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_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, /* 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 */
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
, int is_sync
)
225 * Use the per-process queue, for read requests and syncronous writes
227 if (!(rw
& REQ_RW
) || is_sync
)
230 return CFQ_KEY_ASYNC
;
234 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
235 * We choose the request that is closest to the head right now. Distance
236 * behind the head is penalized and only allowed to a certain extent.
238 static struct request
*
239 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
)
241 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
242 unsigned long back_max
;
243 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
244 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
245 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
247 if (rq1
== NULL
|| rq1
== rq2
)
252 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
254 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
256 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
258 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
264 last
= cfqd
->last_sector
;
267 * by definition, 1KiB is 2 sectors
269 back_max
= cfqd
->cfq_back_max
* 2;
272 * Strict one way elevator _except_ in the case where we allow
273 * short backward seeks which are biased as twice the cost of a
274 * similar forward seek.
278 else if (s1
+ back_max
>= last
)
279 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
281 wrap
|= CFQ_RQ1_WRAP
;
285 else if (s2
+ back_max
>= last
)
286 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
288 wrap
|= CFQ_RQ2_WRAP
;
290 /* Found required data */
293 * By doing switch() on the bit mask "wrap" we avoid having to
294 * check two variables for all permutations: --> faster!
297 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
313 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
316 * Since both rqs are wrapped,
317 * start with the one that's further behind head
318 * (--> only *one* back seek required),
319 * since back seek takes more time than forward.
329 * would be nice to take fifo expire time into account as well
331 static struct request
*
332 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
333 struct request
*last
)
335 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
336 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
337 struct request
*next
= NULL
, *prev
= NULL
;
339 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
342 prev
= rb_entry_rq(rbprev
);
345 next
= rb_entry_rq(rbnext
);
347 rbnext
= rb_first(&cfqq
->sort_list
);
348 if (rbnext
&& rbnext
!= &last
->rb_node
)
349 next
= rb_entry_rq(rbnext
);
352 return cfq_choose_req(cfqd
, next
, prev
);
355 static void cfq_resort_rr_list(struct cfq_queue
*cfqq
, int preempted
)
357 struct cfq_data
*cfqd
= cfqq
->cfqd
;
358 struct list_head
*list
;
361 * Resorting requires the cfqq to be on the RR list already.
363 if (!cfq_cfqq_on_rr(cfqq
))
366 list_del(&cfqq
->cfq_list
);
368 if (cfq_class_rt(cfqq
))
369 list
= &cfqd
->cur_rr
;
370 else if (cfq_class_idle(cfqq
))
371 list
= &cfqd
->idle_rr
;
374 * if cfqq has requests in flight, don't allow it to be
375 * found in cfq_set_active_queue before it has finished them.
376 * this is done to increase fairness between a process that
377 * has lots of io pending vs one that only generates one
378 * sporadically or synchronously
380 if (cfq_cfqq_dispatched(cfqq
))
381 list
= &cfqd
->busy_rr
;
383 list
= &cfqd
->rr_list
[cfqq
->ioprio
];
387 * If this queue was preempted or is new (never been serviced), let
388 * it be added first for fairness but beind other new queues.
389 * Otherwise, just add to the back of the list.
391 if (preempted
|| cfq_cfqq_queue_new(cfqq
)) {
392 struct list_head
*n
= list
;
393 struct cfq_queue
*__cfqq
;
395 while (n
->next
!= list
) {
396 __cfqq
= list_entry_cfqq(n
->next
);
397 if (!cfq_cfqq_queue_new(__cfqq
))
406 list_add_tail(&cfqq
->cfq_list
, list
);
410 * add to busy list of queues for service, trying to be fair in ordering
411 * the pending list according to last request service
414 cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
416 BUG_ON(cfq_cfqq_on_rr(cfqq
));
417 cfq_mark_cfqq_on_rr(cfqq
);
420 cfq_resort_rr_list(cfqq
, 0);
424 cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
426 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
427 cfq_clear_cfqq_on_rr(cfqq
);
428 list_del_init(&cfqq
->cfq_list
);
430 BUG_ON(!cfqd
->busy_queues
);
435 * rb tree support functions
437 static inline void cfq_del_rq_rb(struct request
*rq
)
439 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
440 struct cfq_data
*cfqd
= cfqq
->cfqd
;
441 const int sync
= rq_is_sync(rq
);
443 BUG_ON(!cfqq
->queued
[sync
]);
444 cfqq
->queued
[sync
]--;
446 elv_rb_del(&cfqq
->sort_list
, rq
);
448 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
449 cfq_del_cfqq_rr(cfqd
, cfqq
);
452 static void cfq_add_rq_rb(struct request
*rq
)
454 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
455 struct cfq_data
*cfqd
= cfqq
->cfqd
;
456 struct request
*__alias
;
458 cfqq
->queued
[rq_is_sync(rq
)]++;
461 * looks a little odd, but the first insert might return an alias.
462 * if that happens, put the alias on the dispatch list
464 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
465 cfq_dispatch_insert(cfqd
->queue
, __alias
);
467 if (!cfq_cfqq_on_rr(cfqq
))
468 cfq_add_cfqq_rr(cfqd
, cfqq
);
472 cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
474 elv_rb_del(&cfqq
->sort_list
, rq
);
475 cfqq
->queued
[rq_is_sync(rq
)]--;
479 static struct request
*
480 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
482 struct task_struct
*tsk
= current
;
483 pid_t key
= cfq_queue_pid(tsk
, bio_data_dir(bio
), bio_sync(bio
));
484 struct cfq_queue
*cfqq
;
486 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
488 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
490 return elv_rb_find(&cfqq
->sort_list
, sector
);
496 static void cfq_activate_request(request_queue_t
*q
, struct request
*rq
)
498 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
500 cfqd
->rq_in_driver
++;
503 * If the depth is larger 1, it really could be queueing. But lets
504 * make the mark a little higher - idling could still be good for
505 * low queueing, and a low queueing number could also just indicate
506 * a SCSI mid layer like behaviour where limit+1 is often seen.
508 if (!cfqd
->hw_tag
&& cfqd
->rq_in_driver
> 4)
512 static void cfq_deactivate_request(request_queue_t
*q
, struct request
*rq
)
514 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
516 WARN_ON(!cfqd
->rq_in_driver
);
517 cfqd
->rq_in_driver
--;
520 static void cfq_remove_request(struct request
*rq
)
522 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
524 if (cfqq
->next_rq
== rq
)
525 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
527 list_del_init(&rq
->queuelist
);
530 if (rq_is_meta(rq
)) {
531 WARN_ON(!cfqq
->meta_pending
);
532 cfqq
->meta_pending
--;
537 cfq_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
539 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
540 struct request
*__rq
;
542 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
543 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
545 return ELEVATOR_FRONT_MERGE
;
548 return ELEVATOR_NO_MERGE
;
551 static void cfq_merged_request(request_queue_t
*q
, struct request
*req
,
554 if (type
== ELEVATOR_FRONT_MERGE
) {
555 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
557 cfq_reposition_rq_rb(cfqq
, req
);
562 cfq_merged_requests(request_queue_t
*q
, struct request
*rq
,
563 struct request
*next
)
566 * reposition in fifo if next is older than rq
568 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
569 time_before(next
->start_time
, rq
->start_time
))
570 list_move(&rq
->queuelist
, &next
->queuelist
);
572 cfq_remove_request(next
);
575 static int cfq_allow_merge(request_queue_t
*q
, struct request
*rq
,
578 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
579 const int rw
= bio_data_dir(bio
);
580 struct cfq_queue
*cfqq
;
584 * Disallow merge of a sync bio into an async request.
586 if ((bio_data_dir(bio
) == READ
|| bio_sync(bio
)) && !rq_is_sync(rq
))
590 * Lookup the cfqq that this bio will be queued with. Allow
591 * merge only if rq is queued there.
593 key
= cfq_queue_pid(current
, rw
, bio_sync(bio
));
594 cfqq
= cfq_find_cfq_hash(cfqd
, key
, current
->ioprio
);
596 if (cfqq
== RQ_CFQQ(rq
))
603 __cfq_set_active_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
607 * stop potential idle class queues waiting service
609 del_timer(&cfqd
->idle_class_timer
);
611 cfqq
->slice_start
= jiffies
;
613 cfqq
->slice_left
= 0;
614 cfq_clear_cfqq_must_alloc_slice(cfqq
);
615 cfq_clear_cfqq_fifo_expire(cfqq
);
618 cfqd
->active_queue
= cfqq
;
622 * current cfqq expired its slice (or was too idle), select new one
625 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
628 unsigned long now
= jiffies
;
630 if (cfq_cfqq_wait_request(cfqq
))
631 del_timer(&cfqd
->idle_slice_timer
);
633 if (!preempted
&& !cfq_cfqq_dispatched(cfqq
))
634 cfq_schedule_dispatch(cfqd
);
636 cfq_clear_cfqq_must_dispatch(cfqq
);
637 cfq_clear_cfqq_wait_request(cfqq
);
638 cfq_clear_cfqq_queue_new(cfqq
);
641 * store what was left of this slice, if the queue idled out
644 if (time_after(cfqq
->slice_end
, now
))
645 cfqq
->slice_left
= cfqq
->slice_end
- now
;
647 cfqq
->slice_left
= 0;
649 cfq_resort_rr_list(cfqq
, preempted
);
651 if (cfqq
== cfqd
->active_queue
)
652 cfqd
->active_queue
= NULL
;
654 if (cfqd
->active_cic
) {
655 put_io_context(cfqd
->active_cic
->ioc
);
656 cfqd
->active_cic
= NULL
;
659 cfqd
->dispatch_slice
= 0;
662 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, int preempted
)
664 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
667 __cfq_slice_expired(cfqd
, cfqq
, preempted
);
680 static int cfq_get_next_prio_level(struct cfq_data
*cfqd
)
689 for (p
= cfqd
->cur_prio
; p
<= cfqd
->cur_end_prio
; p
++) {
690 if (!list_empty(&cfqd
->rr_list
[p
])) {
699 if (++cfqd
->cur_end_prio
== CFQ_PRIO_LISTS
) {
700 cfqd
->cur_end_prio
= 0;
707 if (unlikely(prio
== -1))
710 BUG_ON(prio
>= CFQ_PRIO_LISTS
);
712 list_splice_init(&cfqd
->rr_list
[prio
], &cfqd
->cur_rr
);
714 cfqd
->cur_prio
= prio
+ 1;
715 if (cfqd
->cur_prio
> cfqd
->cur_end_prio
) {
716 cfqd
->cur_end_prio
= cfqd
->cur_prio
;
719 if (cfqd
->cur_end_prio
== CFQ_PRIO_LISTS
) {
721 cfqd
->cur_end_prio
= 0;
727 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
)
729 struct cfq_queue
*cfqq
= NULL
;
731 if (!list_empty(&cfqd
->cur_rr
) || cfq_get_next_prio_level(cfqd
) != -1) {
733 * if current list is non-empty, grab first entry. if it is
734 * empty, get next prio level and grab first entry then if any
737 cfqq
= list_entry_cfqq(cfqd
->cur_rr
.next
);
738 } else if (!list_empty(&cfqd
->busy_rr
)) {
740 * If no new queues are available, check if the busy list has
741 * some before falling back to idle io.
743 cfqq
= list_entry_cfqq(cfqd
->busy_rr
.next
);
744 } else if (!list_empty(&cfqd
->idle_rr
)) {
746 * if we have idle queues and no rt or be queues had pending
747 * requests, either allow immediate service if the grace period
748 * has passed or arm the idle grace timer
750 unsigned long end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
752 if (time_after_eq(jiffies
, end
))
753 cfqq
= list_entry_cfqq(cfqd
->idle_rr
.next
);
755 mod_timer(&cfqd
->idle_class_timer
, end
);
758 __cfq_set_active_queue(cfqd
, cfqq
);
762 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
764 static int cfq_arm_slice_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
767 struct cfq_io_context
*cic
;
770 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
771 WARN_ON(cfqq
!= cfqd
->active_queue
);
774 * idle is disabled, either manually or by past process history
776 if (!cfqd
->cfq_slice_idle
)
778 if (!cfq_cfqq_idle_window(cfqq
))
781 * task has exited, don't wait
783 cic
= cfqd
->active_cic
;
784 if (!cic
|| !cic
->ioc
->task
)
787 cfq_mark_cfqq_must_dispatch(cfqq
);
788 cfq_mark_cfqq_wait_request(cfqq
);
790 sl
= min(cfqq
->slice_end
- 1, (unsigned long) cfqd
->cfq_slice_idle
);
793 * we don't want to idle for seeks, but we do want to allow
794 * fair distribution of slice time for a process doing back-to-back
795 * seeks. so allow a little bit of time for him to submit a new rq
797 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
798 sl
= min(sl
, msecs_to_jiffies(2));
800 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
804 static void cfq_dispatch_insert(request_queue_t
*q
, struct request
*rq
)
806 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
807 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
809 cfq_remove_request(rq
);
810 cfqq
->on_dispatch
[rq_is_sync(rq
)]++;
811 elv_dispatch_sort(q
, rq
);
813 rq
= list_entry(q
->queue_head
.prev
, struct request
, queuelist
);
814 cfqd
->last_sector
= rq
->sector
+ rq
->nr_sectors
;
818 * return expired entry, or NULL to just start from scratch in rbtree
820 static inline struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
822 struct cfq_data
*cfqd
= cfqq
->cfqd
;
826 if (cfq_cfqq_fifo_expire(cfqq
))
828 if (list_empty(&cfqq
->fifo
))
831 fifo
= cfq_cfqq_class_sync(cfqq
);
832 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
834 if (time_after(jiffies
, rq
->start_time
+ cfqd
->cfq_fifo_expire
[fifo
])) {
835 cfq_mark_cfqq_fifo_expire(cfqq
);
843 * Scale schedule slice based on io priority. Use the sync time slice only
844 * if a queue is marked sync and has sync io queued. A sync queue with async
845 * io only, should not get full sync slice length.
848 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
850 const int base_slice
= cfqd
->cfq_slice
[cfq_cfqq_sync(cfqq
)];
852 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
854 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - cfqq
->ioprio
));
858 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
860 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
864 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
866 const int base_rq
= cfqd
->cfq_slice_async_rq
;
868 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
870 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
874 * get next queue for service
876 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
878 unsigned long now
= jiffies
;
879 struct cfq_queue
*cfqq
;
881 cfqq
= cfqd
->active_queue
;
888 if (!cfq_cfqq_must_dispatch(cfqq
) && time_after(now
, cfqq
->slice_end
))
892 * if queue has requests, dispatch one. if not, check if
893 * enough slice is left to wait for one
895 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
897 else if (cfq_cfqq_dispatched(cfqq
)) {
900 } else if (cfq_cfqq_class_sync(cfqq
)) {
901 if (cfq_arm_slice_timer(cfqd
, cfqq
))
906 cfq_slice_expired(cfqd
, 0);
908 cfqq
= cfq_set_active_queue(cfqd
);
914 __cfq_dispatch_requests(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
919 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
925 * follow expired path, else get first next available
927 if ((rq
= cfq_check_fifo(cfqq
)) == NULL
)
931 * finally, insert request into driver dispatch list
933 cfq_dispatch_insert(cfqd
->queue
, rq
);
935 cfqd
->dispatch_slice
++;
938 if (!cfqd
->active_cic
) {
939 atomic_inc(&RQ_CIC(rq
)->ioc
->refcount
);
940 cfqd
->active_cic
= RQ_CIC(rq
);
943 if (RB_EMPTY_ROOT(&cfqq
->sort_list
))
946 } while (dispatched
< max_dispatch
);
949 * if slice end isn't set yet, set it.
951 if (!cfqq
->slice_end
)
952 cfq_set_prio_slice(cfqd
, cfqq
);
955 * expire an async queue immediately if it has used up its slice. idle
956 * queue always expire after 1 dispatch round.
958 if ((!cfq_cfqq_sync(cfqq
) &&
959 cfqd
->dispatch_slice
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
960 cfq_class_idle(cfqq
) ||
961 !cfq_cfqq_idle_window(cfqq
))
962 cfq_slice_expired(cfqd
, 0);
968 cfq_forced_dispatch_cfqqs(struct list_head
*list
)
970 struct cfq_queue
*cfqq
, *next
;
974 list_for_each_entry_safe(cfqq
, next
, list
, cfq_list
) {
975 while (cfqq
->next_rq
) {
976 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
979 BUG_ON(!list_empty(&cfqq
->fifo
));
986 cfq_forced_dispatch(struct cfq_data
*cfqd
)
988 int i
, dispatched
= 0;
990 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
991 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->rr_list
[i
]);
993 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->busy_rr
);
994 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->cur_rr
);
995 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->idle_rr
);
997 cfq_slice_expired(cfqd
, 0);
999 BUG_ON(cfqd
->busy_queues
);
1005 cfq_dispatch_requests(request_queue_t
*q
, int force
)
1007 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1008 struct cfq_queue
*cfqq
, *prev_cfqq
;
1011 if (!cfqd
->busy_queues
)
1014 if (unlikely(force
))
1015 return cfq_forced_dispatch(cfqd
);
1019 while ((cfqq
= cfq_select_queue(cfqd
)) != NULL
) {
1023 * Don't repeat dispatch from the previous queue.
1025 if (prev_cfqq
== cfqq
)
1028 cfq_clear_cfqq_must_dispatch(cfqq
);
1029 cfq_clear_cfqq_wait_request(cfqq
);
1030 del_timer(&cfqd
->idle_slice_timer
);
1032 max_dispatch
= cfqd
->cfq_quantum
;
1033 if (cfq_class_idle(cfqq
))
1036 dispatched
+= __cfq_dispatch_requests(cfqd
, cfqq
, max_dispatch
);
1039 * If the dispatch cfqq has idling enabled and is still
1040 * the active queue, break out.
1042 if (cfq_cfqq_idle_window(cfqq
) && cfqd
->active_queue
)
1052 * task holds one reference to the queue, dropped when task exits. each rq
1053 * in-flight on this queue also holds a reference, dropped when rq is freed.
1055 * queue lock must be held here.
1057 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1059 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1061 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1063 if (!atomic_dec_and_test(&cfqq
->ref
))
1066 BUG_ON(rb_first(&cfqq
->sort_list
));
1067 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1068 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1070 if (unlikely(cfqd
->active_queue
== cfqq
))
1071 __cfq_slice_expired(cfqd
, cfqq
, 0);
1074 * it's on the empty list and still hashed
1076 list_del(&cfqq
->cfq_list
);
1077 hlist_del(&cfqq
->cfq_hash
);
1078 kmem_cache_free(cfq_pool
, cfqq
);
1081 static struct cfq_queue
*
1082 __cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned int prio
,
1085 struct hlist_head
*hash_list
= &cfqd
->cfq_hash
[hashval
];
1086 struct hlist_node
*entry
;
1087 struct cfq_queue
*__cfqq
;
1089 hlist_for_each_entry(__cfqq
, entry
, hash_list
, cfq_hash
) {
1090 const unsigned short __p
= IOPRIO_PRIO_VALUE(__cfqq
->org_ioprio_class
, __cfqq
->org_ioprio
);
1092 if (__cfqq
->key
== key
&& (__p
== prio
|| !prio
))
1099 static struct cfq_queue
*
1100 cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned short prio
)
1102 return __cfq_find_cfq_hash(cfqd
, key
, prio
, hash_long(key
, CFQ_QHASH_SHIFT
));
1105 static void cfq_free_io_context(struct io_context
*ioc
)
1107 struct cfq_io_context
*__cic
;
1111 while ((n
= rb_first(&ioc
->cic_root
)) != NULL
) {
1112 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1113 rb_erase(&__cic
->rb_node
, &ioc
->cic_root
);
1114 kmem_cache_free(cfq_ioc_pool
, __cic
);
1118 elv_ioc_count_mod(ioc_count
, -freed
);
1120 if (ioc_gone
&& !elv_ioc_count_read(ioc_count
))
1124 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1126 if (unlikely(cfqq
== cfqd
->active_queue
))
1127 __cfq_slice_expired(cfqd
, cfqq
, 0);
1129 cfq_put_queue(cfqq
);
1132 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1133 struct cfq_io_context
*cic
)
1135 list_del_init(&cic
->queue_list
);
1139 if (cic
->cfqq
[ASYNC
]) {
1140 cfq_exit_cfqq(cfqd
, cic
->cfqq
[ASYNC
]);
1141 cic
->cfqq
[ASYNC
] = NULL
;
1144 if (cic
->cfqq
[SYNC
]) {
1145 cfq_exit_cfqq(cfqd
, cic
->cfqq
[SYNC
]);
1146 cic
->cfqq
[SYNC
] = NULL
;
1152 * Called with interrupts disabled
1154 static void cfq_exit_single_io_context(struct cfq_io_context
*cic
)
1156 struct cfq_data
*cfqd
= cic
->key
;
1159 request_queue_t
*q
= cfqd
->queue
;
1161 spin_lock_irq(q
->queue_lock
);
1162 __cfq_exit_single_io_context(cfqd
, cic
);
1163 spin_unlock_irq(q
->queue_lock
);
1167 static void cfq_exit_io_context(struct io_context
*ioc
)
1169 struct cfq_io_context
*__cic
;
1173 * put the reference this task is holding to the various queues
1176 n
= rb_first(&ioc
->cic_root
);
1178 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1180 cfq_exit_single_io_context(__cic
);
1185 static struct cfq_io_context
*
1186 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1188 struct cfq_io_context
*cic
;
1190 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
, cfqd
->queue
->node
);
1192 memset(cic
, 0, sizeof(*cic
));
1193 cic
->last_end_request
= jiffies
;
1194 INIT_LIST_HEAD(&cic
->queue_list
);
1195 cic
->dtor
= cfq_free_io_context
;
1196 cic
->exit
= cfq_exit_io_context
;
1197 elv_ioc_count_inc(ioc_count
);
1203 static void cfq_init_prio_data(struct cfq_queue
*cfqq
)
1205 struct task_struct
*tsk
= current
;
1208 if (!cfq_cfqq_prio_changed(cfqq
))
1211 ioprio_class
= IOPRIO_PRIO_CLASS(tsk
->ioprio
);
1212 switch (ioprio_class
) {
1214 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1215 case IOPRIO_CLASS_NONE
:
1217 * no prio set, place us in the middle of the BE classes
1219 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1220 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1222 case IOPRIO_CLASS_RT
:
1223 cfqq
->ioprio
= task_ioprio(tsk
);
1224 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1226 case IOPRIO_CLASS_BE
:
1227 cfqq
->ioprio
= task_ioprio(tsk
);
1228 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1230 case IOPRIO_CLASS_IDLE
:
1231 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1233 cfq_clear_cfqq_idle_window(cfqq
);
1238 * keep track of original prio settings in case we have to temporarily
1239 * elevate the priority of this queue
1241 cfqq
->org_ioprio
= cfqq
->ioprio
;
1242 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1244 cfq_resort_rr_list(cfqq
, 0);
1245 cfq_clear_cfqq_prio_changed(cfqq
);
1248 static inline void changed_ioprio(struct cfq_io_context
*cic
)
1250 struct cfq_data
*cfqd
= cic
->key
;
1251 struct cfq_queue
*cfqq
;
1252 unsigned long flags
;
1254 if (unlikely(!cfqd
))
1257 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1259 cfqq
= cic
->cfqq
[ASYNC
];
1261 struct cfq_queue
*new_cfqq
;
1262 new_cfqq
= cfq_get_queue(cfqd
, CFQ_KEY_ASYNC
, cic
->ioc
->task
,
1265 cic
->cfqq
[ASYNC
] = new_cfqq
;
1266 cfq_put_queue(cfqq
);
1270 cfqq
= cic
->cfqq
[SYNC
];
1272 cfq_mark_cfqq_prio_changed(cfqq
);
1274 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1277 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1279 struct cfq_io_context
*cic
;
1282 ioc
->ioprio_changed
= 0;
1284 n
= rb_first(&ioc
->cic_root
);
1286 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1288 changed_ioprio(cic
);
1293 static struct cfq_queue
*
1294 cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
,
1297 const int hashval
= hash_long(key
, CFQ_QHASH_SHIFT
);
1298 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1299 unsigned short ioprio
;
1302 ioprio
= tsk
->ioprio
;
1303 cfqq
= __cfq_find_cfq_hash(cfqd
, key
, ioprio
, hashval
);
1309 } else if (gfp_mask
& __GFP_WAIT
) {
1311 * Inform the allocator of the fact that we will
1312 * just repeat this allocation if it fails, to allow
1313 * the allocator to do whatever it needs to attempt to
1316 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1317 new_cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
|__GFP_NOFAIL
, cfqd
->queue
->node
);
1318 spin_lock_irq(cfqd
->queue
->queue_lock
);
1321 cfqq
= kmem_cache_alloc_node(cfq_pool
, gfp_mask
, cfqd
->queue
->node
);
1326 memset(cfqq
, 0, sizeof(*cfqq
));
1328 INIT_HLIST_NODE(&cfqq
->cfq_hash
);
1329 INIT_LIST_HEAD(&cfqq
->cfq_list
);
1330 INIT_LIST_HEAD(&cfqq
->fifo
);
1333 hlist_add_head(&cfqq
->cfq_hash
, &cfqd
->cfq_hash
[hashval
]);
1334 atomic_set(&cfqq
->ref
, 0);
1337 * set ->slice_left to allow preemption for a new process
1339 cfqq
->slice_left
= 2 * cfqd
->cfq_slice_idle
;
1340 cfq_mark_cfqq_idle_window(cfqq
);
1341 cfq_mark_cfqq_prio_changed(cfqq
);
1342 cfq_mark_cfqq_queue_new(cfqq
);
1343 cfq_init_prio_data(cfqq
);
1347 kmem_cache_free(cfq_pool
, new_cfqq
);
1349 atomic_inc(&cfqq
->ref
);
1351 WARN_ON((gfp_mask
& __GFP_WAIT
) && !cfqq
);
1356 cfq_drop_dead_cic(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1358 WARN_ON(!list_empty(&cic
->queue_list
));
1359 rb_erase(&cic
->rb_node
, &ioc
->cic_root
);
1360 kmem_cache_free(cfq_ioc_pool
, cic
);
1361 elv_ioc_count_dec(ioc_count
);
1364 static struct cfq_io_context
*
1365 cfq_cic_rb_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1368 struct cfq_io_context
*cic
;
1369 void *k
, *key
= cfqd
;
1372 n
= ioc
->cic_root
.rb_node
;
1374 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1375 /* ->key must be copied to avoid race with cfq_exit_queue() */
1378 cfq_drop_dead_cic(ioc
, cic
);
1394 cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1395 struct cfq_io_context
*cic
)
1398 struct rb_node
*parent
;
1399 struct cfq_io_context
*__cic
;
1400 unsigned long flags
;
1408 p
= &ioc
->cic_root
.rb_node
;
1411 __cic
= rb_entry(parent
, struct cfq_io_context
, rb_node
);
1412 /* ->key must be copied to avoid race with cfq_exit_queue() */
1415 cfq_drop_dead_cic(ioc
, __cic
);
1421 else if (cic
->key
> k
)
1422 p
= &(*p
)->rb_right
;
1427 rb_link_node(&cic
->rb_node
, parent
, p
);
1428 rb_insert_color(&cic
->rb_node
, &ioc
->cic_root
);
1430 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1431 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1432 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1436 * Setup general io context and cfq io context. There can be several cfq
1437 * io contexts per general io context, if this process is doing io to more
1438 * than one device managed by cfq.
1440 static struct cfq_io_context
*
1441 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1443 struct io_context
*ioc
= NULL
;
1444 struct cfq_io_context
*cic
;
1446 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1448 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
1452 cic
= cfq_cic_rb_lookup(cfqd
, ioc
);
1456 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1460 cfq_cic_link(cfqd
, ioc
, cic
);
1462 smp_read_barrier_depends();
1463 if (unlikely(ioc
->ioprio_changed
))
1464 cfq_ioc_set_ioprio(ioc
);
1468 put_io_context(ioc
);
1473 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1475 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
1476 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1478 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1479 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1480 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1484 cfq_update_io_seektime(struct cfq_io_context
*cic
, struct request
*rq
)
1489 if (cic
->last_request_pos
< rq
->sector
)
1490 sdist
= rq
->sector
- cic
->last_request_pos
;
1492 sdist
= cic
->last_request_pos
- rq
->sector
;
1495 * Don't allow the seek distance to get too large from the
1496 * odd fragment, pagein, etc
1498 if (cic
->seek_samples
<= 60) /* second&third seek */
1499 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1501 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1503 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1504 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1505 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1506 do_div(total
, cic
->seek_samples
);
1507 cic
->seek_mean
= (sector_t
)total
;
1511 * Disable idle window if the process thinks too long or seeks so much that
1515 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1516 struct cfq_io_context
*cic
)
1518 int enable_idle
= cfq_cfqq_idle_window(cfqq
);
1520 if (!cic
->ioc
->task
|| !cfqd
->cfq_slice_idle
||
1521 (cfqd
->hw_tag
&& CIC_SEEKY(cic
)))
1523 else if (sample_valid(cic
->ttime_samples
)) {
1524 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
1531 cfq_mark_cfqq_idle_window(cfqq
);
1533 cfq_clear_cfqq_idle_window(cfqq
);
1538 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1539 * no or if we aren't sure, a 1 will cause a preempt.
1542 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
1545 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1547 if (cfq_class_idle(new_cfqq
))
1553 if (cfq_class_idle(cfqq
))
1555 if (!cfq_cfqq_wait_request(new_cfqq
))
1558 * if it doesn't have slice left, forget it
1560 if (new_cfqq
->slice_left
< cfqd
->cfq_slice_idle
)
1563 * if the new request is sync, but the currently running queue is
1564 * not, let the sync request have priority.
1566 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
1569 * So both queues are sync. Let the new request get disk time if
1570 * it's a metadata request and the current queue is doing regular IO.
1572 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
1579 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1580 * let it have half of its nominal slice.
1582 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1584 cfq_slice_expired(cfqd
, 1);
1586 if (!cfqq
->slice_left
)
1587 cfqq
->slice_left
= cfq_prio_to_slice(cfqd
, cfqq
) / 2;
1590 * Put the new queue at the front of the of the current list,
1591 * so we know that it will be selected next.
1593 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1594 list_move(&cfqq
->cfq_list
, &cfqd
->cur_rr
);
1596 cfqq
->slice_end
= cfqq
->slice_left
+ jiffies
;
1600 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1601 * something we should do about it
1604 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1607 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1610 cfqq
->meta_pending
++;
1613 * check if this request is a better next-serve candidate)) {
1615 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
1616 BUG_ON(!cfqq
->next_rq
);
1619 * we never wait for an async request and we don't allow preemption
1620 * of an async request. so just return early
1622 if (!rq_is_sync(rq
)) {
1624 * sync process issued an async request, if it's waiting
1625 * then expire it and kick rq handling.
1627 if (cic
== cfqd
->active_cic
&&
1628 del_timer(&cfqd
->idle_slice_timer
)) {
1629 cfq_slice_expired(cfqd
, 0);
1630 blk_start_queueing(cfqd
->queue
);
1635 cfq_update_io_thinktime(cfqd
, cic
);
1636 cfq_update_io_seektime(cic
, rq
);
1637 cfq_update_idle_window(cfqd
, cfqq
, cic
);
1639 cic
->last_request_pos
= rq
->sector
+ rq
->nr_sectors
;
1641 if (cfqq
== cfqd
->active_queue
) {
1643 * if we are waiting for a request for this queue, let it rip
1644 * immediately and flag that we must not expire this queue
1647 if (cfq_cfqq_wait_request(cfqq
)) {
1648 cfq_mark_cfqq_must_dispatch(cfqq
);
1649 del_timer(&cfqd
->idle_slice_timer
);
1650 blk_start_queueing(cfqd
->queue
);
1652 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
1654 * not the active queue - expire current slice if it is
1655 * idle and has expired it's mean thinktime or this new queue
1656 * has some old slice time left and is of higher priority
1658 cfq_preempt_queue(cfqd
, cfqq
);
1659 cfq_mark_cfqq_must_dispatch(cfqq
);
1660 blk_start_queueing(cfqd
->queue
);
1664 static void cfq_insert_request(request_queue_t
*q
, struct request
*rq
)
1666 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1667 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1669 cfq_init_prio_data(cfqq
);
1673 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
1675 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
1678 static void cfq_completed_request(request_queue_t
*q
, struct request
*rq
)
1680 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1681 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1682 const int sync
= rq_is_sync(rq
);
1687 WARN_ON(!cfqd
->rq_in_driver
);
1688 WARN_ON(!cfqq
->on_dispatch
[sync
]);
1689 cfqd
->rq_in_driver
--;
1690 cfqq
->on_dispatch
[sync
]--;
1692 if (!cfq_class_idle(cfqq
))
1693 cfqd
->last_end_request
= now
;
1695 cfq_resort_rr_list(cfqq
, 0);
1698 RQ_CIC(rq
)->last_end_request
= now
;
1701 * If this is the active queue, check if it needs to be expired,
1702 * or if we want to idle in case it has no pending requests.
1704 if (cfqd
->active_queue
== cfqq
) {
1705 if (time_after(now
, cfqq
->slice_end
))
1706 cfq_slice_expired(cfqd
, 0);
1707 else if (sync
&& RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1708 if (!cfq_arm_slice_timer(cfqd
, cfqq
))
1709 cfq_schedule_dispatch(cfqd
);
1715 * we temporarily boost lower priority queues if they are holding fs exclusive
1716 * resources. they are boosted to normal prio (CLASS_BE/4)
1718 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
1720 const int ioprio_class
= cfqq
->ioprio_class
;
1721 const int ioprio
= cfqq
->ioprio
;
1723 if (has_fs_excl()) {
1725 * boost idle prio on transactions that would lock out other
1726 * users of the filesystem
1728 if (cfq_class_idle(cfqq
))
1729 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1730 if (cfqq
->ioprio
> IOPRIO_NORM
)
1731 cfqq
->ioprio
= IOPRIO_NORM
;
1734 * check if we need to unboost the queue
1736 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
1737 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
1738 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
1739 cfqq
->ioprio
= cfqq
->org_ioprio
;
1743 * refile between round-robin lists if we moved the priority class
1745 if ((ioprio_class
!= cfqq
->ioprio_class
|| ioprio
!= cfqq
->ioprio
))
1746 cfq_resort_rr_list(cfqq
, 0);
1749 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
1751 if ((cfq_cfqq_wait_request(cfqq
) || cfq_cfqq_must_alloc(cfqq
)) &&
1752 !cfq_cfqq_must_alloc_slice(cfqq
)) {
1753 cfq_mark_cfqq_must_alloc_slice(cfqq
);
1754 return ELV_MQUEUE_MUST
;
1757 return ELV_MQUEUE_MAY
;
1760 static int cfq_may_queue(request_queue_t
*q
, int rw
)
1762 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1763 struct task_struct
*tsk
= current
;
1764 struct cfq_queue
*cfqq
;
1767 key
= cfq_queue_pid(tsk
, rw
, rw
& REQ_RW_SYNC
);
1770 * don't force setup of a queue from here, as a call to may_queue
1771 * does not necessarily imply that a request actually will be queued.
1772 * so just lookup a possibly existing queue, or return 'may queue'
1775 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
1777 cfq_init_prio_data(cfqq
);
1778 cfq_prio_boost(cfqq
);
1780 return __cfq_may_queue(cfqq
);
1783 return ELV_MQUEUE_MAY
;
1787 * queue lock held here
1789 static void cfq_put_request(struct request
*rq
)
1791 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1794 const int rw
= rq_data_dir(rq
);
1796 BUG_ON(!cfqq
->allocated
[rw
]);
1797 cfqq
->allocated
[rw
]--;
1799 put_io_context(RQ_CIC(rq
)->ioc
);
1801 rq
->elevator_private
= NULL
;
1802 rq
->elevator_private2
= NULL
;
1804 cfq_put_queue(cfqq
);
1809 * Allocate cfq data structures associated with this request.
1812 cfq_set_request(request_queue_t
*q
, struct request
*rq
, gfp_t gfp_mask
)
1814 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1815 struct task_struct
*tsk
= current
;
1816 struct cfq_io_context
*cic
;
1817 const int rw
= rq_data_dir(rq
);
1818 const int is_sync
= rq_is_sync(rq
);
1819 pid_t key
= cfq_queue_pid(tsk
, rw
, is_sync
);
1820 struct cfq_queue
*cfqq
;
1821 unsigned long flags
;
1823 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1825 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
1827 spin_lock_irqsave(q
->queue_lock
, flags
);
1832 if (!cic
->cfqq
[is_sync
]) {
1833 cfqq
= cfq_get_queue(cfqd
, key
, tsk
, gfp_mask
);
1837 cic
->cfqq
[is_sync
] = cfqq
;
1839 cfqq
= cic
->cfqq
[is_sync
];
1841 cfqq
->allocated
[rw
]++;
1842 cfq_clear_cfqq_must_alloc(cfqq
);
1843 atomic_inc(&cfqq
->ref
);
1845 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1847 rq
->elevator_private
= cic
;
1848 rq
->elevator_private2
= cfqq
;
1853 put_io_context(cic
->ioc
);
1855 cfq_schedule_dispatch(cfqd
);
1856 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1860 static void cfq_kick_queue(struct work_struct
*work
)
1862 struct cfq_data
*cfqd
=
1863 container_of(work
, struct cfq_data
, unplug_work
);
1864 request_queue_t
*q
= cfqd
->queue
;
1865 unsigned long flags
;
1867 spin_lock_irqsave(q
->queue_lock
, flags
);
1868 blk_start_queueing(q
);
1869 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1873 * Timer running if the active_queue is currently idling inside its time slice
1875 static void cfq_idle_slice_timer(unsigned long data
)
1877 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1878 struct cfq_queue
*cfqq
;
1879 unsigned long flags
;
1881 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1883 if ((cfqq
= cfqd
->active_queue
) != NULL
) {
1884 unsigned long now
= jiffies
;
1889 if (time_after(now
, cfqq
->slice_end
))
1893 * only expire and reinvoke request handler, if there are
1894 * other queues with pending requests
1896 if (!cfqd
->busy_queues
)
1900 * not expired and it has a request pending, let it dispatch
1902 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1903 cfq_mark_cfqq_must_dispatch(cfqq
);
1908 cfq_slice_expired(cfqd
, 0);
1910 cfq_schedule_dispatch(cfqd
);
1912 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1916 * Timer running if an idle class queue is waiting for service
1918 static void cfq_idle_class_timer(unsigned long data
)
1920 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1921 unsigned long flags
, end
;
1923 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1926 * race with a non-idle queue, reset timer
1928 end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
1929 if (!time_after_eq(jiffies
, end
))
1930 mod_timer(&cfqd
->idle_class_timer
, end
);
1932 cfq_schedule_dispatch(cfqd
);
1934 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1937 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
1939 del_timer_sync(&cfqd
->idle_slice_timer
);
1940 del_timer_sync(&cfqd
->idle_class_timer
);
1941 blk_sync_queue(cfqd
->queue
);
1944 static void cfq_exit_queue(elevator_t
*e
)
1946 struct cfq_data
*cfqd
= e
->elevator_data
;
1947 request_queue_t
*q
= cfqd
->queue
;
1949 cfq_shutdown_timer_wq(cfqd
);
1951 spin_lock_irq(q
->queue_lock
);
1953 if (cfqd
->active_queue
)
1954 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
1956 while (!list_empty(&cfqd
->cic_list
)) {
1957 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
1958 struct cfq_io_context
,
1961 __cfq_exit_single_io_context(cfqd
, cic
);
1964 spin_unlock_irq(q
->queue_lock
);
1966 cfq_shutdown_timer_wq(cfqd
);
1968 kfree(cfqd
->cfq_hash
);
1972 static void *cfq_init_queue(request_queue_t
*q
)
1974 struct cfq_data
*cfqd
;
1977 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
, q
->node
);
1981 memset(cfqd
, 0, sizeof(*cfqd
));
1983 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
1984 INIT_LIST_HEAD(&cfqd
->rr_list
[i
]);
1986 INIT_LIST_HEAD(&cfqd
->busy_rr
);
1987 INIT_LIST_HEAD(&cfqd
->cur_rr
);
1988 INIT_LIST_HEAD(&cfqd
->idle_rr
);
1989 INIT_LIST_HEAD(&cfqd
->cic_list
);
1991 cfqd
->cfq_hash
= kmalloc_node(sizeof(struct hlist_head
) * CFQ_QHASH_ENTRIES
, GFP_KERNEL
, q
->node
);
1992 if (!cfqd
->cfq_hash
)
1995 for (i
= 0; i
< CFQ_QHASH_ENTRIES
; i
++)
1996 INIT_HLIST_HEAD(&cfqd
->cfq_hash
[i
]);
2000 init_timer(&cfqd
->idle_slice_timer
);
2001 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2002 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2004 init_timer(&cfqd
->idle_class_timer
);
2005 cfqd
->idle_class_timer
.function
= cfq_idle_class_timer
;
2006 cfqd
->idle_class_timer
.data
= (unsigned long) cfqd
;
2008 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
2010 cfqd
->cfq_quantum
= cfq_quantum
;
2011 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2012 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2013 cfqd
->cfq_back_max
= cfq_back_max
;
2014 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2015 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2016 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2017 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2018 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2026 static void cfq_slab_kill(void)
2029 kmem_cache_destroy(cfq_pool
);
2031 kmem_cache_destroy(cfq_ioc_pool
);
2034 static int __init
cfq_slab_setup(void)
2036 cfq_pool
= kmem_cache_create("cfq_pool", sizeof(struct cfq_queue
), 0, 0,
2041 cfq_ioc_pool
= kmem_cache_create("cfq_ioc_pool",
2042 sizeof(struct cfq_io_context
), 0, 0, NULL
, NULL
);
2053 * sysfs parts below -->
2057 cfq_var_show(unsigned int var
, char *page
)
2059 return sprintf(page
, "%d\n", var
);
2063 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2065 char *p
= (char *) page
;
2067 *var
= simple_strtoul(p
, &p
, 10);
2071 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2072 static ssize_t __FUNC(elevator_t *e, char *page) \
2074 struct cfq_data *cfqd = e->elevator_data; \
2075 unsigned int __data = __VAR; \
2077 __data = jiffies_to_msecs(__data); \
2078 return cfq_var_show(__data, (page)); \
2080 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2081 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2082 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2083 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2084 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2085 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2086 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2087 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2088 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2089 #undef SHOW_FUNCTION
2091 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2092 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2094 struct cfq_data *cfqd = e->elevator_data; \
2095 unsigned int __data; \
2096 int ret = cfq_var_store(&__data, (page), count); \
2097 if (__data < (MIN)) \
2099 else if (__data > (MAX)) \
2102 *(__PTR) = msecs_to_jiffies(__data); \
2104 *(__PTR) = __data; \
2107 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2108 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1, UINT_MAX
, 1);
2109 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1, UINT_MAX
, 1);
2110 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2111 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1, UINT_MAX
, 0);
2112 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2113 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2114 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2115 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1, UINT_MAX
, 0);
2116 #undef STORE_FUNCTION
2118 #define CFQ_ATTR(name) \
2119 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2121 static struct elv_fs_entry cfq_attrs
[] = {
2123 CFQ_ATTR(fifo_expire_sync
),
2124 CFQ_ATTR(fifo_expire_async
),
2125 CFQ_ATTR(back_seek_max
),
2126 CFQ_ATTR(back_seek_penalty
),
2127 CFQ_ATTR(slice_sync
),
2128 CFQ_ATTR(slice_async
),
2129 CFQ_ATTR(slice_async_rq
),
2130 CFQ_ATTR(slice_idle
),
2134 static struct elevator_type iosched_cfq
= {
2136 .elevator_merge_fn
= cfq_merge
,
2137 .elevator_merged_fn
= cfq_merged_request
,
2138 .elevator_merge_req_fn
= cfq_merged_requests
,
2139 .elevator_allow_merge_fn
= cfq_allow_merge
,
2140 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2141 .elevator_add_req_fn
= cfq_insert_request
,
2142 .elevator_activate_req_fn
= cfq_activate_request
,
2143 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2144 .elevator_queue_empty_fn
= cfq_queue_empty
,
2145 .elevator_completed_req_fn
= cfq_completed_request
,
2146 .elevator_former_req_fn
= elv_rb_former_request
,
2147 .elevator_latter_req_fn
= elv_rb_latter_request
,
2148 .elevator_set_req_fn
= cfq_set_request
,
2149 .elevator_put_req_fn
= cfq_put_request
,
2150 .elevator_may_queue_fn
= cfq_may_queue
,
2151 .elevator_init_fn
= cfq_init_queue
,
2152 .elevator_exit_fn
= cfq_exit_queue
,
2153 .trim
= cfq_free_io_context
,
2155 .elevator_attrs
= cfq_attrs
,
2156 .elevator_name
= "cfq",
2157 .elevator_owner
= THIS_MODULE
,
2160 static int __init
cfq_init(void)
2165 * could be 0 on HZ < 1000 setups
2167 if (!cfq_slice_async
)
2168 cfq_slice_async
= 1;
2169 if (!cfq_slice_idle
)
2172 if (cfq_slab_setup())
2175 ret
= elv_register(&iosched_cfq
);
2182 static void __exit
cfq_exit(void)
2184 DECLARE_COMPLETION_ONSTACK(all_gone
);
2185 elv_unregister(&iosched_cfq
);
2186 ioc_gone
= &all_gone
;
2187 /* ioc_gone's update must be visible before reading ioc_count */
2189 if (elv_ioc_count_read(ioc_count
))
2190 wait_for_completion(ioc_gone
);
2195 module_init(cfq_init
);
2196 module_exit(cfq_exit
);
2198 MODULE_AUTHOR("Jens Axboe");
2199 MODULE_LICENSE("GPL");
2200 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");