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)
49 #define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private)
50 #define RQ_CFQQ(rq) ((rq)->elevator_private2)
52 static struct kmem_cache
*cfq_pool
;
53 static struct kmem_cache
*cfq_ioc_pool
;
55 static DEFINE_PER_CPU(unsigned long, ioc_count
);
56 static struct completion
*ioc_gone
;
58 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
59 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
60 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
65 #define cfq_cfqq_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
67 #define sample_valid(samples) ((samples) > 80)
70 * Most of our rbtree usage is for sorting with min extraction, so
71 * if we cache the leftmost node we don't have to walk down the tree
72 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
73 * move this into the elevator for the rq sorting as well.
79 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
82 * Per block device queue structure
85 request_queue_t
*queue
;
88 * rr list of queues with requests and the count of them
90 struct cfq_rb_root service_tree
;
91 unsigned int busy_queues
;
96 struct hlist_head
*cfq_hash
;
102 * idle window management
104 struct timer_list idle_slice_timer
;
105 struct work_struct unplug_work
;
107 struct cfq_queue
*active_queue
;
108 struct cfq_io_context
*active_cic
;
109 unsigned int dispatch_slice
;
111 struct timer_list idle_class_timer
;
113 sector_t last_position
;
114 unsigned long last_end_request
;
117 * tunables, see top of file
119 unsigned int cfq_quantum
;
120 unsigned int cfq_fifo_expire
[2];
121 unsigned int cfq_back_penalty
;
122 unsigned int cfq_back_max
;
123 unsigned int cfq_slice
[2];
124 unsigned int cfq_slice_async_rq
;
125 unsigned int cfq_slice_idle
;
127 struct list_head cic_list
;
129 sector_t new_seek_mean
;
134 * Per process-grouping structure
137 /* reference count */
139 /* parent cfq_data */
140 struct cfq_data
*cfqd
;
141 /* cfqq lookup hash */
142 struct hlist_node cfq_hash
;
145 /* service_tree member */
146 struct rb_node rb_node
;
147 /* service_tree key */
148 unsigned long rb_key
;
149 /* sorted list of pending requests */
150 struct rb_root sort_list
;
151 /* if fifo isn't expired, next request to serve */
152 struct request
*next_rq
;
153 /* requests queued in sort_list */
155 /* currently allocated requests */
157 /* pending metadata requests */
159 /* fifo list of requests in sort_list */
160 struct list_head fifo
;
162 unsigned long slice_end
;
165 /* number of requests that are on the dispatch list or inside driver */
168 /* io prio of this group */
169 unsigned short ioprio
, org_ioprio
;
170 unsigned short ioprio_class
, org_ioprio_class
;
172 /* various state flags, see below */
175 sector_t last_request_pos
;
178 enum cfqq_state_flags
{
179 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
180 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
181 CFQ_CFQQ_FLAG_must_alloc
, /* must be allowed rq alloc */
182 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
183 CFQ_CFQQ_FLAG_must_dispatch
, /* must dispatch, even if expired */
184 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
185 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
186 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
187 CFQ_CFQQ_FLAG_queue_new
, /* queue never been serviced */
188 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
191 #define CFQ_CFQQ_FNS(name) \
192 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
194 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
196 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
198 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
200 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
202 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
206 CFQ_CFQQ_FNS(wait_request
);
207 CFQ_CFQQ_FNS(must_alloc
);
208 CFQ_CFQQ_FNS(must_alloc_slice
);
209 CFQ_CFQQ_FNS(must_dispatch
);
210 CFQ_CFQQ_FNS(fifo_expire
);
211 CFQ_CFQQ_FNS(idle_window
);
212 CFQ_CFQQ_FNS(prio_changed
);
213 CFQ_CFQQ_FNS(queue_new
);
214 CFQ_CFQQ_FNS(slice_new
);
217 static struct cfq_queue
*cfq_find_cfq_hash(struct cfq_data
*, unsigned int, unsigned short);
218 static void cfq_dispatch_insert(request_queue_t
*, struct request
*);
219 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, unsigned int, struct task_struct
*, gfp_t
);
222 * scheduler run of queue, if there are requests pending and no one in the
223 * driver that will restart queueing
225 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
227 if (cfqd
->busy_queues
)
228 kblockd_schedule_work(&cfqd
->unplug_work
);
231 static int cfq_queue_empty(request_queue_t
*q
)
233 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
235 return !cfqd
->busy_queues
;
238 static inline pid_t
cfq_queue_pid(struct task_struct
*task
, int rw
, int is_sync
)
241 * Use the per-process queue, for read requests and syncronous writes
243 if (!(rw
& REQ_RW
) || is_sync
)
246 return CFQ_KEY_ASYNC
;
250 * Scale schedule slice based on io priority. Use the sync time slice only
251 * if a queue is marked sync and has sync io queued. A sync queue with async
252 * io only, should not get full sync slice length.
254 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, int sync
,
257 const int base_slice
= cfqd
->cfq_slice
[sync
];
259 WARN_ON(prio
>= IOPRIO_BE_NR
);
261 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
265 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
267 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
271 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
273 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
277 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
278 * isn't valid until the first request from the dispatch is activated
279 * and the slice time set.
281 static inline int cfq_slice_used(struct cfq_queue
*cfqq
)
283 if (cfq_cfqq_slice_new(cfqq
))
285 if (time_before(jiffies
, cfqq
->slice_end
))
292 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
293 * We choose the request that is closest to the head right now. Distance
294 * behind the head is penalized and only allowed to a certain extent.
296 static struct request
*
297 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
)
299 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
300 unsigned long back_max
;
301 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
302 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
303 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
305 if (rq1
== NULL
|| rq1
== rq2
)
310 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
312 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
314 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
316 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
322 last
= cfqd
->last_position
;
325 * by definition, 1KiB is 2 sectors
327 back_max
= cfqd
->cfq_back_max
* 2;
330 * Strict one way elevator _except_ in the case where we allow
331 * short backward seeks which are biased as twice the cost of a
332 * similar forward seek.
336 else if (s1
+ back_max
>= last
)
337 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
339 wrap
|= CFQ_RQ1_WRAP
;
343 else if (s2
+ back_max
>= last
)
344 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
346 wrap
|= CFQ_RQ2_WRAP
;
348 /* Found required data */
351 * By doing switch() on the bit mask "wrap" we avoid having to
352 * check two variables for all permutations: --> faster!
355 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
371 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
374 * Since both rqs are wrapped,
375 * start with the one that's further behind head
376 * (--> only *one* back seek required),
377 * since back seek takes more time than forward.
387 * The below is leftmost cache rbtree addon
389 static struct rb_node
*cfq_rb_first(struct cfq_rb_root
*root
)
392 root
->left
= rb_first(&root
->rb
);
397 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
402 rb_erase(n
, &root
->rb
);
407 * would be nice to take fifo expire time into account as well
409 static struct request
*
410 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
411 struct request
*last
)
413 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
414 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
415 struct request
*next
= NULL
, *prev
= NULL
;
417 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
420 prev
= rb_entry_rq(rbprev
);
423 next
= rb_entry_rq(rbnext
);
425 rbnext
= rb_first(&cfqq
->sort_list
);
426 if (rbnext
&& rbnext
!= &last
->rb_node
)
427 next
= rb_entry_rq(rbnext
);
430 return cfq_choose_req(cfqd
, next
, prev
);
433 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
434 struct cfq_queue
*cfqq
)
437 * just an approximation, should be ok.
439 return (cfqd
->busy_queues
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
440 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
444 * The cfqd->service_tree holds all pending cfq_queue's that have
445 * requests waiting to be processed. It is sorted in the order that
446 * we will service the queues.
448 static void cfq_service_tree_add(struct cfq_data
*cfqd
,
449 struct cfq_queue
*cfqq
, int add_front
)
451 struct rb_node
**p
= &cfqd
->service_tree
.rb
.rb_node
;
452 struct rb_node
*parent
= NULL
;
453 unsigned long rb_key
;
457 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
458 rb_key
+= cfqq
->slice_resid
;
459 cfqq
->slice_resid
= 0;
463 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
465 * same position, nothing more to do
467 if (rb_key
== cfqq
->rb_key
)
470 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
475 struct cfq_queue
*__cfqq
;
479 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
482 * sort RT queues first, we always want to give
483 * preference to them. IDLE queues goes to the back.
484 * after that, sort on the next service time.
486 if (cfq_class_rt(cfqq
) > cfq_class_rt(__cfqq
))
488 else if (cfq_class_rt(cfqq
) < cfq_class_rt(__cfqq
))
490 else if (cfq_class_idle(cfqq
) < cfq_class_idle(__cfqq
))
492 else if (cfq_class_idle(cfqq
) > cfq_class_idle(__cfqq
))
494 else if (rb_key
< __cfqq
->rb_key
)
499 if (n
== &(*p
)->rb_right
)
506 cfqd
->service_tree
.left
= &cfqq
->rb_node
;
508 cfqq
->rb_key
= rb_key
;
509 rb_link_node(&cfqq
->rb_node
, parent
, p
);
510 rb_insert_color(&cfqq
->rb_node
, &cfqd
->service_tree
.rb
);
514 * Update cfqq's position in the service tree.
516 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
519 * Resorting requires the cfqq to be on the RR list already.
521 if (cfq_cfqq_on_rr(cfqq
))
522 cfq_service_tree_add(cfqd
, cfqq
, 0);
526 * add to busy list of queues for service, trying to be fair in ordering
527 * the pending list according to last request service
530 cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
532 BUG_ON(cfq_cfqq_on_rr(cfqq
));
533 cfq_mark_cfqq_on_rr(cfqq
);
536 cfq_resort_rr_list(cfqd
, cfqq
);
540 * Called when the cfqq no longer has requests pending, remove it from
544 cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
546 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
547 cfq_clear_cfqq_on_rr(cfqq
);
549 if (!RB_EMPTY_NODE(&cfqq
->rb_node
))
550 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
552 BUG_ON(!cfqd
->busy_queues
);
557 * rb tree support functions
559 static inline void cfq_del_rq_rb(struct request
*rq
)
561 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
562 struct cfq_data
*cfqd
= cfqq
->cfqd
;
563 const int sync
= rq_is_sync(rq
);
565 BUG_ON(!cfqq
->queued
[sync
]);
566 cfqq
->queued
[sync
]--;
568 elv_rb_del(&cfqq
->sort_list
, rq
);
570 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
571 cfq_del_cfqq_rr(cfqd
, cfqq
);
574 static void cfq_add_rq_rb(struct request
*rq
)
576 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
577 struct cfq_data
*cfqd
= cfqq
->cfqd
;
578 struct request
*__alias
;
580 cfqq
->queued
[rq_is_sync(rq
)]++;
583 * looks a little odd, but the first insert might return an alias.
584 * if that happens, put the alias on the dispatch list
586 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
587 cfq_dispatch_insert(cfqd
->queue
, __alias
);
589 if (!cfq_cfqq_on_rr(cfqq
))
590 cfq_add_cfqq_rr(cfqd
, cfqq
);
593 * check if this request is a better next-serve candidate
595 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
596 BUG_ON(!cfqq
->next_rq
);
600 cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
602 elv_rb_del(&cfqq
->sort_list
, rq
);
603 cfqq
->queued
[rq_is_sync(rq
)]--;
607 static struct request
*
608 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
610 struct task_struct
*tsk
= current
;
611 pid_t key
= cfq_queue_pid(tsk
, bio_data_dir(bio
), bio_sync(bio
));
612 struct cfq_queue
*cfqq
;
614 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
616 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
618 return elv_rb_find(&cfqq
->sort_list
, sector
);
624 static void cfq_activate_request(request_queue_t
*q
, struct request
*rq
)
626 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
628 cfqd
->rq_in_driver
++;
631 * If the depth is larger 1, it really could be queueing. But lets
632 * make the mark a little higher - idling could still be good for
633 * low queueing, and a low queueing number could also just indicate
634 * a SCSI mid layer like behaviour where limit+1 is often seen.
636 if (!cfqd
->hw_tag
&& cfqd
->rq_in_driver
> 4)
639 cfqd
->last_position
= rq
->hard_sector
+ rq
->hard_nr_sectors
;
642 static void cfq_deactivate_request(request_queue_t
*q
, struct request
*rq
)
644 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
646 WARN_ON(!cfqd
->rq_in_driver
);
647 cfqd
->rq_in_driver
--;
650 static void cfq_remove_request(struct request
*rq
)
652 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
654 if (cfqq
->next_rq
== rq
)
655 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
657 list_del_init(&rq
->queuelist
);
660 if (rq_is_meta(rq
)) {
661 WARN_ON(!cfqq
->meta_pending
);
662 cfqq
->meta_pending
--;
666 static int cfq_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
668 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
669 struct request
*__rq
;
671 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
672 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
674 return ELEVATOR_FRONT_MERGE
;
677 return ELEVATOR_NO_MERGE
;
680 static void cfq_merged_request(request_queue_t
*q
, struct request
*req
,
683 if (type
== ELEVATOR_FRONT_MERGE
) {
684 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
686 cfq_reposition_rq_rb(cfqq
, req
);
691 cfq_merged_requests(request_queue_t
*q
, struct request
*rq
,
692 struct request
*next
)
695 * reposition in fifo if next is older than rq
697 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
698 time_before(next
->start_time
, rq
->start_time
))
699 list_move(&rq
->queuelist
, &next
->queuelist
);
701 cfq_remove_request(next
);
704 static int cfq_allow_merge(request_queue_t
*q
, struct request
*rq
,
707 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
708 const int rw
= bio_data_dir(bio
);
709 struct cfq_queue
*cfqq
;
713 * Disallow merge of a sync bio into an async request.
715 if ((bio_data_dir(bio
) == READ
|| bio_sync(bio
)) && !rq_is_sync(rq
))
719 * Lookup the cfqq that this bio will be queued with. Allow
720 * merge only if rq is queued there.
722 key
= cfq_queue_pid(current
, rw
, bio_sync(bio
));
723 cfqq
= cfq_find_cfq_hash(cfqd
, key
, current
->ioprio
);
725 if (cfqq
== RQ_CFQQ(rq
))
732 __cfq_set_active_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
736 * stop potential idle class queues waiting service
738 del_timer(&cfqd
->idle_class_timer
);
741 cfq_clear_cfqq_must_alloc_slice(cfqq
);
742 cfq_clear_cfqq_fifo_expire(cfqq
);
743 cfq_mark_cfqq_slice_new(cfqq
);
744 cfq_clear_cfqq_queue_new(cfqq
);
747 cfqd
->active_queue
= cfqq
;
751 * current cfqq expired its slice (or was too idle), select new one
754 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
757 if (cfq_cfqq_wait_request(cfqq
))
758 del_timer(&cfqd
->idle_slice_timer
);
760 cfq_clear_cfqq_must_dispatch(cfqq
);
761 cfq_clear_cfqq_wait_request(cfqq
);
764 * store what was left of this slice, if the queue idled/timed out
766 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
))
767 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
769 cfq_resort_rr_list(cfqd
, cfqq
);
771 if (cfqq
== cfqd
->active_queue
)
772 cfqd
->active_queue
= NULL
;
774 if (cfqd
->active_cic
) {
775 put_io_context(cfqd
->active_cic
->ioc
);
776 cfqd
->active_cic
= NULL
;
779 cfqd
->dispatch_slice
= 0;
782 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, int timed_out
)
784 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
787 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
791 * Get next queue for service. Unless we have a queue preemption,
792 * we'll simply select the first cfqq in the service tree.
794 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
796 struct cfq_queue
*cfqq
;
799 if (RB_EMPTY_ROOT(&cfqd
->service_tree
.rb
))
802 n
= cfq_rb_first(&cfqd
->service_tree
);
803 cfqq
= rb_entry(n
, struct cfq_queue
, rb_node
);
805 if (cfq_class_idle(cfqq
)) {
809 * if we have idle queues and no rt or be queues had
810 * pending requests, either allow immediate service if
811 * the grace period has passed or arm the idle grace
814 end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
815 if (time_before(jiffies
, end
)) {
816 mod_timer(&cfqd
->idle_class_timer
, end
);
825 * Get and set a new active queue for service.
827 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
)
829 struct cfq_queue
*cfqq
;
831 cfqq
= cfq_get_next_queue(cfqd
);
832 __cfq_set_active_queue(cfqd
, cfqq
);
836 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
839 if (rq
->sector
>= cfqd
->last_position
)
840 return rq
->sector
- cfqd
->last_position
;
842 return cfqd
->last_position
- rq
->sector
;
845 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct request
*rq
)
847 struct cfq_io_context
*cic
= cfqd
->active_cic
;
849 if (!sample_valid(cic
->seek_samples
))
852 return cfq_dist_from_last(cfqd
, rq
) <= cic
->seek_mean
;
855 static int cfq_close_cooperator(struct cfq_data
*cfq_data
,
856 struct cfq_queue
*cfqq
)
859 * We should notice if some of the queues are cooperating, eg
860 * working closely on the same area of the disk. In that case,
861 * we can group them together and don't waste time idling.
866 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
868 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
870 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
871 struct cfq_io_context
*cic
;
874 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
875 WARN_ON(cfq_cfqq_slice_new(cfqq
));
878 * idle is disabled, either manually or by past process history
880 if (!cfqd
->cfq_slice_idle
|| !cfq_cfqq_idle_window(cfqq
))
884 * task has exited, don't wait
886 cic
= cfqd
->active_cic
;
887 if (!cic
|| !cic
->ioc
->task
)
891 * See if this prio level has a good candidate
893 if (cfq_close_cooperator(cfqd
, cfqq
) &&
894 (sample_valid(cic
->ttime_samples
) && cic
->ttime_mean
> 2))
897 cfq_mark_cfqq_must_dispatch(cfqq
);
898 cfq_mark_cfqq_wait_request(cfqq
);
901 * we don't want to idle for seeks, but we do want to allow
902 * fair distribution of slice time for a process doing back-to-back
903 * seeks. so allow a little bit of time for him to submit a new rq
905 sl
= cfqd
->cfq_slice_idle
;
906 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
907 sl
= min(sl
, msecs_to_jiffies(CFQ_MIN_TT
));
909 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
913 * Move request from internal lists to the request queue dispatch list.
915 static void cfq_dispatch_insert(request_queue_t
*q
, struct request
*rq
)
917 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
919 cfq_remove_request(rq
);
921 elv_dispatch_sort(q
, rq
);
925 * return expired entry, or NULL to just start from scratch in rbtree
927 static inline struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
929 struct cfq_data
*cfqd
= cfqq
->cfqd
;
933 if (cfq_cfqq_fifo_expire(cfqq
))
936 cfq_mark_cfqq_fifo_expire(cfqq
);
938 if (list_empty(&cfqq
->fifo
))
941 fifo
= cfq_cfqq_sync(cfqq
);
942 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
944 if (time_before(jiffies
, rq
->start_time
+ cfqd
->cfq_fifo_expire
[fifo
]))
951 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
953 const int base_rq
= cfqd
->cfq_slice_async_rq
;
955 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
957 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
961 * Select a queue for service. If we have a current active queue,
962 * check whether to continue servicing it, or retrieve and set a new one.
964 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
966 struct cfq_queue
*cfqq
;
968 cfqq
= cfqd
->active_queue
;
973 * The active queue has run out of time, expire it and select new.
975 if (cfq_slice_used(cfqq
))
979 * The active queue has requests and isn't expired, allow it to
982 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
986 * No requests pending. If the active queue still has requests in
987 * flight or is idling for a new request, allow either of these
988 * conditions to happen (or time out) before selecting a new queue.
990 if (cfqq
->dispatched
|| timer_pending(&cfqd
->idle_slice_timer
)) {
996 cfq_slice_expired(cfqd
, 0);
998 cfqq
= cfq_set_active_queue(cfqd
);
1004 * Dispatch some requests from cfqq, moving them to the request queue
1008 __cfq_dispatch_requests(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1013 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
1019 * follow expired path, else get first next available
1021 if ((rq
= cfq_check_fifo(cfqq
)) == NULL
)
1025 * finally, insert request into driver dispatch list
1027 cfq_dispatch_insert(cfqd
->queue
, rq
);
1029 cfqd
->dispatch_slice
++;
1032 if (!cfqd
->active_cic
) {
1033 atomic_inc(&RQ_CIC(rq
)->ioc
->refcount
);
1034 cfqd
->active_cic
= RQ_CIC(rq
);
1037 if (RB_EMPTY_ROOT(&cfqq
->sort_list
))
1040 } while (dispatched
< max_dispatch
);
1043 * expire an async queue immediately if it has used up its slice. idle
1044 * queue always expire after 1 dispatch round.
1046 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
1047 cfqd
->dispatch_slice
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
1048 cfq_class_idle(cfqq
))) {
1049 cfqq
->slice_end
= jiffies
+ 1;
1050 cfq_slice_expired(cfqd
, 0);
1056 static inline int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
1060 while (cfqq
->next_rq
) {
1061 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
1065 BUG_ON(!list_empty(&cfqq
->fifo
));
1070 * Drain our current requests. Used for barriers and when switching
1071 * io schedulers on-the-fly.
1073 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
1078 while ((n
= cfq_rb_first(&cfqd
->service_tree
)) != NULL
) {
1079 struct cfq_queue
*cfqq
= rb_entry(n
, struct cfq_queue
, rb_node
);
1081 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1084 cfq_slice_expired(cfqd
, 0);
1086 BUG_ON(cfqd
->busy_queues
);
1091 static int cfq_dispatch_requests(request_queue_t
*q
, int force
)
1093 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1094 struct cfq_queue
*cfqq
;
1097 if (!cfqd
->busy_queues
)
1100 if (unlikely(force
))
1101 return cfq_forced_dispatch(cfqd
);
1104 while ((cfqq
= cfq_select_queue(cfqd
)) != NULL
) {
1107 if (cfqd
->busy_queues
> 1) {
1109 * So we have dispatched before in this round, if the
1110 * next queue has idling enabled (must be sync), don't
1111 * allow it service until the previous have completed.
1113 if (cfqd
->rq_in_driver
&& cfq_cfqq_idle_window(cfqq
) &&
1116 if (cfqq
->dispatched
>= cfqd
->cfq_quantum
)
1120 cfq_clear_cfqq_must_dispatch(cfqq
);
1121 cfq_clear_cfqq_wait_request(cfqq
);
1122 del_timer(&cfqd
->idle_slice_timer
);
1124 max_dispatch
= cfqd
->cfq_quantum
;
1125 if (cfq_class_idle(cfqq
))
1128 dispatched
+= __cfq_dispatch_requests(cfqd
, cfqq
, max_dispatch
);
1135 * task holds one reference to the queue, dropped when task exits. each rq
1136 * in-flight on this queue also holds a reference, dropped when rq is freed.
1138 * queue lock must be held here.
1140 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1142 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1144 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1146 if (!atomic_dec_and_test(&cfqq
->ref
))
1149 BUG_ON(rb_first(&cfqq
->sort_list
));
1150 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1151 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1153 if (unlikely(cfqd
->active_queue
== cfqq
)) {
1154 __cfq_slice_expired(cfqd
, cfqq
, 0);
1155 cfq_schedule_dispatch(cfqd
);
1159 * it's on the empty list and still hashed
1161 hlist_del(&cfqq
->cfq_hash
);
1162 kmem_cache_free(cfq_pool
, cfqq
);
1165 static struct cfq_queue
*
1166 __cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned int prio
,
1169 struct hlist_head
*hash_list
= &cfqd
->cfq_hash
[hashval
];
1170 struct hlist_node
*entry
;
1171 struct cfq_queue
*__cfqq
;
1173 hlist_for_each_entry(__cfqq
, entry
, hash_list
, cfq_hash
) {
1174 const unsigned short __p
= IOPRIO_PRIO_VALUE(__cfqq
->org_ioprio_class
, __cfqq
->org_ioprio
);
1176 if (__cfqq
->key
== key
&& (__p
== prio
|| !prio
))
1183 static struct cfq_queue
*
1184 cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned short prio
)
1186 return __cfq_find_cfq_hash(cfqd
, key
, prio
, hash_long(key
, CFQ_QHASH_SHIFT
));
1189 static void cfq_free_io_context(struct io_context
*ioc
)
1191 struct cfq_io_context
*__cic
;
1195 while ((n
= rb_first(&ioc
->cic_root
)) != NULL
) {
1196 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1197 rb_erase(&__cic
->rb_node
, &ioc
->cic_root
);
1198 kmem_cache_free(cfq_ioc_pool
, __cic
);
1202 elv_ioc_count_mod(ioc_count
, -freed
);
1204 if (ioc_gone
&& !elv_ioc_count_read(ioc_count
))
1208 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1210 if (unlikely(cfqq
== cfqd
->active_queue
)) {
1211 __cfq_slice_expired(cfqd
, cfqq
, 0);
1212 cfq_schedule_dispatch(cfqd
);
1215 cfq_put_queue(cfqq
);
1218 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1219 struct cfq_io_context
*cic
)
1221 list_del_init(&cic
->queue_list
);
1225 if (cic
->cfqq
[ASYNC
]) {
1226 cfq_exit_cfqq(cfqd
, cic
->cfqq
[ASYNC
]);
1227 cic
->cfqq
[ASYNC
] = NULL
;
1230 if (cic
->cfqq
[SYNC
]) {
1231 cfq_exit_cfqq(cfqd
, cic
->cfqq
[SYNC
]);
1232 cic
->cfqq
[SYNC
] = NULL
;
1236 static void cfq_exit_single_io_context(struct cfq_io_context
*cic
)
1238 struct cfq_data
*cfqd
= cic
->key
;
1241 request_queue_t
*q
= cfqd
->queue
;
1243 spin_lock_irq(q
->queue_lock
);
1244 __cfq_exit_single_io_context(cfqd
, cic
);
1245 spin_unlock_irq(q
->queue_lock
);
1250 * The process that ioc belongs to has exited, we need to clean up
1251 * and put the internal structures we have that belongs to that process.
1253 static void cfq_exit_io_context(struct io_context
*ioc
)
1255 struct cfq_io_context
*__cic
;
1259 * put the reference this task is holding to the various queues
1262 n
= rb_first(&ioc
->cic_root
);
1264 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1266 cfq_exit_single_io_context(__cic
);
1271 static struct cfq_io_context
*
1272 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1274 struct cfq_io_context
*cic
;
1276 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
, cfqd
->queue
->node
);
1278 memset(cic
, 0, sizeof(*cic
));
1279 cic
->last_end_request
= jiffies
;
1280 INIT_LIST_HEAD(&cic
->queue_list
);
1281 cic
->dtor
= cfq_free_io_context
;
1282 cic
->exit
= cfq_exit_io_context
;
1283 elv_ioc_count_inc(ioc_count
);
1289 static void cfq_init_prio_data(struct cfq_queue
*cfqq
)
1291 struct task_struct
*tsk
= current
;
1294 if (!cfq_cfqq_prio_changed(cfqq
))
1297 ioprio_class
= IOPRIO_PRIO_CLASS(tsk
->ioprio
);
1298 switch (ioprio_class
) {
1300 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1301 case IOPRIO_CLASS_NONE
:
1303 * no prio set, place us in the middle of the BE classes
1305 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1306 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1308 case IOPRIO_CLASS_RT
:
1309 cfqq
->ioprio
= task_ioprio(tsk
);
1310 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1312 case IOPRIO_CLASS_BE
:
1313 cfqq
->ioprio
= task_ioprio(tsk
);
1314 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1316 case IOPRIO_CLASS_IDLE
:
1317 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1319 cfq_clear_cfqq_idle_window(cfqq
);
1324 * keep track of original prio settings in case we have to temporarily
1325 * elevate the priority of this queue
1327 cfqq
->org_ioprio
= cfqq
->ioprio
;
1328 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1329 cfq_clear_cfqq_prio_changed(cfqq
);
1332 static inline void changed_ioprio(struct cfq_io_context
*cic
)
1334 struct cfq_data
*cfqd
= cic
->key
;
1335 struct cfq_queue
*cfqq
;
1336 unsigned long flags
;
1338 if (unlikely(!cfqd
))
1341 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1343 cfqq
= cic
->cfqq
[ASYNC
];
1345 struct cfq_queue
*new_cfqq
;
1346 new_cfqq
= cfq_get_queue(cfqd
, CFQ_KEY_ASYNC
, cic
->ioc
->task
,
1349 cic
->cfqq
[ASYNC
] = new_cfqq
;
1350 cfq_put_queue(cfqq
);
1354 cfqq
= cic
->cfqq
[SYNC
];
1356 cfq_mark_cfqq_prio_changed(cfqq
);
1358 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1361 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1363 struct cfq_io_context
*cic
;
1366 ioc
->ioprio_changed
= 0;
1368 n
= rb_first(&ioc
->cic_root
);
1370 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1372 changed_ioprio(cic
);
1377 static struct cfq_queue
*
1378 cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
,
1381 const int hashval
= hash_long(key
, CFQ_QHASH_SHIFT
);
1382 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1383 unsigned short ioprio
;
1386 ioprio
= tsk
->ioprio
;
1387 cfqq
= __cfq_find_cfq_hash(cfqd
, key
, ioprio
, hashval
);
1393 } else if (gfp_mask
& __GFP_WAIT
) {
1395 * Inform the allocator of the fact that we will
1396 * just repeat this allocation if it fails, to allow
1397 * the allocator to do whatever it needs to attempt to
1400 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1401 new_cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
|__GFP_NOFAIL
, cfqd
->queue
->node
);
1402 spin_lock_irq(cfqd
->queue
->queue_lock
);
1405 cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
, cfqd
->queue
->node
);
1410 memset(cfqq
, 0, sizeof(*cfqq
));
1412 INIT_HLIST_NODE(&cfqq
->cfq_hash
);
1413 RB_CLEAR_NODE(&cfqq
->rb_node
);
1414 INIT_LIST_HEAD(&cfqq
->fifo
);
1417 hlist_add_head(&cfqq
->cfq_hash
, &cfqd
->cfq_hash
[hashval
]);
1418 atomic_set(&cfqq
->ref
, 0);
1421 if (key
!= CFQ_KEY_ASYNC
)
1422 cfq_mark_cfqq_idle_window(cfqq
);
1424 cfq_mark_cfqq_prio_changed(cfqq
);
1425 cfq_mark_cfqq_queue_new(cfqq
);
1426 cfq_init_prio_data(cfqq
);
1430 kmem_cache_free(cfq_pool
, new_cfqq
);
1432 atomic_inc(&cfqq
->ref
);
1434 WARN_ON((gfp_mask
& __GFP_WAIT
) && !cfqq
);
1439 * We drop cfq io contexts lazily, so we may find a dead one.
1442 cfq_drop_dead_cic(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1444 WARN_ON(!list_empty(&cic
->queue_list
));
1445 rb_erase(&cic
->rb_node
, &ioc
->cic_root
);
1446 kmem_cache_free(cfq_ioc_pool
, cic
);
1447 elv_ioc_count_dec(ioc_count
);
1450 static struct cfq_io_context
*
1451 cfq_cic_rb_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1454 struct cfq_io_context
*cic
;
1455 void *k
, *key
= cfqd
;
1458 n
= ioc
->cic_root
.rb_node
;
1460 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1461 /* ->key must be copied to avoid race with cfq_exit_queue() */
1464 cfq_drop_dead_cic(ioc
, cic
);
1480 cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1481 struct cfq_io_context
*cic
)
1484 struct rb_node
*parent
;
1485 struct cfq_io_context
*__cic
;
1486 unsigned long flags
;
1494 p
= &ioc
->cic_root
.rb_node
;
1497 __cic
= rb_entry(parent
, struct cfq_io_context
, rb_node
);
1498 /* ->key must be copied to avoid race with cfq_exit_queue() */
1501 cfq_drop_dead_cic(ioc
, __cic
);
1507 else if (cic
->key
> k
)
1508 p
= &(*p
)->rb_right
;
1513 rb_link_node(&cic
->rb_node
, parent
, p
);
1514 rb_insert_color(&cic
->rb_node
, &ioc
->cic_root
);
1516 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1517 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1518 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1522 * Setup general io context and cfq io context. There can be several cfq
1523 * io contexts per general io context, if this process is doing io to more
1524 * than one device managed by cfq.
1526 static struct cfq_io_context
*
1527 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1529 struct io_context
*ioc
= NULL
;
1530 struct cfq_io_context
*cic
;
1532 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1534 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
1538 cic
= cfq_cic_rb_lookup(cfqd
, ioc
);
1542 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1546 cfq_cic_link(cfqd
, ioc
, cic
);
1548 smp_read_barrier_depends();
1549 if (unlikely(ioc
->ioprio_changed
))
1550 cfq_ioc_set_ioprio(ioc
);
1554 put_io_context(ioc
);
1559 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1561 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
1562 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1564 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1565 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1566 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1570 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
1576 if (cic
->last_request_pos
< rq
->sector
)
1577 sdist
= rq
->sector
- cic
->last_request_pos
;
1579 sdist
= cic
->last_request_pos
- rq
->sector
;
1581 if (!cic
->seek_samples
) {
1582 cfqd
->new_seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1583 cfqd
->new_seek_mean
= cfqd
->new_seek_total
/ 256;
1587 * Don't allow the seek distance to get too large from the
1588 * odd fragment, pagein, etc
1590 if (cic
->seek_samples
<= 60) /* second&third seek */
1591 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1593 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1595 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1596 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1597 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1598 do_div(total
, cic
->seek_samples
);
1599 cic
->seek_mean
= (sector_t
)total
;
1603 * Disable idle window if the process thinks too long or seeks so much that
1607 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1608 struct cfq_io_context
*cic
)
1610 int enable_idle
= cfq_cfqq_idle_window(cfqq
);
1612 if (!cic
->ioc
->task
|| !cfqd
->cfq_slice_idle
||
1613 (cfqd
->hw_tag
&& CIC_SEEKY(cic
)))
1615 else if (sample_valid(cic
->ttime_samples
)) {
1616 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
1623 cfq_mark_cfqq_idle_window(cfqq
);
1625 cfq_clear_cfqq_idle_window(cfqq
);
1629 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1630 * no or if we aren't sure, a 1 will cause a preempt.
1633 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
1636 struct cfq_queue
*cfqq
;
1638 cfqq
= cfqd
->active_queue
;
1642 if (cfq_slice_used(cfqq
))
1645 if (cfq_class_idle(new_cfqq
))
1648 if (cfq_class_idle(cfqq
))
1652 * if the new request is sync, but the currently running queue is
1653 * not, let the sync request have priority.
1655 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
1659 * So both queues are sync. Let the new request get disk time if
1660 * it's a metadata request and the current queue is doing regular IO.
1662 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
1665 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
1669 * if this request is as-good as one we would expect from the
1670 * current cfqq, let it preempt
1672 if (cfq_rq_close(cfqd
, rq
))
1679 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1680 * let it have half of its nominal slice.
1682 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1684 cfq_slice_expired(cfqd
, 1);
1687 * Put the new queue at the front of the of the current list,
1688 * so we know that it will be selected next.
1690 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1692 cfq_service_tree_add(cfqd
, cfqq
, 1);
1694 cfqq
->slice_end
= 0;
1695 cfq_mark_cfqq_slice_new(cfqq
);
1699 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1700 * something we should do about it
1703 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1706 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1709 cfqq
->meta_pending
++;
1711 cfq_update_io_thinktime(cfqd
, cic
);
1712 cfq_update_io_seektime(cfqd
, cic
, rq
);
1713 cfq_update_idle_window(cfqd
, cfqq
, cic
);
1715 cic
->last_request_pos
= rq
->sector
+ rq
->nr_sectors
;
1716 cfqq
->last_request_pos
= cic
->last_request_pos
;
1718 if (cfqq
== cfqd
->active_queue
) {
1720 * if we are waiting for a request for this queue, let it rip
1721 * immediately and flag that we must not expire this queue
1724 if (cfq_cfqq_wait_request(cfqq
)) {
1725 cfq_mark_cfqq_must_dispatch(cfqq
);
1726 del_timer(&cfqd
->idle_slice_timer
);
1727 blk_start_queueing(cfqd
->queue
);
1729 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
1731 * not the active queue - expire current slice if it is
1732 * idle and has expired it's mean thinktime or this new queue
1733 * has some old slice time left and is of higher priority
1735 cfq_preempt_queue(cfqd
, cfqq
);
1736 cfq_mark_cfqq_must_dispatch(cfqq
);
1737 blk_start_queueing(cfqd
->queue
);
1741 static void cfq_insert_request(request_queue_t
*q
, struct request
*rq
)
1743 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1744 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1746 cfq_init_prio_data(cfqq
);
1750 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
1752 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
1755 static void cfq_completed_request(request_queue_t
*q
, struct request
*rq
)
1757 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1758 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1759 const int sync
= rq_is_sync(rq
);
1764 WARN_ON(!cfqd
->rq_in_driver
);
1765 WARN_ON(!cfqq
->dispatched
);
1766 cfqd
->rq_in_driver
--;
1769 if (!cfq_class_idle(cfqq
))
1770 cfqd
->last_end_request
= now
;
1773 RQ_CIC(rq
)->last_end_request
= now
;
1776 * If this is the active queue, check if it needs to be expired,
1777 * or if we want to idle in case it has no pending requests.
1779 if (cfqd
->active_queue
== cfqq
) {
1780 if (cfq_cfqq_slice_new(cfqq
)) {
1781 cfq_set_prio_slice(cfqd
, cfqq
);
1782 cfq_clear_cfqq_slice_new(cfqq
);
1784 if (cfq_slice_used(cfqq
))
1785 cfq_slice_expired(cfqd
, 1);
1786 else if (sync
&& RB_EMPTY_ROOT(&cfqq
->sort_list
))
1787 cfq_arm_slice_timer(cfqd
);
1790 if (!cfqd
->rq_in_driver
)
1791 cfq_schedule_dispatch(cfqd
);
1795 * we temporarily boost lower priority queues if they are holding fs exclusive
1796 * resources. they are boosted to normal prio (CLASS_BE/4)
1798 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
1800 if (has_fs_excl()) {
1802 * boost idle prio on transactions that would lock out other
1803 * users of the filesystem
1805 if (cfq_class_idle(cfqq
))
1806 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1807 if (cfqq
->ioprio
> IOPRIO_NORM
)
1808 cfqq
->ioprio
= IOPRIO_NORM
;
1811 * check if we need to unboost the queue
1813 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
1814 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
1815 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
1816 cfqq
->ioprio
= cfqq
->org_ioprio
;
1820 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
1822 if ((cfq_cfqq_wait_request(cfqq
) || cfq_cfqq_must_alloc(cfqq
)) &&
1823 !cfq_cfqq_must_alloc_slice(cfqq
)) {
1824 cfq_mark_cfqq_must_alloc_slice(cfqq
);
1825 return ELV_MQUEUE_MUST
;
1828 return ELV_MQUEUE_MAY
;
1831 static int cfq_may_queue(request_queue_t
*q
, int rw
)
1833 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1834 struct task_struct
*tsk
= current
;
1835 struct cfq_queue
*cfqq
;
1838 key
= cfq_queue_pid(tsk
, rw
, rw
& REQ_RW_SYNC
);
1841 * don't force setup of a queue from here, as a call to may_queue
1842 * does not necessarily imply that a request actually will be queued.
1843 * so just lookup a possibly existing queue, or return 'may queue'
1846 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
1848 cfq_init_prio_data(cfqq
);
1849 cfq_prio_boost(cfqq
);
1851 return __cfq_may_queue(cfqq
);
1854 return ELV_MQUEUE_MAY
;
1858 * queue lock held here
1860 static void cfq_put_request(struct request
*rq
)
1862 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1865 const int rw
= rq_data_dir(rq
);
1867 BUG_ON(!cfqq
->allocated
[rw
]);
1868 cfqq
->allocated
[rw
]--;
1870 put_io_context(RQ_CIC(rq
)->ioc
);
1872 rq
->elevator_private
= NULL
;
1873 rq
->elevator_private2
= NULL
;
1875 cfq_put_queue(cfqq
);
1880 * Allocate cfq data structures associated with this request.
1883 cfq_set_request(request_queue_t
*q
, struct request
*rq
, gfp_t gfp_mask
)
1885 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1886 struct task_struct
*tsk
= current
;
1887 struct cfq_io_context
*cic
;
1888 const int rw
= rq_data_dir(rq
);
1889 const int is_sync
= rq_is_sync(rq
);
1890 pid_t key
= cfq_queue_pid(tsk
, rw
, is_sync
);
1891 struct cfq_queue
*cfqq
;
1892 unsigned long flags
;
1894 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1896 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
1898 spin_lock_irqsave(q
->queue_lock
, flags
);
1903 if (!cic
->cfqq
[is_sync
]) {
1904 cfqq
= cfq_get_queue(cfqd
, key
, tsk
, gfp_mask
);
1908 cic
->cfqq
[is_sync
] = cfqq
;
1910 cfqq
= cic
->cfqq
[is_sync
];
1912 cfqq
->allocated
[rw
]++;
1913 cfq_clear_cfqq_must_alloc(cfqq
);
1914 atomic_inc(&cfqq
->ref
);
1916 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1918 rq
->elevator_private
= cic
;
1919 rq
->elevator_private2
= cfqq
;
1924 put_io_context(cic
->ioc
);
1926 cfq_schedule_dispatch(cfqd
);
1927 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1931 static void cfq_kick_queue(struct work_struct
*work
)
1933 struct cfq_data
*cfqd
=
1934 container_of(work
, struct cfq_data
, unplug_work
);
1935 request_queue_t
*q
= cfqd
->queue
;
1936 unsigned long flags
;
1938 spin_lock_irqsave(q
->queue_lock
, flags
);
1939 blk_start_queueing(q
);
1940 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1944 * Timer running if the active_queue is currently idling inside its time slice
1946 static void cfq_idle_slice_timer(unsigned long data
)
1948 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1949 struct cfq_queue
*cfqq
;
1950 unsigned long flags
;
1953 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1955 if ((cfqq
= cfqd
->active_queue
) != NULL
) {
1961 if (cfq_slice_used(cfqq
))
1965 * only expire and reinvoke request handler, if there are
1966 * other queues with pending requests
1968 if (!cfqd
->busy_queues
)
1972 * not expired and it has a request pending, let it dispatch
1974 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1975 cfq_mark_cfqq_must_dispatch(cfqq
);
1980 cfq_slice_expired(cfqd
, timed_out
);
1982 cfq_schedule_dispatch(cfqd
);
1984 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1988 * Timer running if an idle class queue is waiting for service
1990 static void cfq_idle_class_timer(unsigned long data
)
1992 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1993 unsigned long flags
, end
;
1995 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1998 * race with a non-idle queue, reset timer
2000 end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
2001 if (!time_after_eq(jiffies
, end
))
2002 mod_timer(&cfqd
->idle_class_timer
, end
);
2004 cfq_schedule_dispatch(cfqd
);
2006 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2009 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
2011 del_timer_sync(&cfqd
->idle_slice_timer
);
2012 del_timer_sync(&cfqd
->idle_class_timer
);
2013 blk_sync_queue(cfqd
->queue
);
2016 static void cfq_exit_queue(elevator_t
*e
)
2018 struct cfq_data
*cfqd
= e
->elevator_data
;
2019 request_queue_t
*q
= cfqd
->queue
;
2021 cfq_shutdown_timer_wq(cfqd
);
2023 spin_lock_irq(q
->queue_lock
);
2025 if (cfqd
->active_queue
)
2026 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
2028 while (!list_empty(&cfqd
->cic_list
)) {
2029 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
2030 struct cfq_io_context
,
2033 __cfq_exit_single_io_context(cfqd
, cic
);
2036 spin_unlock_irq(q
->queue_lock
);
2038 cfq_shutdown_timer_wq(cfqd
);
2040 kfree(cfqd
->cfq_hash
);
2044 static void *cfq_init_queue(request_queue_t
*q
)
2046 struct cfq_data
*cfqd
;
2049 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
, q
->node
);
2053 memset(cfqd
, 0, sizeof(*cfqd
));
2055 cfqd
->service_tree
= CFQ_RB_ROOT
;
2056 INIT_LIST_HEAD(&cfqd
->cic_list
);
2058 cfqd
->cfq_hash
= kmalloc_node(sizeof(struct hlist_head
) * CFQ_QHASH_ENTRIES
, GFP_KERNEL
, q
->node
);
2059 if (!cfqd
->cfq_hash
)
2062 for (i
= 0; i
< CFQ_QHASH_ENTRIES
; i
++)
2063 INIT_HLIST_HEAD(&cfqd
->cfq_hash
[i
]);
2067 init_timer(&cfqd
->idle_slice_timer
);
2068 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2069 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2071 init_timer(&cfqd
->idle_class_timer
);
2072 cfqd
->idle_class_timer
.function
= cfq_idle_class_timer
;
2073 cfqd
->idle_class_timer
.data
= (unsigned long) cfqd
;
2075 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
2077 cfqd
->cfq_quantum
= cfq_quantum
;
2078 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2079 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2080 cfqd
->cfq_back_max
= cfq_back_max
;
2081 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2082 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2083 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2084 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2085 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2093 static void cfq_slab_kill(void)
2096 kmem_cache_destroy(cfq_pool
);
2098 kmem_cache_destroy(cfq_ioc_pool
);
2101 static int __init
cfq_slab_setup(void)
2103 cfq_pool
= kmem_cache_create("cfq_pool", sizeof(struct cfq_queue
), 0, 0,
2108 cfq_ioc_pool
= kmem_cache_create("cfq_ioc_pool",
2109 sizeof(struct cfq_io_context
), 0, 0, NULL
, NULL
);
2120 * sysfs parts below -->
2123 cfq_var_show(unsigned int var
, char *page
)
2125 return sprintf(page
, "%d\n", var
);
2129 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2131 char *p
= (char *) page
;
2133 *var
= simple_strtoul(p
, &p
, 10);
2137 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2138 static ssize_t __FUNC(elevator_t *e, char *page) \
2140 struct cfq_data *cfqd = e->elevator_data; \
2141 unsigned int __data = __VAR; \
2143 __data = jiffies_to_msecs(__data); \
2144 return cfq_var_show(__data, (page)); \
2146 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2147 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2148 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2149 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2150 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2151 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2152 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2153 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2154 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2155 #undef SHOW_FUNCTION
2157 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2158 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2160 struct cfq_data *cfqd = e->elevator_data; \
2161 unsigned int __data; \
2162 int ret = cfq_var_store(&__data, (page), count); \
2163 if (__data < (MIN)) \
2165 else if (__data > (MAX)) \
2168 *(__PTR) = msecs_to_jiffies(__data); \
2170 *(__PTR) = __data; \
2173 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2174 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1, UINT_MAX
, 1);
2175 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1, UINT_MAX
, 1);
2176 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2177 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1, UINT_MAX
, 0);
2178 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2179 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2180 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2181 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1, UINT_MAX
, 0);
2182 #undef STORE_FUNCTION
2184 #define CFQ_ATTR(name) \
2185 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2187 static struct elv_fs_entry cfq_attrs
[] = {
2189 CFQ_ATTR(fifo_expire_sync
),
2190 CFQ_ATTR(fifo_expire_async
),
2191 CFQ_ATTR(back_seek_max
),
2192 CFQ_ATTR(back_seek_penalty
),
2193 CFQ_ATTR(slice_sync
),
2194 CFQ_ATTR(slice_async
),
2195 CFQ_ATTR(slice_async_rq
),
2196 CFQ_ATTR(slice_idle
),
2200 static struct elevator_type iosched_cfq
= {
2202 .elevator_merge_fn
= cfq_merge
,
2203 .elevator_merged_fn
= cfq_merged_request
,
2204 .elevator_merge_req_fn
= cfq_merged_requests
,
2205 .elevator_allow_merge_fn
= cfq_allow_merge
,
2206 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2207 .elevator_add_req_fn
= cfq_insert_request
,
2208 .elevator_activate_req_fn
= cfq_activate_request
,
2209 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2210 .elevator_queue_empty_fn
= cfq_queue_empty
,
2211 .elevator_completed_req_fn
= cfq_completed_request
,
2212 .elevator_former_req_fn
= elv_rb_former_request
,
2213 .elevator_latter_req_fn
= elv_rb_latter_request
,
2214 .elevator_set_req_fn
= cfq_set_request
,
2215 .elevator_put_req_fn
= cfq_put_request
,
2216 .elevator_may_queue_fn
= cfq_may_queue
,
2217 .elevator_init_fn
= cfq_init_queue
,
2218 .elevator_exit_fn
= cfq_exit_queue
,
2219 .trim
= cfq_free_io_context
,
2221 .elevator_attrs
= cfq_attrs
,
2222 .elevator_name
= "cfq",
2223 .elevator_owner
= THIS_MODULE
,
2226 static int __init
cfq_init(void)
2231 * could be 0 on HZ < 1000 setups
2233 if (!cfq_slice_async
)
2234 cfq_slice_async
= 1;
2235 if (!cfq_slice_idle
)
2238 if (cfq_slab_setup())
2241 ret
= elv_register(&iosched_cfq
);
2248 static void __exit
cfq_exit(void)
2250 DECLARE_COMPLETION_ONSTACK(all_gone
);
2251 elv_unregister(&iosched_cfq
);
2252 ioc_gone
= &all_gone
;
2253 /* ioc_gone's update must be visible before reading ioc_count */
2255 if (elv_ioc_count_read(ioc_count
))
2256 wait_for_completion(ioc_gone
);
2261 module_init(cfq_init
);
2262 module_exit(cfq_exit
);
2264 MODULE_AUTHOR("Jens Axboe");
2265 MODULE_LICENSE("GPL");
2266 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");