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 struct kmem_cache
*cfq_pool
;
47 static struct kmem_cache
*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_end
;
150 unsigned long service_last
;
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, /* on round-robin busy list */
166 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
167 CFQ_CFQQ_FLAG_must_alloc
, /* must be allowed rq alloc */
168 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
169 CFQ_CFQQ_FLAG_must_dispatch
, /* must dispatch, even if expired */
170 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
171 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
172 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
173 CFQ_CFQQ_FLAG_queue_new
, /* queue never been serviced */
174 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
177 #define CFQ_CFQQ_FNS(name) \
178 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
180 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
182 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
184 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
186 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
188 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
192 CFQ_CFQQ_FNS(wait_request
);
193 CFQ_CFQQ_FNS(must_alloc
);
194 CFQ_CFQQ_FNS(must_alloc_slice
);
195 CFQ_CFQQ_FNS(must_dispatch
);
196 CFQ_CFQQ_FNS(fifo_expire
);
197 CFQ_CFQQ_FNS(idle_window
);
198 CFQ_CFQQ_FNS(prio_changed
);
199 CFQ_CFQQ_FNS(queue_new
);
200 CFQ_CFQQ_FNS(slice_new
);
203 static struct cfq_queue
*cfq_find_cfq_hash(struct cfq_data
*, unsigned int, unsigned short);
204 static void cfq_dispatch_insert(request_queue_t
*, struct request
*);
205 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
, gfp_t gfp_mask
);
208 * scheduler run of queue, if there are requests pending and no one in the
209 * driver that will restart queueing
211 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
213 if (cfqd
->busy_queues
)
214 kblockd_schedule_work(&cfqd
->unplug_work
);
217 static int cfq_queue_empty(request_queue_t
*q
)
219 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
221 return !cfqd
->busy_queues
;
224 static inline pid_t
cfq_queue_pid(struct task_struct
*task
, int rw
, int is_sync
)
227 * Use the per-process queue, for read requests and syncronous writes
229 if (!(rw
& REQ_RW
) || is_sync
)
232 return CFQ_KEY_ASYNC
;
236 * Scale schedule slice based on io priority. Use the sync time slice only
237 * if a queue is marked sync and has sync io queued. A sync queue with async
238 * io only, should not get full sync slice length.
241 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
243 const int base_slice
= cfqd
->cfq_slice
[cfq_cfqq_sync(cfqq
)];
245 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
247 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - cfqq
->ioprio
));
251 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
253 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
254 cfqq
->slice_end
+= cfqq
->slice_resid
;
257 * Don't carry over residual for more than one slice, we only want
258 * to slightly correct the fairness. Carrying over forever would
259 * easily introduce oscillations.
261 cfqq
->slice_resid
= 0;
265 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
266 * isn't valid until the first request from the dispatch is activated
267 * and the slice time set.
269 static inline int cfq_slice_used(struct cfq_queue
*cfqq
)
271 if (cfq_cfqq_slice_new(cfqq
))
273 if (time_before(jiffies
, cfqq
->slice_end
))
280 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
281 * We choose the request that is closest to the head right now. Distance
282 * behind the head is penalized and only allowed to a certain extent.
284 static struct request
*
285 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
)
287 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
288 unsigned long back_max
;
289 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
290 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
291 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
293 if (rq1
== NULL
|| rq1
== rq2
)
298 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
300 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
302 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
304 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
310 last
= cfqd
->last_sector
;
313 * by definition, 1KiB is 2 sectors
315 back_max
= cfqd
->cfq_back_max
* 2;
318 * Strict one way elevator _except_ in the case where we allow
319 * short backward seeks which are biased as twice the cost of a
320 * similar forward seek.
324 else if (s1
+ back_max
>= last
)
325 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
327 wrap
|= CFQ_RQ1_WRAP
;
331 else if (s2
+ back_max
>= last
)
332 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
334 wrap
|= CFQ_RQ2_WRAP
;
336 /* Found required data */
339 * By doing switch() on the bit mask "wrap" we avoid having to
340 * check two variables for all permutations: --> faster!
343 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
359 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
362 * Since both rqs are wrapped,
363 * start with the one that's further behind head
364 * (--> only *one* back seek required),
365 * since back seek takes more time than forward.
375 * would be nice to take fifo expire time into account as well
377 static struct request
*
378 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
379 struct request
*last
)
381 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
382 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
383 struct request
*next
= NULL
, *prev
= NULL
;
385 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
388 prev
= rb_entry_rq(rbprev
);
391 next
= rb_entry_rq(rbnext
);
393 rbnext
= rb_first(&cfqq
->sort_list
);
394 if (rbnext
&& rbnext
!= &last
->rb_node
)
395 next
= rb_entry_rq(rbnext
);
398 return cfq_choose_req(cfqd
, next
, prev
);
401 static void cfq_resort_rr_list(struct cfq_queue
*cfqq
, int preempted
)
403 struct cfq_data
*cfqd
= cfqq
->cfqd
;
404 struct list_head
*list
, *n
;
405 struct cfq_queue
*__cfqq
;
408 * Resorting requires the cfqq to be on the RR list already.
410 if (!cfq_cfqq_on_rr(cfqq
))
413 list_del(&cfqq
->cfq_list
);
415 if (cfq_class_rt(cfqq
))
416 list
= &cfqd
->cur_rr
;
417 else if (cfq_class_idle(cfqq
))
418 list
= &cfqd
->idle_rr
;
421 * if cfqq has requests in flight, don't allow it to be
422 * found in cfq_set_active_queue before it has finished them.
423 * this is done to increase fairness between a process that
424 * has lots of io pending vs one that only generates one
425 * sporadically or synchronously
427 if (cfq_cfqq_dispatched(cfqq
))
428 list
= &cfqd
->busy_rr
;
430 list
= &cfqd
->rr_list
[cfqq
->ioprio
];
433 if (preempted
|| cfq_cfqq_queue_new(cfqq
)) {
435 * If this queue was preempted or is new (never been serviced),
436 * let it be added first for fairness but beind other new
440 while (n
->next
!= list
) {
441 __cfqq
= list_entry_cfqq(n
->next
);
442 if (!cfq_cfqq_queue_new(__cfqq
))
447 list_add_tail(&cfqq
->cfq_list
, n
);
448 } else if (!cfq_cfqq_class_sync(cfqq
)) {
450 * async queue always goes to the end. this wont be overly
451 * unfair to writes, as the sort of the sync queue wont be
452 * allowed to pass the async queue again.
454 list_add_tail(&cfqq
->cfq_list
, list
);
457 * sort by last service, but don't cross a new or async
458 * queue. we don't cross a new queue because it hasn't been
459 * service before, and we don't cross an async queue because
460 * it gets added to the end on expire.
463 while ((n
= n
->prev
) != list
) {
464 struct cfq_queue
*__cfqq
= list_entry_cfqq(n
);
466 if (!cfq_cfqq_class_sync(cfqq
) || !__cfqq
->service_last
)
468 if (time_before(__cfqq
->service_last
, cfqq
->service_last
))
471 list_add(&cfqq
->cfq_list
, n
);
476 * add to busy list of queues for service, trying to be fair in ordering
477 * the pending list according to last request service
480 cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
482 BUG_ON(cfq_cfqq_on_rr(cfqq
));
483 cfq_mark_cfqq_on_rr(cfqq
);
486 cfq_resort_rr_list(cfqq
, 0);
490 cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
492 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
493 cfq_clear_cfqq_on_rr(cfqq
);
494 list_del_init(&cfqq
->cfq_list
);
496 BUG_ON(!cfqd
->busy_queues
);
501 * rb tree support functions
503 static inline void cfq_del_rq_rb(struct request
*rq
)
505 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
506 struct cfq_data
*cfqd
= cfqq
->cfqd
;
507 const int sync
= rq_is_sync(rq
);
509 BUG_ON(!cfqq
->queued
[sync
]);
510 cfqq
->queued
[sync
]--;
512 elv_rb_del(&cfqq
->sort_list
, rq
);
514 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
515 cfq_del_cfqq_rr(cfqd
, cfqq
);
518 static void cfq_add_rq_rb(struct request
*rq
)
520 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
521 struct cfq_data
*cfqd
= cfqq
->cfqd
;
522 struct request
*__alias
;
524 cfqq
->queued
[rq_is_sync(rq
)]++;
527 * looks a little odd, but the first insert might return an alias.
528 * if that happens, put the alias on the dispatch list
530 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
531 cfq_dispatch_insert(cfqd
->queue
, __alias
);
533 if (!cfq_cfqq_on_rr(cfqq
))
534 cfq_add_cfqq_rr(cfqd
, cfqq
);
537 * check if this request is a better next-serve candidate
539 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
540 BUG_ON(!cfqq
->next_rq
);
544 cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
546 elv_rb_del(&cfqq
->sort_list
, rq
);
547 cfqq
->queued
[rq_is_sync(rq
)]--;
551 static struct request
*
552 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
554 struct task_struct
*tsk
= current
;
555 pid_t key
= cfq_queue_pid(tsk
, bio_data_dir(bio
), bio_sync(bio
));
556 struct cfq_queue
*cfqq
;
558 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
560 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
562 return elv_rb_find(&cfqq
->sort_list
, sector
);
568 static void cfq_activate_request(request_queue_t
*q
, struct request
*rq
)
570 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
572 cfqd
->rq_in_driver
++;
575 * If the depth is larger 1, it really could be queueing. But lets
576 * make the mark a little higher - idling could still be good for
577 * low queueing, and a low queueing number could also just indicate
578 * a SCSI mid layer like behaviour where limit+1 is often seen.
580 if (!cfqd
->hw_tag
&& cfqd
->rq_in_driver
> 4)
584 static void cfq_deactivate_request(request_queue_t
*q
, struct request
*rq
)
586 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
588 WARN_ON(!cfqd
->rq_in_driver
);
589 cfqd
->rq_in_driver
--;
592 static void cfq_remove_request(struct request
*rq
)
594 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
596 if (cfqq
->next_rq
== rq
)
597 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
599 list_del_init(&rq
->queuelist
);
602 if (rq_is_meta(rq
)) {
603 WARN_ON(!cfqq
->meta_pending
);
604 cfqq
->meta_pending
--;
609 cfq_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
611 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
612 struct request
*__rq
;
614 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
615 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
617 return ELEVATOR_FRONT_MERGE
;
620 return ELEVATOR_NO_MERGE
;
623 static void cfq_merged_request(request_queue_t
*q
, struct request
*req
,
626 if (type
== ELEVATOR_FRONT_MERGE
) {
627 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
629 cfq_reposition_rq_rb(cfqq
, req
);
634 cfq_merged_requests(request_queue_t
*q
, struct request
*rq
,
635 struct request
*next
)
638 * reposition in fifo if next is older than rq
640 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
641 time_before(next
->start_time
, rq
->start_time
))
642 list_move(&rq
->queuelist
, &next
->queuelist
);
644 cfq_remove_request(next
);
647 static int cfq_allow_merge(request_queue_t
*q
, struct request
*rq
,
650 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
651 const int rw
= bio_data_dir(bio
);
652 struct cfq_queue
*cfqq
;
656 * Disallow merge of a sync bio into an async request.
658 if ((bio_data_dir(bio
) == READ
|| bio_sync(bio
)) && !rq_is_sync(rq
))
662 * Lookup the cfqq that this bio will be queued with. Allow
663 * merge only if rq is queued there.
665 key
= cfq_queue_pid(current
, rw
, bio_sync(bio
));
666 cfqq
= cfq_find_cfq_hash(cfqd
, key
, current
->ioprio
);
668 if (cfqq
== RQ_CFQQ(rq
))
675 __cfq_set_active_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
679 * stop potential idle class queues waiting service
681 del_timer(&cfqd
->idle_class_timer
);
684 cfq_clear_cfqq_must_alloc_slice(cfqq
);
685 cfq_clear_cfqq_fifo_expire(cfqq
);
686 cfq_mark_cfqq_slice_new(cfqq
);
689 cfqd
->active_queue
= cfqq
;
693 * current cfqq expired its slice (or was too idle), select new one
696 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
697 int preempted
, int timed_out
)
699 if (cfq_cfqq_wait_request(cfqq
))
700 del_timer(&cfqd
->idle_slice_timer
);
702 cfq_clear_cfqq_must_dispatch(cfqq
);
703 cfq_clear_cfqq_wait_request(cfqq
);
704 cfq_clear_cfqq_queue_new(cfqq
);
707 * store what was left of this slice, if the queue idled out
710 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
))
711 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
713 cfq_resort_rr_list(cfqq
, preempted
);
715 if (cfqq
== cfqd
->active_queue
)
716 cfqd
->active_queue
= NULL
;
718 if (cfqd
->active_cic
) {
719 put_io_context(cfqd
->active_cic
->ioc
);
720 cfqd
->active_cic
= NULL
;
723 cfqd
->dispatch_slice
= 0;
726 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, int preempted
,
729 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
732 __cfq_slice_expired(cfqd
, cfqq
, preempted
, timed_out
);
745 static int cfq_get_next_prio_level(struct cfq_data
*cfqd
)
754 for (p
= cfqd
->cur_prio
; p
<= cfqd
->cur_end_prio
; p
++) {
755 if (!list_empty(&cfqd
->rr_list
[p
])) {
764 if (++cfqd
->cur_end_prio
== CFQ_PRIO_LISTS
) {
765 cfqd
->cur_end_prio
= 0;
772 if (unlikely(prio
== -1))
775 BUG_ON(prio
>= CFQ_PRIO_LISTS
);
777 list_splice_init(&cfqd
->rr_list
[prio
], &cfqd
->cur_rr
);
779 cfqd
->cur_prio
= prio
+ 1;
780 if (cfqd
->cur_prio
> cfqd
->cur_end_prio
) {
781 cfqd
->cur_end_prio
= cfqd
->cur_prio
;
784 if (cfqd
->cur_end_prio
== CFQ_PRIO_LISTS
) {
786 cfqd
->cur_end_prio
= 0;
792 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
)
794 struct cfq_queue
*cfqq
= NULL
;
796 if (!list_empty(&cfqd
->cur_rr
) || cfq_get_next_prio_level(cfqd
) != -1) {
798 * if current list is non-empty, grab first entry. if it is
799 * empty, get next prio level and grab first entry then if any
802 cfqq
= list_entry_cfqq(cfqd
->cur_rr
.next
);
803 } else if (!list_empty(&cfqd
->busy_rr
)) {
805 * If no new queues are available, check if the busy list has
806 * some before falling back to idle io.
808 cfqq
= list_entry_cfqq(cfqd
->busy_rr
.next
);
809 } else if (!list_empty(&cfqd
->idle_rr
)) {
811 * if we have idle queues and no rt or be queues had pending
812 * requests, either allow immediate service if the grace period
813 * has passed or arm the idle grace timer
815 unsigned long end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
817 if (time_after_eq(jiffies
, end
))
818 cfqq
= list_entry_cfqq(cfqd
->idle_rr
.next
);
820 mod_timer(&cfqd
->idle_class_timer
, end
);
823 __cfq_set_active_queue(cfqd
, cfqq
);
827 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
829 static int cfq_arm_slice_timer(struct cfq_data
*cfqd
)
831 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
832 struct cfq_io_context
*cic
;
835 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
838 * idle is disabled, either manually or by past process history
840 if (!cfqd
->cfq_slice_idle
)
842 if (!cfq_cfqq_idle_window(cfqq
))
845 * task has exited, don't wait
847 cic
= cfqd
->active_cic
;
848 if (!cic
|| !cic
->ioc
->task
)
851 cfq_mark_cfqq_must_dispatch(cfqq
);
852 cfq_mark_cfqq_wait_request(cfqq
);
854 sl
= min(cfqq
->slice_end
- 1, (unsigned long) cfqd
->cfq_slice_idle
);
857 * we don't want to idle for seeks, but we do want to allow
858 * fair distribution of slice time for a process doing back-to-back
859 * seeks. so allow a little bit of time for him to submit a new rq
861 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
862 sl
= min(sl
, msecs_to_jiffies(2));
864 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
868 static void cfq_dispatch_insert(request_queue_t
*q
, struct request
*rq
)
870 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
871 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
873 cfq_remove_request(rq
);
874 cfqq
->on_dispatch
[rq_is_sync(rq
)]++;
875 elv_dispatch_sort(q
, rq
);
877 rq
= list_entry(q
->queue_head
.prev
, struct request
, queuelist
);
878 cfqd
->last_sector
= rq
->sector
+ rq
->nr_sectors
;
882 * return expired entry, or NULL to just start from scratch in rbtree
884 static inline struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
886 struct cfq_data
*cfqd
= cfqq
->cfqd
;
890 if (cfq_cfqq_fifo_expire(cfqq
))
893 cfq_mark_cfqq_fifo_expire(cfqq
);
895 if (list_empty(&cfqq
->fifo
))
898 fifo
= cfq_cfqq_class_sync(cfqq
);
899 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
901 if (time_after(jiffies
, rq
->start_time
+ cfqd
->cfq_fifo_expire
[fifo
]))
908 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
910 const int base_rq
= cfqd
->cfq_slice_async_rq
;
912 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
914 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
918 * get next queue for service
920 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
922 struct cfq_queue
*cfqq
;
924 cfqq
= cfqd
->active_queue
;
931 if (!cfq_cfqq_must_dispatch(cfqq
) && cfq_slice_used(cfqq
))
935 * if queue has requests, dispatch one. if not, check if
936 * enough slice is left to wait for one
938 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
940 else if (cfq_cfqq_slice_new(cfqq
) || cfq_cfqq_dispatched(cfqq
)) {
943 } else if (cfq_cfqq_class_sync(cfqq
)) {
944 if (cfq_arm_slice_timer(cfqd
))
949 cfq_slice_expired(cfqd
, 0, 0);
951 cfqq
= cfq_set_active_queue(cfqd
);
957 __cfq_dispatch_requests(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
962 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
968 * follow expired path, else get first next available
970 if ((rq
= cfq_check_fifo(cfqq
)) == NULL
)
974 * finally, insert request into driver dispatch list
976 cfq_dispatch_insert(cfqd
->queue
, rq
);
978 cfqd
->dispatch_slice
++;
981 if (!cfqd
->active_cic
) {
982 atomic_inc(&RQ_CIC(rq
)->ioc
->refcount
);
983 cfqd
->active_cic
= RQ_CIC(rq
);
986 if (RB_EMPTY_ROOT(&cfqq
->sort_list
))
989 } while (dispatched
< max_dispatch
);
992 * expire an async queue immediately if it has used up its slice. idle
993 * queue always expire after 1 dispatch round.
995 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
996 cfqd
->dispatch_slice
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
997 cfq_class_idle(cfqq
))) {
998 cfqq
->slice_end
= jiffies
+ 1;
999 cfq_slice_expired(cfqd
, 0, 0);
1006 cfq_forced_dispatch_cfqqs(struct list_head
*list
)
1008 struct cfq_queue
*cfqq
, *next
;
1012 list_for_each_entry_safe(cfqq
, next
, list
, cfq_list
) {
1013 while (cfqq
->next_rq
) {
1014 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
1017 BUG_ON(!list_empty(&cfqq
->fifo
));
1024 cfq_forced_dispatch(struct cfq_data
*cfqd
)
1026 int i
, dispatched
= 0;
1028 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
1029 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->rr_list
[i
]);
1031 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->busy_rr
);
1032 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->cur_rr
);
1033 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->idle_rr
);
1035 cfq_slice_expired(cfqd
, 0, 0);
1037 BUG_ON(cfqd
->busy_queues
);
1043 cfq_dispatch_requests(request_queue_t
*q
, int force
)
1045 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1046 struct cfq_queue
*cfqq
, *prev_cfqq
;
1049 if (!cfqd
->busy_queues
)
1052 if (unlikely(force
))
1053 return cfq_forced_dispatch(cfqd
);
1057 while ((cfqq
= cfq_select_queue(cfqd
)) != NULL
) {
1060 if (cfqd
->busy_queues
> 1) {
1062 * Don't repeat dispatch from the previous queue.
1064 if (prev_cfqq
== cfqq
)
1068 * So we have dispatched before in this round, if the
1069 * next queue has idling enabled (must be sync), don't
1070 * allow it service until the previous have continued.
1072 if (cfqd
->rq_in_driver
&& cfq_cfqq_idle_window(cfqq
))
1076 cfq_clear_cfqq_must_dispatch(cfqq
);
1077 cfq_clear_cfqq_wait_request(cfqq
);
1078 del_timer(&cfqd
->idle_slice_timer
);
1080 max_dispatch
= cfqd
->cfq_quantum
;
1081 if (cfq_class_idle(cfqq
))
1084 dispatched
+= __cfq_dispatch_requests(cfqd
, cfqq
, max_dispatch
);
1092 * task holds one reference to the queue, dropped when task exits. each rq
1093 * in-flight on this queue also holds a reference, dropped when rq is freed.
1095 * queue lock must be held here.
1097 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1099 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1101 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1103 if (!atomic_dec_and_test(&cfqq
->ref
))
1106 BUG_ON(rb_first(&cfqq
->sort_list
));
1107 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1108 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1110 if (unlikely(cfqd
->active_queue
== cfqq
)) {
1111 __cfq_slice_expired(cfqd
, cfqq
, 0, 0);
1112 cfq_schedule_dispatch(cfqd
);
1116 * it's on the empty list and still hashed
1118 list_del(&cfqq
->cfq_list
);
1119 hlist_del(&cfqq
->cfq_hash
);
1120 kmem_cache_free(cfq_pool
, cfqq
);
1123 static struct cfq_queue
*
1124 __cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned int prio
,
1127 struct hlist_head
*hash_list
= &cfqd
->cfq_hash
[hashval
];
1128 struct hlist_node
*entry
;
1129 struct cfq_queue
*__cfqq
;
1131 hlist_for_each_entry(__cfqq
, entry
, hash_list
, cfq_hash
) {
1132 const unsigned short __p
= IOPRIO_PRIO_VALUE(__cfqq
->org_ioprio_class
, __cfqq
->org_ioprio
);
1134 if (__cfqq
->key
== key
&& (__p
== prio
|| !prio
))
1141 static struct cfq_queue
*
1142 cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned short prio
)
1144 return __cfq_find_cfq_hash(cfqd
, key
, prio
, hash_long(key
, CFQ_QHASH_SHIFT
));
1147 static void cfq_free_io_context(struct io_context
*ioc
)
1149 struct cfq_io_context
*__cic
;
1153 while ((n
= rb_first(&ioc
->cic_root
)) != NULL
) {
1154 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1155 rb_erase(&__cic
->rb_node
, &ioc
->cic_root
);
1156 kmem_cache_free(cfq_ioc_pool
, __cic
);
1160 elv_ioc_count_mod(ioc_count
, -freed
);
1162 if (ioc_gone
&& !elv_ioc_count_read(ioc_count
))
1166 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1168 if (unlikely(cfqq
== cfqd
->active_queue
)) {
1169 __cfq_slice_expired(cfqd
, cfqq
, 0, 0);
1170 cfq_schedule_dispatch(cfqd
);
1173 cfq_put_queue(cfqq
);
1176 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1177 struct cfq_io_context
*cic
)
1179 list_del_init(&cic
->queue_list
);
1183 if (cic
->cfqq
[ASYNC
]) {
1184 cfq_exit_cfqq(cfqd
, cic
->cfqq
[ASYNC
]);
1185 cic
->cfqq
[ASYNC
] = NULL
;
1188 if (cic
->cfqq
[SYNC
]) {
1189 cfq_exit_cfqq(cfqd
, cic
->cfqq
[SYNC
]);
1190 cic
->cfqq
[SYNC
] = NULL
;
1196 * Called with interrupts disabled
1198 static void cfq_exit_single_io_context(struct cfq_io_context
*cic
)
1200 struct cfq_data
*cfqd
= cic
->key
;
1203 request_queue_t
*q
= cfqd
->queue
;
1205 spin_lock_irq(q
->queue_lock
);
1206 __cfq_exit_single_io_context(cfqd
, cic
);
1207 spin_unlock_irq(q
->queue_lock
);
1211 static void cfq_exit_io_context(struct io_context
*ioc
)
1213 struct cfq_io_context
*__cic
;
1217 * put the reference this task is holding to the various queues
1220 n
= rb_first(&ioc
->cic_root
);
1222 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1224 cfq_exit_single_io_context(__cic
);
1229 static struct cfq_io_context
*
1230 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1232 struct cfq_io_context
*cic
;
1234 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
, cfqd
->queue
->node
);
1236 memset(cic
, 0, sizeof(*cic
));
1237 cic
->last_end_request
= jiffies
;
1238 INIT_LIST_HEAD(&cic
->queue_list
);
1239 cic
->dtor
= cfq_free_io_context
;
1240 cic
->exit
= cfq_exit_io_context
;
1241 elv_ioc_count_inc(ioc_count
);
1247 static void cfq_init_prio_data(struct cfq_queue
*cfqq
)
1249 struct task_struct
*tsk
= current
;
1252 if (!cfq_cfqq_prio_changed(cfqq
))
1255 ioprio_class
= IOPRIO_PRIO_CLASS(tsk
->ioprio
);
1256 switch (ioprio_class
) {
1258 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1259 case IOPRIO_CLASS_NONE
:
1261 * no prio set, place us in the middle of the BE classes
1263 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1264 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1266 case IOPRIO_CLASS_RT
:
1267 cfqq
->ioprio
= task_ioprio(tsk
);
1268 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1270 case IOPRIO_CLASS_BE
:
1271 cfqq
->ioprio
= task_ioprio(tsk
);
1272 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1274 case IOPRIO_CLASS_IDLE
:
1275 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1277 cfq_clear_cfqq_idle_window(cfqq
);
1282 * keep track of original prio settings in case we have to temporarily
1283 * elevate the priority of this queue
1285 cfqq
->org_ioprio
= cfqq
->ioprio
;
1286 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1288 cfq_resort_rr_list(cfqq
, 0);
1289 cfq_clear_cfqq_prio_changed(cfqq
);
1292 static inline void changed_ioprio(struct cfq_io_context
*cic
)
1294 struct cfq_data
*cfqd
= cic
->key
;
1295 struct cfq_queue
*cfqq
;
1296 unsigned long flags
;
1298 if (unlikely(!cfqd
))
1301 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1303 cfqq
= cic
->cfqq
[ASYNC
];
1305 struct cfq_queue
*new_cfqq
;
1306 new_cfqq
= cfq_get_queue(cfqd
, CFQ_KEY_ASYNC
, cic
->ioc
->task
,
1309 cic
->cfqq
[ASYNC
] = new_cfqq
;
1310 cfq_put_queue(cfqq
);
1314 cfqq
= cic
->cfqq
[SYNC
];
1316 cfq_mark_cfqq_prio_changed(cfqq
);
1318 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1321 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1323 struct cfq_io_context
*cic
;
1326 ioc
->ioprio_changed
= 0;
1328 n
= rb_first(&ioc
->cic_root
);
1330 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1332 changed_ioprio(cic
);
1337 static struct cfq_queue
*
1338 cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
,
1341 const int hashval
= hash_long(key
, CFQ_QHASH_SHIFT
);
1342 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1343 unsigned short ioprio
;
1346 ioprio
= tsk
->ioprio
;
1347 cfqq
= __cfq_find_cfq_hash(cfqd
, key
, ioprio
, hashval
);
1353 } else if (gfp_mask
& __GFP_WAIT
) {
1355 * Inform the allocator of the fact that we will
1356 * just repeat this allocation if it fails, to allow
1357 * the allocator to do whatever it needs to attempt to
1360 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1361 new_cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
|__GFP_NOFAIL
, cfqd
->queue
->node
);
1362 spin_lock_irq(cfqd
->queue
->queue_lock
);
1365 cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
, cfqd
->queue
->node
);
1370 memset(cfqq
, 0, sizeof(*cfqq
));
1372 INIT_HLIST_NODE(&cfqq
->cfq_hash
);
1373 INIT_LIST_HEAD(&cfqq
->cfq_list
);
1374 INIT_LIST_HEAD(&cfqq
->fifo
);
1377 hlist_add_head(&cfqq
->cfq_hash
, &cfqd
->cfq_hash
[hashval
]);
1378 atomic_set(&cfqq
->ref
, 0);
1381 if (key
!= CFQ_KEY_ASYNC
)
1382 cfq_mark_cfqq_idle_window(cfqq
);
1384 cfq_mark_cfqq_prio_changed(cfqq
);
1385 cfq_mark_cfqq_queue_new(cfqq
);
1386 cfq_init_prio_data(cfqq
);
1390 kmem_cache_free(cfq_pool
, new_cfqq
);
1392 atomic_inc(&cfqq
->ref
);
1394 WARN_ON((gfp_mask
& __GFP_WAIT
) && !cfqq
);
1399 cfq_drop_dead_cic(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1401 WARN_ON(!list_empty(&cic
->queue_list
));
1402 rb_erase(&cic
->rb_node
, &ioc
->cic_root
);
1403 kmem_cache_free(cfq_ioc_pool
, cic
);
1404 elv_ioc_count_dec(ioc_count
);
1407 static struct cfq_io_context
*
1408 cfq_cic_rb_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1411 struct cfq_io_context
*cic
;
1412 void *k
, *key
= cfqd
;
1415 n
= ioc
->cic_root
.rb_node
;
1417 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1418 /* ->key must be copied to avoid race with cfq_exit_queue() */
1421 cfq_drop_dead_cic(ioc
, cic
);
1437 cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1438 struct cfq_io_context
*cic
)
1441 struct rb_node
*parent
;
1442 struct cfq_io_context
*__cic
;
1443 unsigned long flags
;
1451 p
= &ioc
->cic_root
.rb_node
;
1454 __cic
= rb_entry(parent
, struct cfq_io_context
, rb_node
);
1455 /* ->key must be copied to avoid race with cfq_exit_queue() */
1458 cfq_drop_dead_cic(ioc
, __cic
);
1464 else if (cic
->key
> k
)
1465 p
= &(*p
)->rb_right
;
1470 rb_link_node(&cic
->rb_node
, parent
, p
);
1471 rb_insert_color(&cic
->rb_node
, &ioc
->cic_root
);
1473 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1474 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1475 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1479 * Setup general io context and cfq io context. There can be several cfq
1480 * io contexts per general io context, if this process is doing io to more
1481 * than one device managed by cfq.
1483 static struct cfq_io_context
*
1484 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1486 struct io_context
*ioc
= NULL
;
1487 struct cfq_io_context
*cic
;
1489 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1491 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
1495 cic
= cfq_cic_rb_lookup(cfqd
, ioc
);
1499 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1503 cfq_cic_link(cfqd
, ioc
, cic
);
1505 smp_read_barrier_depends();
1506 if (unlikely(ioc
->ioprio_changed
))
1507 cfq_ioc_set_ioprio(ioc
);
1511 put_io_context(ioc
);
1516 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1518 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
1519 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1521 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1522 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1523 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1527 cfq_update_io_seektime(struct cfq_io_context
*cic
, struct request
*rq
)
1532 if (cic
->last_request_pos
< rq
->sector
)
1533 sdist
= rq
->sector
- cic
->last_request_pos
;
1535 sdist
= cic
->last_request_pos
- rq
->sector
;
1538 * Don't allow the seek distance to get too large from the
1539 * odd fragment, pagein, etc
1541 if (cic
->seek_samples
<= 60) /* second&third seek */
1542 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1544 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1546 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1547 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1548 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1549 do_div(total
, cic
->seek_samples
);
1550 cic
->seek_mean
= (sector_t
)total
;
1554 * Disable idle window if the process thinks too long or seeks so much that
1558 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1559 struct cfq_io_context
*cic
)
1561 int enable_idle
= cfq_cfqq_idle_window(cfqq
);
1563 if (!cic
->ioc
->task
|| !cfqd
->cfq_slice_idle
||
1564 (cfqd
->hw_tag
&& CIC_SEEKY(cic
)))
1566 else if (sample_valid(cic
->ttime_samples
)) {
1567 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
1574 cfq_mark_cfqq_idle_window(cfqq
);
1576 cfq_clear_cfqq_idle_window(cfqq
);
1580 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1581 * no or if we aren't sure, a 1 will cause a preempt.
1584 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
1587 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1589 if (cfq_class_idle(new_cfqq
))
1595 if (cfq_class_idle(cfqq
))
1597 if (!cfq_cfqq_wait_request(new_cfqq
))
1600 * if the new request is sync, but the currently running queue is
1601 * not, let the sync request have priority.
1603 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
1606 * So both queues are sync. Let the new request get disk time if
1607 * it's a metadata request and the current queue is doing regular IO.
1609 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
1616 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1617 * let it have half of its nominal slice.
1619 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1621 cfq_slice_expired(cfqd
, 1, 1);
1624 * Put the new queue at the front of the of the current list,
1625 * so we know that it will be selected next.
1627 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1628 list_move(&cfqq
->cfq_list
, &cfqd
->cur_rr
);
1630 cfqq
->slice_end
= 0;
1631 cfq_mark_cfqq_slice_new(cfqq
);
1635 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1636 * something we should do about it
1639 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1642 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1645 cfqq
->meta_pending
++;
1648 * we never wait for an async request and we don't allow preemption
1649 * of an async request. so just return early
1651 if (!rq_is_sync(rq
)) {
1653 * sync process issued an async request, if it's waiting
1654 * then expire it and kick rq handling.
1656 if (cic
== cfqd
->active_cic
&&
1657 del_timer(&cfqd
->idle_slice_timer
)) {
1658 cfq_slice_expired(cfqd
, 0, 0);
1659 blk_start_queueing(cfqd
->queue
);
1664 cfq_update_io_thinktime(cfqd
, cic
);
1665 cfq_update_io_seektime(cic
, rq
);
1666 cfq_update_idle_window(cfqd
, cfqq
, cic
);
1668 cic
->last_request_pos
= rq
->sector
+ rq
->nr_sectors
;
1670 if (cfqq
== cfqd
->active_queue
) {
1672 * if we are waiting for a request for this queue, let it rip
1673 * immediately and flag that we must not expire this queue
1676 if (cfq_cfqq_wait_request(cfqq
)) {
1677 cfq_mark_cfqq_must_dispatch(cfqq
);
1678 del_timer(&cfqd
->idle_slice_timer
);
1679 blk_start_queueing(cfqd
->queue
);
1681 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
1683 * not the active queue - expire current slice if it is
1684 * idle and has expired it's mean thinktime or this new queue
1685 * has some old slice time left and is of higher priority
1687 cfq_preempt_queue(cfqd
, cfqq
);
1688 cfq_mark_cfqq_must_dispatch(cfqq
);
1689 blk_start_queueing(cfqd
->queue
);
1693 static void cfq_insert_request(request_queue_t
*q
, struct request
*rq
)
1695 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1696 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1698 cfq_init_prio_data(cfqq
);
1702 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
1704 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
1707 static void cfq_completed_request(request_queue_t
*q
, struct request
*rq
)
1709 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1710 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1711 const int sync
= rq_is_sync(rq
);
1716 WARN_ON(!cfqd
->rq_in_driver
);
1717 WARN_ON(!cfqq
->on_dispatch
[sync
]);
1718 cfqd
->rq_in_driver
--;
1719 cfqq
->on_dispatch
[sync
]--;
1720 cfqq
->service_last
= now
;
1722 if (!cfq_class_idle(cfqq
))
1723 cfqd
->last_end_request
= now
;
1725 cfq_resort_rr_list(cfqq
, 0);
1728 RQ_CIC(rq
)->last_end_request
= now
;
1731 * If this is the active queue, check if it needs to be expired,
1732 * or if we want to idle in case it has no pending requests.
1734 if (cfqd
->active_queue
== cfqq
) {
1735 if (cfq_cfqq_slice_new(cfqq
)) {
1736 cfq_set_prio_slice(cfqd
, cfqq
);
1737 cfq_clear_cfqq_slice_new(cfqq
);
1739 if (cfq_slice_used(cfqq
))
1740 cfq_slice_expired(cfqd
, 0, 1);
1741 else if (sync
&& RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1742 if (!cfq_arm_slice_timer(cfqd
))
1743 cfq_schedule_dispatch(cfqd
);
1749 * we temporarily boost lower priority queues if they are holding fs exclusive
1750 * resources. they are boosted to normal prio (CLASS_BE/4)
1752 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
1754 const int ioprio_class
= cfqq
->ioprio_class
;
1755 const int ioprio
= cfqq
->ioprio
;
1757 if (has_fs_excl()) {
1759 * boost idle prio on transactions that would lock out other
1760 * users of the filesystem
1762 if (cfq_class_idle(cfqq
))
1763 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1764 if (cfqq
->ioprio
> IOPRIO_NORM
)
1765 cfqq
->ioprio
= IOPRIO_NORM
;
1768 * check if we need to unboost the queue
1770 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
1771 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
1772 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
1773 cfqq
->ioprio
= cfqq
->org_ioprio
;
1777 * refile between round-robin lists if we moved the priority class
1779 if ((ioprio_class
!= cfqq
->ioprio_class
|| ioprio
!= cfqq
->ioprio
))
1780 cfq_resort_rr_list(cfqq
, 0);
1783 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
1785 if ((cfq_cfqq_wait_request(cfqq
) || cfq_cfqq_must_alloc(cfqq
)) &&
1786 !cfq_cfqq_must_alloc_slice(cfqq
)) {
1787 cfq_mark_cfqq_must_alloc_slice(cfqq
);
1788 return ELV_MQUEUE_MUST
;
1791 return ELV_MQUEUE_MAY
;
1794 static int cfq_may_queue(request_queue_t
*q
, int rw
)
1796 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1797 struct task_struct
*tsk
= current
;
1798 struct cfq_queue
*cfqq
;
1801 key
= cfq_queue_pid(tsk
, rw
, rw
& REQ_RW_SYNC
);
1804 * don't force setup of a queue from here, as a call to may_queue
1805 * does not necessarily imply that a request actually will be queued.
1806 * so just lookup a possibly existing queue, or return 'may queue'
1809 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
1811 cfq_init_prio_data(cfqq
);
1812 cfq_prio_boost(cfqq
);
1814 return __cfq_may_queue(cfqq
);
1817 return ELV_MQUEUE_MAY
;
1821 * queue lock held here
1823 static void cfq_put_request(struct request
*rq
)
1825 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1828 const int rw
= rq_data_dir(rq
);
1830 BUG_ON(!cfqq
->allocated
[rw
]);
1831 cfqq
->allocated
[rw
]--;
1833 put_io_context(RQ_CIC(rq
)->ioc
);
1835 rq
->elevator_private
= NULL
;
1836 rq
->elevator_private2
= NULL
;
1838 cfq_put_queue(cfqq
);
1843 * Allocate cfq data structures associated with this request.
1846 cfq_set_request(request_queue_t
*q
, struct request
*rq
, gfp_t gfp_mask
)
1848 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1849 struct task_struct
*tsk
= current
;
1850 struct cfq_io_context
*cic
;
1851 const int rw
= rq_data_dir(rq
);
1852 const int is_sync
= rq_is_sync(rq
);
1853 pid_t key
= cfq_queue_pid(tsk
, rw
, is_sync
);
1854 struct cfq_queue
*cfqq
;
1855 unsigned long flags
;
1857 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1859 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
1861 spin_lock_irqsave(q
->queue_lock
, flags
);
1866 if (!cic
->cfqq
[is_sync
]) {
1867 cfqq
= cfq_get_queue(cfqd
, key
, tsk
, gfp_mask
);
1871 cic
->cfqq
[is_sync
] = cfqq
;
1873 cfqq
= cic
->cfqq
[is_sync
];
1875 cfqq
->allocated
[rw
]++;
1876 cfq_clear_cfqq_must_alloc(cfqq
);
1877 atomic_inc(&cfqq
->ref
);
1879 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1881 rq
->elevator_private
= cic
;
1882 rq
->elevator_private2
= cfqq
;
1887 put_io_context(cic
->ioc
);
1889 cfq_schedule_dispatch(cfqd
);
1890 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1894 static void cfq_kick_queue(struct work_struct
*work
)
1896 struct cfq_data
*cfqd
=
1897 container_of(work
, struct cfq_data
, unplug_work
);
1898 request_queue_t
*q
= cfqd
->queue
;
1899 unsigned long flags
;
1901 spin_lock_irqsave(q
->queue_lock
, flags
);
1902 blk_start_queueing(q
);
1903 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1907 * Timer running if the active_queue is currently idling inside its time slice
1909 static void cfq_idle_slice_timer(unsigned long data
)
1911 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1912 struct cfq_queue
*cfqq
;
1913 unsigned long flags
;
1916 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1918 if ((cfqq
= cfqd
->active_queue
) != NULL
) {
1924 if (cfq_slice_used(cfqq
))
1928 * only expire and reinvoke request handler, if there are
1929 * other queues with pending requests
1931 if (!cfqd
->busy_queues
)
1935 * not expired and it has a request pending, let it dispatch
1937 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1938 cfq_mark_cfqq_must_dispatch(cfqq
);
1943 cfq_slice_expired(cfqd
, 0, timed_out
);
1945 cfq_schedule_dispatch(cfqd
);
1947 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1951 * Timer running if an idle class queue is waiting for service
1953 static void cfq_idle_class_timer(unsigned long data
)
1955 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1956 unsigned long flags
, end
;
1958 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1961 * race with a non-idle queue, reset timer
1963 end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
1964 if (!time_after_eq(jiffies
, end
))
1965 mod_timer(&cfqd
->idle_class_timer
, end
);
1967 cfq_schedule_dispatch(cfqd
);
1969 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1972 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
1974 del_timer_sync(&cfqd
->idle_slice_timer
);
1975 del_timer_sync(&cfqd
->idle_class_timer
);
1976 blk_sync_queue(cfqd
->queue
);
1979 static void cfq_exit_queue(elevator_t
*e
)
1981 struct cfq_data
*cfqd
= e
->elevator_data
;
1982 request_queue_t
*q
= cfqd
->queue
;
1984 cfq_shutdown_timer_wq(cfqd
);
1986 spin_lock_irq(q
->queue_lock
);
1988 if (cfqd
->active_queue
)
1989 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0, 0);
1991 while (!list_empty(&cfqd
->cic_list
)) {
1992 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
1993 struct cfq_io_context
,
1996 __cfq_exit_single_io_context(cfqd
, cic
);
1999 spin_unlock_irq(q
->queue_lock
);
2001 cfq_shutdown_timer_wq(cfqd
);
2003 kfree(cfqd
->cfq_hash
);
2007 static void *cfq_init_queue(request_queue_t
*q
)
2009 struct cfq_data
*cfqd
;
2012 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
, q
->node
);
2016 memset(cfqd
, 0, sizeof(*cfqd
));
2018 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
2019 INIT_LIST_HEAD(&cfqd
->rr_list
[i
]);
2021 INIT_LIST_HEAD(&cfqd
->busy_rr
);
2022 INIT_LIST_HEAD(&cfqd
->cur_rr
);
2023 INIT_LIST_HEAD(&cfqd
->idle_rr
);
2024 INIT_LIST_HEAD(&cfqd
->cic_list
);
2026 cfqd
->cfq_hash
= kmalloc_node(sizeof(struct hlist_head
) * CFQ_QHASH_ENTRIES
, GFP_KERNEL
, q
->node
);
2027 if (!cfqd
->cfq_hash
)
2030 for (i
= 0; i
< CFQ_QHASH_ENTRIES
; i
++)
2031 INIT_HLIST_HEAD(&cfqd
->cfq_hash
[i
]);
2035 init_timer(&cfqd
->idle_slice_timer
);
2036 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2037 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2039 init_timer(&cfqd
->idle_class_timer
);
2040 cfqd
->idle_class_timer
.function
= cfq_idle_class_timer
;
2041 cfqd
->idle_class_timer
.data
= (unsigned long) cfqd
;
2043 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
2045 cfqd
->cfq_quantum
= cfq_quantum
;
2046 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2047 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2048 cfqd
->cfq_back_max
= cfq_back_max
;
2049 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2050 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2051 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2052 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2053 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2061 static void cfq_slab_kill(void)
2064 kmem_cache_destroy(cfq_pool
);
2066 kmem_cache_destroy(cfq_ioc_pool
);
2069 static int __init
cfq_slab_setup(void)
2071 cfq_pool
= kmem_cache_create("cfq_pool", sizeof(struct cfq_queue
), 0, 0,
2076 cfq_ioc_pool
= kmem_cache_create("cfq_ioc_pool",
2077 sizeof(struct cfq_io_context
), 0, 0, NULL
, NULL
);
2088 * sysfs parts below -->
2092 cfq_var_show(unsigned int var
, char *page
)
2094 return sprintf(page
, "%d\n", var
);
2098 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2100 char *p
= (char *) page
;
2102 *var
= simple_strtoul(p
, &p
, 10);
2106 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2107 static ssize_t __FUNC(elevator_t *e, char *page) \
2109 struct cfq_data *cfqd = e->elevator_data; \
2110 unsigned int __data = __VAR; \
2112 __data = jiffies_to_msecs(__data); \
2113 return cfq_var_show(__data, (page)); \
2115 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2116 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2117 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2118 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2119 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2120 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2121 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2122 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2123 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2124 #undef SHOW_FUNCTION
2126 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2127 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2129 struct cfq_data *cfqd = e->elevator_data; \
2130 unsigned int __data; \
2131 int ret = cfq_var_store(&__data, (page), count); \
2132 if (__data < (MIN)) \
2134 else if (__data > (MAX)) \
2137 *(__PTR) = msecs_to_jiffies(__data); \
2139 *(__PTR) = __data; \
2142 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2143 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1, UINT_MAX
, 1);
2144 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1, UINT_MAX
, 1);
2145 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2146 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1, UINT_MAX
, 0);
2147 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2148 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2149 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2150 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1, UINT_MAX
, 0);
2151 #undef STORE_FUNCTION
2153 #define CFQ_ATTR(name) \
2154 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2156 static struct elv_fs_entry cfq_attrs
[] = {
2158 CFQ_ATTR(fifo_expire_sync
),
2159 CFQ_ATTR(fifo_expire_async
),
2160 CFQ_ATTR(back_seek_max
),
2161 CFQ_ATTR(back_seek_penalty
),
2162 CFQ_ATTR(slice_sync
),
2163 CFQ_ATTR(slice_async
),
2164 CFQ_ATTR(slice_async_rq
),
2165 CFQ_ATTR(slice_idle
),
2169 static struct elevator_type iosched_cfq
= {
2171 .elevator_merge_fn
= cfq_merge
,
2172 .elevator_merged_fn
= cfq_merged_request
,
2173 .elevator_merge_req_fn
= cfq_merged_requests
,
2174 .elevator_allow_merge_fn
= cfq_allow_merge
,
2175 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2176 .elevator_add_req_fn
= cfq_insert_request
,
2177 .elevator_activate_req_fn
= cfq_activate_request
,
2178 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2179 .elevator_queue_empty_fn
= cfq_queue_empty
,
2180 .elevator_completed_req_fn
= cfq_completed_request
,
2181 .elevator_former_req_fn
= elv_rb_former_request
,
2182 .elevator_latter_req_fn
= elv_rb_latter_request
,
2183 .elevator_set_req_fn
= cfq_set_request
,
2184 .elevator_put_req_fn
= cfq_put_request
,
2185 .elevator_may_queue_fn
= cfq_may_queue
,
2186 .elevator_init_fn
= cfq_init_queue
,
2187 .elevator_exit_fn
= cfq_exit_queue
,
2188 .trim
= cfq_free_io_context
,
2190 .elevator_attrs
= cfq_attrs
,
2191 .elevator_name
= "cfq",
2192 .elevator_owner
= THIS_MODULE
,
2195 static int __init
cfq_init(void)
2200 * could be 0 on HZ < 1000 setups
2202 if (!cfq_slice_async
)
2203 cfq_slice_async
= 1;
2204 if (!cfq_slice_idle
)
2207 if (cfq_slab_setup())
2210 ret
= elv_register(&iosched_cfq
);
2217 static void __exit
cfq_exit(void)
2219 DECLARE_COMPLETION_ONSTACK(all_gone
);
2220 elv_unregister(&iosched_cfq
);
2221 ioc_gone
= &all_gone
;
2222 /* ioc_gone's update must be visible before reading ioc_count */
2224 if (elv_ioc_count_read(ioc_count
))
2225 wait_for_completion(ioc_gone
);
2230 module_init(cfq_init
);
2231 module_exit(cfq_exit
);
2233 MODULE_AUTHOR("Jens Axboe");
2234 MODULE_LICENSE("GPL");
2235 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");