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/rbtree.h>
13 #include <linux/ioprio.h>
18 /* max queue in one round of service */
19 static const int cfq_quantum
= 4;
20 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
21 /* maximum backwards seek, in KiB */
22 static const int cfq_back_max
= 16 * 1024;
23 /* penalty of a backwards seek */
24 static const int cfq_back_penalty
= 2;
25 static const int cfq_slice_sync
= HZ
/ 10;
26 static int cfq_slice_async
= HZ
/ 25;
27 static const int cfq_slice_async_rq
= 2;
28 static int cfq_slice_idle
= HZ
/ 125;
31 * offset from end of service tree
33 #define CFQ_IDLE_DELAY (HZ / 5)
36 * below this threshold, we consider thinktime immediate
38 #define CFQ_MIN_TT (2)
40 #define CFQ_SLICE_SCALE (5)
43 ((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 sample_valid(samples) ((samples) > 80)
62 * Most of our rbtree usage is for sorting with min extraction, so
63 * if we cache the leftmost node we don't have to walk down the tree
64 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
65 * move this into the elevator for the rq sorting as well.
71 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, }
74 * Per block device queue structure
77 struct request_queue
*queue
;
80 * rr list of queues with requests and the count of them
82 struct cfq_rb_root service_tree
;
83 unsigned int busy_queues
;
90 * idle window management
92 struct timer_list idle_slice_timer
;
93 struct work_struct unplug_work
;
95 struct cfq_queue
*active_queue
;
96 struct cfq_io_context
*active_cic
;
99 * async queue for each priority case
101 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
102 struct cfq_queue
*async_idle_cfqq
;
104 sector_t last_position
;
105 unsigned long last_end_request
;
108 * tunables, see top of file
110 unsigned int cfq_quantum
;
111 unsigned int cfq_fifo_expire
[2];
112 unsigned int cfq_back_penalty
;
113 unsigned int cfq_back_max
;
114 unsigned int cfq_slice
[2];
115 unsigned int cfq_slice_async_rq
;
116 unsigned int cfq_slice_idle
;
118 struct list_head cic_list
;
122 * Per process-grouping structure
125 /* reference count */
127 /* parent cfq_data */
128 struct cfq_data
*cfqd
;
129 /* service_tree member */
130 struct rb_node rb_node
;
131 /* service_tree key */
132 unsigned long rb_key
;
133 /* sorted list of pending requests */
134 struct rb_root sort_list
;
135 /* if fifo isn't expired, next request to serve */
136 struct request
*next_rq
;
137 /* requests queued in sort_list */
139 /* currently allocated requests */
141 /* pending metadata requests */
143 /* fifo list of requests in sort_list */
144 struct list_head fifo
;
146 unsigned long slice_end
;
149 /* number of requests that are on the dispatch list or inside driver */
152 /* io prio of this group */
153 unsigned short ioprio
, org_ioprio
;
154 unsigned short ioprio_class
, org_ioprio_class
;
156 /* various state flags, see below */
160 enum cfqq_state_flags
{
161 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
162 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
163 CFQ_CFQQ_FLAG_must_alloc
, /* must be allowed rq alloc */
164 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
165 CFQ_CFQQ_FLAG_must_dispatch
, /* must dispatch, even if expired */
166 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
167 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
168 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
169 CFQ_CFQQ_FLAG_queue_new
, /* queue never been serviced */
170 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
171 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
174 #define CFQ_CFQQ_FNS(name) \
175 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
177 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
179 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
181 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
183 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
185 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
189 CFQ_CFQQ_FNS(wait_request
);
190 CFQ_CFQQ_FNS(must_alloc
);
191 CFQ_CFQQ_FNS(must_alloc_slice
);
192 CFQ_CFQQ_FNS(must_dispatch
);
193 CFQ_CFQQ_FNS(fifo_expire
);
194 CFQ_CFQQ_FNS(idle_window
);
195 CFQ_CFQQ_FNS(prio_changed
);
196 CFQ_CFQQ_FNS(queue_new
);
197 CFQ_CFQQ_FNS(slice_new
);
201 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
202 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, int,
203 struct io_context
*, gfp_t
);
204 static struct cfq_io_context
*cfq_cic_lookup(struct cfq_data
*,
205 struct io_context
*);
207 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_context
*cic
,
210 return cic
->cfqq
[!!is_sync
];
213 static inline void cic_set_cfqq(struct cfq_io_context
*cic
,
214 struct cfq_queue
*cfqq
, int is_sync
)
216 cic
->cfqq
[!!is_sync
] = cfqq
;
220 * We regard a request as SYNC, if it's either a read or has the SYNC bit
221 * set (in which case it could also be direct WRITE).
223 static inline int cfq_bio_sync(struct bio
*bio
)
225 if (bio_data_dir(bio
) == READ
|| bio_sync(bio
))
232 * scheduler run of queue, if there are requests pending and no one in the
233 * driver that will restart queueing
235 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
237 if (cfqd
->busy_queues
)
238 kblockd_schedule_work(&cfqd
->unplug_work
);
241 static int cfq_queue_empty(struct request_queue
*q
)
243 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
245 return !cfqd
->busy_queues
;
249 * Scale schedule slice based on io priority. Use the sync time slice only
250 * if a queue is marked sync and has sync io queued. A sync queue with async
251 * io only, should not get full sync slice length.
253 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, int sync
,
256 const int base_slice
= cfqd
->cfq_slice
[sync
];
258 WARN_ON(prio
>= IOPRIO_BE_NR
);
260 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
264 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
266 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
270 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
272 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
276 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
277 * isn't valid until the first request from the dispatch is activated
278 * and the slice time set.
280 static inline int cfq_slice_used(struct cfq_queue
*cfqq
)
282 if (cfq_cfqq_slice_new(cfqq
))
284 if (time_before(jiffies
, cfqq
->slice_end
))
291 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
292 * We choose the request that is closest to the head right now. Distance
293 * behind the head is penalized and only allowed to a certain extent.
295 static struct request
*
296 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
)
298 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
299 unsigned long back_max
;
300 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
301 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
302 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
304 if (rq1
== NULL
|| rq1
== rq2
)
309 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
311 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
313 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
315 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
321 last
= cfqd
->last_position
;
324 * by definition, 1KiB is 2 sectors
326 back_max
= cfqd
->cfq_back_max
* 2;
329 * Strict one way elevator _except_ in the case where we allow
330 * short backward seeks which are biased as twice the cost of a
331 * similar forward seek.
335 else if (s1
+ back_max
>= last
)
336 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
338 wrap
|= CFQ_RQ1_WRAP
;
342 else if (s2
+ back_max
>= last
)
343 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
345 wrap
|= CFQ_RQ2_WRAP
;
347 /* Found required data */
350 * By doing switch() on the bit mask "wrap" we avoid having to
351 * check two variables for all permutations: --> faster!
354 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
370 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
373 * Since both rqs are wrapped,
374 * start with the one that's further behind head
375 * (--> only *one* back seek required),
376 * since back seek takes more time than forward.
386 * The below is leftmost cache rbtree addon
388 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
391 root
->left
= rb_first(&root
->rb
);
394 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
399 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
404 rb_erase(n
, &root
->rb
);
409 * would be nice to take fifo expire time into account as well
411 static struct request
*
412 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
413 struct request
*last
)
415 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
416 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
417 struct request
*next
= NULL
, *prev
= NULL
;
419 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
422 prev
= rb_entry_rq(rbprev
);
425 next
= rb_entry_rq(rbnext
);
427 rbnext
= rb_first(&cfqq
->sort_list
);
428 if (rbnext
&& rbnext
!= &last
->rb_node
)
429 next
= rb_entry_rq(rbnext
);
432 return cfq_choose_req(cfqd
, next
, prev
);
435 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
436 struct cfq_queue
*cfqq
)
439 * just an approximation, should be ok.
441 return (cfqd
->busy_queues
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
442 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
446 * The cfqd->service_tree holds all pending cfq_queue's that have
447 * requests waiting to be processed. It is sorted in the order that
448 * we will service the queues.
450 static void cfq_service_tree_add(struct cfq_data
*cfqd
,
451 struct cfq_queue
*cfqq
, int add_front
)
453 struct rb_node
**p
, *parent
;
454 struct cfq_queue
*__cfqq
;
455 unsigned long rb_key
;
458 if (cfq_class_idle(cfqq
)) {
459 rb_key
= CFQ_IDLE_DELAY
;
460 parent
= rb_last(&cfqd
->service_tree
.rb
);
461 if (parent
&& parent
!= &cfqq
->rb_node
) {
462 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
463 rb_key
+= __cfqq
->rb_key
;
466 } else if (!add_front
) {
467 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
468 rb_key
+= cfqq
->slice_resid
;
469 cfqq
->slice_resid
= 0;
473 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
475 * same position, nothing more to do
477 if (rb_key
== cfqq
->rb_key
)
480 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
485 p
= &cfqd
->service_tree
.rb
.rb_node
;
490 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
493 * sort RT queues first, we always want to give
494 * preference to them. IDLE queues goes to the back.
495 * after that, sort on the next service time.
497 if (cfq_class_rt(cfqq
) > cfq_class_rt(__cfqq
))
499 else if (cfq_class_rt(cfqq
) < cfq_class_rt(__cfqq
))
501 else if (cfq_class_idle(cfqq
) < cfq_class_idle(__cfqq
))
503 else if (cfq_class_idle(cfqq
) > cfq_class_idle(__cfqq
))
505 else if (rb_key
< __cfqq
->rb_key
)
510 if (n
== &(*p
)->rb_right
)
517 cfqd
->service_tree
.left
= &cfqq
->rb_node
;
519 cfqq
->rb_key
= rb_key
;
520 rb_link_node(&cfqq
->rb_node
, parent
, p
);
521 rb_insert_color(&cfqq
->rb_node
, &cfqd
->service_tree
.rb
);
525 * Update cfqq's position in the service tree.
527 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
530 * Resorting requires the cfqq to be on the RR list already.
532 if (cfq_cfqq_on_rr(cfqq
))
533 cfq_service_tree_add(cfqd
, cfqq
, 0);
537 * add to busy list of queues for service, trying to be fair in ordering
538 * the pending list according to last request service
540 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
542 BUG_ON(cfq_cfqq_on_rr(cfqq
));
543 cfq_mark_cfqq_on_rr(cfqq
);
546 cfq_resort_rr_list(cfqd
, cfqq
);
550 * Called when the cfqq no longer has requests pending, remove it from
553 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
555 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
556 cfq_clear_cfqq_on_rr(cfqq
);
558 if (!RB_EMPTY_NODE(&cfqq
->rb_node
))
559 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
561 BUG_ON(!cfqd
->busy_queues
);
566 * rb tree support functions
568 static void cfq_del_rq_rb(struct request
*rq
)
570 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
571 struct cfq_data
*cfqd
= cfqq
->cfqd
;
572 const int sync
= rq_is_sync(rq
);
574 BUG_ON(!cfqq
->queued
[sync
]);
575 cfqq
->queued
[sync
]--;
577 elv_rb_del(&cfqq
->sort_list
, rq
);
579 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
580 cfq_del_cfqq_rr(cfqd
, cfqq
);
583 static void cfq_add_rq_rb(struct request
*rq
)
585 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
586 struct cfq_data
*cfqd
= cfqq
->cfqd
;
587 struct request
*__alias
;
589 cfqq
->queued
[rq_is_sync(rq
)]++;
592 * looks a little odd, but the first insert might return an alias.
593 * if that happens, put the alias on the dispatch list
595 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
596 cfq_dispatch_insert(cfqd
->queue
, __alias
);
598 if (!cfq_cfqq_on_rr(cfqq
))
599 cfq_add_cfqq_rr(cfqd
, cfqq
);
602 * check if this request is a better next-serve candidate
604 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
605 BUG_ON(!cfqq
->next_rq
);
608 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
610 elv_rb_del(&cfqq
->sort_list
, rq
);
611 cfqq
->queued
[rq_is_sync(rq
)]--;
615 static struct request
*
616 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
618 struct task_struct
*tsk
= current
;
619 struct cfq_io_context
*cic
;
620 struct cfq_queue
*cfqq
;
622 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
626 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
628 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
630 return elv_rb_find(&cfqq
->sort_list
, sector
);
636 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
638 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
640 cfqd
->rq_in_driver
++;
643 * If the depth is larger 1, it really could be queueing. But lets
644 * make the mark a little higher - idling could still be good for
645 * low queueing, and a low queueing number could also just indicate
646 * a SCSI mid layer like behaviour where limit+1 is often seen.
648 if (!cfqd
->hw_tag
&& cfqd
->rq_in_driver
> 4)
651 cfqd
->last_position
= rq
->hard_sector
+ rq
->hard_nr_sectors
;
654 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
656 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
658 WARN_ON(!cfqd
->rq_in_driver
);
659 cfqd
->rq_in_driver
--;
662 static void cfq_remove_request(struct request
*rq
)
664 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
666 if (cfqq
->next_rq
== rq
)
667 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
669 list_del_init(&rq
->queuelist
);
672 if (rq_is_meta(rq
)) {
673 WARN_ON(!cfqq
->meta_pending
);
674 cfqq
->meta_pending
--;
678 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
681 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
682 struct request
*__rq
;
684 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
685 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
687 return ELEVATOR_FRONT_MERGE
;
690 return ELEVATOR_NO_MERGE
;
693 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
696 if (type
== ELEVATOR_FRONT_MERGE
) {
697 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
699 cfq_reposition_rq_rb(cfqq
, req
);
704 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
705 struct request
*next
)
708 * reposition in fifo if next is older than rq
710 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
711 time_before(next
->start_time
, rq
->start_time
))
712 list_move(&rq
->queuelist
, &next
->queuelist
);
714 cfq_remove_request(next
);
717 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
720 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
721 struct cfq_io_context
*cic
;
722 struct cfq_queue
*cfqq
;
725 * Disallow merge of a sync bio into an async request.
727 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
731 * Lookup the cfqq that this bio will be queued with. Allow
732 * merge only if rq is queued there.
734 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
738 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
739 if (cfqq
== RQ_CFQQ(rq
))
745 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
746 struct cfq_queue
*cfqq
)
750 cfq_clear_cfqq_must_alloc_slice(cfqq
);
751 cfq_clear_cfqq_fifo_expire(cfqq
);
752 cfq_mark_cfqq_slice_new(cfqq
);
753 cfq_clear_cfqq_queue_new(cfqq
);
756 cfqd
->active_queue
= cfqq
;
760 * current cfqq expired its slice (or was too idle), select new one
763 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
766 if (cfq_cfqq_wait_request(cfqq
))
767 del_timer(&cfqd
->idle_slice_timer
);
769 cfq_clear_cfqq_must_dispatch(cfqq
);
770 cfq_clear_cfqq_wait_request(cfqq
);
773 * store what was left of this slice, if the queue idled/timed out
775 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
))
776 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
778 cfq_resort_rr_list(cfqd
, cfqq
);
780 if (cfqq
== cfqd
->active_queue
)
781 cfqd
->active_queue
= NULL
;
783 if (cfqd
->active_cic
) {
784 put_io_context(cfqd
->active_cic
->ioc
);
785 cfqd
->active_cic
= NULL
;
789 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, int timed_out
)
791 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
794 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
798 * Get next queue for service. Unless we have a queue preemption,
799 * we'll simply select the first cfqq in the service tree.
801 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
803 if (RB_EMPTY_ROOT(&cfqd
->service_tree
.rb
))
806 return cfq_rb_first(&cfqd
->service_tree
);
810 * Get and set a new active queue for service.
812 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
)
814 struct cfq_queue
*cfqq
;
816 cfqq
= cfq_get_next_queue(cfqd
);
817 __cfq_set_active_queue(cfqd
, cfqq
);
821 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
824 if (rq
->sector
>= cfqd
->last_position
)
825 return rq
->sector
- cfqd
->last_position
;
827 return cfqd
->last_position
- rq
->sector
;
830 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct request
*rq
)
832 struct cfq_io_context
*cic
= cfqd
->active_cic
;
834 if (!sample_valid(cic
->seek_samples
))
837 return cfq_dist_from_last(cfqd
, rq
) <= cic
->seek_mean
;
840 static int cfq_close_cooperator(struct cfq_data
*cfq_data
,
841 struct cfq_queue
*cfqq
)
844 * We should notice if some of the queues are cooperating, eg
845 * working closely on the same area of the disk. In that case,
846 * we can group them together and don't waste time idling.
851 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
853 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
855 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
856 struct cfq_io_context
*cic
;
859 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
860 WARN_ON(cfq_cfqq_slice_new(cfqq
));
863 * idle is disabled, either manually or by past process history
865 if (!cfqd
->cfq_slice_idle
|| !cfq_cfqq_idle_window(cfqq
))
869 * task has exited, don't wait
871 cic
= cfqd
->active_cic
;
872 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
876 * See if this prio level has a good candidate
878 if (cfq_close_cooperator(cfqd
, cfqq
) &&
879 (sample_valid(cic
->ttime_samples
) && cic
->ttime_mean
> 2))
882 cfq_mark_cfqq_must_dispatch(cfqq
);
883 cfq_mark_cfqq_wait_request(cfqq
);
886 * we don't want to idle for seeks, but we do want to allow
887 * fair distribution of slice time for a process doing back-to-back
888 * seeks. so allow a little bit of time for him to submit a new rq
890 sl
= cfqd
->cfq_slice_idle
;
891 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
892 sl
= min(sl
, msecs_to_jiffies(CFQ_MIN_TT
));
894 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
898 * Move request from internal lists to the request queue dispatch list.
900 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
902 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
903 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
905 cfq_remove_request(rq
);
907 elv_dispatch_sort(q
, rq
);
909 if (cfq_cfqq_sync(cfqq
))
914 * return expired entry, or NULL to just start from scratch in rbtree
916 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
918 struct cfq_data
*cfqd
= cfqq
->cfqd
;
922 if (cfq_cfqq_fifo_expire(cfqq
))
925 cfq_mark_cfqq_fifo_expire(cfqq
);
927 if (list_empty(&cfqq
->fifo
))
930 fifo
= cfq_cfqq_sync(cfqq
);
931 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
933 if (time_before(jiffies
, rq
->start_time
+ cfqd
->cfq_fifo_expire
[fifo
]))
940 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
942 const int base_rq
= cfqd
->cfq_slice_async_rq
;
944 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
946 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
950 * Select a queue for service. If we have a current active queue,
951 * check whether to continue servicing it, or retrieve and set a new one.
953 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
955 struct cfq_queue
*cfqq
;
957 cfqq
= cfqd
->active_queue
;
962 * The active queue has run out of time, expire it and select new.
964 if (cfq_slice_used(cfqq
))
968 * The active queue has requests and isn't expired, allow it to
971 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
975 * No requests pending. If the active queue still has requests in
976 * flight or is idling for a new request, allow either of these
977 * conditions to happen (or time out) before selecting a new queue.
979 if (timer_pending(&cfqd
->idle_slice_timer
) ||
980 (cfqq
->dispatched
&& cfq_cfqq_idle_window(cfqq
))) {
986 cfq_slice_expired(cfqd
, 0);
988 cfqq
= cfq_set_active_queue(cfqd
);
994 * Dispatch some requests from cfqq, moving them to the request queue
998 __cfq_dispatch_requests(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1003 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
1009 * follow expired path, else get first next available
1011 rq
= cfq_check_fifo(cfqq
);
1016 * finally, insert request into driver dispatch list
1018 cfq_dispatch_insert(cfqd
->queue
, rq
);
1022 if (!cfqd
->active_cic
) {
1023 atomic_inc(&RQ_CIC(rq
)->ioc
->refcount
);
1024 cfqd
->active_cic
= RQ_CIC(rq
);
1027 if (RB_EMPTY_ROOT(&cfqq
->sort_list
))
1030 } while (dispatched
< max_dispatch
);
1033 * expire an async queue immediately if it has used up its slice. idle
1034 * queue always expire after 1 dispatch round.
1036 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
1037 dispatched
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
1038 cfq_class_idle(cfqq
))) {
1039 cfqq
->slice_end
= jiffies
+ 1;
1040 cfq_slice_expired(cfqd
, 0);
1046 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
1050 while (cfqq
->next_rq
) {
1051 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
1055 BUG_ON(!list_empty(&cfqq
->fifo
));
1060 * Drain our current requests. Used for barriers and when switching
1061 * io schedulers on-the-fly.
1063 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
1065 struct cfq_queue
*cfqq
;
1068 while ((cfqq
= cfq_rb_first(&cfqd
->service_tree
)) != NULL
)
1069 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1071 cfq_slice_expired(cfqd
, 0);
1073 BUG_ON(cfqd
->busy_queues
);
1078 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
1080 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1081 struct cfq_queue
*cfqq
;
1084 if (!cfqd
->busy_queues
)
1087 if (unlikely(force
))
1088 return cfq_forced_dispatch(cfqd
);
1091 while ((cfqq
= cfq_select_queue(cfqd
)) != NULL
) {
1094 max_dispatch
= cfqd
->cfq_quantum
;
1095 if (cfq_class_idle(cfqq
))
1098 if (cfqq
->dispatched
>= max_dispatch
) {
1099 if (cfqd
->busy_queues
> 1)
1101 if (cfqq
->dispatched
>= 4 * max_dispatch
)
1105 if (cfqd
->sync_flight
&& !cfq_cfqq_sync(cfqq
))
1108 cfq_clear_cfqq_must_dispatch(cfqq
);
1109 cfq_clear_cfqq_wait_request(cfqq
);
1110 del_timer(&cfqd
->idle_slice_timer
);
1112 dispatched
+= __cfq_dispatch_requests(cfqd
, cfqq
, max_dispatch
);
1119 * task holds one reference to the queue, dropped when task exits. each rq
1120 * in-flight on this queue also holds a reference, dropped when rq is freed.
1122 * queue lock must be held here.
1124 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1126 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1128 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1130 if (!atomic_dec_and_test(&cfqq
->ref
))
1133 BUG_ON(rb_first(&cfqq
->sort_list
));
1134 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1135 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1137 if (unlikely(cfqd
->active_queue
== cfqq
)) {
1138 __cfq_slice_expired(cfqd
, cfqq
, 0);
1139 cfq_schedule_dispatch(cfqd
);
1142 kmem_cache_free(cfq_pool
, cfqq
);
1146 * Call func for each cic attached to this ioc.
1149 call_for_each_cic(struct io_context
*ioc
,
1150 void (*func
)(struct io_context
*, struct cfq_io_context
*))
1152 struct cfq_io_context
*cic
;
1153 struct hlist_node
*n
;
1156 hlist_for_each_entry_rcu(cic
, n
, &ioc
->cic_list
, cic_list
)
1161 static void cfq_cic_free_rcu(struct rcu_head
*head
)
1163 struct cfq_io_context
*cic
;
1165 cic
= container_of(head
, struct cfq_io_context
, rcu_head
);
1167 kmem_cache_free(cfq_ioc_pool
, cic
);
1168 elv_ioc_count_dec(ioc_count
);
1170 if (ioc_gone
&& !elv_ioc_count_read(ioc_count
))
1174 static void cfq_cic_free(struct cfq_io_context
*cic
)
1176 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
1179 static void cic_free_func(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1181 unsigned long flags
;
1183 BUG_ON(!cic
->dead_key
);
1185 spin_lock_irqsave(&ioc
->lock
, flags
);
1186 radix_tree_delete(&ioc
->radix_root
, cic
->dead_key
);
1187 hlist_del_rcu(&cic
->cic_list
);
1188 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1193 static void cfq_free_io_context(struct io_context
*ioc
)
1196 * ioc->refcount is zero here, or we are called from elv_unregister(),
1197 * so no more cic's are allowed to be linked into this ioc. So it
1198 * should be ok to iterate over the known list, we will see all cic's
1199 * since no new ones are added.
1201 call_for_each_cic(ioc
, cic_free_func
);
1204 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1206 if (unlikely(cfqq
== cfqd
->active_queue
)) {
1207 __cfq_slice_expired(cfqd
, cfqq
, 0);
1208 cfq_schedule_dispatch(cfqd
);
1211 cfq_put_queue(cfqq
);
1214 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1215 struct cfq_io_context
*cic
)
1217 list_del_init(&cic
->queue_list
);
1220 * Make sure key == NULL is seen for dead queues
1223 cic
->dead_key
= (unsigned long) cic
->key
;
1226 if (cic
->cfqq
[ASYNC
]) {
1227 cfq_exit_cfqq(cfqd
, cic
->cfqq
[ASYNC
]);
1228 cic
->cfqq
[ASYNC
] = NULL
;
1231 if (cic
->cfqq
[SYNC
]) {
1232 cfq_exit_cfqq(cfqd
, cic
->cfqq
[SYNC
]);
1233 cic
->cfqq
[SYNC
] = NULL
;
1237 static void cfq_exit_single_io_context(struct io_context
*ioc
,
1238 struct cfq_io_context
*cic
)
1240 struct cfq_data
*cfqd
= cic
->key
;
1243 struct request_queue
*q
= cfqd
->queue
;
1244 unsigned long flags
;
1246 spin_lock_irqsave(q
->queue_lock
, flags
);
1247 __cfq_exit_single_io_context(cfqd
, cic
);
1248 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1253 * The process that ioc belongs to has exited, we need to clean up
1254 * and put the internal structures we have that belongs to that process.
1256 static void cfq_exit_io_context(struct io_context
*ioc
)
1258 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
1259 call_for_each_cic(ioc
, cfq_exit_single_io_context
);
1262 static struct cfq_io_context
*
1263 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1265 struct cfq_io_context
*cic
;
1267 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
1270 cic
->last_end_request
= jiffies
;
1271 INIT_LIST_HEAD(&cic
->queue_list
);
1272 INIT_HLIST_NODE(&cic
->cic_list
);
1273 cic
->dtor
= cfq_free_io_context
;
1274 cic
->exit
= cfq_exit_io_context
;
1275 elv_ioc_count_inc(ioc_count
);
1281 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
1283 struct task_struct
*tsk
= current
;
1286 if (!cfq_cfqq_prio_changed(cfqq
))
1289 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
1290 switch (ioprio_class
) {
1292 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1293 case IOPRIO_CLASS_NONE
:
1295 * no prio set, place us in the middle of the BE classes
1297 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1298 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1300 case IOPRIO_CLASS_RT
:
1301 cfqq
->ioprio
= task_ioprio(ioc
);
1302 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1304 case IOPRIO_CLASS_BE
:
1305 cfqq
->ioprio
= task_ioprio(ioc
);
1306 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1308 case IOPRIO_CLASS_IDLE
:
1309 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1311 cfq_clear_cfqq_idle_window(cfqq
);
1316 * keep track of original prio settings in case we have to temporarily
1317 * elevate the priority of this queue
1319 cfqq
->org_ioprio
= cfqq
->ioprio
;
1320 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1321 cfq_clear_cfqq_prio_changed(cfqq
);
1324 static void changed_ioprio(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1326 struct cfq_data
*cfqd
= cic
->key
;
1327 struct cfq_queue
*cfqq
;
1328 unsigned long flags
;
1330 if (unlikely(!cfqd
))
1333 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1335 cfqq
= cic
->cfqq
[ASYNC
];
1337 struct cfq_queue
*new_cfqq
;
1338 new_cfqq
= cfq_get_queue(cfqd
, ASYNC
, cic
->ioc
, GFP_ATOMIC
);
1340 cic
->cfqq
[ASYNC
] = new_cfqq
;
1341 cfq_put_queue(cfqq
);
1345 cfqq
= cic
->cfqq
[SYNC
];
1347 cfq_mark_cfqq_prio_changed(cfqq
);
1349 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1352 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1354 call_for_each_cic(ioc
, changed_ioprio
);
1355 ioc
->ioprio_changed
= 0;
1358 static struct cfq_queue
*
1359 cfq_find_alloc_queue(struct cfq_data
*cfqd
, int is_sync
,
1360 struct io_context
*ioc
, gfp_t gfp_mask
)
1362 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1363 struct cfq_io_context
*cic
;
1366 cic
= cfq_cic_lookup(cfqd
, ioc
);
1367 /* cic always exists here */
1368 cfqq
= cic_to_cfqq(cic
, is_sync
);
1374 } else if (gfp_mask
& __GFP_WAIT
) {
1376 * Inform the allocator of the fact that we will
1377 * just repeat this allocation if it fails, to allow
1378 * the allocator to do whatever it needs to attempt to
1381 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1382 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
1383 gfp_mask
| __GFP_NOFAIL
| __GFP_ZERO
,
1385 spin_lock_irq(cfqd
->queue
->queue_lock
);
1388 cfqq
= kmem_cache_alloc_node(cfq_pool
,
1389 gfp_mask
| __GFP_ZERO
,
1395 RB_CLEAR_NODE(&cfqq
->rb_node
);
1396 INIT_LIST_HEAD(&cfqq
->fifo
);
1398 atomic_set(&cfqq
->ref
, 0);
1401 cfq_mark_cfqq_prio_changed(cfqq
);
1402 cfq_mark_cfqq_queue_new(cfqq
);
1404 cfq_init_prio_data(cfqq
, ioc
);
1407 if (!cfq_class_idle(cfqq
))
1408 cfq_mark_cfqq_idle_window(cfqq
);
1409 cfq_mark_cfqq_sync(cfqq
);
1414 kmem_cache_free(cfq_pool
, new_cfqq
);
1417 WARN_ON((gfp_mask
& __GFP_WAIT
) && !cfqq
);
1421 static struct cfq_queue
**
1422 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
1424 switch (ioprio_class
) {
1425 case IOPRIO_CLASS_RT
:
1426 return &cfqd
->async_cfqq
[0][ioprio
];
1427 case IOPRIO_CLASS_BE
:
1428 return &cfqd
->async_cfqq
[1][ioprio
];
1429 case IOPRIO_CLASS_IDLE
:
1430 return &cfqd
->async_idle_cfqq
;
1436 static struct cfq_queue
*
1437 cfq_get_queue(struct cfq_data
*cfqd
, int is_sync
, struct io_context
*ioc
,
1440 const int ioprio
= task_ioprio(ioc
);
1441 const int ioprio_class
= task_ioprio_class(ioc
);
1442 struct cfq_queue
**async_cfqq
= NULL
;
1443 struct cfq_queue
*cfqq
= NULL
;
1446 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
1451 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
1457 * pin the queue now that it's allocated, scheduler exit will prune it
1459 if (!is_sync
&& !(*async_cfqq
)) {
1460 atomic_inc(&cfqq
->ref
);
1464 atomic_inc(&cfqq
->ref
);
1469 * We drop cfq io contexts lazily, so we may find a dead one.
1472 cfq_drop_dead_cic(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1473 struct cfq_io_context
*cic
)
1475 unsigned long flags
;
1477 WARN_ON(!list_empty(&cic
->queue_list
));
1479 spin_lock_irqsave(&ioc
->lock
, flags
);
1481 if (ioc
->ioc_data
== cic
)
1482 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
1484 radix_tree_delete(&ioc
->radix_root
, (unsigned long) cfqd
);
1485 hlist_del_rcu(&cic
->cic_list
);
1486 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1491 static struct cfq_io_context
*
1492 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1494 struct cfq_io_context
*cic
;
1501 * we maintain a last-hit cache, to avoid browsing over the tree
1503 cic
= rcu_dereference(ioc
->ioc_data
);
1504 if (cic
&& cic
->key
== cfqd
)
1509 cic
= radix_tree_lookup(&ioc
->radix_root
, (unsigned long) cfqd
);
1513 /* ->key must be copied to avoid race with cfq_exit_queue() */
1516 cfq_drop_dead_cic(cfqd
, ioc
, cic
);
1520 rcu_assign_pointer(ioc
->ioc_data
, cic
);
1528 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1529 * the process specific cfq io context when entered from the block layer.
1530 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1532 static int cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1533 struct cfq_io_context
*cic
, gfp_t gfp_mask
)
1535 unsigned long flags
;
1538 ret
= radix_tree_preload(gfp_mask
);
1543 spin_lock_irqsave(&ioc
->lock
, flags
);
1544 ret
= radix_tree_insert(&ioc
->radix_root
,
1545 (unsigned long) cfqd
, cic
);
1547 hlist_add_head_rcu(&cic
->cic_list
, &ioc
->cic_list
);
1548 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1550 radix_tree_preload_end();
1553 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1554 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1555 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1560 printk(KERN_ERR
"cfq: cic link failed!\n");
1566 * Setup general io context and cfq io context. There can be several cfq
1567 * io contexts per general io context, if this process is doing io to more
1568 * than one device managed by cfq.
1570 static struct cfq_io_context
*
1571 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1573 struct io_context
*ioc
= NULL
;
1574 struct cfq_io_context
*cic
;
1576 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1578 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
1582 cic
= cfq_cic_lookup(cfqd
, ioc
);
1586 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1590 if (cfq_cic_link(cfqd
, ioc
, cic
, gfp_mask
))
1594 smp_read_barrier_depends();
1595 if (unlikely(ioc
->ioprio_changed
))
1596 cfq_ioc_set_ioprio(ioc
);
1602 put_io_context(ioc
);
1607 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1609 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
1610 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1612 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1613 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1614 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1618 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
1624 if (cic
->last_request_pos
< rq
->sector
)
1625 sdist
= rq
->sector
- cic
->last_request_pos
;
1627 sdist
= cic
->last_request_pos
- rq
->sector
;
1630 * Don't allow the seek distance to get too large from the
1631 * odd fragment, pagein, etc
1633 if (cic
->seek_samples
<= 60) /* second&third seek */
1634 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1636 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1638 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1639 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1640 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1641 do_div(total
, cic
->seek_samples
);
1642 cic
->seek_mean
= (sector_t
)total
;
1646 * Disable idle window if the process thinks too long or seeks so much that
1650 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1651 struct cfq_io_context
*cic
)
1656 * Don't idle for async or idle io prio class
1658 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
1661 enable_idle
= cfq_cfqq_idle_window(cfqq
);
1663 if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
1664 (cfqd
->hw_tag
&& CIC_SEEKY(cic
)))
1666 else if (sample_valid(cic
->ttime_samples
)) {
1667 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
1674 cfq_mark_cfqq_idle_window(cfqq
);
1676 cfq_clear_cfqq_idle_window(cfqq
);
1680 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1681 * no or if we aren't sure, a 1 will cause a preempt.
1684 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
1687 struct cfq_queue
*cfqq
;
1689 cfqq
= cfqd
->active_queue
;
1693 if (cfq_slice_used(cfqq
))
1696 if (cfq_class_idle(new_cfqq
))
1699 if (cfq_class_idle(cfqq
))
1703 * if the new request is sync, but the currently running queue is
1704 * not, let the sync request have priority.
1706 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
1710 * So both queues are sync. Let the new request get disk time if
1711 * it's a metadata request and the current queue is doing regular IO.
1713 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
1716 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
1720 * if this request is as-good as one we would expect from the
1721 * current cfqq, let it preempt
1723 if (cfq_rq_close(cfqd
, rq
))
1730 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1731 * let it have half of its nominal slice.
1733 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1735 cfq_slice_expired(cfqd
, 1);
1738 * Put the new queue at the front of the of the current list,
1739 * so we know that it will be selected next.
1741 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1743 cfq_service_tree_add(cfqd
, cfqq
, 1);
1745 cfqq
->slice_end
= 0;
1746 cfq_mark_cfqq_slice_new(cfqq
);
1750 * Called when a new fs request (rq) is added (to cfqq). Check if there's
1751 * something we should do about it
1754 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1757 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1760 cfqq
->meta_pending
++;
1762 cfq_update_io_thinktime(cfqd
, cic
);
1763 cfq_update_io_seektime(cfqd
, cic
, rq
);
1764 cfq_update_idle_window(cfqd
, cfqq
, cic
);
1766 cic
->last_request_pos
= rq
->sector
+ rq
->nr_sectors
;
1768 if (cfqq
== cfqd
->active_queue
) {
1770 * if we are waiting for a request for this queue, let it rip
1771 * immediately and flag that we must not expire this queue
1774 if (cfq_cfqq_wait_request(cfqq
)) {
1775 cfq_mark_cfqq_must_dispatch(cfqq
);
1776 del_timer(&cfqd
->idle_slice_timer
);
1777 blk_start_queueing(cfqd
->queue
);
1779 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
1781 * not the active queue - expire current slice if it is
1782 * idle and has expired it's mean thinktime or this new queue
1783 * has some old slice time left and is of higher priority
1785 cfq_preempt_queue(cfqd
, cfqq
);
1786 cfq_mark_cfqq_must_dispatch(cfqq
);
1787 blk_start_queueing(cfqd
->queue
);
1791 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
1793 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1794 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1796 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
1800 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
1802 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
1805 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
1807 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1808 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1809 const int sync
= rq_is_sync(rq
);
1814 WARN_ON(!cfqd
->rq_in_driver
);
1815 WARN_ON(!cfqq
->dispatched
);
1816 cfqd
->rq_in_driver
--;
1819 if (cfq_cfqq_sync(cfqq
))
1820 cfqd
->sync_flight
--;
1822 if (!cfq_class_idle(cfqq
))
1823 cfqd
->last_end_request
= now
;
1826 RQ_CIC(rq
)->last_end_request
= now
;
1829 * If this is the active queue, check if it needs to be expired,
1830 * or if we want to idle in case it has no pending requests.
1832 if (cfqd
->active_queue
== cfqq
) {
1833 if (cfq_cfqq_slice_new(cfqq
)) {
1834 cfq_set_prio_slice(cfqd
, cfqq
);
1835 cfq_clear_cfqq_slice_new(cfqq
);
1837 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
1838 cfq_slice_expired(cfqd
, 1);
1839 else if (sync
&& RB_EMPTY_ROOT(&cfqq
->sort_list
))
1840 cfq_arm_slice_timer(cfqd
);
1843 if (!cfqd
->rq_in_driver
)
1844 cfq_schedule_dispatch(cfqd
);
1848 * we temporarily boost lower priority queues if they are holding fs exclusive
1849 * resources. they are boosted to normal prio (CLASS_BE/4)
1851 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
1853 if (has_fs_excl()) {
1855 * boost idle prio on transactions that would lock out other
1856 * users of the filesystem
1858 if (cfq_class_idle(cfqq
))
1859 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1860 if (cfqq
->ioprio
> IOPRIO_NORM
)
1861 cfqq
->ioprio
= IOPRIO_NORM
;
1864 * check if we need to unboost the queue
1866 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
1867 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
1868 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
1869 cfqq
->ioprio
= cfqq
->org_ioprio
;
1873 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
1875 if ((cfq_cfqq_wait_request(cfqq
) || cfq_cfqq_must_alloc(cfqq
)) &&
1876 !cfq_cfqq_must_alloc_slice(cfqq
)) {
1877 cfq_mark_cfqq_must_alloc_slice(cfqq
);
1878 return ELV_MQUEUE_MUST
;
1881 return ELV_MQUEUE_MAY
;
1884 static int cfq_may_queue(struct request_queue
*q
, int rw
)
1886 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1887 struct task_struct
*tsk
= current
;
1888 struct cfq_io_context
*cic
;
1889 struct cfq_queue
*cfqq
;
1892 * don't force setup of a queue from here, as a call to may_queue
1893 * does not necessarily imply that a request actually will be queued.
1894 * so just lookup a possibly existing queue, or return 'may queue'
1897 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
1899 return ELV_MQUEUE_MAY
;
1901 cfqq
= cic_to_cfqq(cic
, rw
& REQ_RW_SYNC
);
1903 cfq_init_prio_data(cfqq
, cic
->ioc
);
1904 cfq_prio_boost(cfqq
);
1906 return __cfq_may_queue(cfqq
);
1909 return ELV_MQUEUE_MAY
;
1913 * queue lock held here
1915 static void cfq_put_request(struct request
*rq
)
1917 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1920 const int rw
= rq_data_dir(rq
);
1922 BUG_ON(!cfqq
->allocated
[rw
]);
1923 cfqq
->allocated
[rw
]--;
1925 put_io_context(RQ_CIC(rq
)->ioc
);
1927 rq
->elevator_private
= NULL
;
1928 rq
->elevator_private2
= NULL
;
1930 cfq_put_queue(cfqq
);
1935 * Allocate cfq data structures associated with this request.
1938 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
1940 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1941 struct cfq_io_context
*cic
;
1942 const int rw
= rq_data_dir(rq
);
1943 const int is_sync
= rq_is_sync(rq
);
1944 struct cfq_queue
*cfqq
;
1945 unsigned long flags
;
1947 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1949 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
1951 spin_lock_irqsave(q
->queue_lock
, flags
);
1956 cfqq
= cic_to_cfqq(cic
, is_sync
);
1958 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
1963 cic_set_cfqq(cic
, cfqq
, is_sync
);
1966 cfqq
->allocated
[rw
]++;
1967 cfq_clear_cfqq_must_alloc(cfqq
);
1968 atomic_inc(&cfqq
->ref
);
1970 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1972 rq
->elevator_private
= cic
;
1973 rq
->elevator_private2
= cfqq
;
1978 put_io_context(cic
->ioc
);
1980 cfq_schedule_dispatch(cfqd
);
1981 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1985 static void cfq_kick_queue(struct work_struct
*work
)
1987 struct cfq_data
*cfqd
=
1988 container_of(work
, struct cfq_data
, unplug_work
);
1989 struct request_queue
*q
= cfqd
->queue
;
1990 unsigned long flags
;
1992 spin_lock_irqsave(q
->queue_lock
, flags
);
1993 blk_start_queueing(q
);
1994 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1998 * Timer running if the active_queue is currently idling inside its time slice
2000 static void cfq_idle_slice_timer(unsigned long data
)
2002 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
2003 struct cfq_queue
*cfqq
;
2004 unsigned long flags
;
2007 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2009 cfqq
= cfqd
->active_queue
;
2016 if (cfq_slice_used(cfqq
))
2020 * only expire and reinvoke request handler, if there are
2021 * other queues with pending requests
2023 if (!cfqd
->busy_queues
)
2027 * not expired and it has a request pending, let it dispatch
2029 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
2030 cfq_mark_cfqq_must_dispatch(cfqq
);
2035 cfq_slice_expired(cfqd
, timed_out
);
2037 cfq_schedule_dispatch(cfqd
);
2039 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2042 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
2044 del_timer_sync(&cfqd
->idle_slice_timer
);
2045 kblockd_flush_work(&cfqd
->unplug_work
);
2048 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
2052 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
2053 if (cfqd
->async_cfqq
[0][i
])
2054 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
2055 if (cfqd
->async_cfqq
[1][i
])
2056 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
2059 if (cfqd
->async_idle_cfqq
)
2060 cfq_put_queue(cfqd
->async_idle_cfqq
);
2063 static void cfq_exit_queue(elevator_t
*e
)
2065 struct cfq_data
*cfqd
= e
->elevator_data
;
2066 struct request_queue
*q
= cfqd
->queue
;
2068 cfq_shutdown_timer_wq(cfqd
);
2070 spin_lock_irq(q
->queue_lock
);
2072 if (cfqd
->active_queue
)
2073 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
2075 while (!list_empty(&cfqd
->cic_list
)) {
2076 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
2077 struct cfq_io_context
,
2080 __cfq_exit_single_io_context(cfqd
, cic
);
2083 cfq_put_async_queues(cfqd
);
2085 spin_unlock_irq(q
->queue_lock
);
2087 cfq_shutdown_timer_wq(cfqd
);
2092 static void *cfq_init_queue(struct request_queue
*q
)
2094 struct cfq_data
*cfqd
;
2096 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
2100 cfqd
->service_tree
= CFQ_RB_ROOT
;
2101 INIT_LIST_HEAD(&cfqd
->cic_list
);
2105 init_timer(&cfqd
->idle_slice_timer
);
2106 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2107 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2109 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
2111 cfqd
->last_end_request
= jiffies
;
2112 cfqd
->cfq_quantum
= cfq_quantum
;
2113 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2114 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2115 cfqd
->cfq_back_max
= cfq_back_max
;
2116 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2117 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2118 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2119 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2120 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2125 static void cfq_slab_kill(void)
2128 kmem_cache_destroy(cfq_pool
);
2130 kmem_cache_destroy(cfq_ioc_pool
);
2133 static int __init
cfq_slab_setup(void)
2135 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
2139 cfq_ioc_pool
= KMEM_CACHE(cfq_io_context
, 0);
2150 * sysfs parts below -->
2153 cfq_var_show(unsigned int var
, char *page
)
2155 return sprintf(page
, "%d\n", var
);
2159 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2161 char *p
= (char *) page
;
2163 *var
= simple_strtoul(p
, &p
, 10);
2167 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2168 static ssize_t __FUNC(elevator_t *e, char *page) \
2170 struct cfq_data *cfqd = e->elevator_data; \
2171 unsigned int __data = __VAR; \
2173 __data = jiffies_to_msecs(__data); \
2174 return cfq_var_show(__data, (page)); \
2176 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2177 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2178 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2179 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2180 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2181 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2182 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2183 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2184 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2185 #undef SHOW_FUNCTION
2187 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2188 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \
2190 struct cfq_data *cfqd = e->elevator_data; \
2191 unsigned int __data; \
2192 int ret = cfq_var_store(&__data, (page), count); \
2193 if (__data < (MIN)) \
2195 else if (__data > (MAX)) \
2198 *(__PTR) = msecs_to_jiffies(__data); \
2200 *(__PTR) = __data; \
2203 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2204 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
2206 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
2208 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2209 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
2211 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2212 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2213 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2214 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
2216 #undef STORE_FUNCTION
2218 #define CFQ_ATTR(name) \
2219 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2221 static struct elv_fs_entry cfq_attrs
[] = {
2223 CFQ_ATTR(fifo_expire_sync
),
2224 CFQ_ATTR(fifo_expire_async
),
2225 CFQ_ATTR(back_seek_max
),
2226 CFQ_ATTR(back_seek_penalty
),
2227 CFQ_ATTR(slice_sync
),
2228 CFQ_ATTR(slice_async
),
2229 CFQ_ATTR(slice_async_rq
),
2230 CFQ_ATTR(slice_idle
),
2234 static struct elevator_type iosched_cfq
= {
2236 .elevator_merge_fn
= cfq_merge
,
2237 .elevator_merged_fn
= cfq_merged_request
,
2238 .elevator_merge_req_fn
= cfq_merged_requests
,
2239 .elevator_allow_merge_fn
= cfq_allow_merge
,
2240 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2241 .elevator_add_req_fn
= cfq_insert_request
,
2242 .elevator_activate_req_fn
= cfq_activate_request
,
2243 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2244 .elevator_queue_empty_fn
= cfq_queue_empty
,
2245 .elevator_completed_req_fn
= cfq_completed_request
,
2246 .elevator_former_req_fn
= elv_rb_former_request
,
2247 .elevator_latter_req_fn
= elv_rb_latter_request
,
2248 .elevator_set_req_fn
= cfq_set_request
,
2249 .elevator_put_req_fn
= cfq_put_request
,
2250 .elevator_may_queue_fn
= cfq_may_queue
,
2251 .elevator_init_fn
= cfq_init_queue
,
2252 .elevator_exit_fn
= cfq_exit_queue
,
2253 .trim
= cfq_free_io_context
,
2255 .elevator_attrs
= cfq_attrs
,
2256 .elevator_name
= "cfq",
2257 .elevator_owner
= THIS_MODULE
,
2260 static int __init
cfq_init(void)
2263 * could be 0 on HZ < 1000 setups
2265 if (!cfq_slice_async
)
2266 cfq_slice_async
= 1;
2267 if (!cfq_slice_idle
)
2270 if (cfq_slab_setup())
2273 elv_register(&iosched_cfq
);
2278 static void __exit
cfq_exit(void)
2280 DECLARE_COMPLETION_ONSTACK(all_gone
);
2281 elv_unregister(&iosched_cfq
);
2282 ioc_gone
= &all_gone
;
2283 /* ioc_gone's update must be visible before reading ioc_count */
2285 if (elv_ioc_count_read(ioc_count
))
2286 wait_for_completion(ioc_gone
);
2290 module_init(cfq_init
);
2291 module_exit(cfq_exit
);
2293 MODULE_AUTHOR("Jens Axboe");
2294 MODULE_LICENSE("GPL");
2295 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");