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;
29 #define CFQ_IDLE_GRACE (HZ / 10)
30 #define CFQ_SLICE_SCALE (5)
32 #define CFQ_KEY_ASYNC (0)
35 * for the hash of cfqq inside the cfqd
37 #define CFQ_QHASH_SHIFT 6
38 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
39 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
41 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
43 #define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private)
44 #define RQ_CFQQ(rq) ((rq)->elevator_private2)
46 static kmem_cache_t
*cfq_pool
;
47 static kmem_cache_t
*cfq_ioc_pool
;
49 static DEFINE_PER_CPU(unsigned long, ioc_count
);
50 static struct completion
*ioc_gone
;
52 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
53 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
54 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
59 #define cfq_cfqq_dispatched(cfqq) \
60 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
62 #define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
64 #define cfq_cfqq_sync(cfqq) \
65 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
67 #define sample_valid(samples) ((samples) > 80)
70 * Per block device queue structure
73 request_queue_t
*queue
;
76 * rr list of queues with requests and the count of them
78 struct list_head rr_list
[CFQ_PRIO_LISTS
];
79 struct list_head busy_rr
;
80 struct list_head cur_rr
;
81 struct list_head idle_rr
;
82 unsigned int busy_queues
;
87 struct hlist_head
*cfq_hash
;
93 * idle window management
95 struct timer_list idle_slice_timer
;
96 struct work_struct unplug_work
;
98 struct cfq_queue
*active_queue
;
99 struct cfq_io_context
*active_cic
;
100 int cur_prio
, cur_end_prio
;
101 unsigned int dispatch_slice
;
103 struct timer_list idle_class_timer
;
105 sector_t last_sector
;
106 unsigned long last_end_request
;
109 * tunables, see top of file
111 unsigned int cfq_quantum
;
112 unsigned int cfq_fifo_expire
[2];
113 unsigned int cfq_back_penalty
;
114 unsigned int cfq_back_max
;
115 unsigned int cfq_slice
[2];
116 unsigned int cfq_slice_async_rq
;
117 unsigned int cfq_slice_idle
;
119 struct list_head cic_list
;
123 * Per process-grouping structure
126 /* reference count */
128 /* parent cfq_data */
129 struct cfq_data
*cfqd
;
130 /* cfqq lookup hash */
131 struct hlist_node cfq_hash
;
134 /* member of the rr/busy/cur/idle cfqd list */
135 struct list_head cfq_list
;
136 /* sorted list of pending requests */
137 struct rb_root sort_list
;
138 /* if fifo isn't expired, next request to serve */
139 struct request
*next_rq
;
140 /* requests queued in sort_list */
142 /* currently allocated requests */
144 /* pending metadata requests */
146 /* fifo list of requests in sort_list */
147 struct list_head fifo
;
149 unsigned long slice_start
;
150 unsigned long slice_end
;
151 unsigned long slice_left
;
153 /* number of requests that are on the dispatch list */
156 /* io prio of this group */
157 unsigned short ioprio
, org_ioprio
;
158 unsigned short ioprio_class
, org_ioprio_class
;
160 /* various state flags, see below */
164 enum cfqq_state_flags
{
165 CFQ_CFQQ_FLAG_on_rr
= 0,
166 CFQ_CFQQ_FLAG_wait_request
,
167 CFQ_CFQQ_FLAG_must_alloc
,
168 CFQ_CFQQ_FLAG_must_alloc_slice
,
169 CFQ_CFQQ_FLAG_must_dispatch
,
170 CFQ_CFQQ_FLAG_fifo_expire
,
171 CFQ_CFQQ_FLAG_idle_window
,
172 CFQ_CFQQ_FLAG_prio_changed
,
173 CFQ_CFQQ_FLAG_queue_new
,
176 #define CFQ_CFQQ_FNS(name) \
177 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
179 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
181 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
183 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
185 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
187 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
191 CFQ_CFQQ_FNS(wait_request
);
192 CFQ_CFQQ_FNS(must_alloc
);
193 CFQ_CFQQ_FNS(must_alloc_slice
);
194 CFQ_CFQQ_FNS(must_dispatch
);
195 CFQ_CFQQ_FNS(fifo_expire
);
196 CFQ_CFQQ_FNS(idle_window
);
197 CFQ_CFQQ_FNS(prio_changed
);
198 CFQ_CFQQ_FNS(queue_new
);
201 static struct cfq_queue
*cfq_find_cfq_hash(struct cfq_data
*, unsigned int, unsigned short);
202 static void cfq_dispatch_insert(request_queue_t
*, struct request
*);
203 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
, gfp_t gfp_mask
);
206 * scheduler run of queue, if there are requests pending and no one in the
207 * driver that will restart queueing
209 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
211 if (cfqd
->busy_queues
)
212 kblockd_schedule_work(&cfqd
->unplug_work
);
215 static int cfq_queue_empty(request_queue_t
*q
)
217 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
219 return !cfqd
->busy_queues
;
222 static inline pid_t
cfq_queue_pid(struct task_struct
*task
, int rw
)
224 if (rw
== READ
|| rw
== WRITE_SYNC
)
227 return CFQ_KEY_ASYNC
;
231 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
232 * We choose the request that is closest to the head right now. Distance
233 * behind the head is penalized and only allowed to a certain extent.
235 static struct request
*
236 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
)
238 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
239 unsigned long back_max
;
240 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
241 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
242 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
244 if (rq1
== NULL
|| rq1
== rq2
)
249 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
251 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
253 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
255 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
261 last
= cfqd
->last_sector
;
264 * by definition, 1KiB is 2 sectors
266 back_max
= cfqd
->cfq_back_max
* 2;
269 * Strict one way elevator _except_ in the case where we allow
270 * short backward seeks which are biased as twice the cost of a
271 * similar forward seek.
275 else if (s1
+ back_max
>= last
)
276 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
278 wrap
|= CFQ_RQ1_WRAP
;
282 else if (s2
+ back_max
>= last
)
283 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
285 wrap
|= CFQ_RQ2_WRAP
;
287 /* Found required data */
290 * By doing switch() on the bit mask "wrap" we avoid having to
291 * check two variables for all permutations: --> faster!
294 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
310 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
313 * Since both rqs are wrapped,
314 * start with the one that's further behind head
315 * (--> only *one* back seek required),
316 * since back seek takes more time than forward.
326 * would be nice to take fifo expire time into account as well
328 static struct request
*
329 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
330 struct request
*last
)
332 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
333 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
334 struct request
*next
= NULL
, *prev
= NULL
;
336 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
339 prev
= rb_entry_rq(rbprev
);
342 next
= rb_entry_rq(rbnext
);
344 rbnext
= rb_first(&cfqq
->sort_list
);
345 if (rbnext
&& rbnext
!= &last
->rb_node
)
346 next
= rb_entry_rq(rbnext
);
349 return cfq_choose_req(cfqd
, next
, prev
);
352 static void cfq_resort_rr_list(struct cfq_queue
*cfqq
, int preempted
)
354 struct cfq_data
*cfqd
= cfqq
->cfqd
;
355 struct list_head
*list
;
357 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
359 list_del(&cfqq
->cfq_list
);
361 if (cfq_class_rt(cfqq
))
362 list
= &cfqd
->cur_rr
;
363 else if (cfq_class_idle(cfqq
))
364 list
= &cfqd
->idle_rr
;
367 * if cfqq has requests in flight, don't allow it to be
368 * found in cfq_set_active_queue before it has finished them.
369 * this is done to increase fairness between a process that
370 * has lots of io pending vs one that only generates one
371 * sporadically or synchronously
373 if (cfq_cfqq_dispatched(cfqq
))
374 list
= &cfqd
->busy_rr
;
376 list
= &cfqd
->rr_list
[cfqq
->ioprio
];
380 * If this queue was preempted or is new (never been serviced), let
381 * it be added first for fairness but beind other new queues.
382 * Otherwise, just add to the back of the list.
384 if (preempted
|| cfq_cfqq_queue_new(cfqq
)) {
385 struct list_head
*n
= list
;
386 struct cfq_queue
*__cfqq
;
388 while (n
->next
!= list
) {
389 __cfqq
= list_entry_cfqq(n
->next
);
390 if (!cfq_cfqq_queue_new(__cfqq
))
399 list_add_tail(&cfqq
->cfq_list
, list
);
403 * add to busy list of queues for service, trying to be fair in ordering
404 * the pending list according to last request service
407 cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
409 BUG_ON(cfq_cfqq_on_rr(cfqq
));
410 cfq_mark_cfqq_on_rr(cfqq
);
413 cfq_resort_rr_list(cfqq
, 0);
417 cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
419 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
420 cfq_clear_cfqq_on_rr(cfqq
);
421 list_del_init(&cfqq
->cfq_list
);
423 BUG_ON(!cfqd
->busy_queues
);
428 * rb tree support functions
430 static inline void cfq_del_rq_rb(struct request
*rq
)
432 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
433 struct cfq_data
*cfqd
= cfqq
->cfqd
;
434 const int sync
= rq_is_sync(rq
);
436 BUG_ON(!cfqq
->queued
[sync
]);
437 cfqq
->queued
[sync
]--;
439 elv_rb_del(&cfqq
->sort_list
, rq
);
441 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
442 cfq_del_cfqq_rr(cfqd
, cfqq
);
445 static void cfq_add_rq_rb(struct request
*rq
)
447 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
448 struct cfq_data
*cfqd
= cfqq
->cfqd
;
449 struct request
*__alias
;
451 cfqq
->queued
[rq_is_sync(rq
)]++;
454 * looks a little odd, but the first insert might return an alias.
455 * if that happens, put the alias on the dispatch list
457 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
458 cfq_dispatch_insert(cfqd
->queue
, __alias
);
462 cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
464 elv_rb_del(&cfqq
->sort_list
, rq
);
465 cfqq
->queued
[rq_is_sync(rq
)]--;
469 static struct request
*
470 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
472 struct task_struct
*tsk
= current
;
473 pid_t key
= cfq_queue_pid(tsk
, bio_data_dir(bio
));
474 struct cfq_queue
*cfqq
;
476 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
478 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
480 return elv_rb_find(&cfqq
->sort_list
, sector
);
486 static void cfq_activate_request(request_queue_t
*q
, struct request
*rq
)
488 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
490 cfqd
->rq_in_driver
++;
493 * If the depth is larger 1, it really could be queueing. But lets
494 * make the mark a little higher - idling could still be good for
495 * low queueing, and a low queueing number could also just indicate
496 * a SCSI mid layer like behaviour where limit+1 is often seen.
498 if (!cfqd
->hw_tag
&& cfqd
->rq_in_driver
> 4)
502 static void cfq_deactivate_request(request_queue_t
*q
, struct request
*rq
)
504 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
506 WARN_ON(!cfqd
->rq_in_driver
);
507 cfqd
->rq_in_driver
--;
510 static void cfq_remove_request(struct request
*rq
)
512 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
514 if (cfqq
->next_rq
== rq
)
515 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
517 list_del_init(&rq
->queuelist
);
520 if (rq_is_meta(rq
)) {
521 WARN_ON(!cfqq
->meta_pending
);
522 cfqq
->meta_pending
--;
527 cfq_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
529 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
530 struct request
*__rq
;
532 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
533 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
535 return ELEVATOR_FRONT_MERGE
;
538 return ELEVATOR_NO_MERGE
;
541 static void cfq_merged_request(request_queue_t
*q
, struct request
*req
,
544 if (type
== ELEVATOR_FRONT_MERGE
) {
545 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
547 cfq_reposition_rq_rb(cfqq
, req
);
552 cfq_merged_requests(request_queue_t
*q
, struct request
*rq
,
553 struct request
*next
)
556 * reposition in fifo if next is older than rq
558 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
559 time_before(next
->start_time
, rq
->start_time
))
560 list_move(&rq
->queuelist
, &next
->queuelist
);
562 cfq_remove_request(next
);
566 __cfq_set_active_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
570 * stop potential idle class queues waiting service
572 del_timer(&cfqd
->idle_class_timer
);
574 cfqq
->slice_start
= jiffies
;
576 cfqq
->slice_left
= 0;
577 cfq_clear_cfqq_must_alloc_slice(cfqq
);
578 cfq_clear_cfqq_fifo_expire(cfqq
);
581 cfqd
->active_queue
= cfqq
;
585 * current cfqq expired its slice (or was too idle), select new one
588 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
591 unsigned long now
= jiffies
;
593 if (cfq_cfqq_wait_request(cfqq
))
594 del_timer(&cfqd
->idle_slice_timer
);
596 if (!preempted
&& !cfq_cfqq_dispatched(cfqq
))
597 cfq_schedule_dispatch(cfqd
);
599 cfq_clear_cfqq_must_dispatch(cfqq
);
600 cfq_clear_cfqq_wait_request(cfqq
);
601 cfq_clear_cfqq_queue_new(cfqq
);
604 * store what was left of this slice, if the queue idled out
607 if (time_after(cfqq
->slice_end
, now
))
608 cfqq
->slice_left
= cfqq
->slice_end
- now
;
610 cfqq
->slice_left
= 0;
612 if (cfq_cfqq_on_rr(cfqq
))
613 cfq_resort_rr_list(cfqq
, preempted
);
615 if (cfqq
== cfqd
->active_queue
)
616 cfqd
->active_queue
= NULL
;
618 if (cfqd
->active_cic
) {
619 put_io_context(cfqd
->active_cic
->ioc
);
620 cfqd
->active_cic
= NULL
;
623 cfqd
->dispatch_slice
= 0;
626 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, int preempted
)
628 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
631 __cfq_slice_expired(cfqd
, cfqq
, preempted
);
644 static int cfq_get_next_prio_level(struct cfq_data
*cfqd
)
653 for (p
= cfqd
->cur_prio
; p
<= cfqd
->cur_end_prio
; p
++) {
654 if (!list_empty(&cfqd
->rr_list
[p
])) {
663 if (++cfqd
->cur_end_prio
== CFQ_PRIO_LISTS
) {
664 cfqd
->cur_end_prio
= 0;
671 if (unlikely(prio
== -1))
674 BUG_ON(prio
>= CFQ_PRIO_LISTS
);
676 list_splice_init(&cfqd
->rr_list
[prio
], &cfqd
->cur_rr
);
678 cfqd
->cur_prio
= prio
+ 1;
679 if (cfqd
->cur_prio
> cfqd
->cur_end_prio
) {
680 cfqd
->cur_end_prio
= cfqd
->cur_prio
;
683 if (cfqd
->cur_end_prio
== CFQ_PRIO_LISTS
) {
685 cfqd
->cur_end_prio
= 0;
691 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
)
693 struct cfq_queue
*cfqq
= NULL
;
695 if (!list_empty(&cfqd
->cur_rr
) || cfq_get_next_prio_level(cfqd
) != -1) {
697 * if current list is non-empty, grab first entry. if it is
698 * empty, get next prio level and grab first entry then if any
701 cfqq
= list_entry_cfqq(cfqd
->cur_rr
.next
);
702 } else if (!list_empty(&cfqd
->busy_rr
)) {
704 * If no new queues are available, check if the busy list has
705 * some before falling back to idle io.
707 cfqq
= list_entry_cfqq(cfqd
->busy_rr
.next
);
708 } else if (!list_empty(&cfqd
->idle_rr
)) {
710 * if we have idle queues and no rt or be queues had pending
711 * requests, either allow immediate service if the grace period
712 * has passed or arm the idle grace timer
714 unsigned long end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
716 if (time_after_eq(jiffies
, end
))
717 cfqq
= list_entry_cfqq(cfqd
->idle_rr
.next
);
719 mod_timer(&cfqd
->idle_class_timer
, end
);
722 __cfq_set_active_queue(cfqd
, cfqq
);
726 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
728 static int cfq_arm_slice_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
731 struct cfq_io_context
*cic
;
734 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
735 WARN_ON(cfqq
!= cfqd
->active_queue
);
738 * idle is disabled, either manually or by past process history
740 if (!cfqd
->cfq_slice_idle
)
742 if (!cfq_cfqq_idle_window(cfqq
))
745 * task has exited, don't wait
747 cic
= cfqd
->active_cic
;
748 if (!cic
|| !cic
->ioc
->task
)
751 cfq_mark_cfqq_must_dispatch(cfqq
);
752 cfq_mark_cfqq_wait_request(cfqq
);
754 sl
= min(cfqq
->slice_end
- 1, (unsigned long) cfqd
->cfq_slice_idle
);
757 * we don't want to idle for seeks, but we do want to allow
758 * fair distribution of slice time for a process doing back-to-back
759 * seeks. so allow a little bit of time for him to submit a new rq
761 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
762 sl
= min(sl
, msecs_to_jiffies(2));
764 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
768 static void cfq_dispatch_insert(request_queue_t
*q
, struct request
*rq
)
770 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
771 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
773 cfq_remove_request(rq
);
774 cfqq
->on_dispatch
[rq_is_sync(rq
)]++;
775 elv_dispatch_sort(q
, rq
);
777 rq
= list_entry(q
->queue_head
.prev
, struct request
, queuelist
);
778 cfqd
->last_sector
= rq
->sector
+ rq
->nr_sectors
;
782 * return expired entry, or NULL to just start from scratch in rbtree
784 static inline struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
786 struct cfq_data
*cfqd
= cfqq
->cfqd
;
790 if (cfq_cfqq_fifo_expire(cfqq
))
792 if (list_empty(&cfqq
->fifo
))
795 fifo
= cfq_cfqq_class_sync(cfqq
);
796 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
798 if (time_after(jiffies
, rq
->start_time
+ cfqd
->cfq_fifo_expire
[fifo
])) {
799 cfq_mark_cfqq_fifo_expire(cfqq
);
807 * Scale schedule slice based on io priority. Use the sync time slice only
808 * if a queue is marked sync and has sync io queued. A sync queue with async
809 * io only, should not get full sync slice length.
812 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
814 const int base_slice
= cfqd
->cfq_slice
[cfq_cfqq_sync(cfqq
)];
816 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
818 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - cfqq
->ioprio
));
822 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
824 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
828 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
830 const int base_rq
= cfqd
->cfq_slice_async_rq
;
832 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
834 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
838 * get next queue for service
840 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
842 unsigned long now
= jiffies
;
843 struct cfq_queue
*cfqq
;
845 cfqq
= cfqd
->active_queue
;
852 if (!cfq_cfqq_must_dispatch(cfqq
) && time_after(now
, cfqq
->slice_end
))
856 * if queue has requests, dispatch one. if not, check if
857 * enough slice is left to wait for one
859 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
861 else if (cfq_cfqq_dispatched(cfqq
)) {
864 } else if (cfq_cfqq_class_sync(cfqq
)) {
865 if (cfq_arm_slice_timer(cfqd
, cfqq
))
870 cfq_slice_expired(cfqd
, 0);
872 cfqq
= cfq_set_active_queue(cfqd
);
878 __cfq_dispatch_requests(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
883 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
889 * follow expired path, else get first next available
891 if ((rq
= cfq_check_fifo(cfqq
)) == NULL
)
895 * finally, insert request into driver dispatch list
897 cfq_dispatch_insert(cfqd
->queue
, rq
);
899 cfqd
->dispatch_slice
++;
902 if (!cfqd
->active_cic
) {
903 atomic_inc(&RQ_CIC(rq
)->ioc
->refcount
);
904 cfqd
->active_cic
= RQ_CIC(rq
);
907 if (RB_EMPTY_ROOT(&cfqq
->sort_list
))
910 } while (dispatched
< max_dispatch
);
913 * if slice end isn't set yet, set it.
915 if (!cfqq
->slice_end
)
916 cfq_set_prio_slice(cfqd
, cfqq
);
919 * expire an async queue immediately if it has used up its slice. idle
920 * queue always expire after 1 dispatch round.
922 if ((!cfq_cfqq_sync(cfqq
) &&
923 cfqd
->dispatch_slice
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
924 cfq_class_idle(cfqq
) ||
925 !cfq_cfqq_idle_window(cfqq
))
926 cfq_slice_expired(cfqd
, 0);
932 cfq_forced_dispatch_cfqqs(struct list_head
*list
)
934 struct cfq_queue
*cfqq
, *next
;
938 list_for_each_entry_safe(cfqq
, next
, list
, cfq_list
) {
939 while (cfqq
->next_rq
) {
940 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
943 BUG_ON(!list_empty(&cfqq
->fifo
));
950 cfq_forced_dispatch(struct cfq_data
*cfqd
)
952 int i
, dispatched
= 0;
954 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
955 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->rr_list
[i
]);
957 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->busy_rr
);
958 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->cur_rr
);
959 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->idle_rr
);
961 cfq_slice_expired(cfqd
, 0);
963 BUG_ON(cfqd
->busy_queues
);
969 cfq_dispatch_requests(request_queue_t
*q
, int force
)
971 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
972 struct cfq_queue
*cfqq
, *prev_cfqq
;
975 if (!cfqd
->busy_queues
)
979 return cfq_forced_dispatch(cfqd
);
983 while ((cfqq
= cfq_select_queue(cfqd
)) != NULL
) {
987 * Don't repeat dispatch from the previous queue.
989 if (prev_cfqq
== cfqq
)
992 cfq_clear_cfqq_must_dispatch(cfqq
);
993 cfq_clear_cfqq_wait_request(cfqq
);
994 del_timer(&cfqd
->idle_slice_timer
);
996 max_dispatch
= cfqd
->cfq_quantum
;
997 if (cfq_class_idle(cfqq
))
1000 dispatched
+= __cfq_dispatch_requests(cfqd
, cfqq
, max_dispatch
);
1003 * If the dispatch cfqq has idling enabled and is still
1004 * the active queue, break out.
1006 if (cfq_cfqq_idle_window(cfqq
) && cfqd
->active_queue
)
1016 * task holds one reference to the queue, dropped when task exits. each rq
1017 * in-flight on this queue also holds a reference, dropped when rq is freed.
1019 * queue lock must be held here.
1021 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1023 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1025 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1027 if (!atomic_dec_and_test(&cfqq
->ref
))
1030 BUG_ON(rb_first(&cfqq
->sort_list
));
1031 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1032 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1034 if (unlikely(cfqd
->active_queue
== cfqq
))
1035 __cfq_slice_expired(cfqd
, cfqq
, 0);
1038 * it's on the empty list and still hashed
1040 list_del(&cfqq
->cfq_list
);
1041 hlist_del(&cfqq
->cfq_hash
);
1042 kmem_cache_free(cfq_pool
, cfqq
);
1045 static struct cfq_queue
*
1046 __cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned int prio
,
1049 struct hlist_head
*hash_list
= &cfqd
->cfq_hash
[hashval
];
1050 struct hlist_node
*entry
;
1051 struct cfq_queue
*__cfqq
;
1053 hlist_for_each_entry(__cfqq
, entry
, hash_list
, cfq_hash
) {
1054 const unsigned short __p
= IOPRIO_PRIO_VALUE(__cfqq
->org_ioprio_class
, __cfqq
->org_ioprio
);
1056 if (__cfqq
->key
== key
&& (__p
== prio
|| !prio
))
1063 static struct cfq_queue
*
1064 cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned short prio
)
1066 return __cfq_find_cfq_hash(cfqd
, key
, prio
, hash_long(key
, CFQ_QHASH_SHIFT
));
1069 static void cfq_free_io_context(struct io_context
*ioc
)
1071 struct cfq_io_context
*__cic
;
1075 while ((n
= rb_first(&ioc
->cic_root
)) != NULL
) {
1076 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1077 rb_erase(&__cic
->rb_node
, &ioc
->cic_root
);
1078 kmem_cache_free(cfq_ioc_pool
, __cic
);
1082 elv_ioc_count_mod(ioc_count
, -freed
);
1084 if (ioc_gone
&& !elv_ioc_count_read(ioc_count
))
1088 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1090 if (unlikely(cfqq
== cfqd
->active_queue
))
1091 __cfq_slice_expired(cfqd
, cfqq
, 0);
1093 cfq_put_queue(cfqq
);
1096 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1097 struct cfq_io_context
*cic
)
1099 list_del_init(&cic
->queue_list
);
1103 if (cic
->cfqq
[ASYNC
]) {
1104 cfq_exit_cfqq(cfqd
, cic
->cfqq
[ASYNC
]);
1105 cic
->cfqq
[ASYNC
] = NULL
;
1108 if (cic
->cfqq
[SYNC
]) {
1109 cfq_exit_cfqq(cfqd
, cic
->cfqq
[SYNC
]);
1110 cic
->cfqq
[SYNC
] = NULL
;
1116 * Called with interrupts disabled
1118 static void cfq_exit_single_io_context(struct cfq_io_context
*cic
)
1120 struct cfq_data
*cfqd
= cic
->key
;
1123 request_queue_t
*q
= cfqd
->queue
;
1125 spin_lock_irq(q
->queue_lock
);
1126 __cfq_exit_single_io_context(cfqd
, cic
);
1127 spin_unlock_irq(q
->queue_lock
);
1131 static void cfq_exit_io_context(struct io_context
*ioc
)
1133 struct cfq_io_context
*__cic
;
1137 * put the reference this task is holding to the various queues
1140 n
= rb_first(&ioc
->cic_root
);
1142 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1144 cfq_exit_single_io_context(__cic
);
1149 static struct cfq_io_context
*
1150 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1152 struct cfq_io_context
*cic
;
1154 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
, cfqd
->queue
->node
);
1156 memset(cic
, 0, sizeof(*cic
));
1157 cic
->last_end_request
= jiffies
;
1158 INIT_LIST_HEAD(&cic
->queue_list
);
1159 cic
->dtor
= cfq_free_io_context
;
1160 cic
->exit
= cfq_exit_io_context
;
1161 elv_ioc_count_inc(ioc_count
);
1167 static void cfq_init_prio_data(struct cfq_queue
*cfqq
)
1169 struct task_struct
*tsk
= current
;
1172 if (!cfq_cfqq_prio_changed(cfqq
))
1175 ioprio_class
= IOPRIO_PRIO_CLASS(tsk
->ioprio
);
1176 switch (ioprio_class
) {
1178 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1179 case IOPRIO_CLASS_NONE
:
1181 * no prio set, place us in the middle of the BE classes
1183 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1184 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1186 case IOPRIO_CLASS_RT
:
1187 cfqq
->ioprio
= task_ioprio(tsk
);
1188 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1190 case IOPRIO_CLASS_BE
:
1191 cfqq
->ioprio
= task_ioprio(tsk
);
1192 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1194 case IOPRIO_CLASS_IDLE
:
1195 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1197 cfq_clear_cfqq_idle_window(cfqq
);
1202 * keep track of original prio settings in case we have to temporarily
1203 * elevate the priority of this queue
1205 cfqq
->org_ioprio
= cfqq
->ioprio
;
1206 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1208 if (cfq_cfqq_on_rr(cfqq
))
1209 cfq_resort_rr_list(cfqq
, 0);
1211 cfq_clear_cfqq_prio_changed(cfqq
);
1214 static inline void changed_ioprio(struct cfq_io_context
*cic
)
1216 struct cfq_data
*cfqd
= cic
->key
;
1217 struct cfq_queue
*cfqq
;
1219 if (unlikely(!cfqd
))
1222 spin_lock(cfqd
->queue
->queue_lock
);
1224 cfqq
= cic
->cfqq
[ASYNC
];
1226 struct cfq_queue
*new_cfqq
;
1227 new_cfqq
= cfq_get_queue(cfqd
, CFQ_KEY_ASYNC
, cic
->ioc
->task
,
1230 cic
->cfqq
[ASYNC
] = new_cfqq
;
1231 cfq_put_queue(cfqq
);
1235 cfqq
= cic
->cfqq
[SYNC
];
1237 cfq_mark_cfqq_prio_changed(cfqq
);
1239 spin_unlock(cfqd
->queue
->queue_lock
);
1242 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1244 struct cfq_io_context
*cic
;
1247 ioc
->ioprio_changed
= 0;
1249 n
= rb_first(&ioc
->cic_root
);
1251 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1253 changed_ioprio(cic
);
1258 static struct cfq_queue
*
1259 cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
,
1262 const int hashval
= hash_long(key
, CFQ_QHASH_SHIFT
);
1263 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1264 unsigned short ioprio
;
1267 ioprio
= tsk
->ioprio
;
1268 cfqq
= __cfq_find_cfq_hash(cfqd
, key
, ioprio
, hashval
);
1274 } else if (gfp_mask
& __GFP_WAIT
) {
1276 * Inform the allocator of the fact that we will
1277 * just repeat this allocation if it fails, to allow
1278 * the allocator to do whatever it needs to attempt to
1281 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1282 new_cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
|__GFP_NOFAIL
, cfqd
->queue
->node
);
1283 spin_lock_irq(cfqd
->queue
->queue_lock
);
1286 cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
, cfqd
->queue
->node
);
1291 memset(cfqq
, 0, sizeof(*cfqq
));
1293 INIT_HLIST_NODE(&cfqq
->cfq_hash
);
1294 INIT_LIST_HEAD(&cfqq
->cfq_list
);
1295 INIT_LIST_HEAD(&cfqq
->fifo
);
1298 hlist_add_head(&cfqq
->cfq_hash
, &cfqd
->cfq_hash
[hashval
]);
1299 atomic_set(&cfqq
->ref
, 0);
1302 * set ->slice_left to allow preemption for a new process
1304 cfqq
->slice_left
= 2 * cfqd
->cfq_slice_idle
;
1305 cfq_mark_cfqq_idle_window(cfqq
);
1306 cfq_mark_cfqq_prio_changed(cfqq
);
1307 cfq_mark_cfqq_queue_new(cfqq
);
1308 cfq_init_prio_data(cfqq
);
1312 kmem_cache_free(cfq_pool
, new_cfqq
);
1314 atomic_inc(&cfqq
->ref
);
1316 WARN_ON((gfp_mask
& __GFP_WAIT
) && !cfqq
);
1321 cfq_drop_dead_cic(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1323 WARN_ON(!list_empty(&cic
->queue_list
));
1324 rb_erase(&cic
->rb_node
, &ioc
->cic_root
);
1325 kmem_cache_free(cfq_ioc_pool
, cic
);
1326 elv_ioc_count_dec(ioc_count
);
1329 static struct cfq_io_context
*
1330 cfq_cic_rb_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1333 struct cfq_io_context
*cic
;
1334 void *k
, *key
= cfqd
;
1337 n
= ioc
->cic_root
.rb_node
;
1339 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1340 /* ->key must be copied to avoid race with cfq_exit_queue() */
1343 cfq_drop_dead_cic(ioc
, cic
);
1359 cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1360 struct cfq_io_context
*cic
)
1363 struct rb_node
*parent
;
1364 struct cfq_io_context
*__cic
;
1365 unsigned long flags
;
1373 p
= &ioc
->cic_root
.rb_node
;
1376 __cic
= rb_entry(parent
, struct cfq_io_context
, rb_node
);
1377 /* ->key must be copied to avoid race with cfq_exit_queue() */
1380 cfq_drop_dead_cic(ioc
, __cic
);
1386 else if (cic
->key
> k
)
1387 p
= &(*p
)->rb_right
;
1392 rb_link_node(&cic
->rb_node
, parent
, p
);
1393 rb_insert_color(&cic
->rb_node
, &ioc
->cic_root
);
1395 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1396 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1397 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1401 * Setup general io context and cfq io context. There can be several cfq
1402 * io contexts per general io context, if this process is doing io to more
1403 * than one device managed by cfq.
1405 static struct cfq_io_context
*
1406 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1408 struct io_context
*ioc
= NULL
;
1409 struct cfq_io_context
*cic
;
1411 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1413 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
1417 cic
= cfq_cic_rb_lookup(cfqd
, ioc
);
1421 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1425 cfq_cic_link(cfqd
, ioc
, cic
);
1427 smp_read_barrier_depends();
1428 if (unlikely(ioc
->ioprio_changed
))
1429 cfq_ioc_set_ioprio(ioc
);
1433 put_io_context(ioc
);
1438 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1440 unsigned long elapsed
, ttime
;
1443 * if this context already has stuff queued, thinktime is from
1444 * last queue not last end
1447 if (time_after(cic
->last_end_request
, cic
->last_queue
))
1448 elapsed
= jiffies
- cic
->last_end_request
;
1450 elapsed
= jiffies
- cic
->last_queue
;
1452 elapsed
= jiffies
- cic
->last_end_request
;
1455 ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1457 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1458 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1459 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1463 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
1469 if (cic
->last_request_pos
< rq
->sector
)
1470 sdist
= rq
->sector
- cic
->last_request_pos
;
1472 sdist
= cic
->last_request_pos
- rq
->sector
;
1475 * Don't allow the seek distance to get too large from the
1476 * odd fragment, pagein, etc
1478 if (cic
->seek_samples
<= 60) /* second&third seek */
1479 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1481 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1483 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1484 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1485 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1486 do_div(total
, cic
->seek_samples
);
1487 cic
->seek_mean
= (sector_t
)total
;
1491 * Disable idle window if the process thinks too long or seeks so much that
1495 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1496 struct cfq_io_context
*cic
)
1498 int enable_idle
= cfq_cfqq_idle_window(cfqq
);
1500 if (!cic
->ioc
->task
|| !cfqd
->cfq_slice_idle
||
1501 (cfqd
->hw_tag
&& CIC_SEEKY(cic
)))
1503 else if (sample_valid(cic
->ttime_samples
)) {
1504 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
1511 cfq_mark_cfqq_idle_window(cfqq
);
1513 cfq_clear_cfqq_idle_window(cfqq
);
1518 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1519 * no or if we aren't sure, a 1 will cause a preempt.
1522 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
1525 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1527 if (cfq_class_idle(new_cfqq
))
1533 if (cfq_class_idle(cfqq
))
1535 if (!cfq_cfqq_wait_request(new_cfqq
))
1538 * if it doesn't have slice left, forget it
1540 if (new_cfqq
->slice_left
< cfqd
->cfq_slice_idle
)
1543 * if the new request is sync, but the currently running queue is
1544 * not, let the sync request have priority.
1546 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
1549 * So both queues are sync. Let the new request get disk time if
1550 * it's a metadata request and the current queue is doing regular IO.
1552 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
1559 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1560 * let it have half of its nominal slice.
1562 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1564 cfq_slice_expired(cfqd
, 1);
1566 if (!cfqq
->slice_left
)
1567 cfqq
->slice_left
= cfq_prio_to_slice(cfqd
, cfqq
) / 2;
1570 * Put the new queue at the front of the of the current list,
1571 * so we know that it will be selected next.
1573 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1574 list_move(&cfqq
->cfq_list
, &cfqd
->cur_rr
);
1576 cfqq
->slice_end
= cfqq
->slice_left
+ jiffies
;
1580 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1581 * something we should do about it
1584 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1587 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1590 cfqq
->meta_pending
++;
1593 * check if this request is a better next-serve candidate)) {
1595 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
1596 BUG_ON(!cfqq
->next_rq
);
1599 * we never wait for an async request and we don't allow preemption
1600 * of an async request. so just return early
1602 if (!rq_is_sync(rq
)) {
1604 * sync process issued an async request, if it's waiting
1605 * then expire it and kick rq handling.
1607 if (cic
== cfqd
->active_cic
&&
1608 del_timer(&cfqd
->idle_slice_timer
)) {
1609 cfq_slice_expired(cfqd
, 0);
1610 blk_start_queueing(cfqd
->queue
);
1615 cfq_update_io_thinktime(cfqd
, cic
);
1616 cfq_update_io_seektime(cfqd
, cic
, rq
);
1617 cfq_update_idle_window(cfqd
, cfqq
, cic
);
1619 cic
->last_queue
= jiffies
;
1620 cic
->last_request_pos
= rq
->sector
+ rq
->nr_sectors
;
1622 if (cfqq
== cfqd
->active_queue
) {
1624 * if we are waiting for a request for this queue, let it rip
1625 * immediately and flag that we must not expire this queue
1628 if (cfq_cfqq_wait_request(cfqq
)) {
1629 cfq_mark_cfqq_must_dispatch(cfqq
);
1630 del_timer(&cfqd
->idle_slice_timer
);
1631 blk_start_queueing(cfqd
->queue
);
1633 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
1635 * not the active queue - expire current slice if it is
1636 * idle and has expired it's mean thinktime or this new queue
1637 * has some old slice time left and is of higher priority
1639 cfq_preempt_queue(cfqd
, cfqq
);
1640 cfq_mark_cfqq_must_dispatch(cfqq
);
1641 blk_start_queueing(cfqd
->queue
);
1645 static void cfq_insert_request(request_queue_t
*q
, struct request
*rq
)
1647 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1648 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1650 cfq_init_prio_data(cfqq
);
1654 if (!cfq_cfqq_on_rr(cfqq
))
1655 cfq_add_cfqq_rr(cfqd
, cfqq
);
1657 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
1659 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
1662 static void cfq_completed_request(request_queue_t
*q
, struct request
*rq
)
1664 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1665 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1666 const int sync
= rq_is_sync(rq
);
1671 WARN_ON(!cfqd
->rq_in_driver
);
1672 WARN_ON(!cfqq
->on_dispatch
[sync
]);
1673 cfqd
->rq_in_driver
--;
1674 cfqq
->on_dispatch
[sync
]--;
1676 if (!cfq_class_idle(cfqq
))
1677 cfqd
->last_end_request
= now
;
1679 if (!cfq_cfqq_dispatched(cfqq
) && cfq_cfqq_on_rr(cfqq
))
1680 cfq_resort_rr_list(cfqq
, 0);
1683 RQ_CIC(rq
)->last_end_request
= now
;
1686 * If this is the active queue, check if it needs to be expired,
1687 * or if we want to idle in case it has no pending requests.
1689 if (cfqd
->active_queue
== cfqq
) {
1690 if (time_after(now
, cfqq
->slice_end
))
1691 cfq_slice_expired(cfqd
, 0);
1692 else if (sync
&& RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1693 if (!cfq_arm_slice_timer(cfqd
, cfqq
))
1694 cfq_schedule_dispatch(cfqd
);
1700 * we temporarily boost lower priority queues if they are holding fs exclusive
1701 * resources. they are boosted to normal prio (CLASS_BE/4)
1703 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
1705 const int ioprio_class
= cfqq
->ioprio_class
;
1706 const int ioprio
= cfqq
->ioprio
;
1708 if (has_fs_excl()) {
1710 * boost idle prio on transactions that would lock out other
1711 * users of the filesystem
1713 if (cfq_class_idle(cfqq
))
1714 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1715 if (cfqq
->ioprio
> IOPRIO_NORM
)
1716 cfqq
->ioprio
= IOPRIO_NORM
;
1719 * check if we need to unboost the queue
1721 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
1722 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
1723 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
1724 cfqq
->ioprio
= cfqq
->org_ioprio
;
1728 * refile between round-robin lists if we moved the priority class
1730 if ((ioprio_class
!= cfqq
->ioprio_class
|| ioprio
!= cfqq
->ioprio
) &&
1731 cfq_cfqq_on_rr(cfqq
))
1732 cfq_resort_rr_list(cfqq
, 0);
1735 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
1737 if ((cfq_cfqq_wait_request(cfqq
) || cfq_cfqq_must_alloc(cfqq
)) &&
1738 !cfq_cfqq_must_alloc_slice(cfqq
)) {
1739 cfq_mark_cfqq_must_alloc_slice(cfqq
);
1740 return ELV_MQUEUE_MUST
;
1743 return ELV_MQUEUE_MAY
;
1746 static int cfq_may_queue(request_queue_t
*q
, int rw
)
1748 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1749 struct task_struct
*tsk
= current
;
1750 struct cfq_queue
*cfqq
;
1753 * don't force setup of a queue from here, as a call to may_queue
1754 * does not necessarily imply that a request actually will be queued.
1755 * so just lookup a possibly existing queue, or return 'may queue'
1758 cfqq
= cfq_find_cfq_hash(cfqd
, cfq_queue_pid(tsk
, rw
), tsk
->ioprio
);
1760 cfq_init_prio_data(cfqq
);
1761 cfq_prio_boost(cfqq
);
1763 return __cfq_may_queue(cfqq
);
1766 return ELV_MQUEUE_MAY
;
1770 * queue lock held here
1772 static void cfq_put_request(request_queue_t
*q
, struct request
*rq
)
1774 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1777 const int rw
= rq_data_dir(rq
);
1779 BUG_ON(!cfqq
->allocated
[rw
]);
1780 cfqq
->allocated
[rw
]--;
1782 put_io_context(RQ_CIC(rq
)->ioc
);
1784 rq
->elevator_private
= NULL
;
1785 rq
->elevator_private2
= NULL
;
1787 cfq_put_queue(cfqq
);
1792 * Allocate cfq data structures associated with this request.
1795 cfq_set_request(request_queue_t
*q
, struct request
*rq
, gfp_t gfp_mask
)
1797 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1798 struct task_struct
*tsk
= current
;
1799 struct cfq_io_context
*cic
;
1800 const int rw
= rq_data_dir(rq
);
1801 pid_t key
= cfq_queue_pid(tsk
, rw
);
1802 struct cfq_queue
*cfqq
;
1803 unsigned long flags
;
1804 int is_sync
= key
!= CFQ_KEY_ASYNC
;
1806 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1808 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
1810 spin_lock_irqsave(q
->queue_lock
, flags
);
1815 if (!cic
->cfqq
[is_sync
]) {
1816 cfqq
= cfq_get_queue(cfqd
, key
, tsk
, gfp_mask
);
1820 cic
->cfqq
[is_sync
] = cfqq
;
1822 cfqq
= cic
->cfqq
[is_sync
];
1824 cfqq
->allocated
[rw
]++;
1825 cfq_clear_cfqq_must_alloc(cfqq
);
1826 atomic_inc(&cfqq
->ref
);
1828 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1830 rq
->elevator_private
= cic
;
1831 rq
->elevator_private2
= cfqq
;
1836 put_io_context(cic
->ioc
);
1838 cfq_schedule_dispatch(cfqd
);
1839 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1843 static void cfq_kick_queue(void *data
)
1845 request_queue_t
*q
= data
;
1846 unsigned long flags
;
1848 spin_lock_irqsave(q
->queue_lock
, flags
);
1849 blk_start_queueing(q
);
1850 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1854 * Timer running if the active_queue is currently idling inside its time slice
1856 static void cfq_idle_slice_timer(unsigned long data
)
1858 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1859 struct cfq_queue
*cfqq
;
1860 unsigned long flags
;
1862 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1864 if ((cfqq
= cfqd
->active_queue
) != NULL
) {
1865 unsigned long now
= jiffies
;
1870 if (time_after(now
, cfqq
->slice_end
))
1874 * only expire and reinvoke request handler, if there are
1875 * other queues with pending requests
1877 if (!cfqd
->busy_queues
)
1881 * not expired and it has a request pending, let it dispatch
1883 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1884 cfq_mark_cfqq_must_dispatch(cfqq
);
1889 cfq_slice_expired(cfqd
, 0);
1891 cfq_schedule_dispatch(cfqd
);
1893 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1897 * Timer running if an idle class queue is waiting for service
1899 static void cfq_idle_class_timer(unsigned long data
)
1901 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1902 unsigned long flags
, end
;
1904 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1907 * race with a non-idle queue, reset timer
1909 end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
1910 if (!time_after_eq(jiffies
, end
))
1911 mod_timer(&cfqd
->idle_class_timer
, end
);
1913 cfq_schedule_dispatch(cfqd
);
1915 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1918 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
1920 del_timer_sync(&cfqd
->idle_slice_timer
);
1921 del_timer_sync(&cfqd
->idle_class_timer
);
1922 blk_sync_queue(cfqd
->queue
);
1925 static void cfq_exit_queue(elevator_t
*e
)
1927 struct cfq_data
*cfqd
= e
->elevator_data
;
1928 request_queue_t
*q
= cfqd
->queue
;
1930 cfq_shutdown_timer_wq(cfqd
);
1932 spin_lock_irq(q
->queue_lock
);
1934 if (cfqd
->active_queue
)
1935 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
1937 while (!list_empty(&cfqd
->cic_list
)) {
1938 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
1939 struct cfq_io_context
,
1942 __cfq_exit_single_io_context(cfqd
, cic
);
1945 spin_unlock_irq(q
->queue_lock
);
1947 cfq_shutdown_timer_wq(cfqd
);
1949 kfree(cfqd
->cfq_hash
);
1953 static void *cfq_init_queue(request_queue_t
*q
, elevator_t
*e
)
1955 struct cfq_data
*cfqd
;
1958 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
, q
->node
);
1962 memset(cfqd
, 0, sizeof(*cfqd
));
1964 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
1965 INIT_LIST_HEAD(&cfqd
->rr_list
[i
]);
1967 INIT_LIST_HEAD(&cfqd
->busy_rr
);
1968 INIT_LIST_HEAD(&cfqd
->cur_rr
);
1969 INIT_LIST_HEAD(&cfqd
->idle_rr
);
1970 INIT_LIST_HEAD(&cfqd
->cic_list
);
1972 cfqd
->cfq_hash
= kmalloc_node(sizeof(struct hlist_head
) * CFQ_QHASH_ENTRIES
, GFP_KERNEL
, q
->node
);
1973 if (!cfqd
->cfq_hash
)
1976 for (i
= 0; i
< CFQ_QHASH_ENTRIES
; i
++)
1977 INIT_HLIST_HEAD(&cfqd
->cfq_hash
[i
]);
1981 init_timer(&cfqd
->idle_slice_timer
);
1982 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
1983 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
1985 init_timer(&cfqd
->idle_class_timer
);
1986 cfqd
->idle_class_timer
.function
= cfq_idle_class_timer
;
1987 cfqd
->idle_class_timer
.data
= (unsigned long) cfqd
;
1989 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
, q
);
1991 cfqd
->cfq_quantum
= cfq_quantum
;
1992 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
1993 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
1994 cfqd
->cfq_back_max
= cfq_back_max
;
1995 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
1996 cfqd
->cfq_slice
[0] = cfq_slice_async
;
1997 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
1998 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
1999 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2007 static void cfq_slab_kill(void)
2010 kmem_cache_destroy(cfq_pool
);
2012 kmem_cache_destroy(cfq_ioc_pool
);
2015 static int __init
cfq_slab_setup(void)
2017 cfq_pool
= kmem_cache_create("cfq_pool", sizeof(struct cfq_queue
), 0, 0,
2022 cfq_ioc_pool
= kmem_cache_create("cfq_ioc_pool",
2023 sizeof(struct cfq_io_context
), 0, 0, NULL
, NULL
);
2034 * sysfs parts below -->
2038 cfq_var_show(unsigned int var
, char *page
)
2040 return sprintf(page
, "%d\n", var
);
2044 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2046 char *p
= (char *) page
;
2048 *var
= simple_strtoul(p
, &p
, 10);
2052 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2053 static ssize_t __FUNC(elevator_t *e, char *page) \
2055 struct cfq_data *cfqd = e->elevator_data; \
2056 unsigned int __data = __VAR; \
2058 __data = jiffies_to_msecs(__data); \
2059 return cfq_var_show(__data, (page)); \
2061 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2062 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2063 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2064 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2065 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2066 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2067 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2068 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2069 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2070 #undef SHOW_FUNCTION
2072 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2073 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2075 struct cfq_data *cfqd = e->elevator_data; \
2076 unsigned int __data; \
2077 int ret = cfq_var_store(&__data, (page), count); \
2078 if (__data < (MIN)) \
2080 else if (__data > (MAX)) \
2083 *(__PTR) = msecs_to_jiffies(__data); \
2085 *(__PTR) = __data; \
2088 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2089 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1, UINT_MAX
, 1);
2090 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1, UINT_MAX
, 1);
2091 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2092 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1, UINT_MAX
, 0);
2093 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2094 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2095 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2096 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1, UINT_MAX
, 0);
2097 #undef STORE_FUNCTION
2099 #define CFQ_ATTR(name) \
2100 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2102 static struct elv_fs_entry cfq_attrs
[] = {
2104 CFQ_ATTR(fifo_expire_sync
),
2105 CFQ_ATTR(fifo_expire_async
),
2106 CFQ_ATTR(back_seek_max
),
2107 CFQ_ATTR(back_seek_penalty
),
2108 CFQ_ATTR(slice_sync
),
2109 CFQ_ATTR(slice_async
),
2110 CFQ_ATTR(slice_async_rq
),
2111 CFQ_ATTR(slice_idle
),
2115 static struct elevator_type iosched_cfq
= {
2117 .elevator_merge_fn
= cfq_merge
,
2118 .elevator_merged_fn
= cfq_merged_request
,
2119 .elevator_merge_req_fn
= cfq_merged_requests
,
2120 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2121 .elevator_add_req_fn
= cfq_insert_request
,
2122 .elevator_activate_req_fn
= cfq_activate_request
,
2123 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2124 .elevator_queue_empty_fn
= cfq_queue_empty
,
2125 .elevator_completed_req_fn
= cfq_completed_request
,
2126 .elevator_former_req_fn
= elv_rb_former_request
,
2127 .elevator_latter_req_fn
= elv_rb_latter_request
,
2128 .elevator_set_req_fn
= cfq_set_request
,
2129 .elevator_put_req_fn
= cfq_put_request
,
2130 .elevator_may_queue_fn
= cfq_may_queue
,
2131 .elevator_init_fn
= cfq_init_queue
,
2132 .elevator_exit_fn
= cfq_exit_queue
,
2133 .trim
= cfq_free_io_context
,
2135 .elevator_attrs
= cfq_attrs
,
2136 .elevator_name
= "cfq",
2137 .elevator_owner
= THIS_MODULE
,
2140 static int __init
cfq_init(void)
2145 * could be 0 on HZ < 1000 setups
2147 if (!cfq_slice_async
)
2148 cfq_slice_async
= 1;
2149 if (!cfq_slice_idle
)
2152 if (cfq_slab_setup())
2155 ret
= elv_register(&iosched_cfq
);
2162 static void __exit
cfq_exit(void)
2164 DECLARE_COMPLETION_ONSTACK(all_gone
);
2165 elv_unregister(&iosched_cfq
);
2166 ioc_gone
= &all_gone
;
2167 /* ioc_gone's update must be visible before reading ioc_count */
2169 if (elv_ioc_count_read(ioc_count
))
2170 wait_for_completion(ioc_gone
);
2175 module_init(cfq_init
);
2176 module_exit(cfq_exit
);
2178 MODULE_AUTHOR("Jens Axboe");
2179 MODULE_LICENSE("GPL");
2180 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");