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@suse.de>
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/hash.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
19 static const int cfq_quantum
= 4; /* max queue in one round of service */
20 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
21 static const int cfq_back_max
= 16 * 1024; /* maximum backwards seek, in KiB */
22 static const int cfq_back_penalty
= 2; /* penalty of a backwards seek */
24 static const int cfq_slice_sync
= HZ
/ 10;
25 static int cfq_slice_async
= HZ
/ 25;
26 static const int cfq_slice_async_rq
= 2;
27 static int cfq_slice_idle
= HZ
/ 125;
29 #define CFQ_IDLE_GRACE (HZ / 10)
30 #define CFQ_SLICE_SCALE (5)
32 #define CFQ_KEY_ASYNC (0)
35 * for the hash of cfqq inside the cfqd
37 #define CFQ_QHASH_SHIFT 6
38 #define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT)
39 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
41 #define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
43 #define RQ_CIC(rq) ((struct cfq_io_context*)(rq)->elevator_private)
44 #define RQ_CFQQ(rq) ((rq)->elevator_private2)
46 static kmem_cache_t
*cfq_pool
;
47 static kmem_cache_t
*cfq_ioc_pool
;
49 static atomic_t ioc_count
= ATOMIC_INIT(0);
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
;
85 * non-ordered list of empty cfqq's
87 struct list_head empty_list
;
92 struct hlist_head
*cfq_hash
;
98 * idle window management
100 struct timer_list idle_slice_timer
;
101 struct work_struct unplug_work
;
103 struct cfq_queue
*active_queue
;
104 struct cfq_io_context
*active_cic
;
105 int cur_prio
, cur_end_prio
;
106 unsigned int dispatch_slice
;
108 struct timer_list idle_class_timer
;
110 sector_t last_sector
;
111 unsigned long last_end_request
;
114 * tunables, see top of file
116 unsigned int cfq_quantum
;
117 unsigned int cfq_fifo_expire
[2];
118 unsigned int cfq_back_penalty
;
119 unsigned int cfq_back_max
;
120 unsigned int cfq_slice
[2];
121 unsigned int cfq_slice_async_rq
;
122 unsigned int cfq_slice_idle
;
124 struct list_head cic_list
;
128 * Per process-grouping structure
131 /* reference count */
133 /* parent cfq_data */
134 struct cfq_data
*cfqd
;
135 /* cfqq lookup hash */
136 struct hlist_node cfq_hash
;
139 /* on either rr or empty list of cfqd */
140 struct list_head cfq_list
;
141 /* sorted list of pending requests */
142 struct rb_root sort_list
;
143 /* if fifo isn't expired, next request to serve */
144 struct request
*next_rq
;
145 /* requests queued in sort_list */
147 /* currently allocated requests */
149 /* fifo list of requests in sort_list */
150 struct list_head fifo
;
152 unsigned long slice_start
;
153 unsigned long slice_end
;
154 unsigned long slice_left
;
155 unsigned long service_last
;
157 /* number of requests that are on the dispatch list */
160 /* io prio of this group */
161 unsigned short ioprio
, org_ioprio
;
162 unsigned short ioprio_class
, org_ioprio_class
;
164 /* various state flags, see below */
168 enum cfqq_state_flags
{
169 CFQ_CFQQ_FLAG_on_rr
= 0,
170 CFQ_CFQQ_FLAG_wait_request
,
171 CFQ_CFQQ_FLAG_must_alloc
,
172 CFQ_CFQQ_FLAG_must_alloc_slice
,
173 CFQ_CFQQ_FLAG_must_dispatch
,
174 CFQ_CFQQ_FLAG_fifo_expire
,
175 CFQ_CFQQ_FLAG_idle_window
,
176 CFQ_CFQQ_FLAG_prio_changed
,
179 #define CFQ_CFQQ_FNS(name) \
180 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
182 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
184 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
186 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
188 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
190 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
194 CFQ_CFQQ_FNS(wait_request
);
195 CFQ_CFQQ_FNS(must_alloc
);
196 CFQ_CFQQ_FNS(must_alloc_slice
);
197 CFQ_CFQQ_FNS(must_dispatch
);
198 CFQ_CFQQ_FNS(fifo_expire
);
199 CFQ_CFQQ_FNS(idle_window
);
200 CFQ_CFQQ_FNS(prio_changed
);
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
)
226 if (rw
== READ
|| rw
== WRITE_SYNC
)
229 return CFQ_KEY_ASYNC
;
233 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
234 * We choose the request that is closest to the head right now. Distance
235 * behind the head is penalized and only allowed to a certain extent.
237 static struct request
*
238 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
)
240 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
241 unsigned long back_max
;
242 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
243 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
244 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
246 if (rq1
== NULL
|| rq1
== rq2
)
251 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
253 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
259 last
= cfqd
->last_sector
;
262 * by definition, 1KiB is 2 sectors
264 back_max
= cfqd
->cfq_back_max
* 2;
267 * Strict one way elevator _except_ in the case where we allow
268 * short backward seeks which are biased as twice the cost of a
269 * similar forward seek.
273 else if (s1
+ back_max
>= last
)
274 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
276 wrap
|= CFQ_RQ1_WRAP
;
280 else if (s2
+ back_max
>= last
)
281 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
283 wrap
|= CFQ_RQ2_WRAP
;
285 /* Found required data */
288 * By doing switch() on the bit mask "wrap" we avoid having to
289 * check two variables for all permutations: --> faster!
292 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
308 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
311 * Since both rqs are wrapped,
312 * start with the one that's further behind head
313 * (--> only *one* back seek required),
314 * since back seek takes more time than forward.
324 * would be nice to take fifo expire time into account as well
326 static struct request
*
327 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
328 struct request
*last
)
330 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
331 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
332 struct request
*next
= NULL
, *prev
= NULL
;
334 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
337 prev
= rb_entry_rq(rbprev
);
340 next
= rb_entry_rq(rbnext
);
342 rbnext
= rb_first(&cfqq
->sort_list
);
343 if (rbnext
&& rbnext
!= &last
->rb_node
)
344 next
= rb_entry_rq(rbnext
);
347 return cfq_choose_req(cfqd
, next
, prev
);
350 static void cfq_resort_rr_list(struct cfq_queue
*cfqq
, int preempted
)
352 struct cfq_data
*cfqd
= cfqq
->cfqd
;
353 struct list_head
*list
, *entry
;
355 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
357 list_del(&cfqq
->cfq_list
);
359 if (cfq_class_rt(cfqq
))
360 list
= &cfqd
->cur_rr
;
361 else if (cfq_class_idle(cfqq
))
362 list
= &cfqd
->idle_rr
;
365 * if cfqq has requests in flight, don't allow it to be
366 * found in cfq_set_active_queue before it has finished them.
367 * this is done to increase fairness between a process that
368 * has lots of io pending vs one that only generates one
369 * sporadically or synchronously
371 if (cfq_cfqq_dispatched(cfqq
))
372 list
= &cfqd
->busy_rr
;
374 list
= &cfqd
->rr_list
[cfqq
->ioprio
];
378 * if queue was preempted, just add to front to be fair. busy_rr
379 * isn't sorted, but insert at the back for fairness.
381 if (preempted
|| list
== &cfqd
->busy_rr
) {
385 list_add_tail(&cfqq
->cfq_list
, list
);
390 * sort by when queue was last serviced
393 while ((entry
= entry
->prev
) != list
) {
394 struct cfq_queue
*__cfqq
= list_entry_cfqq(entry
);
396 if (!__cfqq
->service_last
)
398 if (time_before(__cfqq
->service_last
, cfqq
->service_last
))
402 list_add(&cfqq
->cfq_list
, entry
);
406 * add to busy list of queues for service, trying to be fair in ordering
407 * the pending list according to last request service
410 cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
412 BUG_ON(cfq_cfqq_on_rr(cfqq
));
413 cfq_mark_cfqq_on_rr(cfqq
);
416 cfq_resort_rr_list(cfqq
, 0);
420 cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
422 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
423 cfq_clear_cfqq_on_rr(cfqq
);
424 list_move(&cfqq
->cfq_list
, &cfqd
->empty_list
);
426 BUG_ON(!cfqd
->busy_queues
);
431 * rb tree support functions
433 static inline void cfq_del_rq_rb(struct request
*rq
)
435 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
436 struct cfq_data
*cfqd
= cfqq
->cfqd
;
437 const int sync
= rq_is_sync(rq
);
439 BUG_ON(!cfqq
->queued
[sync
]);
440 cfqq
->queued
[sync
]--;
442 elv_rb_del(&cfqq
->sort_list
, rq
);
444 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
445 cfq_del_cfqq_rr(cfqd
, cfqq
);
448 static void cfq_add_rq_rb(struct request
*rq
)
450 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
451 struct cfq_data
*cfqd
= cfqq
->cfqd
;
452 struct request
*__alias
;
454 cfqq
->queued
[rq_is_sync(rq
)]++;
457 * looks a little odd, but the first insert might return an alias.
458 * if that happens, put the alias on the dispatch list
460 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
461 cfq_dispatch_insert(cfqd
->queue
, __alias
);
465 cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
467 elv_rb_del(&cfqq
->sort_list
, rq
);
468 cfqq
->queued
[rq_is_sync(rq
)]--;
472 static struct request
*
473 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
475 struct task_struct
*tsk
= current
;
476 pid_t key
= cfq_queue_pid(tsk
, bio_data_dir(bio
));
477 struct cfq_queue
*cfqq
;
479 cfqq
= cfq_find_cfq_hash(cfqd
, key
, tsk
->ioprio
);
481 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
483 return elv_rb_find(&cfqq
->sort_list
, sector
);
489 static void cfq_activate_request(request_queue_t
*q
, struct request
*rq
)
491 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
493 cfqd
->rq_in_driver
++;
496 * If the depth is larger 1, it really could be queueing. But lets
497 * make the mark a little higher - idling could still be good for
498 * low queueing, and a low queueing number could also just indicate
499 * a SCSI mid layer like behaviour where limit+1 is often seen.
501 if (!cfqd
->hw_tag
&& cfqd
->rq_in_driver
> 4)
505 static void cfq_deactivate_request(request_queue_t
*q
, struct request
*rq
)
507 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
509 WARN_ON(!cfqd
->rq_in_driver
);
510 cfqd
->rq_in_driver
--;
513 static void cfq_remove_request(struct request
*rq
)
515 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
517 if (cfqq
->next_rq
== rq
)
518 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
520 list_del_init(&rq
->queuelist
);
525 cfq_merge(request_queue_t
*q
, struct request
**req
, struct bio
*bio
)
527 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
528 struct request
*__rq
;
530 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
531 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
533 return ELEVATOR_FRONT_MERGE
;
536 return ELEVATOR_NO_MERGE
;
539 static void cfq_merged_request(request_queue_t
*q
, struct request
*req
,
542 if (type
== ELEVATOR_FRONT_MERGE
) {
543 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
545 cfq_reposition_rq_rb(cfqq
, req
);
550 cfq_merged_requests(request_queue_t
*q
, struct request
*rq
,
551 struct request
*next
)
554 * reposition in fifo if next is older than rq
556 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
557 time_before(next
->start_time
, rq
->start_time
))
558 list_move(&rq
->queuelist
, &next
->queuelist
);
560 cfq_remove_request(next
);
564 __cfq_set_active_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
568 * stop potential idle class queues waiting service
570 del_timer(&cfqd
->idle_class_timer
);
572 cfqq
->slice_start
= jiffies
;
574 cfqq
->slice_left
= 0;
575 cfq_clear_cfqq_must_alloc_slice(cfqq
);
576 cfq_clear_cfqq_fifo_expire(cfqq
);
579 cfqd
->active_queue
= cfqq
;
583 * current cfqq expired its slice (or was too idle), select new one
586 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
589 unsigned long now
= jiffies
;
591 if (cfq_cfqq_wait_request(cfqq
))
592 del_timer(&cfqd
->idle_slice_timer
);
594 if (!preempted
&& !cfq_cfqq_dispatched(cfqq
)) {
595 cfqq
->service_last
= now
;
596 cfq_schedule_dispatch(cfqd
);
599 cfq_clear_cfqq_must_dispatch(cfqq
);
600 cfq_clear_cfqq_wait_request(cfqq
);
603 * store what was left of this slice, if the queue idled out
606 if (time_after(cfqq
->slice_end
, now
))
607 cfqq
->slice_left
= cfqq
->slice_end
- now
;
609 cfqq
->slice_left
= 0;
611 if (cfq_cfqq_on_rr(cfqq
))
612 cfq_resort_rr_list(cfqq
, preempted
);
614 if (cfqq
== cfqd
->active_queue
)
615 cfqd
->active_queue
= NULL
;
617 if (cfqd
->active_cic
) {
618 put_io_context(cfqd
->active_cic
->ioc
);
619 cfqd
->active_cic
= NULL
;
622 cfqd
->dispatch_slice
= 0;
625 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, int preempted
)
627 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
630 __cfq_slice_expired(cfqd
, cfqq
, preempted
);
643 static int cfq_get_next_prio_level(struct cfq_data
*cfqd
)
652 for (p
= cfqd
->cur_prio
; p
<= cfqd
->cur_end_prio
; p
++) {
653 if (!list_empty(&cfqd
->rr_list
[p
])) {
662 if (++cfqd
->cur_end_prio
== CFQ_PRIO_LISTS
) {
663 cfqd
->cur_end_prio
= 0;
670 if (unlikely(prio
== -1))
673 BUG_ON(prio
>= CFQ_PRIO_LISTS
);
675 list_splice_init(&cfqd
->rr_list
[prio
], &cfqd
->cur_rr
);
677 cfqd
->cur_prio
= prio
+ 1;
678 if (cfqd
->cur_prio
> cfqd
->cur_end_prio
) {
679 cfqd
->cur_end_prio
= cfqd
->cur_prio
;
682 if (cfqd
->cur_end_prio
== CFQ_PRIO_LISTS
) {
684 cfqd
->cur_end_prio
= 0;
690 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
)
692 struct cfq_queue
*cfqq
= NULL
;
694 if (!list_empty(&cfqd
->cur_rr
) || cfq_get_next_prio_level(cfqd
) != -1) {
696 * if current list is non-empty, grab first entry. if it is
697 * empty, get next prio level and grab first entry then if any
700 cfqq
= list_entry_cfqq(cfqd
->cur_rr
.next
);
701 } else if (!list_empty(&cfqd
->busy_rr
)) {
703 * If no new queues are available, check if the busy list has
704 * some before falling back to idle io.
706 cfqq
= list_entry_cfqq(cfqd
->busy_rr
.next
);
707 } else if (!list_empty(&cfqd
->idle_rr
)) {
709 * if we have idle queues and no rt or be queues had pending
710 * requests, either allow immediate service if the grace period
711 * has passed or arm the idle grace timer
713 unsigned long end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
715 if (time_after_eq(jiffies
, end
))
716 cfqq
= list_entry_cfqq(cfqd
->idle_rr
.next
);
718 mod_timer(&cfqd
->idle_class_timer
, end
);
721 __cfq_set_active_queue(cfqd
, cfqq
);
725 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
727 static int cfq_arm_slice_timer(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
730 struct cfq_io_context
*cic
;
733 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
734 WARN_ON(cfqq
!= cfqd
->active_queue
);
737 * idle is disabled, either manually or by past process history
739 if (!cfqd
->cfq_slice_idle
)
741 if (!cfq_cfqq_idle_window(cfqq
))
744 * task has exited, don't wait
746 cic
= cfqd
->active_cic
;
747 if (!cic
|| !cic
->ioc
->task
)
750 cfq_mark_cfqq_must_dispatch(cfqq
);
751 cfq_mark_cfqq_wait_request(cfqq
);
753 sl
= min(cfqq
->slice_end
- 1, (unsigned long) cfqd
->cfq_slice_idle
);
756 * we don't want to idle for seeks, but we do want to allow
757 * fair distribution of slice time for a process doing back-to-back
758 * seeks. so allow a little bit of time for him to submit a new rq
760 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
761 sl
= min(sl
, msecs_to_jiffies(2));
763 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
767 static void cfq_dispatch_insert(request_queue_t
*q
, struct request
*rq
)
769 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
770 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
772 cfq_remove_request(rq
);
773 cfqq
->on_dispatch
[rq_is_sync(rq
)]++;
774 elv_dispatch_sort(q
, rq
);
776 rq
= list_entry(q
->queue_head
.prev
, struct request
, queuelist
);
777 cfqd
->last_sector
= rq
->sector
+ rq
->nr_sectors
;
781 * return expired entry, or NULL to just start from scratch in rbtree
783 static inline struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
785 struct cfq_data
*cfqd
= cfqq
->cfqd
;
789 if (cfq_cfqq_fifo_expire(cfqq
))
791 if (list_empty(&cfqq
->fifo
))
794 fifo
= cfq_cfqq_class_sync(cfqq
);
795 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
797 if (time_after(jiffies
, rq
->start_time
+ cfqd
->cfq_fifo_expire
[fifo
])) {
798 cfq_mark_cfqq_fifo_expire(cfqq
);
806 * Scale schedule slice based on io priority. Use the sync time slice only
807 * if a queue is marked sync and has sync io queued. A sync queue with async
808 * io only, should not get full sync slice length.
811 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
813 const int base_slice
= cfqd
->cfq_slice
[cfq_cfqq_sync(cfqq
)];
815 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
817 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - cfqq
->ioprio
));
821 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
823 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
827 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
829 const int base_rq
= cfqd
->cfq_slice_async_rq
;
831 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
833 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
837 * get next queue for service
839 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
841 unsigned long now
= jiffies
;
842 struct cfq_queue
*cfqq
;
844 cfqq
= cfqd
->active_queue
;
851 if (!cfq_cfqq_must_dispatch(cfqq
) && time_after(now
, cfqq
->slice_end
))
855 * if queue has requests, dispatch one. if not, check if
856 * enough slice is left to wait for one
858 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
860 else if (cfq_cfqq_dispatched(cfqq
)) {
863 } else if (cfq_cfqq_class_sync(cfqq
)) {
864 if (cfq_arm_slice_timer(cfqd
, cfqq
))
869 cfq_slice_expired(cfqd
, 0);
871 cfqq
= cfq_set_active_queue(cfqd
);
877 __cfq_dispatch_requests(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
882 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
888 * follow expired path, else get first next available
890 if ((rq
= cfq_check_fifo(cfqq
)) == NULL
)
894 * finally, insert request into driver dispatch list
896 cfq_dispatch_insert(cfqd
->queue
, rq
);
898 cfqd
->dispatch_slice
++;
901 if (!cfqd
->active_cic
) {
902 atomic_inc(&RQ_CIC(rq
)->ioc
->refcount
);
903 cfqd
->active_cic
= RQ_CIC(rq
);
906 if (RB_EMPTY_ROOT(&cfqq
->sort_list
))
909 } while (dispatched
< max_dispatch
);
912 * if slice end isn't set yet, set it.
914 if (!cfqq
->slice_end
)
915 cfq_set_prio_slice(cfqd
, cfqq
);
918 * expire an async queue immediately if it has used up its slice. idle
919 * queue always expire after 1 dispatch round.
921 if ((!cfq_cfqq_sync(cfqq
) &&
922 cfqd
->dispatch_slice
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
923 cfq_class_idle(cfqq
) ||
924 !cfq_cfqq_idle_window(cfqq
))
925 cfq_slice_expired(cfqd
, 0);
931 cfq_forced_dispatch_cfqqs(struct list_head
*list
)
933 struct cfq_queue
*cfqq
, *next
;
937 list_for_each_entry_safe(cfqq
, next
, list
, cfq_list
) {
938 while (cfqq
->next_rq
) {
939 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
942 BUG_ON(!list_empty(&cfqq
->fifo
));
949 cfq_forced_dispatch(struct cfq_data
*cfqd
)
951 int i
, dispatched
= 0;
953 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
954 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->rr_list
[i
]);
956 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->busy_rr
);
957 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->cur_rr
);
958 dispatched
+= cfq_forced_dispatch_cfqqs(&cfqd
->idle_rr
);
960 cfq_slice_expired(cfqd
, 0);
962 BUG_ON(cfqd
->busy_queues
);
968 cfq_dispatch_requests(request_queue_t
*q
, int force
)
970 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
971 struct cfq_queue
*cfqq
, *prev_cfqq
;
974 if (!cfqd
->busy_queues
)
978 return cfq_forced_dispatch(cfqd
);
982 while ((cfqq
= cfq_select_queue(cfqd
)) != NULL
) {
986 * Don't repeat dispatch from the previous queue.
988 if (prev_cfqq
== cfqq
)
991 cfq_clear_cfqq_must_dispatch(cfqq
);
992 cfq_clear_cfqq_wait_request(cfqq
);
993 del_timer(&cfqd
->idle_slice_timer
);
995 max_dispatch
= cfqd
->cfq_quantum
;
996 if (cfq_class_idle(cfqq
))
999 dispatched
+= __cfq_dispatch_requests(cfqd
, cfqq
, max_dispatch
);
1002 * If the dispatch cfqq has idling enabled and is still
1003 * the active queue, break out.
1005 if (cfq_cfqq_idle_window(cfqq
) && cfqd
->active_queue
)
1015 * task holds one reference to the queue, dropped when task exits. each rq
1016 * in-flight on this queue also holds a reference, dropped when rq is freed.
1018 * queue lock must be held here.
1020 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1022 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1024 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1026 if (!atomic_dec_and_test(&cfqq
->ref
))
1029 BUG_ON(rb_first(&cfqq
->sort_list
));
1030 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1031 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1033 if (unlikely(cfqd
->active_queue
== cfqq
))
1034 __cfq_slice_expired(cfqd
, cfqq
, 0);
1037 * it's on the empty list and still hashed
1039 list_del(&cfqq
->cfq_list
);
1040 hlist_del(&cfqq
->cfq_hash
);
1041 kmem_cache_free(cfq_pool
, cfqq
);
1044 static inline struct cfq_queue
*
1045 __cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned int prio
,
1048 struct hlist_head
*hash_list
= &cfqd
->cfq_hash
[hashval
];
1049 struct hlist_node
*entry
;
1050 struct cfq_queue
*__cfqq
;
1052 hlist_for_each_entry(__cfqq
, entry
, hash_list
, cfq_hash
) {
1053 const unsigned short __p
= IOPRIO_PRIO_VALUE(__cfqq
->org_ioprio_class
, __cfqq
->org_ioprio
);
1055 if (__cfqq
->key
== key
&& (__p
== prio
|| !prio
))
1062 static struct cfq_queue
*
1063 cfq_find_cfq_hash(struct cfq_data
*cfqd
, unsigned int key
, unsigned short prio
)
1065 return __cfq_find_cfq_hash(cfqd
, key
, prio
, hash_long(key
, CFQ_QHASH_SHIFT
));
1068 static void cfq_free_io_context(struct io_context
*ioc
)
1070 struct cfq_io_context
*__cic
;
1074 while ((n
= rb_first(&ioc
->cic_root
)) != NULL
) {
1075 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1076 rb_erase(&__cic
->rb_node
, &ioc
->cic_root
);
1077 kmem_cache_free(cfq_ioc_pool
, __cic
);
1081 if (atomic_sub_and_test(freed
, &ioc_count
) && ioc_gone
)
1085 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1087 if (unlikely(cfqq
== cfqd
->active_queue
))
1088 __cfq_slice_expired(cfqd
, cfqq
, 0);
1090 cfq_put_queue(cfqq
);
1093 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1094 struct cfq_io_context
*cic
)
1096 list_del_init(&cic
->queue_list
);
1100 if (cic
->cfqq
[ASYNC
]) {
1101 cfq_exit_cfqq(cfqd
, cic
->cfqq
[ASYNC
]);
1102 cic
->cfqq
[ASYNC
] = NULL
;
1105 if (cic
->cfqq
[SYNC
]) {
1106 cfq_exit_cfqq(cfqd
, cic
->cfqq
[SYNC
]);
1107 cic
->cfqq
[SYNC
] = NULL
;
1113 * Called with interrupts disabled
1115 static void cfq_exit_single_io_context(struct cfq_io_context
*cic
)
1117 struct cfq_data
*cfqd
= cic
->key
;
1120 request_queue_t
*q
= cfqd
->queue
;
1122 spin_lock_irq(q
->queue_lock
);
1123 __cfq_exit_single_io_context(cfqd
, cic
);
1124 spin_unlock_irq(q
->queue_lock
);
1128 static void cfq_exit_io_context(struct io_context
*ioc
)
1130 struct cfq_io_context
*__cic
;
1134 * put the reference this task is holding to the various queues
1137 n
= rb_first(&ioc
->cic_root
);
1139 __cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1141 cfq_exit_single_io_context(__cic
);
1146 static struct cfq_io_context
*
1147 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1149 struct cfq_io_context
*cic
= kmem_cache_alloc(cfq_ioc_pool
, gfp_mask
);
1152 memset(cic
, 0, sizeof(*cic
));
1153 cic
->last_end_request
= jiffies
;
1154 INIT_LIST_HEAD(&cic
->queue_list
);
1155 cic
->dtor
= cfq_free_io_context
;
1156 cic
->exit
= cfq_exit_io_context
;
1157 atomic_inc(&ioc_count
);
1163 static void cfq_init_prio_data(struct cfq_queue
*cfqq
)
1165 struct task_struct
*tsk
= current
;
1168 if (!cfq_cfqq_prio_changed(cfqq
))
1171 ioprio_class
= IOPRIO_PRIO_CLASS(tsk
->ioprio
);
1172 switch (ioprio_class
) {
1174 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1175 case IOPRIO_CLASS_NONE
:
1177 * no prio set, place us in the middle of the BE classes
1179 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1180 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1182 case IOPRIO_CLASS_RT
:
1183 cfqq
->ioprio
= task_ioprio(tsk
);
1184 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1186 case IOPRIO_CLASS_BE
:
1187 cfqq
->ioprio
= task_ioprio(tsk
);
1188 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1190 case IOPRIO_CLASS_IDLE
:
1191 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1193 cfq_clear_cfqq_idle_window(cfqq
);
1198 * keep track of original prio settings in case we have to temporarily
1199 * elevate the priority of this queue
1201 cfqq
->org_ioprio
= cfqq
->ioprio
;
1202 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1204 if (cfq_cfqq_on_rr(cfqq
))
1205 cfq_resort_rr_list(cfqq
, 0);
1207 cfq_clear_cfqq_prio_changed(cfqq
);
1210 static inline void changed_ioprio(struct cfq_io_context
*cic
)
1212 struct cfq_data
*cfqd
= cic
->key
;
1213 struct cfq_queue
*cfqq
;
1215 if (unlikely(!cfqd
))
1218 spin_lock(cfqd
->queue
->queue_lock
);
1220 cfqq
= cic
->cfqq
[ASYNC
];
1222 struct cfq_queue
*new_cfqq
;
1223 new_cfqq
= cfq_get_queue(cfqd
, CFQ_KEY_ASYNC
, cic
->ioc
->task
,
1226 cic
->cfqq
[ASYNC
] = new_cfqq
;
1227 cfq_put_queue(cfqq
);
1231 cfqq
= cic
->cfqq
[SYNC
];
1233 cfq_mark_cfqq_prio_changed(cfqq
);
1235 spin_unlock(cfqd
->queue
->queue_lock
);
1238 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1240 struct cfq_io_context
*cic
;
1243 ioc
->ioprio_changed
= 0;
1245 n
= rb_first(&ioc
->cic_root
);
1247 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1249 changed_ioprio(cic
);
1254 static struct cfq_queue
*
1255 cfq_get_queue(struct cfq_data
*cfqd
, unsigned int key
, struct task_struct
*tsk
,
1258 const int hashval
= hash_long(key
, CFQ_QHASH_SHIFT
);
1259 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1260 unsigned short ioprio
;
1263 ioprio
= tsk
->ioprio
;
1264 cfqq
= __cfq_find_cfq_hash(cfqd
, key
, ioprio
, hashval
);
1270 } else if (gfp_mask
& __GFP_WAIT
) {
1272 * Inform the allocator of the fact that we will
1273 * just repeat this allocation if it fails, to allow
1274 * the allocator to do whatever it needs to attempt to
1277 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1278 new_cfqq
= kmem_cache_alloc(cfq_pool
, gfp_mask
|__GFP_NOFAIL
);
1279 spin_lock_irq(cfqd
->queue
->queue_lock
);
1282 cfqq
= kmem_cache_alloc(cfq_pool
, gfp_mask
);
1287 memset(cfqq
, 0, sizeof(*cfqq
));
1289 INIT_HLIST_NODE(&cfqq
->cfq_hash
);
1290 INIT_LIST_HEAD(&cfqq
->cfq_list
);
1291 INIT_LIST_HEAD(&cfqq
->fifo
);
1294 hlist_add_head(&cfqq
->cfq_hash
, &cfqd
->cfq_hash
[hashval
]);
1295 atomic_set(&cfqq
->ref
, 0);
1297 cfqq
->service_last
= 0;
1299 * set ->slice_left to allow preemption for a new process
1301 cfqq
->slice_left
= 2 * cfqd
->cfq_slice_idle
;
1302 cfq_mark_cfqq_idle_window(cfqq
);
1303 cfq_mark_cfqq_prio_changed(cfqq
);
1304 cfq_init_prio_data(cfqq
);
1308 kmem_cache_free(cfq_pool
, new_cfqq
);
1310 atomic_inc(&cfqq
->ref
);
1312 WARN_ON((gfp_mask
& __GFP_WAIT
) && !cfqq
);
1317 cfq_drop_dead_cic(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1319 WARN_ON(!list_empty(&cic
->queue_list
));
1320 rb_erase(&cic
->rb_node
, &ioc
->cic_root
);
1321 kmem_cache_free(cfq_ioc_pool
, cic
);
1322 atomic_dec(&ioc_count
);
1325 static struct cfq_io_context
*
1326 cfq_cic_rb_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1329 struct cfq_io_context
*cic
;
1330 void *k
, *key
= cfqd
;
1333 n
= ioc
->cic_root
.rb_node
;
1335 cic
= rb_entry(n
, struct cfq_io_context
, rb_node
);
1336 /* ->key must be copied to avoid race with cfq_exit_queue() */
1339 cfq_drop_dead_cic(ioc
, cic
);
1355 cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1356 struct cfq_io_context
*cic
)
1359 struct rb_node
*parent
;
1360 struct cfq_io_context
*__cic
;
1368 p
= &ioc
->cic_root
.rb_node
;
1371 __cic
= rb_entry(parent
, struct cfq_io_context
, rb_node
);
1372 /* ->key must be copied to avoid race with cfq_exit_queue() */
1375 cfq_drop_dead_cic(ioc
, __cic
);
1381 else if (cic
->key
> k
)
1382 p
= &(*p
)->rb_right
;
1387 rb_link_node(&cic
->rb_node
, parent
, p
);
1388 rb_insert_color(&cic
->rb_node
, &ioc
->cic_root
);
1390 spin_lock_irq(cfqd
->queue
->queue_lock
);
1391 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1392 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1396 * Setup general io context and cfq io context. There can be several cfq
1397 * io contexts per general io context, if this process is doing io to more
1398 * than one device managed by cfq.
1400 static struct cfq_io_context
*
1401 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1403 struct io_context
*ioc
= NULL
;
1404 struct cfq_io_context
*cic
;
1406 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1408 ioc
= get_io_context(gfp_mask
);
1412 cic
= cfq_cic_rb_lookup(cfqd
, ioc
);
1416 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1420 cfq_cic_link(cfqd
, ioc
, cic
);
1422 smp_read_barrier_depends();
1423 if (unlikely(ioc
->ioprio_changed
))
1424 cfq_ioc_set_ioprio(ioc
);
1428 put_io_context(ioc
);
1433 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1435 unsigned long elapsed
, ttime
;
1438 * if this context already has stuff queued, thinktime is from
1439 * last queue not last end
1442 if (time_after(cic
->last_end_request
, cic
->last_queue
))
1443 elapsed
= jiffies
- cic
->last_end_request
;
1445 elapsed
= jiffies
- cic
->last_queue
;
1447 elapsed
= jiffies
- cic
->last_end_request
;
1450 ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1452 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1453 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1454 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1458 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
1464 if (cic
->last_request_pos
< rq
->sector
)
1465 sdist
= rq
->sector
- cic
->last_request_pos
;
1467 sdist
= cic
->last_request_pos
- rq
->sector
;
1470 * Don't allow the seek distance to get too large from the
1471 * odd fragment, pagein, etc
1473 if (cic
->seek_samples
<= 60) /* second&third seek */
1474 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1476 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1478 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1479 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1480 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1481 do_div(total
, cic
->seek_samples
);
1482 cic
->seek_mean
= (sector_t
)total
;
1486 * Disable idle window if the process thinks too long or seeks so much that
1490 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1491 struct cfq_io_context
*cic
)
1493 int enable_idle
= cfq_cfqq_idle_window(cfqq
);
1495 if (!cic
->ioc
->task
|| !cfqd
->cfq_slice_idle
||
1496 (cfqd
->hw_tag
&& CIC_SEEKY(cic
)))
1498 else if (sample_valid(cic
->ttime_samples
)) {
1499 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
1506 cfq_mark_cfqq_idle_window(cfqq
);
1508 cfq_clear_cfqq_idle_window(cfqq
);
1513 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1514 * no or if we aren't sure, a 1 will cause a preempt.
1517 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
1520 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1522 if (cfq_class_idle(new_cfqq
))
1528 if (cfq_class_idle(cfqq
))
1530 if (!cfq_cfqq_wait_request(new_cfqq
))
1533 * if it doesn't have slice left, forget it
1535 if (new_cfqq
->slice_left
< cfqd
->cfq_slice_idle
)
1537 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
1544 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1545 * let it have half of its nominal slice.
1547 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1549 struct cfq_queue
*__cfqq
, *next
;
1551 list_for_each_entry_safe(__cfqq
, next
, &cfqd
->cur_rr
, cfq_list
)
1552 cfq_resort_rr_list(__cfqq
, 1);
1554 if (!cfqq
->slice_left
)
1555 cfqq
->slice_left
= cfq_prio_to_slice(cfqd
, cfqq
) / 2;
1557 cfqq
->slice_end
= cfqq
->slice_left
+ jiffies
;
1558 cfq_slice_expired(cfqd
, 1);
1559 __cfq_set_active_queue(cfqd
, cfqq
);
1563 * should really be a ll_rw_blk.c helper
1565 static void cfq_start_queueing(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1567 request_queue_t
*q
= cfqd
->queue
;
1569 if (!blk_queue_plugged(q
))
1572 __generic_unplug_device(q
);
1576 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1577 * something we should do about it
1580 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1583 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1586 * check if this request is a better next-serve candidate)) {
1588 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
1589 BUG_ON(!cfqq
->next_rq
);
1592 * we never wait for an async request and we don't allow preemption
1593 * of an async request. so just return early
1595 if (!rq_is_sync(rq
)) {
1597 * sync process issued an async request, if it's waiting
1598 * then expire it and kick rq handling.
1600 if (cic
== cfqd
->active_cic
&&
1601 del_timer(&cfqd
->idle_slice_timer
)) {
1602 cfq_slice_expired(cfqd
, 0);
1603 cfq_start_queueing(cfqd
, cfqq
);
1608 cfq_update_io_thinktime(cfqd
, cic
);
1609 cfq_update_io_seektime(cfqd
, cic
, rq
);
1610 cfq_update_idle_window(cfqd
, cfqq
, cic
);
1612 cic
->last_queue
= jiffies
;
1613 cic
->last_request_pos
= rq
->sector
+ rq
->nr_sectors
;
1615 if (cfqq
== cfqd
->active_queue
) {
1617 * if we are waiting for a request for this queue, let it rip
1618 * immediately and flag that we must not expire this queue
1621 if (cfq_cfqq_wait_request(cfqq
)) {
1622 cfq_mark_cfqq_must_dispatch(cfqq
);
1623 del_timer(&cfqd
->idle_slice_timer
);
1624 cfq_start_queueing(cfqd
, cfqq
);
1626 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
1628 * not the active queue - expire current slice if it is
1629 * idle and has expired it's mean thinktime or this new queue
1630 * has some old slice time left and is of higher priority
1632 cfq_preempt_queue(cfqd
, cfqq
);
1633 cfq_mark_cfqq_must_dispatch(cfqq
);
1634 cfq_start_queueing(cfqd
, cfqq
);
1638 static void cfq_insert_request(request_queue_t
*q
, struct request
*rq
)
1640 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1641 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1643 cfq_init_prio_data(cfqq
);
1647 if (!cfq_cfqq_on_rr(cfqq
))
1648 cfq_add_cfqq_rr(cfqd
, cfqq
);
1650 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
1652 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
1655 static void cfq_completed_request(request_queue_t
*q
, struct request
*rq
)
1657 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1658 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1659 const int sync
= rq_is_sync(rq
);
1664 WARN_ON(!cfqd
->rq_in_driver
);
1665 WARN_ON(!cfqq
->on_dispatch
[sync
]);
1666 cfqd
->rq_in_driver
--;
1667 cfqq
->on_dispatch
[sync
]--;
1669 if (!cfq_class_idle(cfqq
))
1670 cfqd
->last_end_request
= now
;
1672 if (!cfq_cfqq_dispatched(cfqq
)) {
1673 if (cfq_cfqq_on_rr(cfqq
)) {
1674 cfqq
->service_last
= now
;
1675 cfq_resort_rr_list(cfqq
, 0);
1680 RQ_CIC(rq
)->last_end_request
= now
;
1683 * If this is the active queue, check if it needs to be expired,
1684 * or if we want to idle in case it has no pending requests.
1686 if (cfqd
->active_queue
== cfqq
) {
1687 if (time_after(now
, cfqq
->slice_end
))
1688 cfq_slice_expired(cfqd
, 0);
1689 else if (sync
&& RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1690 if (!cfq_arm_slice_timer(cfqd
, cfqq
))
1691 cfq_schedule_dispatch(cfqd
);
1697 * we temporarily boost lower priority queues if they are holding fs exclusive
1698 * resources. they are boosted to normal prio (CLASS_BE/4)
1700 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
1702 const int ioprio_class
= cfqq
->ioprio_class
;
1703 const int ioprio
= cfqq
->ioprio
;
1705 if (has_fs_excl()) {
1707 * boost idle prio on transactions that would lock out other
1708 * users of the filesystem
1710 if (cfq_class_idle(cfqq
))
1711 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1712 if (cfqq
->ioprio
> IOPRIO_NORM
)
1713 cfqq
->ioprio
= IOPRIO_NORM
;
1716 * check if we need to unboost the queue
1718 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
1719 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
1720 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
1721 cfqq
->ioprio
= cfqq
->org_ioprio
;
1725 * refile between round-robin lists if we moved the priority class
1727 if ((ioprio_class
!= cfqq
->ioprio_class
|| ioprio
!= cfqq
->ioprio
) &&
1728 cfq_cfqq_on_rr(cfqq
))
1729 cfq_resort_rr_list(cfqq
, 0);
1732 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
1734 if ((cfq_cfqq_wait_request(cfqq
) || cfq_cfqq_must_alloc(cfqq
)) &&
1735 !cfq_cfqq_must_alloc_slice(cfqq
)) {
1736 cfq_mark_cfqq_must_alloc_slice(cfqq
);
1737 return ELV_MQUEUE_MUST
;
1740 return ELV_MQUEUE_MAY
;
1743 static int cfq_may_queue(request_queue_t
*q
, int rw
)
1745 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1746 struct task_struct
*tsk
= current
;
1747 struct cfq_queue
*cfqq
;
1750 * don't force setup of a queue from here, as a call to may_queue
1751 * does not necessarily imply that a request actually will be queued.
1752 * so just lookup a possibly existing queue, or return 'may queue'
1755 cfqq
= cfq_find_cfq_hash(cfqd
, cfq_queue_pid(tsk
, rw
), tsk
->ioprio
);
1757 cfq_init_prio_data(cfqq
);
1758 cfq_prio_boost(cfqq
);
1760 return __cfq_may_queue(cfqq
);
1763 return ELV_MQUEUE_MAY
;
1767 * queue lock held here
1769 static void cfq_put_request(request_queue_t
*q
, struct request
*rq
)
1771 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1774 const int rw
= rq_data_dir(rq
);
1776 BUG_ON(!cfqq
->allocated
[rw
]);
1777 cfqq
->allocated
[rw
]--;
1779 put_io_context(RQ_CIC(rq
)->ioc
);
1781 rq
->elevator_private
= NULL
;
1782 rq
->elevator_private2
= NULL
;
1784 cfq_put_queue(cfqq
);
1789 * Allocate cfq data structures associated with this request.
1792 cfq_set_request(request_queue_t
*q
, struct request
*rq
, gfp_t gfp_mask
)
1794 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1795 struct task_struct
*tsk
= current
;
1796 struct cfq_io_context
*cic
;
1797 const int rw
= rq_data_dir(rq
);
1798 pid_t key
= cfq_queue_pid(tsk
, rw
);
1799 struct cfq_queue
*cfqq
;
1800 unsigned long flags
;
1801 int is_sync
= key
!= CFQ_KEY_ASYNC
;
1803 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1805 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
1807 spin_lock_irqsave(q
->queue_lock
, flags
);
1812 if (!cic
->cfqq
[is_sync
]) {
1813 cfqq
= cfq_get_queue(cfqd
, key
, tsk
, gfp_mask
);
1817 cic
->cfqq
[is_sync
] = cfqq
;
1819 cfqq
= cic
->cfqq
[is_sync
];
1821 cfqq
->allocated
[rw
]++;
1822 cfq_clear_cfqq_must_alloc(cfqq
);
1823 atomic_inc(&cfqq
->ref
);
1825 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1827 rq
->elevator_private
= cic
;
1828 rq
->elevator_private2
= cfqq
;
1833 put_io_context(cic
->ioc
);
1835 cfq_schedule_dispatch(cfqd
);
1836 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1840 static void cfq_kick_queue(void *data
)
1842 request_queue_t
*q
= data
;
1843 unsigned long flags
;
1845 spin_lock_irqsave(q
->queue_lock
, flags
);
1848 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1852 * Timer running if the active_queue is currently idling inside its time slice
1854 static void cfq_idle_slice_timer(unsigned long data
)
1856 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1857 struct cfq_queue
*cfqq
;
1858 unsigned long flags
;
1860 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1862 if ((cfqq
= cfqd
->active_queue
) != NULL
) {
1863 unsigned long now
= jiffies
;
1868 if (time_after(now
, cfqq
->slice_end
))
1872 * only expire and reinvoke request handler, if there are
1873 * other queues with pending requests
1875 if (!cfqd
->busy_queues
)
1879 * not expired and it has a request pending, let it dispatch
1881 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1882 cfq_mark_cfqq_must_dispatch(cfqq
);
1887 cfq_slice_expired(cfqd
, 0);
1889 cfq_schedule_dispatch(cfqd
);
1891 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1895 * Timer running if an idle class queue is waiting for service
1897 static void cfq_idle_class_timer(unsigned long data
)
1899 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
1900 unsigned long flags
, end
;
1902 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1905 * race with a non-idle queue, reset timer
1907 end
= cfqd
->last_end_request
+ CFQ_IDLE_GRACE
;
1908 if (!time_after_eq(jiffies
, end
))
1909 mod_timer(&cfqd
->idle_class_timer
, end
);
1911 cfq_schedule_dispatch(cfqd
);
1913 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1916 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
1918 del_timer_sync(&cfqd
->idle_slice_timer
);
1919 del_timer_sync(&cfqd
->idle_class_timer
);
1920 blk_sync_queue(cfqd
->queue
);
1923 static void cfq_exit_queue(elevator_t
*e
)
1925 struct cfq_data
*cfqd
= e
->elevator_data
;
1926 request_queue_t
*q
= cfqd
->queue
;
1928 cfq_shutdown_timer_wq(cfqd
);
1930 spin_lock_irq(q
->queue_lock
);
1932 if (cfqd
->active_queue
)
1933 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
1935 while (!list_empty(&cfqd
->cic_list
)) {
1936 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
1937 struct cfq_io_context
,
1940 __cfq_exit_single_io_context(cfqd
, cic
);
1943 spin_unlock_irq(q
->queue_lock
);
1945 cfq_shutdown_timer_wq(cfqd
);
1947 kfree(cfqd
->cfq_hash
);
1951 static void *cfq_init_queue(request_queue_t
*q
, elevator_t
*e
)
1953 struct cfq_data
*cfqd
;
1956 cfqd
= kmalloc(sizeof(*cfqd
), GFP_KERNEL
);
1960 memset(cfqd
, 0, sizeof(*cfqd
));
1962 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
1963 INIT_LIST_HEAD(&cfqd
->rr_list
[i
]);
1965 INIT_LIST_HEAD(&cfqd
->busy_rr
);
1966 INIT_LIST_HEAD(&cfqd
->cur_rr
);
1967 INIT_LIST_HEAD(&cfqd
->idle_rr
);
1968 INIT_LIST_HEAD(&cfqd
->empty_list
);
1969 INIT_LIST_HEAD(&cfqd
->cic_list
);
1971 cfqd
->cfq_hash
= kmalloc(sizeof(struct hlist_head
) * CFQ_QHASH_ENTRIES
, GFP_KERNEL
);
1972 if (!cfqd
->cfq_hash
)
1975 for (i
= 0; i
< CFQ_QHASH_ENTRIES
; i
++)
1976 INIT_HLIST_HEAD(&cfqd
->cfq_hash
[i
]);
1980 init_timer(&cfqd
->idle_slice_timer
);
1981 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
1982 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
1984 init_timer(&cfqd
->idle_class_timer
);
1985 cfqd
->idle_class_timer
.function
= cfq_idle_class_timer
;
1986 cfqd
->idle_class_timer
.data
= (unsigned long) cfqd
;
1988 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
, q
);
1990 cfqd
->cfq_quantum
= cfq_quantum
;
1991 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
1992 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
1993 cfqd
->cfq_back_max
= cfq_back_max
;
1994 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
1995 cfqd
->cfq_slice
[0] = cfq_slice_async
;
1996 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
1997 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
1998 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2006 static void cfq_slab_kill(void)
2009 kmem_cache_destroy(cfq_pool
);
2011 kmem_cache_destroy(cfq_ioc_pool
);
2014 static int __init
cfq_slab_setup(void)
2016 cfq_pool
= kmem_cache_create("cfq_pool", sizeof(struct cfq_queue
), 0, 0,
2021 cfq_ioc_pool
= kmem_cache_create("cfq_ioc_pool",
2022 sizeof(struct cfq_io_context
), 0, 0, NULL
, NULL
);
2033 * sysfs parts below -->
2037 cfq_var_show(unsigned int var
, char *page
)
2039 return sprintf(page
, "%d\n", var
);
2043 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2045 char *p
= (char *) page
;
2047 *var
= simple_strtoul(p
, &p
, 10);
2051 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2052 static ssize_t __FUNC(elevator_t *e, char *page) \
2054 struct cfq_data *cfqd = e->elevator_data; \
2055 unsigned int __data = __VAR; \
2057 __data = jiffies_to_msecs(__data); \
2058 return cfq_var_show(__data, (page)); \
2060 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2061 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2062 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2063 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2064 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2065 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2066 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2067 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2068 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2069 #undef SHOW_FUNCTION
2071 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2072 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2074 struct cfq_data *cfqd = e->elevator_data; \
2075 unsigned int __data; \
2076 int ret = cfq_var_store(&__data, (page), count); \
2077 if (__data < (MIN)) \
2079 else if (__data > (MAX)) \
2082 *(__PTR) = msecs_to_jiffies(__data); \
2084 *(__PTR) = __data; \
2087 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2088 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1, UINT_MAX
, 1);
2089 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1, UINT_MAX
, 1);
2090 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2091 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1, UINT_MAX
, 0);
2092 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2093 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2094 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2095 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1, UINT_MAX
, 0);
2096 #undef STORE_FUNCTION
2098 #define CFQ_ATTR(name) \
2099 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2101 static struct elv_fs_entry cfq_attrs
[] = {
2103 CFQ_ATTR(fifo_expire_sync
),
2104 CFQ_ATTR(fifo_expire_async
),
2105 CFQ_ATTR(back_seek_max
),
2106 CFQ_ATTR(back_seek_penalty
),
2107 CFQ_ATTR(slice_sync
),
2108 CFQ_ATTR(slice_async
),
2109 CFQ_ATTR(slice_async_rq
),
2110 CFQ_ATTR(slice_idle
),
2114 static struct elevator_type iosched_cfq
= {
2116 .elevator_merge_fn
= cfq_merge
,
2117 .elevator_merged_fn
= cfq_merged_request
,
2118 .elevator_merge_req_fn
= cfq_merged_requests
,
2119 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2120 .elevator_add_req_fn
= cfq_insert_request
,
2121 .elevator_activate_req_fn
= cfq_activate_request
,
2122 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2123 .elevator_queue_empty_fn
= cfq_queue_empty
,
2124 .elevator_completed_req_fn
= cfq_completed_request
,
2125 .elevator_former_req_fn
= elv_rb_former_request
,
2126 .elevator_latter_req_fn
= elv_rb_latter_request
,
2127 .elevator_set_req_fn
= cfq_set_request
,
2128 .elevator_put_req_fn
= cfq_put_request
,
2129 .elevator_may_queue_fn
= cfq_may_queue
,
2130 .elevator_init_fn
= cfq_init_queue
,
2131 .elevator_exit_fn
= cfq_exit_queue
,
2132 .trim
= cfq_free_io_context
,
2134 .elevator_attrs
= cfq_attrs
,
2135 .elevator_name
= "cfq",
2136 .elevator_owner
= THIS_MODULE
,
2139 static int __init
cfq_init(void)
2144 * could be 0 on HZ < 1000 setups
2146 if (!cfq_slice_async
)
2147 cfq_slice_async
= 1;
2148 if (!cfq_slice_idle
)
2151 if (cfq_slab_setup())
2154 ret
= elv_register(&iosched_cfq
);
2161 static void __exit
cfq_exit(void)
2163 DECLARE_COMPLETION(all_gone
);
2164 elv_unregister(&iosched_cfq
);
2165 ioc_gone
= &all_gone
;
2166 /* ioc_gone's update must be visible before reading ioc_count */
2168 if (atomic_read(&ioc_count
))
2169 wait_for_completion(ioc_gone
);
2174 module_init(cfq_init
);
2175 module_exit(cfq_exit
);
2177 MODULE_AUTHOR("Jens Axboe");
2178 MODULE_LICENSE("GPL");
2179 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");