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, cfq_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
;
115 unsigned int seek_samples
;
118 sector_t last_request_pos
;
122 struct cfq_queue
*new_cfqq
;
126 * Per block device queue structure
129 struct request_queue
*queue
;
132 * rr list of queues with requests and the count of them
134 struct cfq_rb_root service_tree
;
137 * Each priority tree is sorted by next_request position. These
138 * trees are used when determining if two or more queues are
139 * interleaving requests (see cfq_close_cooperator).
141 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
143 unsigned int busy_queues
;
149 * queue-depth detection
154 int rq_in_driver_peak
;
157 * idle window management
159 struct timer_list idle_slice_timer
;
160 struct work_struct unplug_work
;
162 struct cfq_queue
*active_queue
;
163 struct cfq_io_context
*active_cic
;
166 * async queue for each priority case
168 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
169 struct cfq_queue
*async_idle_cfqq
;
171 sector_t last_position
;
174 * tunables, see top of file
176 unsigned int cfq_quantum
;
177 unsigned int cfq_fifo_expire
[2];
178 unsigned int cfq_back_penalty
;
179 unsigned int cfq_back_max
;
180 unsigned int cfq_slice
[2];
181 unsigned int cfq_slice_async_rq
;
182 unsigned int cfq_slice_idle
;
183 unsigned int cfq_latency
;
185 struct list_head cic_list
;
188 * Fallback dummy cfqq for extreme OOM conditions
190 struct cfq_queue oom_cfqq
;
192 unsigned long last_end_sync_rq
;
195 enum cfqq_state_flags
{
196 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
197 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
198 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
199 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
200 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
201 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
202 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
203 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
204 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
205 CFQ_CFQQ_FLAG_coop
, /* has done a coop jump of the queue */
208 #define CFQ_CFQQ_FNS(name) \
209 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
211 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
213 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
215 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
217 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
219 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
223 CFQ_CFQQ_FNS(wait_request
);
224 CFQ_CFQQ_FNS(must_dispatch
);
225 CFQ_CFQQ_FNS(must_alloc_slice
);
226 CFQ_CFQQ_FNS(fifo_expire
);
227 CFQ_CFQQ_FNS(idle_window
);
228 CFQ_CFQQ_FNS(prio_changed
);
229 CFQ_CFQQ_FNS(slice_new
);
234 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
235 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
236 #define cfq_log(cfqd, fmt, args...) \
237 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
239 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
240 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, bool,
241 struct io_context
*, gfp_t
);
242 static struct cfq_io_context
*cfq_cic_lookup(struct cfq_data
*,
243 struct io_context
*);
245 static inline int rq_in_driver(struct cfq_data
*cfqd
)
247 return cfqd
->rq_in_driver
[0] + cfqd
->rq_in_driver
[1];
250 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_context
*cic
,
253 return cic
->cfqq
[is_sync
];
256 static inline void cic_set_cfqq(struct cfq_io_context
*cic
,
257 struct cfq_queue
*cfqq
, bool is_sync
)
259 cic
->cfqq
[is_sync
] = cfqq
;
263 * We regard a request as SYNC, if it's either a read or has the SYNC bit
264 * set (in which case it could also be direct WRITE).
266 static inline bool cfq_bio_sync(struct bio
*bio
)
268 return bio_data_dir(bio
) == READ
|| bio_rw_flagged(bio
, BIO_RW_SYNCIO
);
272 * scheduler run of queue, if there are requests pending and no one in the
273 * driver that will restart queueing
275 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
277 if (cfqd
->busy_queues
) {
278 cfq_log(cfqd
, "schedule dispatch");
279 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
283 static int cfq_queue_empty(struct request_queue
*q
)
285 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
287 return !cfqd
->busy_queues
;
291 * Scale schedule slice based on io priority. Use the sync time slice only
292 * if a queue is marked sync and has sync io queued. A sync queue with async
293 * io only, should not get full sync slice length.
295 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
298 const int base_slice
= cfqd
->cfq_slice
[sync
];
300 WARN_ON(prio
>= IOPRIO_BE_NR
);
302 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
306 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
308 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
312 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
314 cfqq
->slice_end
= cfq_prio_to_slice(cfqd
, cfqq
) + jiffies
;
315 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
319 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
320 * isn't valid until the first request from the dispatch is activated
321 * and the slice time set.
323 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
325 if (cfq_cfqq_slice_new(cfqq
))
327 if (time_before(jiffies
, cfqq
->slice_end
))
334 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
335 * We choose the request that is closest to the head right now. Distance
336 * behind the head is penalized and only allowed to a certain extent.
338 static struct request
*
339 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
)
341 sector_t last
, s1
, s2
, d1
= 0, d2
= 0;
342 unsigned long back_max
;
343 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
344 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
345 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
347 if (rq1
== NULL
|| rq1
== rq2
)
352 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
354 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
356 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
358 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
361 s1
= blk_rq_pos(rq1
);
362 s2
= blk_rq_pos(rq2
);
364 last
= cfqd
->last_position
;
367 * by definition, 1KiB is 2 sectors
369 back_max
= cfqd
->cfq_back_max
* 2;
372 * Strict one way elevator _except_ in the case where we allow
373 * short backward seeks which are biased as twice the cost of a
374 * similar forward seek.
378 else if (s1
+ back_max
>= last
)
379 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
381 wrap
|= CFQ_RQ1_WRAP
;
385 else if (s2
+ back_max
>= last
)
386 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
388 wrap
|= CFQ_RQ2_WRAP
;
390 /* Found required data */
393 * By doing switch() on the bit mask "wrap" we avoid having to
394 * check two variables for all permutations: --> faster!
397 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
413 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
416 * Since both rqs are wrapped,
417 * start with the one that's further behind head
418 * (--> only *one* back seek required),
419 * since back seek takes more time than forward.
429 * The below is leftmost cache rbtree addon
431 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
434 root
->left
= rb_first(&root
->rb
);
437 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
442 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
448 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
452 rb_erase_init(n
, &root
->rb
);
456 * would be nice to take fifo expire time into account as well
458 static struct request
*
459 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
460 struct request
*last
)
462 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
463 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
464 struct request
*next
= NULL
, *prev
= NULL
;
466 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
469 prev
= rb_entry_rq(rbprev
);
472 next
= rb_entry_rq(rbnext
);
474 rbnext
= rb_first(&cfqq
->sort_list
);
475 if (rbnext
&& rbnext
!= &last
->rb_node
)
476 next
= rb_entry_rq(rbnext
);
479 return cfq_choose_req(cfqd
, next
, prev
);
482 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
483 struct cfq_queue
*cfqq
)
486 * just an approximation, should be ok.
488 return (cfqd
->busy_queues
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
489 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
493 * The cfqd->service_tree holds all pending cfq_queue's that have
494 * requests waiting to be processed. It is sorted in the order that
495 * we will service the queues.
497 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
500 struct rb_node
**p
, *parent
;
501 struct cfq_queue
*__cfqq
;
502 unsigned long rb_key
;
505 if (cfq_class_idle(cfqq
)) {
506 rb_key
= CFQ_IDLE_DELAY
;
507 parent
= rb_last(&cfqd
->service_tree
.rb
);
508 if (parent
&& parent
!= &cfqq
->rb_node
) {
509 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
510 rb_key
+= __cfqq
->rb_key
;
513 } else if (!add_front
) {
515 * Get our rb key offset. Subtract any residual slice
516 * value carried from last service. A negative resid
517 * count indicates slice overrun, and this should position
518 * the next service time further away in the tree.
520 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
521 rb_key
-= cfqq
->slice_resid
;
522 cfqq
->slice_resid
= 0;
525 __cfqq
= cfq_rb_first(&cfqd
->service_tree
);
526 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
529 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
531 * same position, nothing more to do
533 if (rb_key
== cfqq
->rb_key
)
536 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
541 p
= &cfqd
->service_tree
.rb
.rb_node
;
546 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
549 * sort RT queues first, we always want to give
550 * preference to them. IDLE queues goes to the back.
551 * after that, sort on the next service time.
553 if (cfq_class_rt(cfqq
) > cfq_class_rt(__cfqq
))
555 else if (cfq_class_rt(cfqq
) < cfq_class_rt(__cfqq
))
557 else if (cfq_class_idle(cfqq
) < cfq_class_idle(__cfqq
))
559 else if (cfq_class_idle(cfqq
) > cfq_class_idle(__cfqq
))
561 else if (time_before(rb_key
, __cfqq
->rb_key
))
566 if (n
== &(*p
)->rb_right
)
573 cfqd
->service_tree
.left
= &cfqq
->rb_node
;
575 cfqq
->rb_key
= rb_key
;
576 rb_link_node(&cfqq
->rb_node
, parent
, p
);
577 rb_insert_color(&cfqq
->rb_node
, &cfqd
->service_tree
.rb
);
580 static struct cfq_queue
*
581 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
582 sector_t sector
, struct rb_node
**ret_parent
,
583 struct rb_node
***rb_link
)
585 struct rb_node
**p
, *parent
;
586 struct cfq_queue
*cfqq
= NULL
;
594 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
597 * Sort strictly based on sector. Smallest to the left,
598 * largest to the right.
600 if (sector
> blk_rq_pos(cfqq
->next_rq
))
602 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
610 *ret_parent
= parent
;
616 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
618 struct rb_node
**p
, *parent
;
619 struct cfq_queue
*__cfqq
;
622 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
626 if (cfq_class_idle(cfqq
))
631 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
632 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
633 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
635 rb_link_node(&cfqq
->p_node
, parent
, p
);
636 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
642 * Update cfqq's position in the service tree.
644 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
647 * Resorting requires the cfqq to be on the RR list already.
649 if (cfq_cfqq_on_rr(cfqq
)) {
650 cfq_service_tree_add(cfqd
, cfqq
, 0);
651 cfq_prio_tree_add(cfqd
, cfqq
);
656 * add to busy list of queues for service, trying to be fair in ordering
657 * the pending list according to last request service
659 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
661 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
662 BUG_ON(cfq_cfqq_on_rr(cfqq
));
663 cfq_mark_cfqq_on_rr(cfqq
);
666 cfq_resort_rr_list(cfqd
, cfqq
);
670 * Called when the cfqq no longer has requests pending, remove it from
673 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
675 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
676 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
677 cfq_clear_cfqq_on_rr(cfqq
);
679 if (!RB_EMPTY_NODE(&cfqq
->rb_node
))
680 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
682 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
686 BUG_ON(!cfqd
->busy_queues
);
691 * rb tree support functions
693 static void cfq_del_rq_rb(struct request
*rq
)
695 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
696 struct cfq_data
*cfqd
= cfqq
->cfqd
;
697 const int sync
= rq_is_sync(rq
);
699 BUG_ON(!cfqq
->queued
[sync
]);
700 cfqq
->queued
[sync
]--;
702 elv_rb_del(&cfqq
->sort_list
, rq
);
704 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
705 cfq_del_cfqq_rr(cfqd
, cfqq
);
708 static void cfq_add_rq_rb(struct request
*rq
)
710 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
711 struct cfq_data
*cfqd
= cfqq
->cfqd
;
712 struct request
*__alias
, *prev
;
714 cfqq
->queued
[rq_is_sync(rq
)]++;
717 * looks a little odd, but the first insert might return an alias.
718 * if that happens, put the alias on the dispatch list
720 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
721 cfq_dispatch_insert(cfqd
->queue
, __alias
);
723 if (!cfq_cfqq_on_rr(cfqq
))
724 cfq_add_cfqq_rr(cfqd
, cfqq
);
727 * check if this request is a better next-serve candidate
729 prev
= cfqq
->next_rq
;
730 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
733 * adjust priority tree position, if ->next_rq changes
735 if (prev
!= cfqq
->next_rq
)
736 cfq_prio_tree_add(cfqd
, cfqq
);
738 BUG_ON(!cfqq
->next_rq
);
741 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
743 elv_rb_del(&cfqq
->sort_list
, rq
);
744 cfqq
->queued
[rq_is_sync(rq
)]--;
748 static struct request
*
749 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
751 struct task_struct
*tsk
= current
;
752 struct cfq_io_context
*cic
;
753 struct cfq_queue
*cfqq
;
755 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
759 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
761 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
763 return elv_rb_find(&cfqq
->sort_list
, sector
);
769 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
771 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
773 cfqd
->rq_in_driver
[rq_is_sync(rq
)]++;
774 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
777 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
780 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
782 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
783 const int sync
= rq_is_sync(rq
);
785 WARN_ON(!cfqd
->rq_in_driver
[sync
]);
786 cfqd
->rq_in_driver
[sync
]--;
787 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
791 static void cfq_remove_request(struct request
*rq
)
793 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
795 if (cfqq
->next_rq
== rq
)
796 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
798 list_del_init(&rq
->queuelist
);
801 cfqq
->cfqd
->rq_queued
--;
802 if (rq_is_meta(rq
)) {
803 WARN_ON(!cfqq
->meta_pending
);
804 cfqq
->meta_pending
--;
808 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
811 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
812 struct request
*__rq
;
814 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
815 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
817 return ELEVATOR_FRONT_MERGE
;
820 return ELEVATOR_NO_MERGE
;
823 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
826 if (type
== ELEVATOR_FRONT_MERGE
) {
827 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
829 cfq_reposition_rq_rb(cfqq
, req
);
834 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
835 struct request
*next
)
838 * reposition in fifo if next is older than rq
840 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
841 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
842 list_move(&rq
->queuelist
, &next
->queuelist
);
843 rq_set_fifo_time(rq
, rq_fifo_time(next
));
846 cfq_remove_request(next
);
849 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
852 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
853 struct cfq_io_context
*cic
;
854 struct cfq_queue
*cfqq
;
857 * Disallow merge of a sync bio into an async request.
859 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
863 * Lookup the cfqq that this bio will be queued with. Allow
864 * merge only if rq is queued there.
866 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
870 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
871 return cfqq
== RQ_CFQQ(rq
);
874 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
875 struct cfq_queue
*cfqq
)
878 cfq_log_cfqq(cfqd
, cfqq
, "set_active");
880 cfqq
->slice_dispatch
= 0;
882 cfq_clear_cfqq_wait_request(cfqq
);
883 cfq_clear_cfqq_must_dispatch(cfqq
);
884 cfq_clear_cfqq_must_alloc_slice(cfqq
);
885 cfq_clear_cfqq_fifo_expire(cfqq
);
886 cfq_mark_cfqq_slice_new(cfqq
);
888 del_timer(&cfqd
->idle_slice_timer
);
891 cfqd
->active_queue
= cfqq
;
895 * current cfqq expired its slice (or was too idle), select new one
898 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
901 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
903 if (cfq_cfqq_wait_request(cfqq
))
904 del_timer(&cfqd
->idle_slice_timer
);
906 cfq_clear_cfqq_wait_request(cfqq
);
909 * store what was left of this slice, if the queue idled/timed out
911 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
)) {
912 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
913 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
916 cfq_resort_rr_list(cfqd
, cfqq
);
918 if (cfqq
== cfqd
->active_queue
)
919 cfqd
->active_queue
= NULL
;
921 if (cfqd
->active_cic
) {
922 put_io_context(cfqd
->active_cic
->ioc
);
923 cfqd
->active_cic
= NULL
;
927 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
929 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
932 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
936 * Get next queue for service. Unless we have a queue preemption,
937 * we'll simply select the first cfqq in the service tree.
939 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
941 if (RB_EMPTY_ROOT(&cfqd
->service_tree
.rb
))
944 return cfq_rb_first(&cfqd
->service_tree
);
948 * Get and set a new active queue for service.
950 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
951 struct cfq_queue
*cfqq
)
954 cfqq
= cfq_get_next_queue(cfqd
);
956 cfq_clear_cfqq_coop(cfqq
);
959 __cfq_set_active_queue(cfqd
, cfqq
);
963 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
966 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
967 return blk_rq_pos(rq
) - cfqd
->last_position
;
969 return cfqd
->last_position
- blk_rq_pos(rq
);
972 #define CFQQ_SEEK_THR 8 * 1024
973 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
975 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
978 sector_t sdist
= cfqq
->seek_mean
;
980 if (!sample_valid(cfqq
->seek_samples
))
981 sdist
= CFQQ_SEEK_THR
;
983 return cfq_dist_from_last(cfqd
, rq
) <= sdist
;
986 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
987 struct cfq_queue
*cur_cfqq
)
989 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
990 struct rb_node
*parent
, *node
;
991 struct cfq_queue
*__cfqq
;
992 sector_t sector
= cfqd
->last_position
;
994 if (RB_EMPTY_ROOT(root
))
998 * First, if we find a request starting at the end of the last
999 * request, choose it.
1001 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1006 * If the exact sector wasn't found, the parent of the NULL leaf
1007 * will contain the closest sector.
1009 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1010 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1013 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1014 node
= rb_next(&__cfqq
->p_node
);
1016 node
= rb_prev(&__cfqq
->p_node
);
1020 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1021 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1029 * cur_cfqq - passed in so that we don't decide that the current queue is
1030 * closely cooperating with itself.
1032 * So, basically we're assuming that that cur_cfqq has dispatched at least
1033 * one request, and that cfqd->last_position reflects a position on the disk
1034 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1037 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1038 struct cfq_queue
*cur_cfqq
,
1041 struct cfq_queue
*cfqq
;
1044 * We should notice if some of the queues are cooperating, eg
1045 * working closely on the same area of the disk. In that case,
1046 * we can group them together and don't waste time idling.
1048 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1053 * It only makes sense to merge sync queues.
1055 if (!cfq_cfqq_sync(cfqq
))
1058 if (cfq_cfqq_coop(cfqq
))
1062 cfq_mark_cfqq_coop(cfqq
);
1066 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1068 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1069 struct cfq_io_context
*cic
;
1073 * SSD device without seek penalty, disable idling. But only do so
1074 * for devices that support queuing, otherwise we still have a problem
1075 * with sync vs async workloads.
1077 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
1080 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
1081 WARN_ON(cfq_cfqq_slice_new(cfqq
));
1084 * idle is disabled, either manually or by past process history
1086 if (!cfqd
->cfq_slice_idle
|| !cfq_cfqq_idle_window(cfqq
))
1090 * still requests with the driver, don't idle
1092 if (rq_in_driver(cfqd
))
1096 * task has exited, don't wait
1098 cic
= cfqd
->active_cic
;
1099 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
1103 * If our average think time is larger than the remaining time
1104 * slice, then don't idle. This avoids overrunning the allotted
1107 if (sample_valid(cic
->ttime_samples
) &&
1108 (cfqq
->slice_end
- jiffies
< cic
->ttime_mean
))
1111 cfq_mark_cfqq_wait_request(cfqq
);
1114 * we don't want to idle for seeks, but we do want to allow
1115 * fair distribution of slice time for a process doing back-to-back
1116 * seeks. so allow a little bit of time for him to submit a new rq
1118 sl
= cfqd
->cfq_slice_idle
;
1119 if (sample_valid(cfqq
->seek_samples
) && CFQQ_SEEKY(cfqq
))
1120 sl
= min(sl
, msecs_to_jiffies(CFQ_MIN_TT
));
1122 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
1123 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu", sl
);
1127 * Move request from internal lists to the request queue dispatch list.
1129 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
1131 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1132 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1134 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
1136 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
1137 cfq_remove_request(rq
);
1139 elv_dispatch_sort(q
, rq
);
1141 if (cfq_cfqq_sync(cfqq
))
1142 cfqd
->sync_flight
++;
1146 * return expired entry, or NULL to just start from scratch in rbtree
1148 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
1150 struct request
*rq
= NULL
;
1152 if (cfq_cfqq_fifo_expire(cfqq
))
1155 cfq_mark_cfqq_fifo_expire(cfqq
);
1157 if (list_empty(&cfqq
->fifo
))
1160 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
1161 if (time_before(jiffies
, rq_fifo_time(rq
)))
1164 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
1169 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1171 const int base_rq
= cfqd
->cfq_slice_async_rq
;
1173 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1175 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
1179 * Must be called with the queue_lock held.
1181 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
1183 int process_refs
, io_refs
;
1185 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
1186 process_refs
= atomic_read(&cfqq
->ref
) - io_refs
;
1187 BUG_ON(process_refs
< 0);
1188 return process_refs
;
1191 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
1194 struct cfq_queue
*__cfqq
;
1196 /* Avoid a circular list and skip interim queue merges */
1197 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
1203 process_refs
= cfqq_process_refs(cfqq
);
1205 * If the process for the cfqq has gone away, there is no
1206 * sense in merging the queues.
1208 if (process_refs
== 0)
1211 cfqq
->new_cfqq
= new_cfqq
;
1212 atomic_add(process_refs
, &new_cfqq
->ref
);
1216 * Select a queue for service. If we have a current active queue,
1217 * check whether to continue servicing it, or retrieve and set a new one.
1219 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
1221 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1223 cfqq
= cfqd
->active_queue
;
1228 * The active queue has run out of time, expire it and select new.
1230 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
))
1234 * The active queue has requests and isn't expired, allow it to
1237 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
1241 * If another queue has a request waiting within our mean seek
1242 * distance, let it run. The expire code will check for close
1243 * cooperators and put the close queue at the front of the service
1244 * tree. If possible, merge the expiring queue with the new cfqq.
1246 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
, 0);
1248 if (!cfqq
->new_cfqq
)
1249 cfq_setup_merge(cfqq
, new_cfqq
);
1254 * No requests pending. If the active queue still has requests in
1255 * flight or is idling for a new request, allow either of these
1256 * conditions to happen (or time out) before selecting a new queue.
1258 if (timer_pending(&cfqd
->idle_slice_timer
) ||
1259 (cfqq
->dispatched
&& cfq_cfqq_idle_window(cfqq
))) {
1265 cfq_slice_expired(cfqd
, 0);
1267 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
1272 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
1276 while (cfqq
->next_rq
) {
1277 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
1281 BUG_ON(!list_empty(&cfqq
->fifo
));
1286 * Drain our current requests. Used for barriers and when switching
1287 * io schedulers on-the-fly.
1289 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
1291 struct cfq_queue
*cfqq
;
1294 while ((cfqq
= cfq_rb_first(&cfqd
->service_tree
)) != NULL
)
1295 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1297 cfq_slice_expired(cfqd
, 0);
1299 BUG_ON(cfqd
->busy_queues
);
1301 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
1305 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1307 unsigned int max_dispatch
;
1310 * Drain async requests before we start sync IO
1312 if (cfq_cfqq_idle_window(cfqq
) && cfqd
->rq_in_driver
[BLK_RW_ASYNC
])
1316 * If this is an async queue and we have sync IO in flight, let it wait
1318 if (cfqd
->sync_flight
&& !cfq_cfqq_sync(cfqq
))
1321 max_dispatch
= cfqd
->cfq_quantum
;
1322 if (cfq_class_idle(cfqq
))
1326 * Does this cfqq already have too much IO in flight?
1328 if (cfqq
->dispatched
>= max_dispatch
) {
1330 * idle queue must always only have a single IO in flight
1332 if (cfq_class_idle(cfqq
))
1336 * We have other queues, don't allow more IO from this one
1338 if (cfqd
->busy_queues
> 1)
1342 * Sole queue user, allow bigger slice
1348 * Async queues must wait a bit before being allowed dispatch.
1349 * We also ramp up the dispatch depth gradually for async IO,
1350 * based on the last sync IO we serviced
1352 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
1353 unsigned long last_sync
= jiffies
- cfqd
->last_end_sync_rq
;
1356 depth
= last_sync
/ cfqd
->cfq_slice
[1];
1357 if (!depth
&& !cfqq
->dispatched
)
1359 if (depth
< max_dispatch
)
1360 max_dispatch
= depth
;
1364 * If we're below the current max, allow a dispatch
1366 return cfqq
->dispatched
< max_dispatch
;
1370 * Dispatch a request from cfqq, moving them to the request queue
1373 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1377 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
1379 if (!cfq_may_dispatch(cfqd
, cfqq
))
1383 * follow expired path, else get first next available
1385 rq
= cfq_check_fifo(cfqq
);
1390 * insert request into driver dispatch list
1392 cfq_dispatch_insert(cfqd
->queue
, rq
);
1394 if (!cfqd
->active_cic
) {
1395 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1397 atomic_long_inc(&cic
->ioc
->refcount
);
1398 cfqd
->active_cic
= cic
;
1405 * Find the cfqq that we need to service and move a request from that to the
1408 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
1410 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1411 struct cfq_queue
*cfqq
;
1413 if (!cfqd
->busy_queues
)
1416 if (unlikely(force
))
1417 return cfq_forced_dispatch(cfqd
);
1419 cfqq
= cfq_select_queue(cfqd
);
1424 * Dispatch a request from this cfqq, if it is allowed
1426 if (!cfq_dispatch_request(cfqd
, cfqq
))
1429 cfqq
->slice_dispatch
++;
1430 cfq_clear_cfqq_must_dispatch(cfqq
);
1433 * expire an async queue immediately if it has used up its slice. idle
1434 * queue always expire after 1 dispatch round.
1436 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
1437 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
1438 cfq_class_idle(cfqq
))) {
1439 cfqq
->slice_end
= jiffies
+ 1;
1440 cfq_slice_expired(cfqd
, 0);
1443 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
1448 * task holds one reference to the queue, dropped when task exits. each rq
1449 * in-flight on this queue also holds a reference, dropped when rq is freed.
1451 * queue lock must be held here.
1453 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1455 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1457 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1459 if (!atomic_dec_and_test(&cfqq
->ref
))
1462 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
1463 BUG_ON(rb_first(&cfqq
->sort_list
));
1464 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1465 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1467 if (unlikely(cfqd
->active_queue
== cfqq
)) {
1468 __cfq_slice_expired(cfqd
, cfqq
, 0);
1469 cfq_schedule_dispatch(cfqd
);
1472 kmem_cache_free(cfq_pool
, cfqq
);
1476 * Must always be called with the rcu_read_lock() held
1479 __call_for_each_cic(struct io_context
*ioc
,
1480 void (*func
)(struct io_context
*, struct cfq_io_context
*))
1482 struct cfq_io_context
*cic
;
1483 struct hlist_node
*n
;
1485 hlist_for_each_entry_rcu(cic
, n
, &ioc
->cic_list
, cic_list
)
1490 * Call func for each cic attached to this ioc.
1493 call_for_each_cic(struct io_context
*ioc
,
1494 void (*func
)(struct io_context
*, struct cfq_io_context
*))
1497 __call_for_each_cic(ioc
, func
);
1501 static void cfq_cic_free_rcu(struct rcu_head
*head
)
1503 struct cfq_io_context
*cic
;
1505 cic
= container_of(head
, struct cfq_io_context
, rcu_head
);
1507 kmem_cache_free(cfq_ioc_pool
, cic
);
1508 elv_ioc_count_dec(cfq_ioc_count
);
1512 * CFQ scheduler is exiting, grab exit lock and check
1513 * the pending io context count. If it hits zero,
1514 * complete ioc_gone and set it back to NULL
1516 spin_lock(&ioc_gone_lock
);
1517 if (ioc_gone
&& !elv_ioc_count_read(cfq_ioc_count
)) {
1521 spin_unlock(&ioc_gone_lock
);
1525 static void cfq_cic_free(struct cfq_io_context
*cic
)
1527 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
1530 static void cic_free_func(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1532 unsigned long flags
;
1534 BUG_ON(!cic
->dead_key
);
1536 spin_lock_irqsave(&ioc
->lock
, flags
);
1537 radix_tree_delete(&ioc
->radix_root
, cic
->dead_key
);
1538 hlist_del_rcu(&cic
->cic_list
);
1539 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1545 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1546 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1547 * and ->trim() which is called with the task lock held
1549 static void cfq_free_io_context(struct io_context
*ioc
)
1552 * ioc->refcount is zero here, or we are called from elv_unregister(),
1553 * so no more cic's are allowed to be linked into this ioc. So it
1554 * should be ok to iterate over the known list, we will see all cic's
1555 * since no new ones are added.
1557 __call_for_each_cic(ioc
, cic_free_func
);
1560 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1562 struct cfq_queue
*__cfqq
, *next
;
1564 if (unlikely(cfqq
== cfqd
->active_queue
)) {
1565 __cfq_slice_expired(cfqd
, cfqq
, 0);
1566 cfq_schedule_dispatch(cfqd
);
1570 * If this queue was scheduled to merge with another queue, be
1571 * sure to drop the reference taken on that queue (and others in
1572 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
1574 __cfqq
= cfqq
->new_cfqq
;
1576 if (__cfqq
== cfqq
) {
1577 WARN(1, "cfqq->new_cfqq loop detected\n");
1580 next
= __cfqq
->new_cfqq
;
1581 cfq_put_queue(__cfqq
);
1585 cfq_put_queue(cfqq
);
1588 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1589 struct cfq_io_context
*cic
)
1591 struct io_context
*ioc
= cic
->ioc
;
1593 list_del_init(&cic
->queue_list
);
1596 * Make sure key == NULL is seen for dead queues
1599 cic
->dead_key
= (unsigned long) cic
->key
;
1602 if (ioc
->ioc_data
== cic
)
1603 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
1605 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
1606 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
1607 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
1610 if (cic
->cfqq
[BLK_RW_SYNC
]) {
1611 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
1612 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
1616 static void cfq_exit_single_io_context(struct io_context
*ioc
,
1617 struct cfq_io_context
*cic
)
1619 struct cfq_data
*cfqd
= cic
->key
;
1622 struct request_queue
*q
= cfqd
->queue
;
1623 unsigned long flags
;
1625 spin_lock_irqsave(q
->queue_lock
, flags
);
1628 * Ensure we get a fresh copy of the ->key to prevent
1629 * race between exiting task and queue
1631 smp_read_barrier_depends();
1633 __cfq_exit_single_io_context(cfqd
, cic
);
1635 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1640 * The process that ioc belongs to has exited, we need to clean up
1641 * and put the internal structures we have that belongs to that process.
1643 static void cfq_exit_io_context(struct io_context
*ioc
)
1645 call_for_each_cic(ioc
, cfq_exit_single_io_context
);
1648 static struct cfq_io_context
*
1649 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1651 struct cfq_io_context
*cic
;
1653 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
1656 cic
->last_end_request
= jiffies
;
1657 INIT_LIST_HEAD(&cic
->queue_list
);
1658 INIT_HLIST_NODE(&cic
->cic_list
);
1659 cic
->dtor
= cfq_free_io_context
;
1660 cic
->exit
= cfq_exit_io_context
;
1661 elv_ioc_count_inc(cfq_ioc_count
);
1667 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
1669 struct task_struct
*tsk
= current
;
1672 if (!cfq_cfqq_prio_changed(cfqq
))
1675 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
1676 switch (ioprio_class
) {
1678 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1679 case IOPRIO_CLASS_NONE
:
1681 * no prio set, inherit CPU scheduling settings
1683 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1684 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
1686 case IOPRIO_CLASS_RT
:
1687 cfqq
->ioprio
= task_ioprio(ioc
);
1688 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1690 case IOPRIO_CLASS_BE
:
1691 cfqq
->ioprio
= task_ioprio(ioc
);
1692 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1694 case IOPRIO_CLASS_IDLE
:
1695 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1697 cfq_clear_cfqq_idle_window(cfqq
);
1702 * keep track of original prio settings in case we have to temporarily
1703 * elevate the priority of this queue
1705 cfqq
->org_ioprio
= cfqq
->ioprio
;
1706 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1707 cfq_clear_cfqq_prio_changed(cfqq
);
1710 static void changed_ioprio(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1712 struct cfq_data
*cfqd
= cic
->key
;
1713 struct cfq_queue
*cfqq
;
1714 unsigned long flags
;
1716 if (unlikely(!cfqd
))
1719 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1721 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
1723 struct cfq_queue
*new_cfqq
;
1724 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->ioc
,
1727 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
1728 cfq_put_queue(cfqq
);
1732 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
1734 cfq_mark_cfqq_prio_changed(cfqq
);
1736 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1739 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1741 call_for_each_cic(ioc
, changed_ioprio
);
1742 ioc
->ioprio_changed
= 0;
1745 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1746 pid_t pid
, bool is_sync
)
1748 RB_CLEAR_NODE(&cfqq
->rb_node
);
1749 RB_CLEAR_NODE(&cfqq
->p_node
);
1750 INIT_LIST_HEAD(&cfqq
->fifo
);
1752 atomic_set(&cfqq
->ref
, 0);
1755 cfq_mark_cfqq_prio_changed(cfqq
);
1758 if (!cfq_class_idle(cfqq
))
1759 cfq_mark_cfqq_idle_window(cfqq
);
1760 cfq_mark_cfqq_sync(cfqq
);
1765 static struct cfq_queue
*
1766 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
1767 struct io_context
*ioc
, gfp_t gfp_mask
)
1769 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1770 struct cfq_io_context
*cic
;
1773 cic
= cfq_cic_lookup(cfqd
, ioc
);
1774 /* cic always exists here */
1775 cfqq
= cic_to_cfqq(cic
, is_sync
);
1778 * Always try a new alloc if we fell back to the OOM cfqq
1779 * originally, since it should just be a temporary situation.
1781 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
1786 } else if (gfp_mask
& __GFP_WAIT
) {
1787 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1788 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
1789 gfp_mask
| __GFP_ZERO
,
1791 spin_lock_irq(cfqd
->queue
->queue_lock
);
1795 cfqq
= kmem_cache_alloc_node(cfq_pool
,
1796 gfp_mask
| __GFP_ZERO
,
1801 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
1802 cfq_init_prio_data(cfqq
, ioc
);
1803 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
1805 cfqq
= &cfqd
->oom_cfqq
;
1809 kmem_cache_free(cfq_pool
, new_cfqq
);
1814 static struct cfq_queue
**
1815 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
1817 switch (ioprio_class
) {
1818 case IOPRIO_CLASS_RT
:
1819 return &cfqd
->async_cfqq
[0][ioprio
];
1820 case IOPRIO_CLASS_BE
:
1821 return &cfqd
->async_cfqq
[1][ioprio
];
1822 case IOPRIO_CLASS_IDLE
:
1823 return &cfqd
->async_idle_cfqq
;
1829 static struct cfq_queue
*
1830 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
1833 const int ioprio
= task_ioprio(ioc
);
1834 const int ioprio_class
= task_ioprio_class(ioc
);
1835 struct cfq_queue
**async_cfqq
= NULL
;
1836 struct cfq_queue
*cfqq
= NULL
;
1839 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
1844 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
1847 * pin the queue now that it's allocated, scheduler exit will prune it
1849 if (!is_sync
&& !(*async_cfqq
)) {
1850 atomic_inc(&cfqq
->ref
);
1854 atomic_inc(&cfqq
->ref
);
1859 * We drop cfq io contexts lazily, so we may find a dead one.
1862 cfq_drop_dead_cic(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1863 struct cfq_io_context
*cic
)
1865 unsigned long flags
;
1867 WARN_ON(!list_empty(&cic
->queue_list
));
1869 spin_lock_irqsave(&ioc
->lock
, flags
);
1871 BUG_ON(ioc
->ioc_data
== cic
);
1873 radix_tree_delete(&ioc
->radix_root
, (unsigned long) cfqd
);
1874 hlist_del_rcu(&cic
->cic_list
);
1875 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1880 static struct cfq_io_context
*
1881 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1883 struct cfq_io_context
*cic
;
1884 unsigned long flags
;
1893 * we maintain a last-hit cache, to avoid browsing over the tree
1895 cic
= rcu_dereference(ioc
->ioc_data
);
1896 if (cic
&& cic
->key
== cfqd
) {
1902 cic
= radix_tree_lookup(&ioc
->radix_root
, (unsigned long) cfqd
);
1906 /* ->key must be copied to avoid race with cfq_exit_queue() */
1909 cfq_drop_dead_cic(cfqd
, ioc
, cic
);
1914 spin_lock_irqsave(&ioc
->lock
, flags
);
1915 rcu_assign_pointer(ioc
->ioc_data
, cic
);
1916 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1924 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1925 * the process specific cfq io context when entered from the block layer.
1926 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1928 static int cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1929 struct cfq_io_context
*cic
, gfp_t gfp_mask
)
1931 unsigned long flags
;
1934 ret
= radix_tree_preload(gfp_mask
);
1939 spin_lock_irqsave(&ioc
->lock
, flags
);
1940 ret
= radix_tree_insert(&ioc
->radix_root
,
1941 (unsigned long) cfqd
, cic
);
1943 hlist_add_head_rcu(&cic
->cic_list
, &ioc
->cic_list
);
1944 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1946 radix_tree_preload_end();
1949 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1950 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1951 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1956 printk(KERN_ERR
"cfq: cic link failed!\n");
1962 * Setup general io context and cfq io context. There can be several cfq
1963 * io contexts per general io context, if this process is doing io to more
1964 * than one device managed by cfq.
1966 static struct cfq_io_context
*
1967 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1969 struct io_context
*ioc
= NULL
;
1970 struct cfq_io_context
*cic
;
1972 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1974 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
1978 cic
= cfq_cic_lookup(cfqd
, ioc
);
1982 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1986 if (cfq_cic_link(cfqd
, ioc
, cic
, gfp_mask
))
1990 smp_read_barrier_depends();
1991 if (unlikely(ioc
->ioprio_changed
))
1992 cfq_ioc_set_ioprio(ioc
);
1998 put_io_context(ioc
);
2003 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
2005 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
2006 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
2008 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
2009 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
2010 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
2014 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2020 if (!cfqq
->last_request_pos
)
2022 else if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
2023 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
2025 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
2028 * Don't allow the seek distance to get too large from the
2029 * odd fragment, pagein, etc
2031 if (cfqq
->seek_samples
<= 60) /* second&third seek */
2032 sdist
= min(sdist
, (cfqq
->seek_mean
* 4) + 2*1024*1024);
2034 sdist
= min(sdist
, (cfqq
->seek_mean
* 4) + 2*1024*64);
2036 cfqq
->seek_samples
= (7*cfqq
->seek_samples
+ 256) / 8;
2037 cfqq
->seek_total
= (7*cfqq
->seek_total
+ (u64
)256*sdist
) / 8;
2038 total
= cfqq
->seek_total
+ (cfqq
->seek_samples
/2);
2039 do_div(total
, cfqq
->seek_samples
);
2040 cfqq
->seek_mean
= (sector_t
)total
;
2044 * Disable idle window if the process thinks too long or seeks so much that
2048 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2049 struct cfq_io_context
*cic
)
2051 int old_idle
, enable_idle
;
2054 * Don't idle for async or idle io prio class
2056 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
2059 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
2061 if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
2062 (!cfqd
->cfq_latency
&& cfqd
->hw_tag
&& CFQQ_SEEKY(cfqq
)))
2064 else if (sample_valid(cic
->ttime_samples
)) {
2065 unsigned int slice_idle
= cfqd
->cfq_slice_idle
;
2066 if (sample_valid(cfqq
->seek_samples
) && CFQQ_SEEKY(cfqq
))
2067 slice_idle
= msecs_to_jiffies(CFQ_MIN_TT
);
2068 if (cic
->ttime_mean
> slice_idle
)
2074 if (old_idle
!= enable_idle
) {
2075 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
2077 cfq_mark_cfqq_idle_window(cfqq
);
2079 cfq_clear_cfqq_idle_window(cfqq
);
2084 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2085 * no or if we aren't sure, a 1 will cause a preempt.
2088 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
2091 struct cfq_queue
*cfqq
;
2093 cfqq
= cfqd
->active_queue
;
2097 if (cfq_slice_used(cfqq
))
2100 if (cfq_class_idle(new_cfqq
))
2103 if (cfq_class_idle(cfqq
))
2107 * if the new request is sync, but the currently running queue is
2108 * not, let the sync request have priority.
2110 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
2114 * So both queues are sync. Let the new request get disk time if
2115 * it's a metadata request and the current queue is doing regular IO.
2117 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
2121 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2123 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
2126 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
2130 * if this request is as-good as one we would expect from the
2131 * current cfqq, let it preempt
2133 if (cfq_rq_close(cfqd
, cfqq
, rq
))
2140 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2141 * let it have half of its nominal slice.
2143 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2145 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
2146 cfq_slice_expired(cfqd
, 1);
2149 * Put the new queue at the front of the of the current list,
2150 * so we know that it will be selected next.
2152 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
2154 cfq_service_tree_add(cfqd
, cfqq
, 1);
2156 cfqq
->slice_end
= 0;
2157 cfq_mark_cfqq_slice_new(cfqq
);
2161 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2162 * something we should do about it
2165 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2168 struct cfq_io_context
*cic
= RQ_CIC(rq
);
2172 cfqq
->meta_pending
++;
2174 cfq_update_io_thinktime(cfqd
, cic
);
2175 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
2176 cfq_update_idle_window(cfqd
, cfqq
, cic
);
2178 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2180 if (cfqq
== cfqd
->active_queue
) {
2182 * Remember that we saw a request from this process, but
2183 * don't start queuing just yet. Otherwise we risk seeing lots
2184 * of tiny requests, because we disrupt the normal plugging
2185 * and merging. If the request is already larger than a single
2186 * page, let it rip immediately. For that case we assume that
2187 * merging is already done. Ditto for a busy system that
2188 * has other work pending, don't risk delaying until the
2189 * idle timer unplug to continue working.
2191 if (cfq_cfqq_wait_request(cfqq
)) {
2192 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
2193 cfqd
->busy_queues
> 1) {
2194 del_timer(&cfqd
->idle_slice_timer
);
2195 __blk_run_queue(cfqd
->queue
);
2197 cfq_mark_cfqq_must_dispatch(cfqq
);
2199 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
2201 * not the active queue - expire current slice if it is
2202 * idle and has expired it's mean thinktime or this new queue
2203 * has some old slice time left and is of higher priority or
2204 * this new queue is RT and the current one is BE
2206 cfq_preempt_queue(cfqd
, cfqq
);
2207 __blk_run_queue(cfqd
->queue
);
2211 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
2213 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2214 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2216 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
2217 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
2221 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
2222 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
2224 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
2228 * Update hw_tag based on peak queue depth over 50 samples under
2231 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
2233 if (rq_in_driver(cfqd
) > cfqd
->rq_in_driver_peak
)
2234 cfqd
->rq_in_driver_peak
= rq_in_driver(cfqd
);
2236 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
2237 rq_in_driver(cfqd
) <= CFQ_HW_QUEUE_MIN
)
2240 if (cfqd
->hw_tag_samples
++ < 50)
2243 if (cfqd
->rq_in_driver_peak
>= CFQ_HW_QUEUE_MIN
)
2248 cfqd
->hw_tag_samples
= 0;
2249 cfqd
->rq_in_driver_peak
= 0;
2252 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
2254 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2255 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2256 const int sync
= rq_is_sync(rq
);
2260 cfq_log_cfqq(cfqd
, cfqq
, "complete");
2262 cfq_update_hw_tag(cfqd
);
2264 WARN_ON(!cfqd
->rq_in_driver
[sync
]);
2265 WARN_ON(!cfqq
->dispatched
);
2266 cfqd
->rq_in_driver
[sync
]--;
2269 if (cfq_cfqq_sync(cfqq
))
2270 cfqd
->sync_flight
--;
2273 RQ_CIC(rq
)->last_end_request
= now
;
2274 cfqd
->last_end_sync_rq
= now
;
2278 * If this is the active queue, check if it needs to be expired,
2279 * or if we want to idle in case it has no pending requests.
2281 if (cfqd
->active_queue
== cfqq
) {
2282 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
2284 if (cfq_cfqq_slice_new(cfqq
)) {
2285 cfq_set_prio_slice(cfqd
, cfqq
);
2286 cfq_clear_cfqq_slice_new(cfqq
);
2289 * If there are no requests waiting in this queue, and
2290 * there are other queues ready to issue requests, AND
2291 * those other queues are issuing requests within our
2292 * mean seek distance, give them a chance to run instead
2295 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
2296 cfq_slice_expired(cfqd
, 1);
2297 else if (cfqq_empty
&& !cfq_close_cooperator(cfqd
, cfqq
, 1) &&
2298 sync
&& !rq_noidle(rq
))
2299 cfq_arm_slice_timer(cfqd
);
2302 if (!rq_in_driver(cfqd
))
2303 cfq_schedule_dispatch(cfqd
);
2307 * we temporarily boost lower priority queues if they are holding fs exclusive
2308 * resources. they are boosted to normal prio (CLASS_BE/4)
2310 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
2312 if (has_fs_excl()) {
2314 * boost idle prio on transactions that would lock out other
2315 * users of the filesystem
2317 if (cfq_class_idle(cfqq
))
2318 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2319 if (cfqq
->ioprio
> IOPRIO_NORM
)
2320 cfqq
->ioprio
= IOPRIO_NORM
;
2323 * check if we need to unboost the queue
2325 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
2326 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
2327 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
2328 cfqq
->ioprio
= cfqq
->org_ioprio
;
2332 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
2334 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
2335 cfq_mark_cfqq_must_alloc_slice(cfqq
);
2336 return ELV_MQUEUE_MUST
;
2339 return ELV_MQUEUE_MAY
;
2342 static int cfq_may_queue(struct request_queue
*q
, int rw
)
2344 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2345 struct task_struct
*tsk
= current
;
2346 struct cfq_io_context
*cic
;
2347 struct cfq_queue
*cfqq
;
2350 * don't force setup of a queue from here, as a call to may_queue
2351 * does not necessarily imply that a request actually will be queued.
2352 * so just lookup a possibly existing queue, or return 'may queue'
2355 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
2357 return ELV_MQUEUE_MAY
;
2359 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
2361 cfq_init_prio_data(cfqq
, cic
->ioc
);
2362 cfq_prio_boost(cfqq
);
2364 return __cfq_may_queue(cfqq
);
2367 return ELV_MQUEUE_MAY
;
2371 * queue lock held here
2373 static void cfq_put_request(struct request
*rq
)
2375 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2378 const int rw
= rq_data_dir(rq
);
2380 BUG_ON(!cfqq
->allocated
[rw
]);
2381 cfqq
->allocated
[rw
]--;
2383 put_io_context(RQ_CIC(rq
)->ioc
);
2385 rq
->elevator_private
= NULL
;
2386 rq
->elevator_private2
= NULL
;
2388 cfq_put_queue(cfqq
);
2392 static struct cfq_queue
*
2393 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
2394 struct cfq_queue
*cfqq
)
2396 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
2397 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
2398 cfq_put_queue(cfqq
);
2399 return cic_to_cfqq(cic
, 1);
2403 * Allocate cfq data structures associated with this request.
2406 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
2408 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2409 struct cfq_io_context
*cic
;
2410 const int rw
= rq_data_dir(rq
);
2411 const bool is_sync
= rq_is_sync(rq
);
2412 struct cfq_queue
*cfqq
;
2413 unsigned long flags
;
2415 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2417 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
2419 spin_lock_irqsave(q
->queue_lock
, flags
);
2424 cfqq
= cic_to_cfqq(cic
, is_sync
);
2425 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2426 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
2427 cic_set_cfqq(cic
, cfqq
, is_sync
);
2430 * Check to see if this queue is scheduled to merge with
2431 * another, closely cooperating queue. The merging of
2432 * queues happens here as it must be done in process context.
2433 * The reference on new_cfqq was taken in merge_cfqqs.
2436 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
2439 cfqq
->allocated
[rw
]++;
2440 atomic_inc(&cfqq
->ref
);
2442 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2444 rq
->elevator_private
= cic
;
2445 rq
->elevator_private2
= cfqq
;
2450 put_io_context(cic
->ioc
);
2452 cfq_schedule_dispatch(cfqd
);
2453 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2454 cfq_log(cfqd
, "set_request fail");
2458 static void cfq_kick_queue(struct work_struct
*work
)
2460 struct cfq_data
*cfqd
=
2461 container_of(work
, struct cfq_data
, unplug_work
);
2462 struct request_queue
*q
= cfqd
->queue
;
2464 spin_lock_irq(q
->queue_lock
);
2465 __blk_run_queue(cfqd
->queue
);
2466 spin_unlock_irq(q
->queue_lock
);
2470 * Timer running if the active_queue is currently idling inside its time slice
2472 static void cfq_idle_slice_timer(unsigned long data
)
2474 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
2475 struct cfq_queue
*cfqq
;
2476 unsigned long flags
;
2479 cfq_log(cfqd
, "idle timer fired");
2481 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2483 cfqq
= cfqd
->active_queue
;
2488 * We saw a request before the queue expired, let it through
2490 if (cfq_cfqq_must_dispatch(cfqq
))
2496 if (cfq_slice_used(cfqq
))
2500 * only expire and reinvoke request handler, if there are
2501 * other queues with pending requests
2503 if (!cfqd
->busy_queues
)
2507 * not expired and it has a request pending, let it dispatch
2509 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2513 cfq_slice_expired(cfqd
, timed_out
);
2515 cfq_schedule_dispatch(cfqd
);
2517 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2520 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
2522 del_timer_sync(&cfqd
->idle_slice_timer
);
2523 cancel_work_sync(&cfqd
->unplug_work
);
2526 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
2530 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
2531 if (cfqd
->async_cfqq
[0][i
])
2532 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
2533 if (cfqd
->async_cfqq
[1][i
])
2534 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
2537 if (cfqd
->async_idle_cfqq
)
2538 cfq_put_queue(cfqd
->async_idle_cfqq
);
2541 static void cfq_exit_queue(struct elevator_queue
*e
)
2543 struct cfq_data
*cfqd
= e
->elevator_data
;
2544 struct request_queue
*q
= cfqd
->queue
;
2546 cfq_shutdown_timer_wq(cfqd
);
2548 spin_lock_irq(q
->queue_lock
);
2550 if (cfqd
->active_queue
)
2551 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
2553 while (!list_empty(&cfqd
->cic_list
)) {
2554 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
2555 struct cfq_io_context
,
2558 __cfq_exit_single_io_context(cfqd
, cic
);
2561 cfq_put_async_queues(cfqd
);
2563 spin_unlock_irq(q
->queue_lock
);
2565 cfq_shutdown_timer_wq(cfqd
);
2570 static void *cfq_init_queue(struct request_queue
*q
)
2572 struct cfq_data
*cfqd
;
2575 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
2579 cfqd
->service_tree
= CFQ_RB_ROOT
;
2582 * Not strictly needed (since RB_ROOT just clears the node and we
2583 * zeroed cfqd on alloc), but better be safe in case someone decides
2584 * to add magic to the rb code
2586 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
2587 cfqd
->prio_trees
[i
] = RB_ROOT
;
2590 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2591 * Grab a permanent reference to it, so that the normal code flow
2592 * will not attempt to free it.
2594 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
2595 atomic_inc(&cfqd
->oom_cfqq
.ref
);
2597 INIT_LIST_HEAD(&cfqd
->cic_list
);
2601 init_timer(&cfqd
->idle_slice_timer
);
2602 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2603 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2605 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
2607 cfqd
->cfq_quantum
= cfq_quantum
;
2608 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2609 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2610 cfqd
->cfq_back_max
= cfq_back_max
;
2611 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2612 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2613 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2614 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2615 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2616 cfqd
->cfq_latency
= 1;
2618 cfqd
->last_end_sync_rq
= jiffies
;
2622 static void cfq_slab_kill(void)
2625 * Caller already ensured that pending RCU callbacks are completed,
2626 * so we should have no busy allocations at this point.
2629 kmem_cache_destroy(cfq_pool
);
2631 kmem_cache_destroy(cfq_ioc_pool
);
2634 static int __init
cfq_slab_setup(void)
2636 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
2640 cfq_ioc_pool
= KMEM_CACHE(cfq_io_context
, 0);
2651 * sysfs parts below -->
2654 cfq_var_show(unsigned int var
, char *page
)
2656 return sprintf(page
, "%d\n", var
);
2660 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2662 char *p
= (char *) page
;
2664 *var
= simple_strtoul(p
, &p
, 10);
2668 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2669 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2671 struct cfq_data *cfqd = e->elevator_data; \
2672 unsigned int __data = __VAR; \
2674 __data = jiffies_to_msecs(__data); \
2675 return cfq_var_show(__data, (page)); \
2677 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2678 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2679 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2680 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2681 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2682 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2683 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2684 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2685 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2686 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
2687 #undef SHOW_FUNCTION
2689 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2690 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2692 struct cfq_data *cfqd = e->elevator_data; \
2693 unsigned int __data; \
2694 int ret = cfq_var_store(&__data, (page), count); \
2695 if (__data < (MIN)) \
2697 else if (__data > (MAX)) \
2700 *(__PTR) = msecs_to_jiffies(__data); \
2702 *(__PTR) = __data; \
2705 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2706 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
2708 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
2710 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2711 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
2713 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2714 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2715 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2716 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
2718 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
2719 #undef STORE_FUNCTION
2721 #define CFQ_ATTR(name) \
2722 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2724 static struct elv_fs_entry cfq_attrs
[] = {
2726 CFQ_ATTR(fifo_expire_sync
),
2727 CFQ_ATTR(fifo_expire_async
),
2728 CFQ_ATTR(back_seek_max
),
2729 CFQ_ATTR(back_seek_penalty
),
2730 CFQ_ATTR(slice_sync
),
2731 CFQ_ATTR(slice_async
),
2732 CFQ_ATTR(slice_async_rq
),
2733 CFQ_ATTR(slice_idle
),
2734 CFQ_ATTR(low_latency
),
2738 static struct elevator_type iosched_cfq
= {
2740 .elevator_merge_fn
= cfq_merge
,
2741 .elevator_merged_fn
= cfq_merged_request
,
2742 .elevator_merge_req_fn
= cfq_merged_requests
,
2743 .elevator_allow_merge_fn
= cfq_allow_merge
,
2744 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2745 .elevator_add_req_fn
= cfq_insert_request
,
2746 .elevator_activate_req_fn
= cfq_activate_request
,
2747 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2748 .elevator_queue_empty_fn
= cfq_queue_empty
,
2749 .elevator_completed_req_fn
= cfq_completed_request
,
2750 .elevator_former_req_fn
= elv_rb_former_request
,
2751 .elevator_latter_req_fn
= elv_rb_latter_request
,
2752 .elevator_set_req_fn
= cfq_set_request
,
2753 .elevator_put_req_fn
= cfq_put_request
,
2754 .elevator_may_queue_fn
= cfq_may_queue
,
2755 .elevator_init_fn
= cfq_init_queue
,
2756 .elevator_exit_fn
= cfq_exit_queue
,
2757 .trim
= cfq_free_io_context
,
2759 .elevator_attrs
= cfq_attrs
,
2760 .elevator_name
= "cfq",
2761 .elevator_owner
= THIS_MODULE
,
2764 static int __init
cfq_init(void)
2767 * could be 0 on HZ < 1000 setups
2769 if (!cfq_slice_async
)
2770 cfq_slice_async
= 1;
2771 if (!cfq_slice_idle
)
2774 if (cfq_slab_setup())
2777 elv_register(&iosched_cfq
);
2782 static void __exit
cfq_exit(void)
2784 DECLARE_COMPLETION_ONSTACK(all_gone
);
2785 elv_unregister(&iosched_cfq
);
2786 ioc_gone
= &all_gone
;
2787 /* ioc_gone's update must be visible before reading ioc_count */
2791 * this also protects us from entering cfq_slab_kill() with
2792 * pending RCU callbacks
2794 if (elv_ioc_count_read(cfq_ioc_count
))
2795 wait_for_completion(&all_gone
);
2799 module_init(cfq_init
);
2800 module_exit(cfq_exit
);
2802 MODULE_AUTHOR("Jens Axboe");
2803 MODULE_LICENSE("GPL");
2804 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");