context
*cic
;
882 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
883 WARN_ON(cfq_cfqq_slice_new(cfqq
));
886 * idle is disabled, either manually or by past process history
888 if (!cfqd
->cfq_slice_idle
|| !cfq_cfqq_idle_window(cfqq
))
892 * task has exited, don't wait
894 cic
= cfqd
->active_cic
;
895 if (!cic
|| !cic
->ioc
->task
)
899 * See if this prio level has a good candidate
901 if (cfq_close_cooperator(cfqd
, cfqq
) &&
902 (sample_valid(cic
->ttime_samples
) && cic
->ttime_mean
> 2))
905 cfq_mark_cfqq_must_dispatch(cfqq
);
906 cfq_mark_cfqq_wait_request(cfqq
);
909 * we don't want to idle for seeks, but we do want to allow
910 * fair distribution of slice time for a process doing back-to-back
911 * seeks. so allow a little bit of time for him to submit a new rq
913 sl
= cfqd
->cfq_slice_idle
;
914 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
915 sl
= min(sl
, msecs_to_jiffies(CFQ_MIN_TT
));
917 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
921 * Move request from internal lists to the request queue dispatch list.
923 static void cfq_dispatch_insert(request_queue_t
*q
, struct request
*rq
)
925 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
927 cfq_remove_request(rq
);
929 elv_dispatch_sort(q
, rq
);
933 * return expired entry, or NULL to just start from scratch in rbtree
935 static inline struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
937 struct cfq_data
*cfqd
= cfqq
->cfqd
;
941 if (cfq_cfqq_fifo_expire(cfqq
))
944 cfq_mark_cfqq_fifo_expire(cfqq
);
946 if (list_empty(&cfqq
->fifo
))
949 fifo
= cfq_cfqq_sync(cfqq
);
950 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
952 if (time_before(jiffies
, rq
->start_time
+ cfqd
->cfq_fifo_expire
[fifo
]))
959 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
961 const int base_rq
= cfqd
->cfq_slice_async_rq
;
963 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
965 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
969 * Select a queue for service. If we have a current active queue,
970 * check whether to continue servicing it, or retrieve and set a new one.
972 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
974 struct cfq_queue
*cfqq
;
976 cfqq
= cfqd
->active_queue
;
981 * The active queue has run out of time, expire it and select new.
983 if (cfq_slice_used(cfqq
))
987 * The active queue has requests and isn't expired, allow it to
990 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
994 * No requests pending. If the active queue still has requests in
995 * flight or is idling for a new request, allow either of these
996 * conditions to happen (or time out) before selecting a new queue.
998 if (cfqq
->dispatched
|| timer_pending(&cfqd
->idle_slice_timer
)) {
1004 cfq_slice_expired(cfqd
, 0, 0);
1006 cfqq
= cfq_set_active_queue(cfqd
);
1012 * Dispatch some requests from cfqq, moving them to the request queue
1016 __cfq_dispatch_requests(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1021 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
1027 * follow expired path, else get first next available
1029 if ((rq
= cfq_check_fifo(cfqq
)) == NULL
)
1033 * finally, insert request into driver dispatch list
1035 cfq_dispatch_insert(cfqd
->queue
, rq
);
1037 cfqd
->dispatch_slice
++;
1040 if (!cfqd
->active_cic
) {
1041 atomic_inc(&RQ_CIC(rq
)->ioc
->refcount
);
1042 cfqd
->active_cic
= RQ_CIC(rq
);
1045 if (RB_EMPTY_ROOT(&cfqq
->sort_list
))
1048 } while (dispatched
< max_dispatch
);
1051 * expire an async queue immediately if it has used up its slice. idle
1052 * queue always expire after 1 dispatch round.
1054 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
1055 cfqd
->dispatch_slice
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
1056 cfq_class_idle(cfqq
))) {
1057 cfqq
->slice_end
= jiffies
+ 1;
1058 cfq_slice_expired(cfqd
, 0, 0);
1064 static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
1068 while (cfqq
->next_rq
) {
1069 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
1073 BUG_ON(!list_empty(&cfqq
->fifo
));
1077 static int cfq_forced_dispatch_cfqqs(struct list_head
*list
)
1079 struct cfq_queue
*cfqq
, *next
;
1083 list_for_each_entry_safe(cfqq
, next
, list
, cfq_list
)
1084 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1090 * Drain our current requests. Used for barriers and when switching
1091 * io schedulers on-the-fly.
1093 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
1098 while ((n
= cfq_rb_first(&cfqd
->service_tree
)) != NULL
) {
1099 struct cfq_queue
*cfqq
= rb_entry(n
, struct cfq_queue
, rb_node
);
1101 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1104 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->cur_rr
);
1106 cfq_slice_expired(cfqd
, 0, 0);
1108 BUG_ON(cfqd
->busy_queues
);
1113 static int cfq_dispatch_requests(request_queue_t
*q
, int force
)
1115 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1116 struct cfq_queue
*cfqq
;
1119 if (!cfqd
->busy_queues
)
1122 if (unlikely(force
))
1123 return cfq_forced_dispatch(cfqd
);
1126 while ((cfqq
= cfq_select_queue(cfqd
)) != NULL
) {
1129 if (cfqd
->busy_queues
> 1) {
1131 * So we have dispatched before in this round, if the
1132 * next queue has idling enabled (must be sync), don't
1133 * allow it service until the previous have completed.
1135 if (cfqd
->rq_in_driver
&& cfq_cfqq_idle_window(cfqq
) &&
1138 if (cfqq
->dispatched
>= cfqd
->cfq_quantum
)
1142 cfq_clear_cfqq_must_dispatch(cfqq
);
1143 cfq_clear_cfqq_wait_request(cfqq
);
1144 del_timer(&cfqd
->idle_slice_timer
);
1146 max_dispatch
= cfqd
->cfq_quantum
;
1147 if (cfq_class_idle(cfqq
))
1150 dispatched
+= __cfq_dispatch_requests(cfqd
, cfqq
, max_dispatch
);
1157 * task holds one reference to the queue, dropped when task exits. each rq
1158 * in-flight on this queue also holds a reference, dropped when rq is freed.
1160 * queue lock must be held here.
1162 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1164 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1166 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1168 if (!atomic_dec_and_test(&cfqq
->ref
))
1171 BUG_ON(rb_first(&cfqq
->sort_list
));
1172 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1173 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1175 if (unlikely(cfqd
->active_queue
== cfqq
)) {
1176 __cfq_slice_expired(cfqd
, cfqq
, 0, 0);
1177 cfq_schedule_dispatch(cfqd
);
1181 * it's on the empty list and still hashed
1183 hlist_del(&cfqq
->cfq_hash
);
1184 kmem_cache_free(cfq_pool
, cfqq
);
1187 static struct cfq_queue
*
1188 __cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned int prio
,
1191 struct hlist_head
*hash_list
= &cfqd
->cfq_hash
[hashval
];
1192 struct hlist_node
*entry
;
1193 struct cfq_queue
*__cfqq
;
1195 hlist_for_each_entry(__cfqq
, entry
, hash_list
, cfq_hash
) {
1196 const unsigned short __p
= IOPRIO_PRIO_VALUE(__cfqq
->org_ioprio_class
, __cfqq
->org_ioprio
);
1198 if (__cfqq
->key
== key
&& (__p
== prio
|| !prio
))
1205 static struct cfq_queue
*
1206 cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned short prio
)
1208 return __cfq_find_cfq_hash(cfqd
, key
, prio
, hash_long(key
, CFQ_QHASH_SHIFT
));
1211 static void cfq_free_io_context(struct io_context
*ioc
)
1213 struct cfq_io_context
*__cic
;
1217 while ((n
= rb_first(&ioc
->cic_root
)) != NULL
) {
1218 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1219 rb_erase(&__cic
->rb_node
, &ioc
->cic_root
);
1220 kmem_cache_free(cfq_ioc_pool
, __cic
);
1224 elv_ioc_count_mod(ioc_count
, -freed
);
1226 if (ioc_gone
&& !elv_ioc_count_read(ioc_count
))
1230 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1232 if (unlikely(cfqq
== cfqd
->active_queue
)) {
1233 __cfq_slice_expired(cfqd
, cfqq
, 0, 0);
1234 cfq_schedule_dispatch(cfqd
);
1237 cfq_put_queue(cfqq
);
1240 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1241 struct cfq_io_context
*cic
)
1243 list_del_init(&cic
->queue_list
);
1247 if (cic
->cfqq
[ASYNC
]) {
1248 cfq_exit_cfqq(cfqd
, cic
->cfqq
[ASYNC
]);
1249 cic
->cfqq
[ASYNC
] = NULL
;
1252 if (cic
->cfqq
[SYNC
]) {
1253 cfq_exit_cfqq(cfqd
, cic
->cfqq
[SYNC
]);
1254 cic
->cfqq
[SYNC
] = NULL
;
1258 static void cfq_exit_single_io_context(struct cfq_io_context
*cic
)
1260 struct cfq_data
*cfqd
= cic
->key
;
1263 request_queue_t
*q
= cfqd
->queue
;
1265 spin_lock_irq(q
->queue_lock
);
1266 __cfq_exit_single_io_context(cfqd
, cic
);
1267 spin_unlock_irq(q
->queue_lock
);
1272 * The process that ioc belongs to has exited, we need to clean up
1273 * and put the internal structures we have that belongs to that process.
1275 static void cfq_exit_io_context(struct io_context
*ioc
)
1277 struct cfq_io_context
*__cic
;
1281 * put the reference this task is holding to the various queues
1284 n
= rb_first(&ioc
->cic_root
);
1286 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1288 cfq_exit_single_io_context(__cic
);
1293 static struct cfq_io_context
*
1294 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1296 struct cfq_io_context
*cic
;
1298 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
, cfqd
->queue
->node
);
1300 memset(cic
, 0, sizeof(*cic
));
1301 cic
->last_end_request
= jiffies
;
1302 INIT_LIST_HEAD(&cic
->queue_list
);
1303 cic
->dtor
= cfq_free_io_context
;
1304 cic
->exit
= cfq_exit_io_context
;
1305 elv_ioc_count_inc(ioc_count
);
1311 static void cfq_init_prio_data(struct cfq_queue
*cfqq
)
1313 struct task_struct
*tsk
= current
;
1316 if (!cfq_cfqq_prio_changed(cfqq
))
1319 ioprio_class
= IOPRIO_PRIO_CLASS(tsk
->ioprio
);
1320 switch (ioprio_class
) {
1322 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1323 case IOPRIO_CLASS_NONE
:
1325 * no prio set, place us in the middle of the BE classes
1327 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1328 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1330 case IOPRIO_CLASS_RT
:
1331 cfqq
->ioprio
= task_ioprio(tsk
);
1332 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1334 case IOPRIO_CLASS_BE
:
1335 cfqq
->ioprio
= task_ioprio(tsk
);
1336 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1338 case IOPRIO_CLASS_IDLE
:
1339 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1341 cfq_clear_cfqq_idle_window(cfqq
);
1346 * keep track of original prio settings in case we have to temporarily
1347 * elevate the priority of this queue
1349 cfqq
->org_ioprio
= cfqq
->ioprio
;
1350 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1351 cfq_clear_cfqq_prio_changed(cfqq
);
1354 static inline void changed_ioprio(struct cfq_io_context
*cic
)
1356 struct cfq_data
*cfqd
= cic
->key
;
1357 struct cfq_queue
*cfqq
;
1358 unsigned long flags
;
1360 if (unlikely(!cfqd
))
1363 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1365 cfqq
= cic
->cfqq
[ASYNC
];
1367 struct cfq_queue
*new_cfqq
;
1368 new_cfqq
= cfq_get_queue(cfqd
, CFQ_KEY_ASYNC
, cic
->ioc
->task
,
1371 cic
->cfqq
[ASYNC
] = new_cfqq
;
1372 cfq_put_queue(cfqq
);
1376 cfqq
= cic
->cfqq
[SYNC
];
1378 cfq_mark_cfqq_prio_changed(cfqq
);
1380 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1383 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1385 struct cfq_io_context
*cic
;
1388 ioc
->ioprio_changed
= 0;
1390 n
= rb_first(&ioc
->cic_root
);
1392 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1394 changed_ioprio(cic
);
1399 static struct cfq_queue
*
1400 cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
,
1403 const int hashval
= hash_long(key
, CFQ_QHASH_SHIFT
);
1404 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1405 unsigned short ioprio
;
1408 ioprio
= tsk
->ioprio
;
1409 cfqq
= __cfq_find_cfq_hash(cfqd
, key
, ioprio
, hashval
);
1415 } else if (gfp_mask
& __GFP_WAIT
) {
1417 * Inform the allocator of the fact that we will
1418 * just repeat this allocation if it fails, to allow
1419 * the allocator to do whatever it needs to attempt to
1422 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1423 new_cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
|__GFP_NOFAIL
, cfqd
->queue
->node
);
1424 spin_lock_irq(cfqd
->queue
->queue_lock
);
1427 cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
, cfqd
->queue
->node
);
1432 memset(cfqq
, 0, sizeof(*cfqq
));
1434 INIT_HLIST_NODE(&cfqq
->cfq_hash
);
1435 INIT_LIST_HEAD(&cfqq
->cfq_list
);
1436 RB_CLEAR_NODE(&cfqq
->rb_node
);
1437 INIT_LIST_HEAD(&cfqq
->fifo
);
1440 hlist_add_head(&cfqq
->cfq_hash
, &cfqd
->cfq_hash
[hashval
]);
1441 atomic_set(&cfqq
->ref
, 0);
1444 if (key
!= CFQ_KEY_ASYNC
)
1445 cfq_mark_cfqq_idle_window(cfqq
);
1447 cfq_mark_cfqq_prio_changed(cfqq
);
1448 cfq_mark_cfqq_queue_new(cfqq
);
1449 cfq_init_prio_data(cfqq
);
1453 kmem_cache_free(cfq_pool
, new_cfqq
);
1455 atomic_inc(&cfqq
->ref
);
1457 WARN_ON((gfp_mask
& __GFP_WAIT
) && !cfqq
);
1462 * We drop cfq io contexts lazily, so we may find a dead one.
1465 cfq_drop_dead_cic(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1467 WARN_ON(!list_empty(&cic
->queue_list
));
1468 rb_erase(&cic
->rb_node
, &ioc
->cic_root
);
1469 kmem_cache_free(cfq_ioc_pool
, cic
);
1470 elv_ioc_count_dec(ioc_count
);
1473 static struct cfq_io_context
*
1474 cfq_cic_rb_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1477 struct cfq_io_context
*cic
;
1478 void *k
, *key
= cfqd
;
1481 n
= ioc
->cic_root
.rb_node
;
1483 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1484 /* ->key must be copied to avoid race with cfq_exit_queue() */
1487 cfq_drop_dead_cic(ioc
, cic
);
1503 cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1504 struct cfq_io_context
*cic
)
1507 struct rb_node
*parent
;
1508 struct cfq_io_context
*__cic
;
1509 unsigned long flags
;
1517 p
= &ioc
->cic_root
.rb_node
;
1520 __cic
= rb_entry(parent
, struct cfq_io_context
, rb_node
);
1521 /* ->key must be copied to avoid race with cfq_exit_queue() */
1524 cfq_drop_dead_cic(ioc
, __cic
);
1530 else if (cic
->key
> k
)
1531 p
= &(*p
)->rb_right
;
1536 rb_link_node(&cic
->rb_node
, parent
, p
);
1537 rb_insert_color(&cic
->rb_node
, &ioc
->cic_root
);
1539 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1540 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1541 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1545 * Setup general io context and cfq io context. There can be several cfq
1546 * io contexts per general io context, if this process is doing io to more
1547 * than one device managed by cfq.
1549 static struct cfq_io_context
*
1550 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1552 struct io_context
*ioc
= NULL
;
1553 struct cfq_io_context
*cic
;
1555 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1557 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
1561 cic
= cfq_cic_rb_lookup(cfqd
, ioc
);
1565 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1569 cfq_cic_link(cfqd
, ioc
, cic
);
1571 smp_read_barrier_depends();
1572 if (unlikely(ioc
->ioprio_changed
))
1573 cfq_ioc_set_ioprio(ioc
);
1577 put_io_context(ioc
);
1582 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1584 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
1585 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1587 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1588 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1589 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1593 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
1599 if (cic
->last_request_pos
< rq
->sector
)
1600 sdist
= rq
->sector
- cic
->last_request_pos
;
1602 sdist
= cic
->last_request_pos
- rq
->sector
;
1604 if (!cic
->seek_samples
) {
1605 cfqd
->new_seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1606 cfqd
->new_seek_mean
= cfqd
->new_seek_total
/ 256;
1610 * Don't allow the seek distance to get too large from the
1611 * odd fragment, pagein, etc
1613 if (cic
->seek_samples
<= 60) /* second&third seek */
1614 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1616 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1618 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1619 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1620 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1621 do_div(total
, cic
->seek_samples
);
1622 cic
->seek_mean
= (sector_t
)total
;
1626 * Disable idle window if the process thinks too long or seeks so much that
1630 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1631 struct cfq_io_context
*cic
)
1633 int enable_idle
= cfq_cfqq_idle_window(cfqq
);
1635 if (!cic
->ioc
->task
|| !cfqd
->cfq_slice_idle
||
1636 (cfqd
->hw_tag
&& CIC_SEEKY(cic
)))
1638 else if (sample_valid(cic
->ttime_samples
)) {
1639 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
1646 cfq_mark_cfqq_idle_window(cfqq
);
1648 cfq_clear_cfqq_idle_window(cfqq
);
1652 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1653 * no or if we aren't sure, a 1 will cause a preempt.
1656 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
1659 struct cfq_queue
*cfqq
;
1661 cfqq
= cfqd
->active_queue
;
1665 if (cfq_slice_used(cfqq
))
1668 if (cfq_class_idle(new_cfqq
))
1671 if (cfq_class_idle(cfqq
))
1675 * if the new request is sync, but the currently running queue is
1676 * not, let the sync request have priority.
1678 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
1682 * So both queues are sync. Let the new request get disk time if
1683 * it's a metadata request and the current queue is doing regular IO.
1685 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
1688 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
1692 * if this request is as-good as one we would expect from the
1693 * current cfqq, let it preempt
1695 if (cfq_rq_close(cfqd
, rq
))
1702 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1703 * let it have half of its nominal slice.
1705 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1707 cfq_slice_expired(cfqd
, 1, 1);
1710 * Put the new queue at the front of the of the current list,
1711 * so we know that it will be selected next.
1713 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1714 list_del_init(&cfqq
->cfq_list
);
1715 list_add(&cfqq
->cfq_list
, &cfqd
->cur_rr
);
1717 cfqq
->slice_end
= 0;
1718 cfq_mark_cfqq_slice_new(cfqq
);
1722 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1723 * something we should do about it
1726 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1729 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1732 cfqq
->meta_pending
++;
1734 cfq_update_io_thinktime(cfqd
, cic
);
1735 cfq_update_io_seektime(cfqd
, cic
, rq
);
1736 cfq_update_idle_window(cfqd
, cfqq
, cic
);
1738 cic
->last_request_pos
= rq
->sector
+ rq
->nr_sectors
;
1739 cfqq
->last_request_pos
= cic
->last_request_pos
;
1741 if (cfqq
== cfqd
->active_queue
) {
1743 * if we are waiting for a request for this queue, let it rip
1744 * immediately and flag that we must not expire this queue
1747 if (cfq_cfqq_wait_request(cfqq
)) {
1748 cfq_mark_cfqq_must_dispatch(cfqq
);
1749 del_timer(&cfqd
->idle_slice_timer
);
1750 blk_start_queueing(cfqd
->queue
);
1752 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
1754 * not the active queue - expire current slice if it is
1755 * idle and has expired it's mean thinktime or this new queue
1756 * has some old slice time left and is of higher priority
1758 cfq_preempt_queue(cfqd
, cfqq
);
1759 cfq_mark_cfqq_must_dispatch(cfqq
);
1760 blk_start_queueing(cfqd
->queue
);
1764 static void cfq_insert_request(request_queue_t
*q
, struct request
*rq
)
1766 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1767 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1769 cfq_init_prio_data(cfqq
);
1773 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
1775 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
1778 static void cfq_completed_request(request_queue_t
*q
, struct request
*rq
)
1780 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1781 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1782 const int sync
= rq_is_sync(rq
);
1787 WARN_ON(!cfqd
->rq_in_driver
);
1788 WARN_ON(!cfqq
->dispatched
);
1789 cfqd
->rq_in_driver
--;
1792 if (!cfq_class_idle(cfqq
))
1793 cfqd
->last_end_request
= now
;
1796 RQ_CIC(rq
)->last_end_request
= now
;
1799 * If this is the active queue, check if it needs to be expired,
1800 * or if we want to idle in case it has no pending requests.
1802 if (cfqd
->active_queue
== cfqq
) {
1803 if (cfq_cfqq_slice_new(cfqq
)) {
1804 cfq_set_prio_slice(cfqd
, cfqq
);
1805 cfq_clear_cfqq_slice_new(cfqq
);
1807 if (cfq_slice_used(cfqq
))
1808 cfq_slice_expired(cfqd
, 0, 1);
1809 else if (sync
&& RB_EMPTY_ROOT(&cfqq
->sort_list
))
1810 cfq_arm_slice_timer(cfqd
);
1813 if (!cfqd
->rq_in_driver
)
1814 cfq_schedule_dispatch(cfqd
);
1818 * we temporarily boost lower priority queues if they are holding fs exclusive
1819 * resources. they are boosted to normal prio (CLASS_BE/4)
1821 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
1823 if (has_fs_excl()) {
1825 * boost idle prio on transactions that would lock out other
1826 * users of the filesystem
1828 if (cfq_class_idle(cfqq
))
1829 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1830 if (cfqq
->ioprio
> IOPRIO_NORM
)
1831 cfqq
->ioprio
= IOPRIO_NORM
;
1834 * check if we need to unboost the queue
1836 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
1837 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
1838 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
1839 cfqq
->ioprio
= cfqq
->org_ioprio
;
1843 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
1845 if ((cfq_cfqq_wait_request(cfqq
) || cfq_cfqq_must_alloc(cfqq
)) &&
1846 !cfq_cfqq_must_alloc_slice(cfqq
)) {
1847 cfq_mark_cfqq_must_alloc_slice(cfqq
);
1848 return ELV_MQUEUE_MUST
;
1851 return ELV_MQUEUE_MAY
;
1854 static int cfq_may_queue(request_queue_t
*q
, int rw
)
1856 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1857 struct task_struct
*tsk
= current
;
1858 struct cfq_queue
*cfqq
;
1861 key
= cfq_queue_pid(tsk
, rw
, rw
& REQ_RW_SYNC
);
1864 * don't force setup of a queue from here, as a call to may_queue
1865 * does not necessarily imply that a request actually will be queued.
1866 * so just lookup a possibly existing queue, or return 'may queue'
1869 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
1871 cfq_init_prio_data(cfqq
);
1872 cfq_prio_boost(cfqq
);
1874 return __cfq_may_queue(cfqq
);
1877 return ELV_MQUEUE_MAY
;
1881 * queue lock held here
1883 static void cfq_put_request(struct request
*rq
)
1885 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1888 const int rw
= rq_data_dir(rq
);
1890 BUG_ON(!cfqq
->allocated
[rw
]);
1891 cfqq
->allocated
[rw
]--;
1893 put_io_context(RQ_CIC(rq
)->ioc
);
1895 rq
->elevator_private
= NULL
;
1896 rq
->elevator_private2
= NULL
;
1898 cfq_put_queue(cfqq
);
1903 * Allocate cfq data structures associated with this request.
1906 cfq_set_request(request_queue_t
*q
, struct request
*rq
, gfp_t gfp_mask
)
1908 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1909 struct task_struct
*tsk
= current
;
1910 struct cfq_io_context
*cic
;
1911 const int rw
= rq_data_dir(rq
);
1912 const int is_sync
= rq_is_sync(rq
);
1913 pid_t key
= cfq_queue_pid(tsk
, rw
, is_sync
);
1914 struct cfq_queue
*cfqq
;
1915 unsigned long flags
;
1917 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1919 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
1921 spin_lock_irqsave(q
->queue_lock
, flags
);
1926 if (!cic
->cfqq
[is_sync
]) {
1927 cfqq
= cfq_get_queue(cfqd
, key
, tsk
, gfp_mask
);
1931 cic
->cfqq
[is_sync
] = cfqq
;
1933 cfqq
= cic
->cfqq
[is_sync
];
1935 cfqq
->allocated
[rw
]++;
1936 cfq_clear_cfqq_must_alloc(cfqq
);
1937 atomic_inc(&cfqq
->ref
);
1939 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1941 rq
->elevator_private
= cic
;
1942 rq
->elevator_private2
= cfqq
;
1947 put_io_context(cic
->ioc
);
1949 cfq_schedule_dispatch(cfqd
);
1950 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1954 static void cfq_kick_queue(struct work_struct
*work
)
1956 struct cfq_data
*cfqd
=
1957 container_of(work
, struct cfq_data
, unplug_work
);
1958 request_queue_t
*q
= cfqd
->queue
;
1959 unsigned long flags
;
1961 spin_lock_irqsave(q
->queue_lock
, flags
);
1962 blk_start_queueing(q
);
1963 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1967 * Timer running if the active_queue is currently idling inside its time slice
1969 static void cfq_idle_slice_timer(unsigned long data
)
1971 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1972 struct cfq_queue
*cfqq
;
1973 unsigned long flags
;
1976 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1978 if ((cfqq
= cfqd
->active_queue
) != NULL
) {
1984 if (cfq_slice_used(cfqq
))
1988 * only expire and reinvoke request handler, if there are
1989 * other queues with pending requests
1991 if (!cfqd
->busy_queues
)
1995 * not expired and it has a request pending, let it dispatch
1997 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1998 cfq_mark_cfqq_must_dispatch(cfqq
);
2003 cfq_slice_expired(cfqd
, 0, timed_out
);
2005 cfq_schedule_dispatch(cfqd
);
2007 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2011 * Timer running if an idle class queue is waiting for service
2013 static void cfq_idle_class_timer(unsigned long data
)
2015 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
2016 unsigned long flags
, end
;
2018 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2021 * race with a non-idle queue, reset timer
2023 end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
2024 if (!time_after_eq(jiffies
, end
))
2025 mod_timer(&cfqd
->idle_class_timer
, end
);
2027 cfq_schedule_dispatch(cfqd
);
2029 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2032 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
2034 del_timer_sync(&cfqd
->idle_slice_timer
);
2035 del_timer_sync(&cfqd
->idle_class_timer
);
2036 blk_sync_queue(cfqd
->queue
);
2039 static void cfq_exit_queue(elevator_t
*e
)
2041 struct cfq_data
*cfqd
= e
->elevator_data
;
2042 request_queue_t
*q
= cfqd
->queue
;
2044 cfq_shutdown_timer_wq(cfqd
);
2046 spin_lock_irq(q
->queue_lock
);
2048 if (cfqd
->active_queue
)
2049 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0, 0);
2051 while (!list_empty(&cfqd
->cic_list
)) {
2052 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
2053 struct cfq_io_context
,
2056 __cfq_exit_single_io_context(cfqd
, cic
);
2059 spin_unlock_irq(q
->queue_lock
);
2061 cfq_shutdown_timer_wq(cfqd
);
2063 kfree(cfqd
->cfq_hash
);
2067 static void *cfq_init_queue(request_queue_t
*q
)
2069 struct cfq_data
*cfqd
;
2072 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
, q
->node
);
2076 memset(cfqd
, 0, sizeof(*cfqd
));
2078 cfqd
->service_tree
= CFQ_RB_ROOT
;
2079 INIT_LIST_HEAD(&cfqd
->cur_rr
);
2080 INIT_LIST_HEAD(&cfqd
->cic_list
);
2082 cfqd
->cfq_hash
= kmalloc_node(sizeof(struct hlist_head
) * CFQ_QHASH_ENTRIES
, GFP_KERNEL
, q
->node
);
2083 if (!cfqd
->cfq_hash
)
2086 for (i
= 0; i
< CFQ_QHASH_ENTRIES
; i
++)
2087 INIT_HLIST_HEAD(&cfqd
->cfq_hash
[i
]);
2091 init_timer(&cfqd
->idle_slice_timer
);
2092 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2093 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2095 init_timer(&cfqd
->idle_class_timer
);
2096 cfqd
->idle_class_timer
.function
= cfq_idle_class_timer
;
2097 cfqd
->idle_class_timer
.data
= (unsigned long) cfqd
;
2099 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
2101 cfqd
->cfq_quantum
= cfq_quantum
;
2102 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2103 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2104 cfqd
->cfq_back_max
= cfq_back_max
;
2105 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2106 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2107 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2108 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2109 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2117 static void cfq_slab_kill(void)
2120 kmem_cache_destroy(cfq_pool
);
2122 kmem_cache_destroy(cfq_ioc_pool
);
2125 static int __init
cfq_slab_setup(void)
2127 cfq_pool
= kmem_cache_create("cfq_pool", sizeof(struct cfq_queue
), 0, 0,
2132 cfq_ioc_pool
= kmem_cache_create("cfq_ioc_pool",
2133 sizeof(struct cfq_io_context
), 0, 0, NULL
, NULL
);
2144 * sysfs parts below -->
2147 cfq_var_show(unsigned int var
, char *page
)
2149 return sprintf(page
, "%d\n", var
);
2153 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2155 char *p
= (char *) page
;
2157 *var
= simple_strtoul(p
, &p
, 10);
2161 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2162 static ssize_t __FUNC(elevator_t *e, char *page) \
2164 struct cfq_data *cfqd = e->elevator_data; \
2165 unsigned int __data = __VAR; \
2167 __data = jiffies_to_msecs(__data); \
2168 return cfq_var_show(__data, (page)); \
2170 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2171 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2172 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2173 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2174 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2175 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2176 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2177 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2178 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2179 #undef SHOW_FUNCTION
2181 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2182 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2184 struct cfq_data *cfqd = e->elevator_data; \
2185 unsigned int __data; \
2186 int ret = cfq_var_store(&__data, (page), count); \
2187 if (__data < (MIN)) \
2189 else if (__data > (MAX)) \
2192 *(__PTR) = msecs_to_jiffies(__data); \
2194 *(__PTR) = __data; \
2197 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2198 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1, UINT_MAX
, 1);
2199 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1, UINT_MAX
, 1);
2200 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2201 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1, UINT_MAX
, 0);
2202 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2203 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2204 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2205 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1, UINT_MAX
, 0);
2206 #undef STORE_FUNCTION
2208 #define CFQ_ATTR(name) \
2209 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2211 static struct elv_fs_entry cfq_attrs
[] = {
2213 CFQ_ATTR(fifo_expire_sync
),
2214 CFQ_ATTR(fifo_expire_async
),
2215 CFQ_ATTR(back_seek_max
),
2216 CFQ_ATTR(back_seek_penalty
),
2217 CFQ_ATTR(slice_sync
),
2218 CFQ_ATTR(slice_async
),
2219 CFQ_ATTR(slice_async_rq
),
2220 CFQ_ATTR(slice_idle
),
2224 static struct elevator_type iosched_cfq
= {
2226 .elevator_merge_fn
= cfq_merge
,
2227 .elevator_merged_fn
= cfq_merged_request
,
2228 .elevator_merge_req_fn
= cfq_merged_requests
,
2229 .elevator_allow_merge_fn
= cfq_allow_merge
,
2230 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2231 .elevator_add_req_fn
= cfq_insert_request
,
2232 .elevator_activate_req_fn
= cfq_activate_request
,
2233 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2234 .elevator_queue_empty_fn
= cfq_queue_empty
,
2235 .elevator_completed_req_fn
= cfq_completed_request
,
2236 .elevator_former_req_fn
= elv_rb_former_request
,
2237 .elevator_latter_req_fn
= elv_rb_latter_request
,
2238 .elevator_set_req_fn
= cfq_set_request
,
2239 .elevator_put_req_fn
= cfq_put_request
,
2240 .elevator_may_queue_fn
= cfq_may_queue
,
2241 .elevator_init_fn
= cfq_init_queue
,
2242 .elevator_exit_fn
= cfq_exit_queue
,
2243 .trim
= cfq_free_io_context
,
2245 .elevator_attrs
= cfq_attrs
,
2246 .elevator_name
= "cfq",
2247 .elevator_owner
= THIS_MODULE
,
2250 static int __init
cfq_init(void)
2255 * could be 0 on HZ < 1000 setups
2257 if (!cfq_slice_async
)
2258 cfq_slice_async
= 1;
2259 if (!cfq_slice_idle
)
2262 if (cfq_slab_setup())
2265 ret
= elv_register(&iosched_cfq
);
2272 static void __exit
cfq_exit(void)
2274 DECLARE_COMPLETION_ONSTACK(all_gone
);
2275 elv_unregister(&iosched_cfq
);
2276 ioc_gone
= &all_gone
;
2277 /* ioc_gone's update must be visible before reading ioc_count */
2279 if (elv_ioc_count_read(ioc_count
))
2280 wait_for_completion(ioc_gone
);
2285 module_init(cfq_init
);
2286 module_exit(cfq_exit
);
2288 MODULE_AUTHOR("Jens Axboe");
2289 MODULE_LICENSE("GPL");
2290 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");