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
;
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
;
176 unsigned int cfq_latency
;
178 struct list_head cic_list
;
181 * Fallback dummy cfqq for extreme OOM conditions
183 struct cfq_queue oom_cfqq
;
185 unsigned long last_end_sync_rq
;
188 enum cfqq_state_flags
{
189 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
190 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
191 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
192 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
193 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
194 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
195 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
196 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
197 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
198 CFQ_CFQQ_FLAG_coop
, /* has done a coop jump of the queue */
199 CFQ_CFQQ_FLAG_coop_preempt
, /* coop preempt */
202 #define CFQ_CFQQ_FNS(name) \
203 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
205 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
207 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
209 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
211 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
213 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
217 CFQ_CFQQ_FNS(wait_request
);
218 CFQ_CFQQ_FNS(must_dispatch
);
219 CFQ_CFQQ_FNS(must_alloc_slice
);
220 CFQ_CFQQ_FNS(fifo_expire
);
221 CFQ_CFQQ_FNS(idle_window
);
222 CFQ_CFQQ_FNS(prio_changed
);
223 CFQ_CFQQ_FNS(slice_new
);
226 CFQ_CFQQ_FNS(coop_preempt
);
229 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
230 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
231 #define cfq_log(cfqd, fmt, args...) \
232 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
234 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
235 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, bool,
236 struct io_context
*, gfp_t
);
237 static struct cfq_io_context
*cfq_cic_lookup(struct cfq_data
*,
238 struct io_context
*);
240 static inline int rq_in_driver(struct cfq_data
*cfqd
)
242 return cfqd
->rq_in_driver
[0] + cfqd
->rq_in_driver
[1];
245 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_context
*cic
,
248 return cic
->cfqq
[is_sync
];
251 static inline void cic_set_cfqq(struct cfq_io_context
*cic
,
252 struct cfq_queue
*cfqq
, bool is_sync
)
254 cic
->cfqq
[is_sync
] = cfqq
;
258 * We regard a request as SYNC, if it's either a read or has the SYNC bit
259 * set (in which case it could also be direct WRITE).
261 static inline bool cfq_bio_sync(struct bio
*bio
)
263 return bio_data_dir(bio
) == READ
|| bio_rw_flagged(bio
, BIO_RW_SYNCIO
);
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
, bool 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 bool 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
) {
510 * Get our rb key offset. Subtract any residual slice
511 * value carried from last service. A negative resid
512 * count indicates slice overrun, and this should position
513 * the next service time further away in the tree.
515 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
516 rb_key
-= cfqq
->slice_resid
;
517 cfqq
->slice_resid
= 0;
520 __cfqq
= cfq_rb_first(&cfqd
->service_tree
);
521 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
524 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
526 * same position, nothing more to do
528 if (rb_key
== cfqq
->rb_key
)
531 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
536 p
= &cfqd
->service_tree
.rb
.rb_node
;
541 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
544 * sort RT queues first, we always want to give
545 * preference to them. IDLE queues goes to the back.
546 * after that, sort on the next service time.
548 if (cfq_class_rt(cfqq
) > cfq_class_rt(__cfqq
))
550 else if (cfq_class_rt(cfqq
) < cfq_class_rt(__cfqq
))
552 else if (cfq_class_idle(cfqq
) < cfq_class_idle(__cfqq
))
554 else if (cfq_class_idle(cfqq
) > cfq_class_idle(__cfqq
))
556 else if (time_before(rb_key
, __cfqq
->rb_key
))
561 if (n
== &(*p
)->rb_right
)
568 cfqd
->service_tree
.left
= &cfqq
->rb_node
;
570 cfqq
->rb_key
= rb_key
;
571 rb_link_node(&cfqq
->rb_node
, parent
, p
);
572 rb_insert_color(&cfqq
->rb_node
, &cfqd
->service_tree
.rb
);
575 static struct cfq_queue
*
576 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
577 sector_t sector
, struct rb_node
**ret_parent
,
578 struct rb_node
***rb_link
)
580 struct rb_node
**p
, *parent
;
581 struct cfq_queue
*cfqq
= NULL
;
589 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
592 * Sort strictly based on sector. Smallest to the left,
593 * largest to the right.
595 if (sector
> blk_rq_pos(cfqq
->next_rq
))
597 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
605 *ret_parent
= parent
;
611 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
613 struct rb_node
**p
, *parent
;
614 struct cfq_queue
*__cfqq
;
617 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
621 if (cfq_class_idle(cfqq
))
626 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
627 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
628 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
630 rb_link_node(&cfqq
->p_node
, parent
, p
);
631 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
637 * Update cfqq's position in the service tree.
639 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
642 * Resorting requires the cfqq to be on the RR list already.
644 if (cfq_cfqq_on_rr(cfqq
)) {
645 cfq_service_tree_add(cfqd
, cfqq
, 0);
646 cfq_prio_tree_add(cfqd
, cfqq
);
651 * add to busy list of queues for service, trying to be fair in ordering
652 * the pending list according to last request service
654 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
656 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
657 BUG_ON(cfq_cfqq_on_rr(cfqq
));
658 cfq_mark_cfqq_on_rr(cfqq
);
661 cfq_resort_rr_list(cfqd
, cfqq
);
665 * Called when the cfqq no longer has requests pending, remove it from
668 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
670 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
671 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
672 cfq_clear_cfqq_on_rr(cfqq
);
674 if (!RB_EMPTY_NODE(&cfqq
->rb_node
))
675 cfq_rb_erase(&cfqq
->rb_node
, &cfqd
->service_tree
);
677 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
681 BUG_ON(!cfqd
->busy_queues
);
686 * rb tree support functions
688 static void cfq_del_rq_rb(struct request
*rq
)
690 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
691 struct cfq_data
*cfqd
= cfqq
->cfqd
;
692 const int sync
= rq_is_sync(rq
);
694 BUG_ON(!cfqq
->queued
[sync
]);
695 cfqq
->queued
[sync
]--;
697 elv_rb_del(&cfqq
->sort_list
, rq
);
699 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
700 cfq_del_cfqq_rr(cfqd
, cfqq
);
703 static void cfq_add_rq_rb(struct request
*rq
)
705 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
706 struct cfq_data
*cfqd
= cfqq
->cfqd
;
707 struct request
*__alias
, *prev
;
709 cfqq
->queued
[rq_is_sync(rq
)]++;
712 * looks a little odd, but the first insert might return an alias.
713 * if that happens, put the alias on the dispatch list
715 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
716 cfq_dispatch_insert(cfqd
->queue
, __alias
);
718 if (!cfq_cfqq_on_rr(cfqq
))
719 cfq_add_cfqq_rr(cfqd
, cfqq
);
722 * check if this request is a better next-serve candidate
724 prev
= cfqq
->next_rq
;
725 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
);
728 * adjust priority tree position, if ->next_rq changes
730 if (prev
!= cfqq
->next_rq
)
731 cfq_prio_tree_add(cfqd
, cfqq
);
733 BUG_ON(!cfqq
->next_rq
);
736 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
738 elv_rb_del(&cfqq
->sort_list
, rq
);
739 cfqq
->queued
[rq_is_sync(rq
)]--;
743 static struct request
*
744 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
746 struct task_struct
*tsk
= current
;
747 struct cfq_io_context
*cic
;
748 struct cfq_queue
*cfqq
;
750 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
754 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
756 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
758 return elv_rb_find(&cfqq
->sort_list
, sector
);
764 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
766 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
768 cfqd
->rq_in_driver
[rq_is_sync(rq
)]++;
769 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
772 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
775 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
777 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
778 const int sync
= rq_is_sync(rq
);
780 WARN_ON(!cfqd
->rq_in_driver
[sync
]);
781 cfqd
->rq_in_driver
[sync
]--;
782 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
786 static void cfq_remove_request(struct request
*rq
)
788 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
790 if (cfqq
->next_rq
== rq
)
791 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
793 list_del_init(&rq
->queuelist
);
796 cfqq
->cfqd
->rq_queued
--;
797 if (rq_is_meta(rq
)) {
798 WARN_ON(!cfqq
->meta_pending
);
799 cfqq
->meta_pending
--;
803 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
806 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
807 struct request
*__rq
;
809 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
810 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
812 return ELEVATOR_FRONT_MERGE
;
815 return ELEVATOR_NO_MERGE
;
818 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
821 if (type
== ELEVATOR_FRONT_MERGE
) {
822 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
824 cfq_reposition_rq_rb(cfqq
, req
);
829 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
830 struct request
*next
)
833 * reposition in fifo if next is older than rq
835 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
836 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
837 list_move(&rq
->queuelist
, &next
->queuelist
);
838 rq_set_fifo_time(rq
, rq_fifo_time(next
));
841 cfq_remove_request(next
);
844 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
847 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
848 struct cfq_io_context
*cic
;
849 struct cfq_queue
*cfqq
;
852 * Disallow merge of a sync bio into an async request.
854 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
858 * Lookup the cfqq that this bio will be queued with. Allow
859 * merge only if rq is queued there.
861 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
865 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
866 return cfqq
== RQ_CFQQ(rq
);
869 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
870 struct cfq_queue
*cfqq
)
873 cfq_log_cfqq(cfqd
, cfqq
, "set_active");
875 cfqq
->slice_dispatch
= 0;
877 cfq_clear_cfqq_wait_request(cfqq
);
878 cfq_clear_cfqq_must_dispatch(cfqq
);
879 cfq_clear_cfqq_must_alloc_slice(cfqq
);
880 cfq_clear_cfqq_fifo_expire(cfqq
);
881 cfq_mark_cfqq_slice_new(cfqq
);
883 del_timer(&cfqd
->idle_slice_timer
);
886 cfqd
->active_queue
= cfqq
;
890 * current cfqq expired its slice (or was too idle), select new one
893 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
896 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
898 if (cfq_cfqq_wait_request(cfqq
))
899 del_timer(&cfqd
->idle_slice_timer
);
901 cfq_clear_cfqq_wait_request(cfqq
);
904 * store what was left of this slice, if the queue idled/timed out
906 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
)) {
907 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
908 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
911 cfq_resort_rr_list(cfqd
, cfqq
);
913 if (cfqq
== cfqd
->active_queue
)
914 cfqd
->active_queue
= NULL
;
916 if (cfqd
->active_cic
) {
917 put_io_context(cfqd
->active_cic
->ioc
);
918 cfqd
->active_cic
= NULL
;
922 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
924 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
927 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
931 * Get next queue for service. Unless we have a queue preemption,
932 * we'll simply select the first cfqq in the service tree.
934 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
936 if (RB_EMPTY_ROOT(&cfqd
->service_tree
.rb
))
939 return cfq_rb_first(&cfqd
->service_tree
);
943 * Get and set a new active queue for service.
945 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
946 struct cfq_queue
*cfqq
)
949 cfqq
= cfq_get_next_queue(cfqd
);
950 if (cfqq
&& !cfq_cfqq_coop_preempt(cfqq
))
951 cfq_clear_cfqq_coop(cfqq
);
955 cfq_clear_cfqq_coop_preempt(cfqq
);
957 __cfq_set_active_queue(cfqd
, cfqq
);
961 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
964 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
965 return blk_rq_pos(rq
) - cfqd
->last_position
;
967 return cfqd
->last_position
- blk_rq_pos(rq
);
970 #define CIC_SEEK_THR 8 * 1024
971 #define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
973 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct request
*rq
)
975 struct cfq_io_context
*cic
= cfqd
->active_cic
;
976 sector_t sdist
= cic
->seek_mean
;
978 if (!sample_valid(cic
->seek_samples
))
979 sdist
= CIC_SEEK_THR
;
981 return cfq_dist_from_last(cfqd
, rq
) <= sdist
;
984 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
985 struct cfq_queue
*cur_cfqq
)
987 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
988 struct rb_node
*parent
, *node
;
989 struct cfq_queue
*__cfqq
;
990 sector_t sector
= cfqd
->last_position
;
992 if (RB_EMPTY_ROOT(root
))
996 * First, if we find a request starting at the end of the last
997 * request, choose it.
999 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1004 * If the exact sector wasn't found, the parent of the NULL leaf
1005 * will contain the closest sector.
1007 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1008 if (cfq_rq_close(cfqd
, __cfqq
->next_rq
))
1011 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1012 node
= rb_next(&__cfqq
->p_node
);
1014 node
= rb_prev(&__cfqq
->p_node
);
1018 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1019 if (cfq_rq_close(cfqd
, __cfqq
->next_rq
))
1027 * cur_cfqq - passed in so that we don't decide that the current queue is
1028 * closely cooperating with itself.
1030 * So, basically we're assuming that that cur_cfqq has dispatched at least
1031 * one request, and that cfqd->last_position reflects a position on the disk
1032 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1035 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1036 struct cfq_queue
*cur_cfqq
,
1039 struct cfq_queue
*cfqq
;
1042 * A valid cfq_io_context is necessary to compare requests against
1043 * the seek_mean of the current cfqq.
1045 if (!cfqd
->active_cic
)
1049 * We should notice if some of the queues are cooperating, eg
1050 * working closely on the same area of the disk. In that case,
1051 * we can group them together and don't waste time idling.
1053 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1057 if (cfq_cfqq_coop(cfqq
))
1061 cfq_mark_cfqq_coop(cfqq
);
1065 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1067 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1068 struct cfq_io_context
*cic
;
1072 * SSD device without seek penalty, disable idling. But only do so
1073 * for devices that support queuing, otherwise we still have a problem
1074 * with sync vs async workloads.
1076 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
1079 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
1080 WARN_ON(cfq_cfqq_slice_new(cfqq
));
1083 * idle is disabled, either manually or by past process history
1085 if (!cfqd
->cfq_slice_idle
|| !cfq_cfqq_idle_window(cfqq
))
1089 * still requests with the driver, don't idle
1091 if (rq_in_driver(cfqd
))
1095 * task has exited, don't wait
1097 cic
= cfqd
->active_cic
;
1098 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
1102 * If our average think time is larger than the remaining time
1103 * slice, then don't idle. This avoids overrunning the allotted
1106 if (sample_valid(cic
->ttime_samples
) &&
1107 (cfqq
->slice_end
- jiffies
< cic
->ttime_mean
))
1110 cfq_mark_cfqq_wait_request(cfqq
);
1113 * we don't want to idle for seeks, but we do want to allow
1114 * fair distribution of slice time for a process doing back-to-back
1115 * seeks. so allow a little bit of time for him to submit a new rq
1117 sl
= cfqd
->cfq_slice_idle
;
1118 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
1119 sl
= min(sl
, msecs_to_jiffies(CFQ_MIN_TT
));
1121 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
1122 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu", sl
);
1126 * Move request from internal lists to the request queue dispatch list.
1128 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
1130 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1131 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1133 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
1135 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
1136 cfq_remove_request(rq
);
1138 elv_dispatch_sort(q
, rq
);
1140 if (cfq_cfqq_sync(cfqq
))
1141 cfqd
->sync_flight
++;
1145 * return expired entry, or NULL to just start from scratch in rbtree
1147 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
1149 struct request
*rq
= NULL
;
1151 if (cfq_cfqq_fifo_expire(cfqq
))
1154 cfq_mark_cfqq_fifo_expire(cfqq
);
1156 if (list_empty(&cfqq
->fifo
))
1159 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
1160 if (time_before(jiffies
, rq_fifo_time(rq
)))
1163 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
1168 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1170 const int base_rq
= cfqd
->cfq_slice_async_rq
;
1172 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1174 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
1178 * Select a queue for service. If we have a current active queue,
1179 * check whether to continue servicing it, or retrieve and set a new one.
1181 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
1183 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1185 cfqq
= cfqd
->active_queue
;
1190 * The active queue has run out of time, expire it and select new.
1192 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
))
1196 * The active queue has requests and isn't expired, allow it to
1199 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
1203 * If another queue has a request waiting within our mean seek
1204 * distance, let it run. The expire code will check for close
1205 * cooperators and put the close queue at the front of the service
1208 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
, 0);
1213 * No requests pending. If the active queue still has requests in
1214 * flight or is idling for a new request, allow either of these
1215 * conditions to happen (or time out) before selecting a new queue.
1217 if (timer_pending(&cfqd
->idle_slice_timer
) ||
1218 (cfqq
->dispatched
&& cfq_cfqq_idle_window(cfqq
))) {
1224 cfq_slice_expired(cfqd
, 0);
1226 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
1231 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
1235 while (cfqq
->next_rq
) {
1236 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
1240 BUG_ON(!list_empty(&cfqq
->fifo
));
1245 * Drain our current requests. Used for barriers and when switching
1246 * io schedulers on-the-fly.
1248 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
1250 struct cfq_queue
*cfqq
;
1253 while ((cfqq
= cfq_rb_first(&cfqd
->service_tree
)) != NULL
)
1254 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1256 cfq_slice_expired(cfqd
, 0);
1258 BUG_ON(cfqd
->busy_queues
);
1260 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
1264 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1266 unsigned int max_dispatch
;
1269 * Drain async requests before we start sync IO
1271 if (cfq_cfqq_idle_window(cfqq
) && cfqd
->rq_in_driver
[BLK_RW_ASYNC
])
1275 * If this is an async queue and we have sync IO in flight, let it wait
1277 if (cfqd
->sync_flight
&& !cfq_cfqq_sync(cfqq
))
1280 max_dispatch
= cfqd
->cfq_quantum
;
1281 if (cfq_class_idle(cfqq
))
1285 * Does this cfqq already have too much IO in flight?
1287 if (cfqq
->dispatched
>= max_dispatch
) {
1289 * idle queue must always only have a single IO in flight
1291 if (cfq_class_idle(cfqq
))
1295 * We have other queues, don't allow more IO from this one
1297 if (cfqd
->busy_queues
> 1)
1301 * Sole queue user, allow bigger slice
1307 * Async queues must wait a bit before being allowed dispatch.
1308 * We also ramp up the dispatch depth gradually for async IO,
1309 * based on the last sync IO we serviced
1311 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
1312 unsigned long last_sync
= jiffies
- cfqd
->last_end_sync_rq
;
1315 depth
= last_sync
/ cfqd
->cfq_slice
[1];
1316 if (!depth
&& !cfqq
->dispatched
)
1318 if (depth
< max_dispatch
)
1319 max_dispatch
= depth
;
1323 * If we're below the current max, allow a dispatch
1325 return cfqq
->dispatched
< max_dispatch
;
1329 * Dispatch a request from cfqq, moving them to the request queue
1332 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1336 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
1338 if (!cfq_may_dispatch(cfqd
, cfqq
))
1342 * follow expired path, else get first next available
1344 rq
= cfq_check_fifo(cfqq
);
1349 * insert request into driver dispatch list
1351 cfq_dispatch_insert(cfqd
->queue
, rq
);
1353 if (!cfqd
->active_cic
) {
1354 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1356 atomic_long_inc(&cic
->ioc
->refcount
);
1357 cfqd
->active_cic
= cic
;
1364 * Find the cfqq that we need to service and move a request from that to the
1367 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
1369 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1370 struct cfq_queue
*cfqq
;
1372 if (!cfqd
->busy_queues
)
1375 if (unlikely(force
))
1376 return cfq_forced_dispatch(cfqd
);
1378 cfqq
= cfq_select_queue(cfqd
);
1383 * Dispatch a request from this cfqq, if it is allowed
1385 if (!cfq_dispatch_request(cfqd
, cfqq
))
1388 cfqq
->slice_dispatch
++;
1389 cfq_clear_cfqq_must_dispatch(cfqq
);
1392 * expire an async queue immediately if it has used up its slice. idle
1393 * queue always expire after 1 dispatch round.
1395 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
1396 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
1397 cfq_class_idle(cfqq
))) {
1398 cfqq
->slice_end
= jiffies
+ 1;
1399 cfq_slice_expired(cfqd
, 0);
1402 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
1407 * task holds one reference to the queue, dropped when task exits. each rq
1408 * in-flight on this queue also holds a reference, dropped when rq is freed.
1410 * queue lock must be held here.
1412 static void cfq_put_queue(struct cfq_queue
*cfqq
)
1414 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1416 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
1418 if (!atomic_dec_and_test(&cfqq
->ref
))
1421 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
1422 BUG_ON(rb_first(&cfqq
->sort_list
));
1423 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
1424 BUG_ON(cfq_cfqq_on_rr(cfqq
));
1426 if (unlikely(cfqd
->active_queue
== cfqq
)) {
1427 __cfq_slice_expired(cfqd
, cfqq
, 0);
1428 cfq_schedule_dispatch(cfqd
);
1431 kmem_cache_free(cfq_pool
, cfqq
);
1435 * Must always be called with the rcu_read_lock() held
1438 __call_for_each_cic(struct io_context
*ioc
,
1439 void (*func
)(struct io_context
*, struct cfq_io_context
*))
1441 struct cfq_io_context
*cic
;
1442 struct hlist_node
*n
;
1444 hlist_for_each_entry_rcu(cic
, n
, &ioc
->cic_list
, cic_list
)
1449 * Call func for each cic attached to this ioc.
1452 call_for_each_cic(struct io_context
*ioc
,
1453 void (*func
)(struct io_context
*, struct cfq_io_context
*))
1456 __call_for_each_cic(ioc
, func
);
1460 static void cfq_cic_free_rcu(struct rcu_head
*head
)
1462 struct cfq_io_context
*cic
;
1464 cic
= container_of(head
, struct cfq_io_context
, rcu_head
);
1466 kmem_cache_free(cfq_ioc_pool
, cic
);
1467 elv_ioc_count_dec(cfq_ioc_count
);
1471 * CFQ scheduler is exiting, grab exit lock and check
1472 * the pending io context count. If it hits zero,
1473 * complete ioc_gone and set it back to NULL
1475 spin_lock(&ioc_gone_lock
);
1476 if (ioc_gone
&& !elv_ioc_count_read(cfq_ioc_count
)) {
1480 spin_unlock(&ioc_gone_lock
);
1484 static void cfq_cic_free(struct cfq_io_context
*cic
)
1486 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
1489 static void cic_free_func(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1491 unsigned long flags
;
1493 BUG_ON(!cic
->dead_key
);
1495 spin_lock_irqsave(&ioc
->lock
, flags
);
1496 radix_tree_delete(&ioc
->radix_root
, cic
->dead_key
);
1497 hlist_del_rcu(&cic
->cic_list
);
1498 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1504 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1505 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1506 * and ->trim() which is called with the task lock held
1508 static void cfq_free_io_context(struct io_context
*ioc
)
1511 * ioc->refcount is zero here, or we are called from elv_unregister(),
1512 * so no more cic's are allowed to be linked into this ioc. So it
1513 * should be ok to iterate over the known list, we will see all cic's
1514 * since no new ones are added.
1516 __call_for_each_cic(ioc
, cic_free_func
);
1519 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1521 if (unlikely(cfqq
== cfqd
->active_queue
)) {
1522 __cfq_slice_expired(cfqd
, cfqq
, 0);
1523 cfq_schedule_dispatch(cfqd
);
1526 cfq_put_queue(cfqq
);
1529 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
1530 struct cfq_io_context
*cic
)
1532 struct io_context
*ioc
= cic
->ioc
;
1534 list_del_init(&cic
->queue_list
);
1537 * Make sure key == NULL is seen for dead queues
1540 cic
->dead_key
= (unsigned long) cic
->key
;
1544 if (rcu_dereference(ioc
->ioc_data
) == cic
) {
1546 spin_lock(&ioc
->lock
);
1547 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
1548 spin_unlock(&ioc
->lock
);
1552 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
1553 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
1554 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
1557 if (cic
->cfqq
[BLK_RW_SYNC
]) {
1558 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
1559 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
1563 static void cfq_exit_single_io_context(struct io_context
*ioc
,
1564 struct cfq_io_context
*cic
)
1566 struct cfq_data
*cfqd
= cic
->key
;
1569 struct request_queue
*q
= cfqd
->queue
;
1570 unsigned long flags
;
1572 spin_lock_irqsave(q
->queue_lock
, flags
);
1575 * Ensure we get a fresh copy of the ->key to prevent
1576 * race between exiting task and queue
1578 smp_read_barrier_depends();
1580 __cfq_exit_single_io_context(cfqd
, cic
);
1582 spin_unlock_irqrestore(q
->queue_lock
, flags
);
1587 * The process that ioc belongs to has exited, we need to clean up
1588 * and put the internal structures we have that belongs to that process.
1590 static void cfq_exit_io_context(struct io_context
*ioc
)
1592 call_for_each_cic(ioc
, cfq_exit_single_io_context
);
1595 static struct cfq_io_context
*
1596 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1598 struct cfq_io_context
*cic
;
1600 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
1603 cic
->last_end_request
= jiffies
;
1604 INIT_LIST_HEAD(&cic
->queue_list
);
1605 INIT_HLIST_NODE(&cic
->cic_list
);
1606 cic
->dtor
= cfq_free_io_context
;
1607 cic
->exit
= cfq_exit_io_context
;
1608 elv_ioc_count_inc(cfq_ioc_count
);
1614 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
1616 struct task_struct
*tsk
= current
;
1619 if (!cfq_cfqq_prio_changed(cfqq
))
1622 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
1623 switch (ioprio_class
) {
1625 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
1626 case IOPRIO_CLASS_NONE
:
1628 * no prio set, inherit CPU scheduling settings
1630 cfqq
->ioprio
= task_nice_ioprio(tsk
);
1631 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
1633 case IOPRIO_CLASS_RT
:
1634 cfqq
->ioprio
= task_ioprio(ioc
);
1635 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
1637 case IOPRIO_CLASS_BE
:
1638 cfqq
->ioprio
= task_ioprio(ioc
);
1639 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
1641 case IOPRIO_CLASS_IDLE
:
1642 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
1644 cfq_clear_cfqq_idle_window(cfqq
);
1649 * keep track of original prio settings in case we have to temporarily
1650 * elevate the priority of this queue
1652 cfqq
->org_ioprio
= cfqq
->ioprio
;
1653 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
1654 cfq_clear_cfqq_prio_changed(cfqq
);
1657 static void changed_ioprio(struct io_context
*ioc
, struct cfq_io_context
*cic
)
1659 struct cfq_data
*cfqd
= cic
->key
;
1660 struct cfq_queue
*cfqq
;
1661 unsigned long flags
;
1663 if (unlikely(!cfqd
))
1666 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1668 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
1670 struct cfq_queue
*new_cfqq
;
1671 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->ioc
,
1674 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
1675 cfq_put_queue(cfqq
);
1679 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
1681 cfq_mark_cfqq_prio_changed(cfqq
);
1683 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1686 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
1688 call_for_each_cic(ioc
, changed_ioprio
);
1689 ioc
->ioprio_changed
= 0;
1692 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1693 pid_t pid
, bool is_sync
)
1695 RB_CLEAR_NODE(&cfqq
->rb_node
);
1696 RB_CLEAR_NODE(&cfqq
->p_node
);
1697 INIT_LIST_HEAD(&cfqq
->fifo
);
1699 atomic_set(&cfqq
->ref
, 0);
1702 cfq_mark_cfqq_prio_changed(cfqq
);
1705 if (!cfq_class_idle(cfqq
))
1706 cfq_mark_cfqq_idle_window(cfqq
);
1707 cfq_mark_cfqq_sync(cfqq
);
1712 static struct cfq_queue
*
1713 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
1714 struct io_context
*ioc
, gfp_t gfp_mask
)
1716 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1717 struct cfq_io_context
*cic
;
1720 cic
= cfq_cic_lookup(cfqd
, ioc
);
1721 /* cic always exists here */
1722 cfqq
= cic_to_cfqq(cic
, is_sync
);
1725 * Always try a new alloc if we fell back to the OOM cfqq
1726 * originally, since it should just be a temporary situation.
1728 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
1733 } else if (gfp_mask
& __GFP_WAIT
) {
1734 spin_unlock_irq(cfqd
->queue
->queue_lock
);
1735 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
1736 gfp_mask
| __GFP_ZERO
,
1738 spin_lock_irq(cfqd
->queue
->queue_lock
);
1742 cfqq
= kmem_cache_alloc_node(cfq_pool
,
1743 gfp_mask
| __GFP_ZERO
,
1748 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
1749 cfq_init_prio_data(cfqq
, ioc
);
1750 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
1752 cfqq
= &cfqd
->oom_cfqq
;
1756 kmem_cache_free(cfq_pool
, new_cfqq
);
1761 static struct cfq_queue
**
1762 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
1764 switch (ioprio_class
) {
1765 case IOPRIO_CLASS_RT
:
1766 return &cfqd
->async_cfqq
[0][ioprio
];
1767 case IOPRIO_CLASS_BE
:
1768 return &cfqd
->async_cfqq
[1][ioprio
];
1769 case IOPRIO_CLASS_IDLE
:
1770 return &cfqd
->async_idle_cfqq
;
1776 static struct cfq_queue
*
1777 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
1780 const int ioprio
= task_ioprio(ioc
);
1781 const int ioprio_class
= task_ioprio_class(ioc
);
1782 struct cfq_queue
**async_cfqq
= NULL
;
1783 struct cfq_queue
*cfqq
= NULL
;
1786 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
1791 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
1794 * pin the queue now that it's allocated, scheduler exit will prune it
1796 if (!is_sync
&& !(*async_cfqq
)) {
1797 atomic_inc(&cfqq
->ref
);
1801 atomic_inc(&cfqq
->ref
);
1806 * We drop cfq io contexts lazily, so we may find a dead one.
1809 cfq_drop_dead_cic(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1810 struct cfq_io_context
*cic
)
1812 unsigned long flags
;
1814 WARN_ON(!list_empty(&cic
->queue_list
));
1816 spin_lock_irqsave(&ioc
->lock
, flags
);
1818 BUG_ON(ioc
->ioc_data
== cic
);
1820 radix_tree_delete(&ioc
->radix_root
, (unsigned long) cfqd
);
1821 hlist_del_rcu(&cic
->cic_list
);
1822 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1827 static struct cfq_io_context
*
1828 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
1830 struct cfq_io_context
*cic
;
1831 unsigned long flags
;
1840 * we maintain a last-hit cache, to avoid browsing over the tree
1842 cic
= rcu_dereference(ioc
->ioc_data
);
1843 if (cic
&& cic
->key
== cfqd
) {
1849 cic
= radix_tree_lookup(&ioc
->radix_root
, (unsigned long) cfqd
);
1853 /* ->key must be copied to avoid race with cfq_exit_queue() */
1856 cfq_drop_dead_cic(cfqd
, ioc
, cic
);
1861 spin_lock_irqsave(&ioc
->lock
, flags
);
1862 rcu_assign_pointer(ioc
->ioc_data
, cic
);
1863 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1871 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1872 * the process specific cfq io context when entered from the block layer.
1873 * Also adds the cic to a per-cfqd list, used when this queue is removed.
1875 static int cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
1876 struct cfq_io_context
*cic
, gfp_t gfp_mask
)
1878 unsigned long flags
;
1881 ret
= radix_tree_preload(gfp_mask
);
1886 spin_lock_irqsave(&ioc
->lock
, flags
);
1887 ret
= radix_tree_insert(&ioc
->radix_root
,
1888 (unsigned long) cfqd
, cic
);
1890 hlist_add_head_rcu(&cic
->cic_list
, &ioc
->cic_list
);
1891 spin_unlock_irqrestore(&ioc
->lock
, flags
);
1893 radix_tree_preload_end();
1896 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
1897 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
1898 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
1903 printk(KERN_ERR
"cfq: cic link failed!\n");
1909 * Setup general io context and cfq io context. There can be several cfq
1910 * io contexts per general io context, if this process is doing io to more
1911 * than one device managed by cfq.
1913 static struct cfq_io_context
*
1914 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
1916 struct io_context
*ioc
= NULL
;
1917 struct cfq_io_context
*cic
;
1919 might_sleep_if(gfp_mask
& __GFP_WAIT
);
1921 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
1925 cic
= cfq_cic_lookup(cfqd
, ioc
);
1929 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
1933 if (cfq_cic_link(cfqd
, ioc
, cic
, gfp_mask
))
1937 smp_read_barrier_depends();
1938 if (unlikely(ioc
->ioprio_changed
))
1939 cfq_ioc_set_ioprio(ioc
);
1945 put_io_context(ioc
);
1950 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
1952 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
1953 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
1955 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
1956 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
1957 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
1961 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
1967 if (!cic
->last_request_pos
)
1969 else if (cic
->last_request_pos
< blk_rq_pos(rq
))
1970 sdist
= blk_rq_pos(rq
) - cic
->last_request_pos
;
1972 sdist
= cic
->last_request_pos
- blk_rq_pos(rq
);
1975 * Don't allow the seek distance to get too large from the
1976 * odd fragment, pagein, etc
1978 if (cic
->seek_samples
<= 60) /* second&third seek */
1979 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*1024);
1981 sdist
= min(sdist
, (cic
->seek_mean
* 4) + 2*1024*64);
1983 cic
->seek_samples
= (7*cic
->seek_samples
+ 256) / 8;
1984 cic
->seek_total
= (7*cic
->seek_total
+ (u64
)256*sdist
) / 8;
1985 total
= cic
->seek_total
+ (cic
->seek_samples
/2);
1986 do_div(total
, cic
->seek_samples
);
1987 cic
->seek_mean
= (sector_t
)total
;
1991 * Disable idle window if the process thinks too long or seeks so much that
1995 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1996 struct cfq_io_context
*cic
)
1998 int old_idle
, enable_idle
;
2001 * Don't idle for async or idle io prio class
2003 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
2006 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
2008 if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
2009 (!cfqd
->cfq_latency
&& cfqd
->hw_tag
&& CIC_SEEKY(cic
)))
2011 else if (sample_valid(cic
->ttime_samples
)) {
2012 unsigned int slice_idle
= cfqd
->cfq_slice_idle
;
2013 if (sample_valid(cic
->seek_samples
) && CIC_SEEKY(cic
))
2014 slice_idle
= msecs_to_jiffies(CFQ_MIN_TT
);
2015 if (cic
->ttime_mean
> slice_idle
)
2021 if (old_idle
!= enable_idle
) {
2022 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
2024 cfq_mark_cfqq_idle_window(cfqq
);
2026 cfq_clear_cfqq_idle_window(cfqq
);
2031 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2032 * no or if we aren't sure, a 1 will cause a preempt.
2035 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
2038 struct cfq_queue
*cfqq
;
2040 cfqq
= cfqd
->active_queue
;
2044 if (cfq_slice_used(cfqq
))
2047 if (cfq_class_idle(new_cfqq
))
2050 if (cfq_class_idle(cfqq
))
2054 * if the new request is sync, but the currently running queue is
2055 * not, let the sync request have priority.
2057 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
2061 * So both queues are sync. Let the new request get disk time if
2062 * it's a metadata request and the current queue is doing regular IO.
2064 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
2068 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2070 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
2073 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
2077 * if this request is as-good as one we would expect from the
2078 * current cfqq, let it preempt
2080 if (cfq_rq_close(cfqd
, rq
) && (!cfq_cfqq_coop(new_cfqq
) ||
2081 cfqd
->busy_queues
== 1)) {
2083 * Mark new queue coop_preempt, so its coop flag will not be
2084 * cleared when new queue gets scheduled at the very first time
2086 cfq_mark_cfqq_coop_preempt(new_cfqq
);
2087 cfq_mark_cfqq_coop(new_cfqq
);
2095 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2096 * let it have half of its nominal slice.
2098 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2100 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
2101 cfq_slice_expired(cfqd
, 1);
2104 * Put the new queue at the front of the of the current list,
2105 * so we know that it will be selected next.
2107 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
2109 cfq_service_tree_add(cfqd
, cfqq
, 1);
2111 cfqq
->slice_end
= 0;
2112 cfq_mark_cfqq_slice_new(cfqq
);
2116 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2117 * something we should do about it
2120 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2123 struct cfq_io_context
*cic
= RQ_CIC(rq
);
2127 cfqq
->meta_pending
++;
2129 cfq_update_io_thinktime(cfqd
, cic
);
2130 cfq_update_io_seektime(cfqd
, cic
, rq
);
2131 cfq_update_idle_window(cfqd
, cfqq
, cic
);
2133 cic
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2135 if (cfqq
== cfqd
->active_queue
) {
2137 * Remember that we saw a request from this process, but
2138 * don't start queuing just yet. Otherwise we risk seeing lots
2139 * of tiny requests, because we disrupt the normal plugging
2140 * and merging. If the request is already larger than a single
2141 * page, let it rip immediately. For that case we assume that
2142 * merging is already done. Ditto for a busy system that
2143 * has other work pending, don't risk delaying until the
2144 * idle timer unplug to continue working.
2146 if (cfq_cfqq_wait_request(cfqq
)) {
2147 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
2148 cfqd
->busy_queues
> 1) {
2149 del_timer(&cfqd
->idle_slice_timer
);
2150 __blk_run_queue(cfqd
->queue
);
2152 cfq_mark_cfqq_must_dispatch(cfqq
);
2154 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
2156 * not the active queue - expire current slice if it is
2157 * idle and has expired it's mean thinktime or this new queue
2158 * has some old slice time left and is of higher priority or
2159 * this new queue is RT and the current one is BE
2161 cfq_preempt_queue(cfqd
, cfqq
);
2162 __blk_run_queue(cfqd
->queue
);
2166 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
2168 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2169 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2171 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
2172 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
2176 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
2177 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
2179 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
2183 * Update hw_tag based on peak queue depth over 50 samples under
2186 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
2188 if (rq_in_driver(cfqd
) > cfqd
->rq_in_driver_peak
)
2189 cfqd
->rq_in_driver_peak
= rq_in_driver(cfqd
);
2191 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
2192 rq_in_driver(cfqd
) <= CFQ_HW_QUEUE_MIN
)
2195 if (cfqd
->hw_tag_samples
++ < 50)
2198 if (cfqd
->rq_in_driver_peak
>= CFQ_HW_QUEUE_MIN
)
2203 cfqd
->hw_tag_samples
= 0;
2204 cfqd
->rq_in_driver_peak
= 0;
2207 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
2209 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2210 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2211 const int sync
= rq_is_sync(rq
);
2215 cfq_log_cfqq(cfqd
, cfqq
, "complete");
2217 cfq_update_hw_tag(cfqd
);
2219 WARN_ON(!cfqd
->rq_in_driver
[sync
]);
2220 WARN_ON(!cfqq
->dispatched
);
2221 cfqd
->rq_in_driver
[sync
]--;
2224 if (cfq_cfqq_sync(cfqq
))
2225 cfqd
->sync_flight
--;
2228 RQ_CIC(rq
)->last_end_request
= now
;
2229 cfqd
->last_end_sync_rq
= now
;
2233 * If this is the active queue, check if it needs to be expired,
2234 * or if we want to idle in case it has no pending requests.
2236 if (cfqd
->active_queue
== cfqq
) {
2237 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
2239 if (cfq_cfqq_slice_new(cfqq
)) {
2240 cfq_set_prio_slice(cfqd
, cfqq
);
2241 cfq_clear_cfqq_slice_new(cfqq
);
2244 * If there are no requests waiting in this queue, and
2245 * there are other queues ready to issue requests, AND
2246 * those other queues are issuing requests within our
2247 * mean seek distance, give them a chance to run instead
2250 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
2251 cfq_slice_expired(cfqd
, 1);
2252 else if (cfqq_empty
&& !cfq_close_cooperator(cfqd
, cfqq
, 1) &&
2253 sync
&& !rq_noidle(rq
))
2254 cfq_arm_slice_timer(cfqd
);
2257 if (!rq_in_driver(cfqd
))
2258 cfq_schedule_dispatch(cfqd
);
2262 * we temporarily boost lower priority queues if they are holding fs exclusive
2263 * resources. they are boosted to normal prio (CLASS_BE/4)
2265 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
2267 if (has_fs_excl()) {
2269 * boost idle prio on transactions that would lock out other
2270 * users of the filesystem
2272 if (cfq_class_idle(cfqq
))
2273 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2274 if (cfqq
->ioprio
> IOPRIO_NORM
)
2275 cfqq
->ioprio
= IOPRIO_NORM
;
2278 * check if we need to unboost the queue
2280 if (cfqq
->ioprio_class
!= cfqq
->org_ioprio_class
)
2281 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
2282 if (cfqq
->ioprio
!= cfqq
->org_ioprio
)
2283 cfqq
->ioprio
= cfqq
->org_ioprio
;
2287 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
2289 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
2290 cfq_mark_cfqq_must_alloc_slice(cfqq
);
2291 return ELV_MQUEUE_MUST
;
2294 return ELV_MQUEUE_MAY
;
2297 static int cfq_may_queue(struct request_queue
*q
, int rw
)
2299 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2300 struct task_struct
*tsk
= current
;
2301 struct cfq_io_context
*cic
;
2302 struct cfq_queue
*cfqq
;
2305 * don't force setup of a queue from here, as a call to may_queue
2306 * does not necessarily imply that a request actually will be queued.
2307 * so just lookup a possibly existing queue, or return 'may queue'
2310 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
2312 return ELV_MQUEUE_MAY
;
2314 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
2316 cfq_init_prio_data(cfqq
, cic
->ioc
);
2317 cfq_prio_boost(cfqq
);
2319 return __cfq_may_queue(cfqq
);
2322 return ELV_MQUEUE_MAY
;
2326 * queue lock held here
2328 static void cfq_put_request(struct request
*rq
)
2330 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2333 const int rw
= rq_data_dir(rq
);
2335 BUG_ON(!cfqq
->allocated
[rw
]);
2336 cfqq
->allocated
[rw
]--;
2338 put_io_context(RQ_CIC(rq
)->ioc
);
2340 rq
->elevator_private
= NULL
;
2341 rq
->elevator_private2
= NULL
;
2343 cfq_put_queue(cfqq
);
2348 * Allocate cfq data structures associated with this request.
2351 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
2353 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2354 struct cfq_io_context
*cic
;
2355 const int rw
= rq_data_dir(rq
);
2356 const bool is_sync
= rq_is_sync(rq
);
2357 struct cfq_queue
*cfqq
;
2358 unsigned long flags
;
2360 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2362 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
2364 spin_lock_irqsave(q
->queue_lock
, flags
);
2369 cfqq
= cic_to_cfqq(cic
, is_sync
);
2370 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2371 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
2372 cic_set_cfqq(cic
, cfqq
, is_sync
);
2375 cfqq
->allocated
[rw
]++;
2376 atomic_inc(&cfqq
->ref
);
2378 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2380 rq
->elevator_private
= cic
;
2381 rq
->elevator_private2
= cfqq
;
2386 put_io_context(cic
->ioc
);
2388 cfq_schedule_dispatch(cfqd
);
2389 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2390 cfq_log(cfqd
, "set_request fail");
2394 static void cfq_kick_queue(struct work_struct
*work
)
2396 struct cfq_data
*cfqd
=
2397 container_of(work
, struct cfq_data
, unplug_work
);
2398 struct request_queue
*q
= cfqd
->queue
;
2400 spin_lock_irq(q
->queue_lock
);
2401 __blk_run_queue(cfqd
->queue
);
2402 spin_unlock_irq(q
->queue_lock
);
2406 * Timer running if the active_queue is currently idling inside its time slice
2408 static void cfq_idle_slice_timer(unsigned long data
)
2410 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
2411 struct cfq_queue
*cfqq
;
2412 unsigned long flags
;
2415 cfq_log(cfqd
, "idle timer fired");
2417 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2419 cfqq
= cfqd
->active_queue
;
2424 * We saw a request before the queue expired, let it through
2426 if (cfq_cfqq_must_dispatch(cfqq
))
2432 if (cfq_slice_used(cfqq
))
2436 * only expire and reinvoke request handler, if there are
2437 * other queues with pending requests
2439 if (!cfqd
->busy_queues
)
2443 * not expired and it has a request pending, let it dispatch
2445 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
2449 cfq_slice_expired(cfqd
, timed_out
);
2451 cfq_schedule_dispatch(cfqd
);
2453 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2456 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
2458 del_timer_sync(&cfqd
->idle_slice_timer
);
2459 cancel_work_sync(&cfqd
->unplug_work
);
2462 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
2466 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
2467 if (cfqd
->async_cfqq
[0][i
])
2468 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
2469 if (cfqd
->async_cfqq
[1][i
])
2470 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
2473 if (cfqd
->async_idle_cfqq
)
2474 cfq_put_queue(cfqd
->async_idle_cfqq
);
2477 static void cfq_exit_queue(struct elevator_queue
*e
)
2479 struct cfq_data
*cfqd
= e
->elevator_data
;
2480 struct request_queue
*q
= cfqd
->queue
;
2482 cfq_shutdown_timer_wq(cfqd
);
2484 spin_lock_irq(q
->queue_lock
);
2486 if (cfqd
->active_queue
)
2487 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
2489 while (!list_empty(&cfqd
->cic_list
)) {
2490 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
2491 struct cfq_io_context
,
2494 __cfq_exit_single_io_context(cfqd
, cic
);
2497 cfq_put_async_queues(cfqd
);
2499 spin_unlock_irq(q
->queue_lock
);
2501 cfq_shutdown_timer_wq(cfqd
);
2506 static void *cfq_init_queue(struct request_queue
*q
)
2508 struct cfq_data
*cfqd
;
2511 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
2515 cfqd
->service_tree
= CFQ_RB_ROOT
;
2518 * Not strictly needed (since RB_ROOT just clears the node and we
2519 * zeroed cfqd on alloc), but better be safe in case someone decides
2520 * to add magic to the rb code
2522 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
2523 cfqd
->prio_trees
[i
] = RB_ROOT
;
2526 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2527 * Grab a permanent reference to it, so that the normal code flow
2528 * will not attempt to free it.
2530 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
2531 atomic_inc(&cfqd
->oom_cfqq
.ref
);
2533 INIT_LIST_HEAD(&cfqd
->cic_list
);
2537 init_timer(&cfqd
->idle_slice_timer
);
2538 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
2539 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
2541 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
2543 cfqd
->cfq_quantum
= cfq_quantum
;
2544 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
2545 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
2546 cfqd
->cfq_back_max
= cfq_back_max
;
2547 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
2548 cfqd
->cfq_slice
[0] = cfq_slice_async
;
2549 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
2550 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
2551 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
2552 cfqd
->cfq_latency
= 1;
2554 cfqd
->last_end_sync_rq
= jiffies
;
2558 static void cfq_slab_kill(void)
2561 * Caller already ensured that pending RCU callbacks are completed,
2562 * so we should have no busy allocations at this point.
2565 kmem_cache_destroy(cfq_pool
);
2567 kmem_cache_destroy(cfq_ioc_pool
);
2570 static int __init
cfq_slab_setup(void)
2572 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
2576 cfq_ioc_pool
= KMEM_CACHE(cfq_io_context
, 0);
2587 * sysfs parts below -->
2590 cfq_var_show(unsigned int var
, char *page
)
2592 return sprintf(page
, "%d\n", var
);
2596 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
2598 char *p
= (char *) page
;
2600 *var
= simple_strtoul(p
, &p
, 10);
2604 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
2605 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
2607 struct cfq_data *cfqd = e->elevator_data; \
2608 unsigned int __data = __VAR; \
2610 __data = jiffies_to_msecs(__data); \
2611 return cfq_var_show(__data, (page)); \
2613 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
2614 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
2615 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
2616 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
2617 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
2618 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
2619 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
2620 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
2621 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
2622 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
2623 #undef SHOW_FUNCTION
2625 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
2626 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2628 struct cfq_data *cfqd = e->elevator_data; \
2629 unsigned int __data; \
2630 int ret = cfq_var_store(&__data, (page), count); \
2631 if (__data < (MIN)) \
2633 else if (__data > (MAX)) \
2636 *(__PTR) = msecs_to_jiffies(__data); \
2638 *(__PTR) = __data; \
2641 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
2642 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
2644 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
2646 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
2647 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
2649 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
2650 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
2651 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
2652 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
2654 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
2655 #undef STORE_FUNCTION
2657 #define CFQ_ATTR(name) \
2658 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2660 static struct elv_fs_entry cfq_attrs
[] = {
2662 CFQ_ATTR(fifo_expire_sync
),
2663 CFQ_ATTR(fifo_expire_async
),
2664 CFQ_ATTR(back_seek_max
),
2665 CFQ_ATTR(back_seek_penalty
),
2666 CFQ_ATTR(slice_sync
),
2667 CFQ_ATTR(slice_async
),
2668 CFQ_ATTR(slice_async_rq
),
2669 CFQ_ATTR(slice_idle
),
2670 CFQ_ATTR(low_latency
),
2674 static struct elevator_type iosched_cfq
= {
2676 .elevator_merge_fn
= cfq_merge
,
2677 .elevator_merged_fn
= cfq_merged_request
,
2678 .elevator_merge_req_fn
= cfq_merged_requests
,
2679 .elevator_allow_merge_fn
= cfq_allow_merge
,
2680 .elevator_dispatch_fn
= cfq_dispatch_requests
,
2681 .elevator_add_req_fn
= cfq_insert_request
,
2682 .elevator_activate_req_fn
= cfq_activate_request
,
2683 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
2684 .elevator_queue_empty_fn
= cfq_queue_empty
,
2685 .elevator_completed_req_fn
= cfq_completed_request
,
2686 .elevator_former_req_fn
= elv_rb_former_request
,
2687 .elevator_latter_req_fn
= elv_rb_latter_request
,
2688 .elevator_set_req_fn
= cfq_set_request
,
2689 .elevator_put_req_fn
= cfq_put_request
,
2690 .elevator_may_queue_fn
= cfq_may_queue
,
2691 .elevator_init_fn
= cfq_init_queue
,
2692 .elevator_exit_fn
= cfq_exit_queue
,
2693 .trim
= cfq_free_io_context
,
2695 .elevator_attrs
= cfq_attrs
,
2696 .elevator_name
= "cfq",
2697 .elevator_owner
= THIS_MODULE
,
2700 static int __init
cfq_init(void)
2703 * could be 0 on HZ < 1000 setups
2705 if (!cfq_slice_async
)
2706 cfq_slice_async
= 1;
2707 if (!cfq_slice_idle
)
2710 if (cfq_slab_setup())
2713 elv_register(&iosched_cfq
);
2718 static void __exit
cfq_exit(void)
2720 DECLARE_COMPLETION_ONSTACK(all_gone
);
2721 elv_unregister(&iosched_cfq
);
2722 ioc_gone
= &all_gone
;
2723 /* ioc_gone's update must be visible before reading ioc_count */
2727 * this also protects us from entering cfq_slab_kill() with
2728 * pending RCU callbacks
2730 if (elv_ioc_count_read(cfq_ioc_count
))
2731 wait_for_completion(&all_gone
);
2735 module_init(cfq_init
);
2736 module_exit(cfq_exit
);
2738 MODULE_AUTHOR("Jens Axboe");
2739 MODULE_LICENSE("GPL");
2740 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");