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>
14 #include <linux/blktrace_api.h>
19 /* max queue in one round of service */
20 static const int cfq_quantum
= 4;
21 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max
= 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty
= 2;
26 static const int cfq_slice_sync
= HZ
/ 10;
27 static int cfq_slice_async
= HZ
/ 25;
28 static const int cfq_slice_async_rq
= 2;
29 static int cfq_slice_idle
= HZ
/ 125;
32 * offset from end of service tree
34 #define CFQ_IDLE_DELAY (HZ / 5)
37 * below this threshold, we consider thinktime immediate
39 #define CFQ_MIN_TT (2)
41 #define CFQ_SLICE_SCALE (5)
42 #define CFQ_HW_QUEUE_MIN (5)
45 ((struct cfq_io_context *) (rq)->elevator_private)
46 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
48 static struct kmem_cache
*cfq_pool
;
49 static struct kmem_cache
*cfq_ioc_pool
;
51 static DEFINE_PER_CPU(unsigned long, ioc_count
);
52 static struct completion
*ioc_gone
;
53 static DEFINE_SPINLOCK(ioc_gone_lock
);
55 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #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 process-grouping structure
79 /* various state flags, see below */
82 struct cfq_data
*cfqd
;
83 /* service_tree member */
84 struct rb_node rb_node
;
85 /* service_tree key */
87 /* prio tree member */
88 struct rb_node p_node
;
89 /* prio tree root we belong to, if any */
90 struct rb_root
*p_root
;
91 /* sorted list of pending requests */
92 struct rb_root sort_list
;
93 /* if fifo isn't expired, next request to serve */
94 struct request
*next_rq
;
95 /* requests queued in sort_list */
97 /* currently allocated requests */
99 /* fifo list of requests in sort_list */
100 struct list_head fifo
;
102 unsigned long slice_end
;
104 unsigned int slice_dispatch
;
106 /* pending metadata requests */
108 /* number of requests that are on the dispatch list or inside driver */
111 /* io prio of this group */
112 unsigned short ioprio
, org_ioprio
;
113 unsigned short ioprio_class
, org_ioprio_class
;
119 * Per block device queue structure
122 struct request_queue
*queue
;
125 * rr list of queues with requests and the count of them
127 struct cfq_rb_root service_tree
;
130 * Each priority tree is sorted by next_request position. These
131 * trees are used when determining if two or more queues are
132 * interleaving requests (see cfq_close_cooperator).
134 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
136 unsigned int busy_queues
;
142 * queue-depth detection
147 int rq_in_driver_peak
;
150 * idle window management
152 struct timer_list idle_slice_timer
;
153 struct work_struct unplug_work
;
155 struct cfq_queue
*active_queue
;
156 struct cfq_io_context
*active_cic
;
159 * async queue for each priority case
161 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
162 struct cfq_queue
*async_idle_cfqq
;
164 sector_t last_position
;
167 * tunables, see top of file
169 unsigned int cfq_quantum
;
170 unsigned int cfq_fifo_expire
[2];
171 unsigned int cfq_back_penalty
;
172 unsigned int cfq_back_max
;
173 unsigned int cfq_slice
[2];
174 unsigned int cfq_slice_async_rq
;
175 unsigned int cfq_slice_idle
;
177 struct list_head cic_list
;
180 * Fallback dummy cfqq for extreme OOM conditions
182 struct cfq_queue oom_cfqq
;
185 enum cfqq_state_flags
{
186 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
187 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
188 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
189 CFQ_CFQQ_FLAG_must_alloc
, /* must be allowed rq alloc */
190 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
191 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
192 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
193 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
194 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
195 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
196 CFQ_CFQQ_FLAG_coop
, /* has done a coop jump of the queue */
199 #define CFQ_CFQQ_FNS(name) \
200 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
202 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
204 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
206 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
208 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
210 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
214 CFQ_CFQQ_FNS(wait_request
);
215 CFQ_CFQQ_FNS(must_dispatch
);
216 CFQ_CFQQ_FNS(must_alloc
);
217 CFQ_CFQQ_FNS(must_alloc_slice
);
218 CFQ_CFQQ_FNS(fifo_expire
);
219 CFQ_CFQQ_FNS(idle_window
);
220 CFQ_CFQQ_FNS(prio_changed
);
221 CFQ_CFQQ_FNS(slice_new
);
226 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
227 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
228 #define cfq_log(cfqd, fmt, args...) \
229 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
231 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
232 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, int,
233 struct io_context
*, gfp_t
);
234 static struct cfq_io_context
*cfq_cic_lookup(struct cfq_data
*,
235 struct io_context
*);
237 static inline int rq_in_driver(struct cfq_data
*cfqd
)
239 return cfqd
->rq_in_driver
[0] + cfqd
->rq_in_driver
[1];
242 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_context
*cic
,
245 return cic
->cfqq
[!!is_sync
];
248 static inline void cic_set_cfqq(struct cfq_io_context
*cic
,
249 struct cfq_queue
*cfqq
, int is_sync
)
251 cic
->cfqq
[!!is_sync
] = cfqq
;
255 * We regard a request as SYNC, if it's either a read or has the SYNC bit
256 * set (in which case it could also be direct WRITE).
258 static inline int cfq_bio_sync(struct bio
*bio
)
260 if (bio_data_dir(bio
) == READ
|| bio_sync(bio
))
267 * scheduler run of queue, if there are requests pending and no one in the
268 * driver that will restart queueing
270 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
272 if (cfqd
->busy_queues
) {
273 cfq_log(cfqd
, "schedule dispatch");
274 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
278 static int cfq_queue_empty(struct request_queue
*q
)
280 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
282 return !cfqd
->busy_queues
;
286 * Scale schedule slice based on io priority. Use the sync time slice only
287 * if a queue is marked sync and has sync io queued. A sync queue with async
288 * io only, should not get full sync slice length.
290 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, int sync
,
293 const int base_slice
= cfqd
->cfq_slice
[sync
];
295 WARN_ON(prio
>= IOPRIO_BE_NR
);
297 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
301 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
303 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
307 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
309 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
310 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
314 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
315 * isn't valid until the first request from the dispatch is activated
316 * and the slice time set.
318 static inline int cfq_slice_used(struct cfq_queue
*cfqq
)
320 if (cfq_cfqq_slice_new(cfqq
))
322 if (time_before(jiffies
, cfqq
->slice_end
))
329 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
330 * We choose the request that is closest to the head right now. Distance
331 * behind the head is penalized and only allowed to a certain extent.
333 static struct request
*
334 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
)
336 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
337 unsigned long back_max
;
338 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
339 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
340 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
342 if (rq1
== NULL
|| rq1
== rq2
)
347 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
349 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
351 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
353 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
356 s1
= blk_rq_pos(rq1
);
357 s2
= blk_rq_pos(rq2
);
359 last
= cfqd
->last_position
;
362 * by definition, 1KiB is 2 sectors
364 back_max
= cfqd
->cfq_back_max
* 2;
367 * Strict one way elevator _except_ in the case where we allow
368 * short backward seeks which are biased as twice the cost of a
369 * similar forward seek.
373 else if (s1
+ back_max
>= last
)
374 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
376 wrap
|= CFQ_RQ1_WRAP
;
380 else if (s2
+ back_max
>= last
)
381 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
383 wrap
|= CFQ_RQ2_WRAP
;
385 /* Found required data */
388 * By doing switch() on the bit mask "wrap" we avoid having to
389 * check two variables for all permutations: --> faster!
392 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
408 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
411 * Since both rqs are wrapped,
412 * start with the one that's further behind head
413 * (--> only *one* back seek required),
414 * since back seek takes more time than forward.
424 * The below is leftmost cache rbtree addon
426 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
429 root
->left
= rb_first(&root
->rb
);
432 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
437 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
443 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
447 rb_erase_init(n
, &root
->rb
);
451 * would be nice to take fifo expire time into account as well
453 static struct request
*
454 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
455 struct request
*last
)
457 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
458 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
459 struct request
*next
= NULL
, *prev
= NULL
;
461 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
464 prev
= rb_entry_rq(rbprev
);
467 next
= rb_entry_rq(rbnext
);
469 rbnext
= rb_first(&cfqq
->sort_list
);
470 if (rbnext
&& rbnext
!= &last
->rb_node
)
471 next
= rb_entry_rq(rbnext
);
474 return cfq_choose_req(cfqd
, next
, prev
);
477 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
478 struct cfq_queue
*cfqq
)
481 * just an approximation, should be ok.
483 return (cfqd
->busy_queues
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
484 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
488 * The cfqd->service_tree holds all pending cfq_queue's that have
489 * requests waiting to be processed. It is sorted in the order that
490 * we will service the queues.
492 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
495 struct rb_node
**p
, *parent
;
496 struct cfq_queue
*__cfqq
;
497 unsigned long rb_key
;
500 if (cfq_class_idle(cfqq
)) {
501 rb_key
= CFQ_IDLE_DELAY
;
502 parent
= rb_last(&cfqd
->service_tree
.rb
);
503 if (parent
&& parent
!= &cfqq
->rb_node
) {
504 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
505 rb_key
+= __cfqq
->rb_key
;
508 } else if (!add_front
) {
509 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
510 rb_key
+= cfqq
->slice_resid
;
511 cfqq
->slice_resid
= 0;
515 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
517 * same position, nothing more to do
519 if (rb_key
== cfqq
->rb_key
)
522 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
527 p
= &cfqd
->service_tree
.rb
.rb_node
;
532 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
535 * sort RT queues first, we always want to give
536 * preference to them. IDLE queues goes to the back.
537 * after that, sort on the next service time.
539 if (cfq_class_rt(cfqq
) > cfq_class_rt(__cfqq
))
541 else if (cfq_class_rt(cfqq
) < cfq_class_rt(__cfqq
))
543 else if (cfq_class_idle(cfqq
) < cfq_class_idle(__cfqq
))
545 else if (cfq_class_idle(cfqq
) > cfq_class_idle(__cfqq
))
547 else if (rb_key
< __cfqq
->rb_key
)
552 if (n
== &(*p
)->rb_right
)
559 cfqd
->service_tree
.left
= &cfqq
->rb_node
;
561 cfqq
->rb_key
= rb_key
;
562 rb_link_node(&cfqq
->rb_node
, parent
, p
);
563 rb_insert_color(&cfqq
->rb_node
, &cfqd
->service_tree
.rb
);
566 static struct cfq_queue
*
567 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
568 sector_t sector
, struct rb_node
**ret_parent
,
569 struct rb_node
***rb_link
)
571 struct rb_node
**p
, *parent
;
572 struct cfq_queue
*cfqq
= NULL
;
580 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
583 * Sort strictly based on sector. Smallest to the left,
584 * largest to the right.
586 if (sector
> blk_rq_pos(cfqq
->next_rq
))
588 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
596 *ret_parent
= parent
;
602 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
604 struct rb_node
**p
, *parent
;
605 struct cfq_queue
*__cfqq
;
608 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
612 if (cfq_class_idle(cfqq
))
617 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
618 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
619 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
621 rb_link_node(&cfqq
->p_node
, parent
, p
);
622 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
628 * Update cfqq's position in the service tree.
630 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
633 * Resorting requires the cfqq to be on the RR list already.
635 if (cfq_cfqq_on_rr(cfqq
)) {
636 cfq_service_tree_add(cfqd
, cfqq
, 0);
637 cfq_prio_tree_add(cfqd
, cfqq
);
642 * add to busy list of queues for service, trying to be fair in ordering
643 * the pending list according to last request service
645 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
647 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
648 BUG_ON(cfq_cfqq_on_rr(cfqq
));
649 cfq_mark_cfqq_on_rr(cfqq
);
652 cfq_resort_rr_list(cfqd
, cfqq
);
656 * Called when the cfqq no longer has requests pending, remove it from
659 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
661 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
662 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
663 cfq_clear_cfqq_on_rr(cfqq
);
665 if (!RB_EMPTY_NODE(&cfqq
->rb_node
))
666 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
668 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
672 BUG_ON(!cfqd
->busy_queues
);
677 * rb tree support functions
679 static void cfq_del_rq_rb(struct request
*rq
)
681 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
682 struct cfq_data
*cfqd
= cfqq
->cfqd
;
683 const int sync
= rq_is_sync(rq
);
685 BUG_ON(!cfqq
->queued
[sync
]);
686 cfqq
->queued
[sync
]--;
688 elv_rb_del(&cfqq
->sort_list
, rq
);
690 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
691 cfq_del_cfqq_rr(cfqd
, cfqq
);
694 static void cfq_add_rq_rb(struct request
*rq
)
696 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
697 struct cfq_data
*cfqd
= cfqq
->cfqd
;
698 struct request
*__alias
, *prev
;
700 cfqq
->queued
[rq_is_sync(rq
)]++;
703 * looks a little odd, but the first insert might return an alias.
704 * if that happens, put the alias on the dispatch list
706 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
707 cfq_dispatch_insert(cfqd
->queue
, __alias
);
709 if (!cfq_cfqq_on_rr(cfqq
))
710 cfq_add_cfqq_rr(cfqd
, cfqq
);
713 * check if this request is a better next-serve candidate
715 prev
= cfqq
->next_rq
;
716 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
719 * adjust priority tree position, if ->next_rq changes
721 if (prev
!= cfqq
->next_rq
)
722 cfq_prio_tree_add(cfqd
, cfqq
);
724 BUG_ON(!cfqq
->next_rq
);
727 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
729 elv_rb_del(&cfqq
->sort_list
, rq
);
730 cfqq
->queued
[rq_is_sync(rq
)]--;
734 static struct request
*
735 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
737 struct task_struct
*tsk
= current
;
738 struct cfq_io_context
*cic
;
739 struct cfq_queue
*cfqq
;
741 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
745 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
747 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
749 return elv_rb_find(&cfqq
->sort_list
, sector
);
755 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
757 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
759 cfqd
->rq_in_driver
[rq_is_sync(rq
)]++;
760 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
763 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
766 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
768 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
769 const int sync
= rq_is_sync(rq
);
771 WARN_ON(!cfqd
->rq_in_driver
[sync
]);
772 cfqd
->rq_in_driver
[sync
]--;
773 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
777 static void cfq_remove_request(struct request
*rq
)
779 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
781 if (cfqq
->next_rq
== rq
)
782 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
784 list_del_init(&rq
->queuelist
);
787 cfqq
->cfqd
->rq_queued
--;
788 if (rq_is_meta(rq
)) {
789 WARN_ON(!cfqq
->meta_pending
);
790 cfqq
->meta_pending
--;
794 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
797 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
798 struct request
*__rq
;
800 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
801 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
803 return ELEVATOR_FRONT_MERGE
;
806 return ELEVATOR_NO_MERGE
;
809 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
812 if (type
== ELEVATOR_FRONT_MERGE
) {
813 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
815 cfq_reposition_rq_rb(cfqq
, req
);
820 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
821 struct request
*next
)
824 * reposition in fifo if next is older than rq
826 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
827 time_before(next
->start_time
, rq
->start_time
))
828 list_move(&rq
->queuelist
, &next
->queuelist
);
830 cfq_remove_request(next
);
833 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
836 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
837 struct cfq_io_context
*cic
;
838 struct cfq_queue
*cfqq
;
841 * Disallow merge of a sync bio into an async request.
843 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
847 * Lookup the cfqq that this bio will be queued with. Allow
848 * merge only if rq is queued there.
850 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
854 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
855 if (cfqq
== RQ_CFQQ(rq
))
861 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
862 struct cfq_queue
*cfqq
)
865 cfq_log_cfqq(cfqd
, cfqq
, "set_active");
867 cfqq
->slice_dispatch
= 0;
869 cfq_clear_cfqq_wait_request(cfqq
);
870 cfq_clear_cfqq_must_dispatch(cfqq
);
871 cfq_clear_cfqq_must_alloc_slice(cfqq
);
872 cfq_clear_cfqq_fifo_expire(cfqq
);
873 cfq_mark_cfqq_slice_new(cfqq
);
875 del_timer(&cfqd
->idle_slice_timer
);
878 cfqd
->active_queue
= cfqq
;
882 * current cfqq expired its slice (or was too idle), select new one
885 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
888 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
890 if (cfq_cfqq_wait_request(cfqq
))
891 del_timer(&cfqd
->idle_slice_timer
);
893 cfq_clear_cfqq_wait_request(cfqq
);
896 * store what was left of this slice, if the queue idled/timed out
898 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
)) {
899 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
900 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
903 cfq_resort_rr_list(cfqd
, cfqq
);
905 if (cfqq
== cfqd
->active_queue
)
906 cfqd
->active_queue
= NULL
;
908 if (cfqd
->active_cic
) {
909 put_io_context(cfqd
->active_cic
->ioc
);
910 cfqd
->active_cic
= NULL
;
914 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, int timed_out
)
916 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
919 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
923 * Get next queue for service. Unless we have a queue preemption,
924 * we'll simply select the first cfqq in the service tree.
926 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
928 if (RB_EMPTY_ROOT(&cfqd
->service_tree
.rb
))
931 return cfq_rb_first(&cfqd
->service_tree
);
935 * Get and set a new active queue for service.
937 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
938 struct cfq_queue
*cfqq
)
941 cfqq
= cfq_get_next_queue(cfqd
);
943 cfq_clear_cfqq_coop(cfqq
);
946 __cfq_set_active_queue(cfqd
, cfqq
);
950 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
953 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
954 return blk_rq_pos(rq
) - cfqd
->last_position
;
956 return cfqd
->last_position
- blk_rq_pos(rq
);
959 #define CIC_SEEK_THR 8 * 1024
960 #define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
962 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct request
*rq
)
964 struct cfq_io_context
*cic
= cfqd
->active_cic
;
965 sector_t sdist
= cic
->seek_mean
;
967 if (!sample_valid(cic
->seek_samples
))
968 sdist
= CIC_SEEK_THR
;
970 return cfq_dist_from_last(cfqd
, rq
) <= sdist
;
973 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
974 struct cfq_queue
*cur_cfqq
)
976 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
977 struct rb_node
*parent
, *node
;
978 struct cfq_queue
*__cfqq
;
979 sector_t sector
= cfqd
->last_position
;
981 if (RB_EMPTY_ROOT(root
))
985 * First, if we find a request starting at the end of the last
986 * request, choose it.
988 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
993 * If the exact sector wasn't found, the parent of the NULL leaf
994 * will contain the closest sector.
996 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
997 if (cfq_rq_close(cfqd
, __cfqq
->next_rq
))
1000 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1001 node
= rb_next(&__cfqq
->p_node
);
1003 node
= rb_prev(&__cfqq
->p_node
);
1007 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1008 if (cfq_rq_close(cfqd
, __cfqq
->next_rq
))
1016 * cur_cfqq - passed in so that we don't decide that the current queue is
1017 * closely cooperating with itself.
1019 * So, basically we're assuming that that cur_cfqq has dispatched at least
1020 * one request, and that cfqd->last_position reflects a position on the disk
1021 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1024 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1025 struct cfq_queue
*cur_cfqq
,
1028 struct cfq_queue
*cfqq
;
1031 * A valid cfq_io_context is necessary to compare requests against
1032 * the seek_mean of the current cfqq.
1034 if (!cfqd
->active_cic
)
1038 * We should notice if some of the queues are cooperating, eg
1039 * working closely on the same area of the disk. In that case,
1040 * we can group them together and don't waste time idling.
1042 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1046 if (cfq_cfqq_coop(cfqq
))
1050 cfq_mark_cfqq_coop(cfqq
);
1054 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1056 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1057 struct cfq_io_context
*cic
;
1061 * SSD device without seek penalty, disable idling. But only do so
1062 * for devices that support queuing, otherwise we still have a problem
1063 * with sync vs async workloads.
1065 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
1068 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
1069 WARN_ON(cfq_cfqq_slice_new(cfqq
));
1072 * idle is disabled, either manually or by past process history
1074 if (!cfqd
->cfq_slice_idle
|| !cfq_cfqq_idle_window(cfqq
))
1078 * still requests with the driver, don't idle
1080 if (rq_in_driver(cfqd
))
1084 * task has exited, don't wait
1086 cic
= cfqd
->active_cic
;
1087 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
1090 cfq_mark_cfqq_wait_request(cfqq
);
1093 * we don't want to idle for seeks, but we do want to allow
1094 * fair distribution of slice time for a process doing back-to-back
1095 * seeks. so allow a little bit of time for him to submit a new rq
1097 sl
= cfqd
->cfq_slice_idle
;
1098 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
1099 sl
= min(sl
, msecs_to_jiffies(CFQ_MIN_TT
));
1101 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
1102 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu", sl
);
1106 * Move request from internal lists to the request queue dispatch list.
1108 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
1110 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1111 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1113 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
1115 cfq_remove_request(rq
);
1117 elv_dispatch_sort(q
, rq
);
1119 if (cfq_cfqq_sync(cfqq
))
1120 cfqd
->sync_flight
++;
1124 * return expired entry, or NULL to just start from scratch in rbtree
1126 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
1128 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1132 if (cfq_cfqq_fifo_expire(cfqq
))
1135 cfq_mark_cfqq_fifo_expire(cfqq
);
1137 if (list_empty(&cfqq
->fifo
))
1140 fifo
= cfq_cfqq_sync(cfqq
);
1141 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
1143 if (time_before(jiffies
, rq
->start_time
+ cfqd
->cfq_fifo_expire
[fifo
]))
1146 cfq_log_cfqq(cfqd
, cfqq
, "fifo=%p", rq
);
1151 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1153 const int base_rq
= cfqd
->cfq_slice_async_rq
;
1155 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1157 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
1161 * Select a queue for service. If we have a current active queue,
1162 * check whether to continue servicing it, or retrieve and set a new one.
1164 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
1166 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1168 cfqq
= cfqd
->active_queue
;
1173 * The active queue has run out of time, expire it and select new.
1175 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
))
1179 * The active queue has requests and isn't expired, allow it to
1182 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
1186 * If another queue has a request waiting within our mean seek
1187 * distance, let it run. The expire code will check for close
1188 * cooperators and put the close queue at the front of the service
1191 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
, 0);
1196 * No requests pending. If the active queue still has requests in
1197 * flight or is idling for a new request, allow either of these
1198 * conditions to happen (or time out) before selecting a new queue.
1200 if (timer_pending(&cfqd
->idle_slice_timer
) ||
1201 (cfqq
->dispatched
&& cfq_cfqq_idle_window(cfqq
))) {
1207 cfq_slice_expired(cfqd
, 0);
1209 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
1214 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
1218 while (cfqq
->next_rq
) {
1219 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
1223 BUG_ON(!list_empty(&cfqq
->fifo
));
1228 * Drain our current requests. Used for barriers and when switching
1229 * io schedulers on-the-fly.
1231 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
1233 struct cfq_queue
*cfqq
;
1236 while ((cfqq
= cfq_rb_first(&cfqd
->service_tree
)) != NULL
)
1237 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1239 cfq_slice_expired(cfqd
, 0);
1241 BUG_ON(cfqd
->busy_queues
);
1243 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
1248 * Dispatch a request from cfqq, moving them to the request queue
1251 static void cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1255 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
1258 * follow expired path, else get first next available
1260 rq
= cfq_check_fifo(cfqq
);
1265 * insert request into driver dispatch list
1267 cfq_dispatch_insert(cfqd
->queue
, rq
);
1269 if (!cfqd
->active_cic
) {
1270 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1272 atomic_long_inc(&cic
->ioc
->refcount
);
1273 cfqd
->active_cic
= cic
;
1278 * Find the cfqq that we need to service and move a request from that to the
1281 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
1283 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1284 struct cfq_queue
*cfqq
;
1285 unsigned int max_dispatch
;
1287 if (!cfqd
->busy_queues
)
1290 if (unlikely(force
))
1291 return cfq_forced_dispatch(cfqd
);
1293 cfqq
= cfq_select_queue(cfqd
);
1298 * Drain async requests before we start sync IO
1300 if (cfq_cfqq_idle_window(cfqq
) && cfqd
->rq_in_driver
[BLK_RW_ASYNC
])
1304 * If this is an async queue and we have sync IO in flight, let it wait
1306 if (cfqd
->sync_flight
&& !cfq_cfqq_sync(cfqq
))
1309 max_dispatch
= cfqd
->cfq_quantum
;
1310 if (cfq_class_idle(cfqq
))
1314 * Does this cfqq already have too much IO in flight?
1316 if (cfqq
->dispatched
>= max_dispatch
) {
1318 * idle queue must always only have a single IO in flight
1320 if (cfq_class_idle(cfqq
))
1324 * We have other queues, don't allow more IO from this one
1326 if (cfqd
->busy_queues
> 1)
1330 * we are the only queue, allow up to 4 times of 'quantum'
1332 if (cfqq
->dispatched
>= 4 * max_dispatch
)
1337 * Dispatch a request from this cfqq
1339 cfq_dispatch_request(cfqd
, cfqq
);
1340 cfqq
->slice_dispatch
++;
1341 cfq_clear_cfqq_must_dispatch(cfqq
);
1344 * expire an async queue immediately if it has used up its slice. idle
1345 * queue always expire after 1 dispatch round.
1347 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
1348 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
1349 cfq_class_idle(cfqq
))) {
1350 cfqq
->slice_end
= jiffies
+ 1;
1351 cfq_slice_expired(cfqd
, 0);
1354 cfq_log(cfqd
, "dispatched a request");
1359 * task holds one reference to the queue, dropped when task exits. each rq
1360 * in-flight on this queue also holds a reference, dropped when rq is freed.
1362 * queue lock must be held here.
1364 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1366 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1368 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1370 if (!atomic_dec_and_test(&cfqq
->ref
))
1373 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
1374 BUG_ON(rb_first(&cfqq
->sort_list
));
1375 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1376 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1378 if (unlikely(cfqd
->active_queue
== cfqq
)) {
1379 __cfq_slice_expired(cfqd
, cfqq
, 0);
1380 cfq_schedule_dispatch(cfqd
);
1383 kmem_cache_free(cfq_pool
, cfqq
);
1387 * Must always be called with the rcu_read_lock() held
1390 __call_for_each_cic(struct io_context
*ioc
,
1391 void (*func
)(struct io_context
*, struct cfq_io_context
*))
1393 struct cfq_io_context
*cic
;
1394 struct hlist_node
*n
;
1396 hlist_for_each_entry_rcu(cic
, n
, &ioc
->cic_list
, cic_list
)
1401 * Call func for each cic attached to this ioc.
1404 call_for_each_cic(struct io_context
*ioc
,
1405 void (*func
)(struct io_context
*, struct cfq_io_context
*))
1408 __call_for_each_cic(ioc
, func
);
1412 static void cfq_cic_free_rcu(struct rcu_head
*head
)
1414 struct cfq_io_context
*cic
;
1416 cic
= container_of(head
, struct cfq_io_context
, rcu_head
);
1418 kmem_cache_free(cfq_ioc_pool
, cic
);
1419 elv_ioc_count_dec(ioc_count
);
1423 * CFQ scheduler is exiting, grab exit lock and check
1424 * the pending io context count. If it hits zero,
1425 * complete ioc_gone and set it back to NULL
1427 spin_lock(&ioc_gone_lock
);
1428 if (ioc_gone
&& !elv_ioc_count_read(ioc_count
)) {
1432 spin_unlock(&ioc_gone_lock
);
1436 static void cfq_cic_free(struct cfq_io_context
*cic
)
1438 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
1441 static void cic_free_func(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1443 unsigned long flags
;
1445 BUG_ON(!cic
->dead_key
);
1447 spin_lock_irqsave(&ioc
->lock
, flags
);
1448 radix_tree_delete(&ioc
->radix_root
, cic
->dead_key
);
1449 hlist_del_rcu(&cic
->cic_list
);
1450 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1456 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1457 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1458 * and ->trim() which is called with the task lock held
1460 static void cfq_free_io_context(struct io_context
*ioc
)
1463 * ioc->refcount is zero here, or we are called from elv_unregister(),
1464 * so no more cic's are allowed to be linked into this ioc. So it
1465 * should be ok to iterate over the known list, we will see all cic's
1466 * since no new ones are added.
1468 __call_for_each_cic(ioc
, cic_free_func
);
1471 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1473 if (unlikely(cfqq
== cfqd
->active_queue
)) {
1474 __cfq_slice_expired(cfqd
, cfqq
, 0);
1475 cfq_schedule_dispatch(cfqd
);
1478 cfq_put_queue(cfqq
);
1481 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1482 struct cfq_io_context
*cic
)
1484 struct io_context
*ioc
= cic
->ioc
;
1486 list_del_init(&cic
->queue_list
);
1489 * Make sure key == NULL is seen for dead queues
1492 cic
->dead_key
= (unsigned long) cic
->key
;
1495 if (ioc
->ioc_data
== cic
)
1496 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
1498 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
1499 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
1500 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
1503 if (cic
->cfqq
[BLK_RW_SYNC
]) {
1504 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
1505 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
1509 static void cfq_exit_single_io_context(struct io_context
*ioc
,
1510 struct cfq_io_context
*cic
)
1512 struct cfq_data
*cfqd
= cic
->key
;
1515 struct request_queue
*q
= cfqd
->queue
;
1516 unsigned long flags
;
1518 spin_lock_irqsave(q
->queue_lock
, flags
);
1521 * Ensure we get a fresh copy of the ->key to prevent
1522 * race between exiting task and queue
1524 smp_read_barrier_depends();
1526 __cfq_exit_single_io_context(cfqd
, cic
);
1528 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1533 * The process that ioc belongs to has exited, we need to clean up
1534 * and put the internal structures we have that belongs to that process.
1536 static void cfq_exit_io_context(struct io_context
*ioc
)
1538 call_for_each_cic(ioc
, cfq_exit_single_io_context
);
1541 static struct cfq_io_context
*
1542 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1544 struct cfq_io_context
*cic
;
1546 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
1549 cic
->last_end_request
= jiffies
;
1550 INIT_LIST_HEAD(&cic
->queue_list
);
1551 INIT_HLIST_NODE(&cic
->cic_list
);
1552 cic
->dtor
= cfq_free_io_context
;
1553 cic
->exit
= cfq_exit_io_context
;
1554 elv_ioc_count_inc(ioc_count
);
1560 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
1562 struct task_struct
*tsk
= current
;
1565 if (!cfq_cfqq_prio_changed(cfqq
))
1568 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
1569 switch (ioprio_class
) {
1571 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1572 case IOPRIO_CLASS_NONE
:
1574 * no prio set, inherit CPU scheduling settings
1576 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1577 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
1579 case IOPRIO_CLASS_RT
:
1580 cfqq
->ioprio
= task_ioprio(ioc
);
1581 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1583 case IOPRIO_CLASS_BE
:
1584 cfqq
->ioprio
= task_ioprio(ioc
);
1585 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1587 case IOPRIO_CLASS_IDLE
:
1588 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1590 cfq_clear_cfqq_idle_window(cfqq
);
1595 * keep track of original prio settings in case we have to temporarily
1596 * elevate the priority of this queue
1598 cfqq
->org_ioprio
= cfqq
->ioprio
;
1599 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1600 cfq_clear_cfqq_prio_changed(cfqq
);
1603 static void changed_ioprio(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1605 struct cfq_data
*cfqd
= cic
->key
;
1606 struct cfq_queue
*cfqq
;
1607 unsigned long flags
;
1609 if (unlikely(!cfqd
))
1612 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1614 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
1616 struct cfq_queue
*new_cfqq
;
1617 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->ioc
,
1620 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
1621 cfq_put_queue(cfqq
);
1625 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
1627 cfq_mark_cfqq_prio_changed(cfqq
);
1629 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1632 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1634 call_for_each_cic(ioc
, changed_ioprio
);
1635 ioc
->ioprio_changed
= 0;
1638 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1639 pid_t pid
, int is_sync
)
1641 RB_CLEAR_NODE(&cfqq
->rb_node
);
1642 RB_CLEAR_NODE(&cfqq
->p_node
);
1643 INIT_LIST_HEAD(&cfqq
->fifo
);
1645 atomic_set(&cfqq
->ref
, 0);
1648 cfq_mark_cfqq_prio_changed(cfqq
);
1651 if (!cfq_class_idle(cfqq
))
1652 cfq_mark_cfqq_idle_window(cfqq
);
1653 cfq_mark_cfqq_sync(cfqq
);
1658 static struct cfq_queue
*
1659 cfq_find_alloc_queue(struct cfq_data
*cfqd
, int is_sync
,
1660 struct io_context
*ioc
, gfp_t gfp_mask
)
1662 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1663 struct cfq_io_context
*cic
;
1666 cic
= cfq_cic_lookup(cfqd
, ioc
);
1667 /* cic always exists here */
1668 cfqq
= cic_to_cfqq(cic
, is_sync
);
1671 * Always try a new alloc if we fell back to the OOM cfqq
1672 * originally, since it should just be a temporary situation.
1674 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
1679 } else if (gfp_mask
& __GFP_WAIT
) {
1680 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1681 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
1682 gfp_mask
| __GFP_ZERO
,
1684 spin_lock_irq(cfqd
->queue
->queue_lock
);
1688 cfqq
= kmem_cache_alloc_node(cfq_pool
,
1689 gfp_mask
| __GFP_ZERO
,
1694 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
1695 cfq_init_prio_data(cfqq
, ioc
);
1696 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
1698 cfqq
= &cfqd
->oom_cfqq
;
1702 kmem_cache_free(cfq_pool
, new_cfqq
);
1707 static struct cfq_queue
**
1708 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
1710 switch (ioprio_class
) {
1711 case IOPRIO_CLASS_RT
:
1712 return &cfqd
->async_cfqq
[0][ioprio
];
1713 case IOPRIO_CLASS_BE
:
1714 return &cfqd
->async_cfqq
[1][ioprio
];
1715 case IOPRIO_CLASS_IDLE
:
1716 return &cfqd
->async_idle_cfqq
;
1722 static struct cfq_queue
*
1723 cfq_get_queue(struct cfq_data
*cfqd
, int is_sync
, struct io_context
*ioc
,
1726 const int ioprio
= task_ioprio(ioc
);
1727 const int ioprio_class
= task_ioprio_class(ioc
);
1728 struct cfq_queue
**async_cfqq
= NULL
;
1729 struct cfq_queue
*cfqq
= NULL
;
1732 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
1737 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
1740 * pin the queue now that it's allocated, scheduler exit will prune it
1742 if (!is_sync
&& !(*async_cfqq
)) {
1743 atomic_inc(&cfqq
->ref
);
1747 atomic_inc(&cfqq
->ref
);
1752 * We drop cfq io contexts lazily, so we may find a dead one.
1755 cfq_drop_dead_cic(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1756 struct cfq_io_context
*cic
)
1758 unsigned long flags
;
1760 WARN_ON(!list_empty(&cic
->queue_list
));
1762 spin_lock_irqsave(&ioc
->lock
, flags
);
1764 BUG_ON(ioc
->ioc_data
== cic
);
1766 radix_tree_delete(&ioc
->radix_root
, (unsigned long) cfqd
);
1767 hlist_del_rcu(&cic
->cic_list
);
1768 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1773 static struct cfq_io_context
*
1774 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1776 struct cfq_io_context
*cic
;
1777 unsigned long flags
;
1786 * we maintain a last-hit cache, to avoid browsing over the tree
1788 cic
= rcu_dereference(ioc
->ioc_data
);
1789 if (cic
&& cic
->key
== cfqd
) {
1795 cic
= radix_tree_lookup(&ioc
->radix_root
, (unsigned long) cfqd
);
1799 /* ->key must be copied to avoid race with cfq_exit_queue() */
1802 cfq_drop_dead_cic(cfqd
, ioc
, cic
);
1807 spin_lock_irqsave(&ioc
->lock
, flags
);
1808 rcu_assign_pointer(ioc
->ioc_data
, cic
);
1809 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1817 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1818 * the process specific cfq io context when entered from the block layer.
1819 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1821 static int cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1822 struct cfq_io_context
*cic
, gfp_t gfp_mask
)
1824 unsigned long flags
;
1827 ret
= radix_tree_preload(gfp_mask
);
1832 spin_lock_irqsave(&ioc
->lock
, flags
);
1833 ret
= radix_tree_insert(&ioc
->radix_root
,
1834 (unsigned long) cfqd
, cic
);
1836 hlist_add_head_rcu(&cic
->cic_list
, &ioc
->cic_list
);
1837 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1839 radix_tree_preload_end();
1842 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1843 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1844 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1849 printk(KERN_ERR
"cfq: cic link failed!\n");
1855 * Setup general io context and cfq io context. There can be several cfq
1856 * io contexts per general io context, if this process is doing io to more
1857 * than one device managed by cfq.
1859 static struct cfq_io_context
*
1860 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1862 struct io_context
*ioc
= NULL
;
1863 struct cfq_io_context
*cic
;
1865 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1867 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
1871 cic
= cfq_cic_lookup(cfqd
, ioc
);
1875 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1879 if (cfq_cic_link(cfqd
, ioc
, cic
, gfp_mask
))
1883 smp_read_barrier_depends();
1884 if (unlikely(ioc
->ioprio_changed
))
1885 cfq_ioc_set_ioprio(ioc
);
1891 put_io_context(ioc
);
1896 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1898 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
1899 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1901 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1902 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1903 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1907 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
1913 if (!cic
->last_request_pos
)
1915 else if (cic
->last_request_pos
< blk_rq_pos(rq
))
1916 sdist
= blk_rq_pos(rq
) - cic
->last_request_pos
;
1918 sdist
= cic
->last_request_pos
- blk_rq_pos(rq
);
1921 * Don't allow the seek distance to get too large from the
1922 * odd fragment, pagein, etc
1924 if (cic
->seek_samples
<= 60) /* second&third seek */
1925 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1927 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1929 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1930 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1931 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1932 do_div(total
, cic
->seek_samples
);
1933 cic
->seek_mean
= (sector_t
)total
;
1937 * Disable idle window if the process thinks too long or seeks so much that
1941 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1942 struct cfq_io_context
*cic
)
1944 int old_idle
, enable_idle
;
1947 * Don't idle for async or idle io prio class
1949 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
1952 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
1954 if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
1955 (cfqd
->hw_tag
&& CIC_SEEKY(cic
)))
1957 else if (sample_valid(cic
->ttime_samples
)) {
1958 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
1964 if (old_idle
!= enable_idle
) {
1965 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
1967 cfq_mark_cfqq_idle_window(cfqq
);
1969 cfq_clear_cfqq_idle_window(cfqq
);
1974 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1975 * no or if we aren't sure, a 1 will cause a preempt.
1978 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
1981 struct cfq_queue
*cfqq
;
1983 cfqq
= cfqd
->active_queue
;
1987 if (cfq_slice_used(cfqq
))
1990 if (cfq_class_idle(new_cfqq
))
1993 if (cfq_class_idle(cfqq
))
1997 * if the new request is sync, but the currently running queue is
1998 * not, let the sync request have priority.
2000 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
2004 * So both queues are sync. Let the new request get disk time if
2005 * it's a metadata request and the current queue is doing regular IO.
2007 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
2011 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2013 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
2016 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
2020 * if this request is as-good as one we would expect from the
2021 * current cfqq, let it preempt
2023 if (cfq_rq_close(cfqd
, rq
))
2030 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2031 * let it have half of its nominal slice.
2033 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2035 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
2036 cfq_slice_expired(cfqd
, 1);
2039 * Put the new queue at the front of the of the current list,
2040 * so we know that it will be selected next.
2042 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
2044 cfq_service_tree_add(cfqd
, cfqq
, 1);
2046 cfqq
->slice_end
= 0;
2047 cfq_mark_cfqq_slice_new(cfqq
);
2051 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2052 * something we should do about it
2055 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2058 struct cfq_io_context
*cic
= RQ_CIC(rq
);
2062 cfqq
->meta_pending
++;
2064 cfq_update_io_thinktime(cfqd
, cic
);
2065 cfq_update_io_seektime(cfqd
, cic
, rq
);
2066 cfq_update_idle_window(cfqd
, cfqq
, cic
);
2068 cic
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2070 if (cfqq
== cfqd
->active_queue
) {
2072 * Remember that we saw a request from this process, but
2073 * don't start queuing just yet. Otherwise we risk seeing lots
2074 * of tiny requests, because we disrupt the normal plugging
2075 * and merging. If the request is already larger than a single
2076 * page, let it rip immediately. For that case we assume that
2077 * merging is already done. Ditto for a busy system that
2078 * has other work pending, don't risk delaying until the
2079 * idle timer unplug to continue working.
2081 if (cfq_cfqq_wait_request(cfqq
)) {
2082 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
2083 cfqd
->busy_queues
> 1) {
2084 del_timer(&cfqd
->idle_slice_timer
);
2085 __blk_run_queue(cfqd
->queue
);
2087 cfq_mark_cfqq_must_dispatch(cfqq
);
2089 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
2091 * not the active queue - expire current slice if it is
2092 * idle and has expired it's mean thinktime or this new queue
2093 * has some old slice time left and is of higher priority or
2094 * this new queue is RT and the current one is BE
2096 cfq_preempt_queue(cfqd
, cfqq
);
2097 __blk_run_queue(cfqd
->queue
);
2101 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
2103 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2104 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2106 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
2107 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
2111 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
2113 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
2117 * Update hw_tag based on peak queue depth over 50 samples under
2120 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
2122 if (rq_in_driver(cfqd
) > cfqd
->rq_in_driver_peak
)
2123 cfqd
->rq_in_driver_peak
= rq_in_driver(cfqd
);
2125 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
2126 rq_in_driver(cfqd
) <= CFQ_HW_QUEUE_MIN
)
2129 if (cfqd
->hw_tag_samples
++ < 50)
2132 if (cfqd
->rq_in_driver_peak
>= CFQ_HW_QUEUE_MIN
)
2137 cfqd
->hw_tag_samples
= 0;
2138 cfqd
->rq_in_driver_peak
= 0;
2141 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
2143 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2144 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2145 const int sync
= rq_is_sync(rq
);
2149 cfq_log_cfqq(cfqd
, cfqq
, "complete");
2151 cfq_update_hw_tag(cfqd
);
2153 WARN_ON(!cfqd
->rq_in_driver
[sync
]);
2154 WARN_ON(!cfqq
->dispatched
);
2155 cfqd
->rq_in_driver
[sync
]--;
2158 if (cfq_cfqq_sync(cfqq
))
2159 cfqd
->sync_flight
--;
2162 RQ_CIC(rq
)->last_end_request
= now
;
2165 * If this is the active queue, check if it needs to be expired,
2166 * or if we want to idle in case it has no pending requests.
2168 if (cfqd
->active_queue
== cfqq
) {
2169 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
2171 if (cfq_cfqq_slice_new(cfqq
)) {
2172 cfq_set_prio_slice(cfqd
, cfqq
);
2173 cfq_clear_cfqq_slice_new(cfqq
);
2176 * If there are no requests waiting in this queue, and
2177 * there are other queues ready to issue requests, AND
2178 * those other queues are issuing requests within our
2179 * mean seek distance, give them a chance to run instead
2182 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
2183 cfq_slice_expired(cfqd
, 1);
2184 else if (cfqq_empty
&& !cfq_close_cooperator(cfqd
, cfqq
, 1) &&
2185 sync
&& !rq_noidle(rq
))
2186 cfq_arm_slice_timer(cfqd
);
2189 if (!rq_in_driver(cfqd
))
2190 cfq_schedule_dispatch(cfqd
);
2194 * we temporarily boost lower priority queues if they are holding fs exclusive
2195 * resources. they are boosted to normal prio (CLASS_BE/4)
2197 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
2199 if (has_fs_excl()) {
2201 * boost idle prio on transactions that would lock out other
2202 * users of the filesystem
2204 if (cfq_class_idle(cfqq
))
2205 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2206 if (cfqq
->ioprio
> IOPRIO_NORM
)
2207 cfqq
->ioprio
= IOPRIO_NORM
;
2210 * check if we need to unboost the queue
2212 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
2213 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
2214 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
2215 cfqq
->ioprio
= cfqq
->org_ioprio
;
2219 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
2221 if ((cfq_cfqq_wait_request(cfqq
) || cfq_cfqq_must_alloc(cfqq
)) &&
2222 !cfq_cfqq_must_alloc_slice(cfqq
)) {
2223 cfq_mark_cfqq_must_alloc_slice(cfqq
);
2224 return ELV_MQUEUE_MUST
;
2227 return ELV_MQUEUE_MAY
;
2230 static int cfq_may_queue(struct request_queue
*q
, int rw
)
2232 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2233 struct task_struct
*tsk
= current
;
2234 struct cfq_io_context
*cic
;
2235 struct cfq_queue
*cfqq
;
2238 * don't force setup of a queue from here, as a call to may_queue
2239 * does not necessarily imply that a request actually will be queued.
2240 * so just lookup a possibly existing queue, or return 'may queue'
2243 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
2245 return ELV_MQUEUE_MAY
;
2247 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
2249 cfq_init_prio_data(cfqq
, cic
->ioc
);
2250 cfq_prio_boost(cfqq
);
2252 return __cfq_may_queue(cfqq
);
2255 return ELV_MQUEUE_MAY
;
2259 * queue lock held here
2261 static void cfq_put_request(struct request
*rq
)
2263 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2266 const int rw
= rq_data_dir(rq
);
2268 BUG_ON(!cfqq
->allocated
[rw
]);
2269 cfqq
->allocated
[rw
]--;
2271 put_io_context(RQ_CIC(rq
)->ioc
);
2273 rq
->elevator_private
= NULL
;
2274 rq
->elevator_private2
= NULL
;
2276 cfq_put_queue(cfqq
);
2281 * Allocate cfq data structures associated with this request.
2284 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
2286 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2287 struct cfq_io_context
*cic
;
2288 const int rw
= rq_data_dir(rq
);
2289 const int is_sync
= rq_is_sync(rq
);
2290 struct cfq_queue
*cfqq
;
2291 unsigned long flags
;
2293 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2295 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
2297 spin_lock_irqsave(q
->queue_lock
, flags
);
2302 cfqq
= cic_to_cfqq(cic
, is_sync
);
2303 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2304 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
2305 cic_set_cfqq(cic
, cfqq
, is_sync
);
2308 cfqq
->allocated
[rw
]++;
2309 cfq_clear_cfqq_must_alloc(cfqq
);
2310 atomic_inc(&cfqq
->ref
);
2312 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2314 rq
->elevator_private
= cic
;
2315 rq
->elevator_private2
= cfqq
;
2320 put_io_context(cic
->ioc
);
2322 cfq_schedule_dispatch(cfqd
);
2323 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2324 cfq_log(cfqd
, "set_request fail");
2328 static void cfq_kick_queue(struct work_struct
*work
)
2330 struct cfq_data
*cfqd
=
2331 container_of(work
, struct cfq_data
, unplug_work
);
2332 struct request_queue
*q
= cfqd
->queue
;
2334 spin_lock_irq(q
->queue_lock
);
2335 __blk_run_queue(cfqd
->queue
);
2336 spin_unlock_irq(q
->queue_lock
);
2340 * Timer running if the active_queue is currently idling inside its time slice
2342 static void cfq_idle_slice_timer(unsigned long data
)
2344 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
2345 struct cfq_queue
*cfqq
;
2346 unsigned long flags
;
2349 cfq_log(cfqd
, "idle timer fired");
2351 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2353 cfqq
= cfqd
->active_queue
;
2358 * We saw a request before the queue expired, let it through
2360 if (cfq_cfqq_must_dispatch(cfqq
))
2366 if (cfq_slice_used(cfqq
))
2370 * only expire and reinvoke request handler, if there are
2371 * other queues with pending requests
2373 if (!cfqd
->busy_queues
)
2377 * not expired and it has a request pending, let it dispatch
2379 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2383 cfq_slice_expired(cfqd
, timed_out
);
2385 cfq_schedule_dispatch(cfqd
);
2387 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2390 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
2392 del_timer_sync(&cfqd
->idle_slice_timer
);
2393 cancel_work_sync(&cfqd
->unplug_work
);
2396 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
2400 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
2401 if (cfqd
->async_cfqq
[0][i
])
2402 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
2403 if (cfqd
->async_cfqq
[1][i
])
2404 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
2407 if (cfqd
->async_idle_cfqq
)
2408 cfq_put_queue(cfqd
->async_idle_cfqq
);
2411 static void cfq_exit_queue(struct elevator_queue
*e
)
2413 struct cfq_data
*cfqd
= e
->elevator_data
;
2414 struct request_queue
*q
= cfqd
->queue
;
2416 cfq_shutdown_timer_wq(cfqd
);
2418 spin_lock_irq(q
->queue_lock
);
2420 if (cfqd
->active_queue
)
2421 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
2423 while (!list_empty(&cfqd
->cic_list
)) {
2424 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
2425 struct cfq_io_context
,
2428 __cfq_exit_single_io_context(cfqd
, cic
);
2431 cfq_put_async_queues(cfqd
);
2433 spin_unlock_irq(q
->queue_lock
);
2435 cfq_shutdown_timer_wq(cfqd
);
2440 static void *cfq_init_queue(struct request_queue
*q
)
2442 struct cfq_data
*cfqd
;
2445 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
2449 cfqd
->service_tree
= CFQ_RB_ROOT
;
2452 * Not strictly needed (since RB_ROOT just clears the node and we
2453 * zeroed cfqd on alloc), but better be safe in case someone decides
2454 * to add magic to the rb code
2456 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
2457 cfqd
->prio_trees
[i
] = RB_ROOT
;
2460 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2461 * Grab a permanent reference to it, so that the normal code flow
2462 * will not attempt to free it.
2464 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
2465 atomic_inc(&cfqd
->oom_cfqq
.ref
);
2467 INIT_LIST_HEAD(&cfqd
->cic_list
);
2471 init_timer(&cfqd
->idle_slice_timer
);
2472 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2473 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2475 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
2477 cfqd
->cfq_quantum
= cfq_quantum
;
2478 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2479 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2480 cfqd
->cfq_back_max
= cfq_back_max
;
2481 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2482 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2483 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2484 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2485 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2491 static void cfq_slab_kill(void)
2494 * Caller already ensured that pending RCU callbacks are completed,
2495 * so we should have no busy allocations at this point.
2498 kmem_cache_destroy(cfq_pool
);
2500 kmem_cache_destroy(cfq_ioc_pool
);
2503 static int __init
cfq_slab_setup(void)
2505 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
2509 cfq_ioc_pool
= KMEM_CACHE(cfq_io_context
, 0);
2520 * sysfs parts below -->
2523 cfq_var_show(unsigned int var
, char *page
)
2525 return sprintf(page
, "%d\n", var
);
2529 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2531 char *p
= (char *) page
;
2533 *var
= simple_strtoul(p
, &p
, 10);
2537 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2538 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2540 struct cfq_data *cfqd = e->elevator_data; \
2541 unsigned int __data = __VAR; \
2543 __data = jiffies_to_msecs(__data); \
2544 return cfq_var_show(__data, (page)); \
2546 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2547 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2548 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2549 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2550 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2551 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2552 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2553 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2554 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2555 #undef SHOW_FUNCTION
2557 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2558 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2560 struct cfq_data *cfqd = e->elevator_data; \
2561 unsigned int __data; \
2562 int ret = cfq_var_store(&__data, (page), count); \
2563 if (__data < (MIN)) \
2565 else if (__data > (MAX)) \
2568 *(__PTR) = msecs_to_jiffies(__data); \
2570 *(__PTR) = __data; \
2573 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2574 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
2576 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
2578 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2579 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
2581 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2582 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2583 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2584 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
2586 #undef STORE_FUNCTION
2588 #define CFQ_ATTR(name) \
2589 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2591 static struct elv_fs_entry cfq_attrs
[] = {
2593 CFQ_ATTR(fifo_expire_sync
),
2594 CFQ_ATTR(fifo_expire_async
),
2595 CFQ_ATTR(back_seek_max
),
2596 CFQ_ATTR(back_seek_penalty
),
2597 CFQ_ATTR(slice_sync
),
2598 CFQ_ATTR(slice_async
),
2599 CFQ_ATTR(slice_async_rq
),
2600 CFQ_ATTR(slice_idle
),
2604 static struct elevator_type iosched_cfq
= {
2606 .elevator_merge_fn
= cfq_merge
,
2607 .elevator_merged_fn
= cfq_merged_request
,
2608 .elevator_merge_req_fn
= cfq_merged_requests
,
2609 .elevator_allow_merge_fn
= cfq_allow_merge
,
2610 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2611 .elevator_add_req_fn
= cfq_insert_request
,
2612 .elevator_activate_req_fn
= cfq_activate_request
,
2613 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2614 .elevator_queue_empty_fn
= cfq_queue_empty
,
2615 .elevator_completed_req_fn
= cfq_completed_request
,
2616 .elevator_former_req_fn
= elv_rb_former_request
,
2617 .elevator_latter_req_fn
= elv_rb_latter_request
,
2618 .elevator_set_req_fn
= cfq_set_request
,
2619 .elevator_put_req_fn
= cfq_put_request
,
2620 .elevator_may_queue_fn
= cfq_may_queue
,
2621 .elevator_init_fn
= cfq_init_queue
,
2622 .elevator_exit_fn
= cfq_exit_queue
,
2623 .trim
= cfq_free_io_context
,
2625 .elevator_attrs
= cfq_attrs
,
2626 .elevator_name
= "cfq",
2627 .elevator_owner
= THIS_MODULE
,
2630 static int __init
cfq_init(void)
2633 * could be 0 on HZ < 1000 setups
2635 if (!cfq_slice_async
)
2636 cfq_slice_async
= 1;
2637 if (!cfq_slice_idle
)
2640 if (cfq_slab_setup())
2643 elv_register(&iosched_cfq
);
2648 static void __exit
cfq_exit(void)
2650 DECLARE_COMPLETION_ONSTACK(all_gone
);
2651 elv_unregister(&iosched_cfq
);
2652 ioc_gone
= &all_gone
;
2653 /* ioc_gone's update must be visible before reading ioc_count */
2657 * this also protects us from entering cfq_slab_kill() with
2658 * pending RCU callbacks
2660 if (elv_ioc_count_read(ioc_count
))
2661 wait_for_completion(&all_gone
);
2665 module_init(cfq_init
);
2666 module_exit(cfq_exit
);
2668 MODULE_AUTHOR("Jens Axboe");
2669 MODULE_LICENSE("GPL");
2670 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");