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/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
16 #include "blk-cgroup.h"
21 /* max queue in one round of service */
22 static const int cfq_quantum
= 4;
23 static const int cfq_fifo_expire
[2] = { HZ
/ 4, HZ
/ 8 };
24 /* maximum backwards seek, in KiB */
25 static const int cfq_back_max
= 16 * 1024;
26 /* penalty of a backwards seek */
27 static const int cfq_back_penalty
= 2;
28 static const int cfq_slice_sync
= HZ
/ 10;
29 static int cfq_slice_async
= HZ
/ 25;
30 static const int cfq_slice_async_rq
= 2;
31 static int cfq_slice_idle
= HZ
/ 125;
32 static const int cfq_target_latency
= HZ
* 3/10; /* 300 ms */
33 static const int cfq_hist_divisor
= 4;
36 * offset from end of service tree
38 #define CFQ_IDLE_DELAY (HZ / 5)
41 * below this threshold, we consider thinktime immediate
43 #define CFQ_MIN_TT (2)
46 * Allow merged cfqqs to perform this amount of seeky I/O before
47 * deciding to break the queues up again.
49 #define CFQQ_COOP_TOUT (HZ)
51 #define CFQ_SLICE_SCALE (5)
52 #define CFQ_HW_QUEUE_MIN (5)
53 #define CFQ_SERVICE_SHIFT 12
56 ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
59 static struct kmem_cache
*cfq_pool
;
60 static struct kmem_cache
*cfq_ioc_pool
;
62 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count
);
63 static struct completion
*ioc_gone
;
64 static DEFINE_SPINLOCK(ioc_gone_lock
);
66 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
67 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
68 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
70 #define sample_valid(samples) ((samples) > 80)
71 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
74 * Most of our rbtree usage is for sorting with min extraction, so
75 * if we cache the leftmost node we don't have to walk down the tree
76 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
77 * move this into the elevator for the rq sorting as well.
84 struct rb_node
*active
;
85 unsigned total_weight
;
87 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
90 * Per process-grouping structure
95 /* various state flags, see below */
98 struct cfq_data
*cfqd
;
99 /* service_tree member */
100 struct rb_node rb_node
;
101 /* service_tree key */
102 unsigned long rb_key
;
103 /* prio tree member */
104 struct rb_node p_node
;
105 /* prio tree root we belong to, if any */
106 struct rb_root
*p_root
;
107 /* sorted list of pending requests */
108 struct rb_root sort_list
;
109 /* if fifo isn't expired, next request to serve */
110 struct request
*next_rq
;
111 /* requests queued in sort_list */
113 /* currently allocated requests */
115 /* fifo list of requests in sort_list */
116 struct list_head fifo
;
118 unsigned long slice_end
;
120 unsigned int slice_dispatch
;
122 /* pending metadata requests */
124 /* number of requests that are on the dispatch list or inside driver */
127 /* io prio of this group */
128 unsigned short ioprio
, org_ioprio
;
129 unsigned short ioprio_class
, org_ioprio_class
;
131 unsigned int seek_samples
;
134 sector_t last_request_pos
;
135 unsigned long seeky_start
;
139 struct cfq_rb_root
*service_tree
;
140 struct cfq_queue
*new_cfqq
;
141 struct cfq_group
*cfqg
;
145 * First index in the service_trees.
146 * IDLE is handled separately, so it has negative index
155 * Second index in the service_trees.
159 SYNC_NOIDLE_WORKLOAD
= 1,
163 /* This is per cgroup per device grouping structure */
165 /* group service_tree member */
166 struct rb_node rb_node
;
168 /* group service_tree key */
173 /* number of cfqq currently on this group */
176 /* Per group busy queus average. Useful for workload slice calc. */
177 unsigned int busy_queues_avg
[2];
179 * rr lists of queues with requests, onle rr for each priority class.
180 * Counts are embedded in the cfq_rb_root
182 struct cfq_rb_root service_trees
[2][3];
183 struct cfq_rb_root service_tree_idle
;
187 * Per block device queue structure
190 struct request_queue
*queue
;
191 /* Root service tree for cfq_groups */
192 struct cfq_rb_root grp_service_tree
;
193 struct cfq_group root_group
;
194 /* Number of active cfq groups on group service tree */
198 * The priority currently being served
200 enum wl_prio_t serving_prio
;
201 enum wl_type_t serving_type
;
202 unsigned long workload_expires
;
203 struct cfq_group
*serving_group
;
204 bool noidle_tree_requires_idle
;
207 * Each priority tree is sorted by next_request position. These
208 * trees are used when determining if two or more queues are
209 * interleaving requests (see cfq_close_cooperator).
211 struct rb_root prio_trees
[CFQ_PRIO_LISTS
];
213 unsigned int busy_queues
;
219 * queue-depth detection
225 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
226 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
229 int hw_tag_est_depth
;
230 unsigned int hw_tag_samples
;
233 * idle window management
235 struct timer_list idle_slice_timer
;
236 struct work_struct unplug_work
;
238 struct cfq_queue
*active_queue
;
239 struct cfq_io_context
*active_cic
;
242 * async queue for each priority case
244 struct cfq_queue
*async_cfqq
[2][IOPRIO_BE_NR
];
245 struct cfq_queue
*async_idle_cfqq
;
247 sector_t last_position
;
250 * tunables, see top of file
252 unsigned int cfq_quantum
;
253 unsigned int cfq_fifo_expire
[2];
254 unsigned int cfq_back_penalty
;
255 unsigned int cfq_back_max
;
256 unsigned int cfq_slice
[2];
257 unsigned int cfq_slice_async_rq
;
258 unsigned int cfq_slice_idle
;
259 unsigned int cfq_latency
;
261 struct list_head cic_list
;
264 * Fallback dummy cfqq for extreme OOM conditions
266 struct cfq_queue oom_cfqq
;
268 unsigned long last_end_sync_rq
;
271 static struct cfq_rb_root
*service_tree_for(struct cfq_group
*cfqg
,
274 struct cfq_data
*cfqd
)
279 if (prio
== IDLE_WORKLOAD
)
280 return &cfqg
->service_tree_idle
;
282 return &cfqg
->service_trees
[prio
][type
];
285 enum cfqq_state_flags
{
286 CFQ_CFQQ_FLAG_on_rr
= 0, /* on round-robin busy list */
287 CFQ_CFQQ_FLAG_wait_request
, /* waiting for a request */
288 CFQ_CFQQ_FLAG_must_dispatch
, /* must be allowed a dispatch */
289 CFQ_CFQQ_FLAG_must_alloc_slice
, /* per-slice must_alloc flag */
290 CFQ_CFQQ_FLAG_fifo_expire
, /* FIFO checked in this slice */
291 CFQ_CFQQ_FLAG_idle_window
, /* slice idling enabled */
292 CFQ_CFQQ_FLAG_prio_changed
, /* task priority has changed */
293 CFQ_CFQQ_FLAG_slice_new
, /* no requests dispatched in slice */
294 CFQ_CFQQ_FLAG_sync
, /* synchronous queue */
295 CFQ_CFQQ_FLAG_coop
, /* cfqq is shared */
296 CFQ_CFQQ_FLAG_deep
, /* sync cfqq experienced large depth */
299 #define CFQ_CFQQ_FNS(name) \
300 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
302 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
304 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
306 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
308 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
310 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
314 CFQ_CFQQ_FNS(wait_request
);
315 CFQ_CFQQ_FNS(must_dispatch
);
316 CFQ_CFQQ_FNS(must_alloc_slice
);
317 CFQ_CFQQ_FNS(fifo_expire
);
318 CFQ_CFQQ_FNS(idle_window
);
319 CFQ_CFQQ_FNS(prio_changed
);
320 CFQ_CFQQ_FNS(slice_new
);
326 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
327 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
328 #define cfq_log(cfqd, fmt, args...) \
329 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
331 /* Traverses through cfq group service trees */
332 #define for_each_cfqg_st(cfqg, i, j, st) \
333 for (i = 0; i <= IDLE_WORKLOAD; i++) \
334 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
335 : &cfqg->service_tree_idle; \
336 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
337 (i == IDLE_WORKLOAD && j == 0); \
338 j++, st = i < IDLE_WORKLOAD ? \
339 &cfqg->service_trees[i][j]: NULL) \
342 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
344 if (cfq_class_idle(cfqq
))
345 return IDLE_WORKLOAD
;
346 if (cfq_class_rt(cfqq
))
352 static enum wl_type_t
cfqq_type(struct cfq_queue
*cfqq
)
354 if (!cfq_cfqq_sync(cfqq
))
355 return ASYNC_WORKLOAD
;
356 if (!cfq_cfqq_idle_window(cfqq
))
357 return SYNC_NOIDLE_WORKLOAD
;
358 return SYNC_WORKLOAD
;
361 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl
,
362 struct cfq_data
*cfqd
,
363 struct cfq_group
*cfqg
)
365 if (wl
== IDLE_WORKLOAD
)
366 return cfqg
->service_tree_idle
.count
;
368 return cfqg
->service_trees
[wl
][ASYNC_WORKLOAD
].count
369 + cfqg
->service_trees
[wl
][SYNC_NOIDLE_WORKLOAD
].count
370 + cfqg
->service_trees
[wl
][SYNC_WORKLOAD
].count
;
373 static void cfq_dispatch_insert(struct request_queue
*, struct request
*);
374 static struct cfq_queue
*cfq_get_queue(struct cfq_data
*, bool,
375 struct io_context
*, gfp_t
);
376 static struct cfq_io_context
*cfq_cic_lookup(struct cfq_data
*,
377 struct io_context
*);
379 static inline int rq_in_driver(struct cfq_data
*cfqd
)
381 return cfqd
->rq_in_driver
[0] + cfqd
->rq_in_driver
[1];
384 static inline struct cfq_queue
*cic_to_cfqq(struct cfq_io_context
*cic
,
387 return cic
->cfqq
[is_sync
];
390 static inline void cic_set_cfqq(struct cfq_io_context
*cic
,
391 struct cfq_queue
*cfqq
, bool is_sync
)
393 cic
->cfqq
[is_sync
] = cfqq
;
397 * We regard a request as SYNC, if it's either a read or has the SYNC bit
398 * set (in which case it could also be direct WRITE).
400 static inline bool cfq_bio_sync(struct bio
*bio
)
402 return bio_data_dir(bio
) == READ
|| bio_rw_flagged(bio
, BIO_RW_SYNCIO
);
406 * scheduler run of queue, if there are requests pending and no one in the
407 * driver that will restart queueing
409 static inline void cfq_schedule_dispatch(struct cfq_data
*cfqd
)
411 if (cfqd
->busy_queues
) {
412 cfq_log(cfqd
, "schedule dispatch");
413 kblockd_schedule_work(cfqd
->queue
, &cfqd
->unplug_work
);
417 static int cfq_queue_empty(struct request_queue
*q
)
419 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
421 return !cfqd
->rq_queued
;
425 * Scale schedule slice based on io priority. Use the sync time slice only
426 * if a queue is marked sync and has sync io queued. A sync queue with async
427 * io only, should not get full sync slice length.
429 static inline int cfq_prio_slice(struct cfq_data
*cfqd
, bool sync
,
432 const int base_slice
= cfqd
->cfq_slice
[sync
];
434 WARN_ON(prio
>= IOPRIO_BE_NR
);
436 return base_slice
+ (base_slice
/CFQ_SLICE_SCALE
* (4 - prio
));
440 cfq_prio_to_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
442 return cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
);
445 static inline u64
cfq_scale_slice(unsigned long delta
, struct cfq_group
*cfqg
)
447 u64 d
= delta
<< CFQ_SERVICE_SHIFT
;
449 d
= d
* BLKIO_WEIGHT_DEFAULT
;
450 do_div(d
, cfqg
->weight
);
454 static inline u64
max_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
456 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
458 min_vdisktime
= vdisktime
;
460 return min_vdisktime
;
463 static inline u64
min_vdisktime(u64 min_vdisktime
, u64 vdisktime
)
465 s64 delta
= (s64
)(vdisktime
- min_vdisktime
);
467 min_vdisktime
= vdisktime
;
469 return min_vdisktime
;
472 static void update_min_vdisktime(struct cfq_rb_root
*st
)
474 u64 vdisktime
= st
->min_vdisktime
;
475 struct cfq_group
*cfqg
;
478 cfqg
= rb_entry_cfqg(st
->active
);
479 vdisktime
= cfqg
->vdisktime
;
483 cfqg
= rb_entry_cfqg(st
->left
);
484 vdisktime
= min_vdisktime(vdisktime
, cfqg
->vdisktime
);
487 st
->min_vdisktime
= max_vdisktime(st
->min_vdisktime
, vdisktime
);
491 * get averaged number of queues of RT/BE priority.
492 * average is updated, with a formula that gives more weight to higher numbers,
493 * to quickly follows sudden increases and decrease slowly
496 static inline unsigned cfq_group_get_avg_queues(struct cfq_data
*cfqd
,
497 struct cfq_group
*cfqg
, bool rt
)
499 unsigned min_q
, max_q
;
500 unsigned mult
= cfq_hist_divisor
- 1;
501 unsigned round
= cfq_hist_divisor
/ 2;
502 unsigned busy
= cfq_group_busy_queues_wl(rt
, cfqd
, cfqg
);
504 min_q
= min(cfqg
->busy_queues_avg
[rt
], busy
);
505 max_q
= max(cfqg
->busy_queues_avg
[rt
], busy
);
506 cfqg
->busy_queues_avg
[rt
] = (mult
* max_q
+ min_q
+ round
) /
508 return cfqg
->busy_queues_avg
[rt
];
511 static inline unsigned
512 cfq_group_slice(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
514 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
516 return cfq_target_latency
* cfqg
->weight
/ st
->total_weight
;
520 cfq_set_prio_slice(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
522 unsigned slice
= cfq_prio_to_slice(cfqd
, cfqq
);
523 if (cfqd
->cfq_latency
) {
525 * interested queues (we consider only the ones with the same
526 * priority class in the cfq group)
528 unsigned iq
= cfq_group_get_avg_queues(cfqd
, cfqq
->cfqg
,
530 unsigned sync_slice
= cfqd
->cfq_slice
[1];
531 unsigned expect_latency
= sync_slice
* iq
;
532 unsigned group_slice
= cfq_group_slice(cfqd
, cfqq
->cfqg
);
534 if (expect_latency
> group_slice
) {
535 unsigned base_low_slice
= 2 * cfqd
->cfq_slice_idle
;
536 /* scale low_slice according to IO priority
537 * and sync vs async */
539 min(slice
, base_low_slice
* slice
/ sync_slice
);
540 /* the adapted slice value is scaled to fit all iqs
541 * into the target latency */
542 slice
= max(slice
* group_slice
/ expect_latency
,
546 cfqq
->slice_end
= jiffies
+ slice
;
547 cfq_log_cfqq(cfqd
, cfqq
, "set_slice=%lu", cfqq
->slice_end
- jiffies
);
551 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
552 * isn't valid until the first request from the dispatch is activated
553 * and the slice time set.
555 static inline bool cfq_slice_used(struct cfq_queue
*cfqq
)
557 if (cfq_cfqq_slice_new(cfqq
))
559 if (time_before(jiffies
, cfqq
->slice_end
))
566 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
567 * We choose the request that is closest to the head right now. Distance
568 * behind the head is penalized and only allowed to a certain extent.
570 static struct request
*
571 cfq_choose_req(struct cfq_data
*cfqd
, struct request
*rq1
, struct request
*rq2
, sector_t last
)
573 sector_t s1
, s2
, d1
= 0, d2
= 0;
574 unsigned long back_max
;
575 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
576 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
577 unsigned wrap
= 0; /* bit mask: requests behind the disk head? */
579 if (rq1
== NULL
|| rq1
== rq2
)
584 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
586 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
588 if (rq_is_meta(rq1
) && !rq_is_meta(rq2
))
590 else if (rq_is_meta(rq2
) && !rq_is_meta(rq1
))
593 s1
= blk_rq_pos(rq1
);
594 s2
= blk_rq_pos(rq2
);
597 * by definition, 1KiB is 2 sectors
599 back_max
= cfqd
->cfq_back_max
* 2;
602 * Strict one way elevator _except_ in the case where we allow
603 * short backward seeks which are biased as twice the cost of a
604 * similar forward seek.
608 else if (s1
+ back_max
>= last
)
609 d1
= (last
- s1
) * cfqd
->cfq_back_penalty
;
611 wrap
|= CFQ_RQ1_WRAP
;
615 else if (s2
+ back_max
>= last
)
616 d2
= (last
- s2
) * cfqd
->cfq_back_penalty
;
618 wrap
|= CFQ_RQ2_WRAP
;
620 /* Found required data */
623 * By doing switch() on the bit mask "wrap" we avoid having to
624 * check two variables for all permutations: --> faster!
627 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
643 case (CFQ_RQ1_WRAP
|CFQ_RQ2_WRAP
): /* both rqs wrapped */
646 * Since both rqs are wrapped,
647 * start with the one that's further behind head
648 * (--> only *one* back seek required),
649 * since back seek takes more time than forward.
659 * The below is leftmost cache rbtree addon
661 static struct cfq_queue
*cfq_rb_first(struct cfq_rb_root
*root
)
663 /* Service tree is empty */
668 root
->left
= rb_first(&root
->rb
);
671 return rb_entry(root
->left
, struct cfq_queue
, rb_node
);
676 static struct cfq_group
*cfq_rb_first_group(struct cfq_rb_root
*root
)
679 root
->left
= rb_first(&root
->rb
);
682 return rb_entry_cfqg(root
->left
);
687 static void rb_erase_init(struct rb_node
*n
, struct rb_root
*root
)
693 static void cfq_rb_erase(struct rb_node
*n
, struct cfq_rb_root
*root
)
697 rb_erase_init(n
, &root
->rb
);
702 * would be nice to take fifo expire time into account as well
704 static struct request
*
705 cfq_find_next_rq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
706 struct request
*last
)
708 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
709 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
710 struct request
*next
= NULL
, *prev
= NULL
;
712 BUG_ON(RB_EMPTY_NODE(&last
->rb_node
));
715 prev
= rb_entry_rq(rbprev
);
718 next
= rb_entry_rq(rbnext
);
720 rbnext
= rb_first(&cfqq
->sort_list
);
721 if (rbnext
&& rbnext
!= &last
->rb_node
)
722 next
= rb_entry_rq(rbnext
);
725 return cfq_choose_req(cfqd
, next
, prev
, blk_rq_pos(last
));
728 static unsigned long cfq_slice_offset(struct cfq_data
*cfqd
,
729 struct cfq_queue
*cfqq
)
732 * just an approximation, should be ok.
734 return (cfqq
->cfqg
->nr_cfqq
- 1) * (cfq_prio_slice(cfqd
, 1, 0) -
735 cfq_prio_slice(cfqd
, cfq_cfqq_sync(cfqq
), cfqq
->ioprio
));
739 cfqg_key(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
741 return cfqg
->vdisktime
- st
->min_vdisktime
;
745 __cfq_group_service_tree_add(struct cfq_rb_root
*st
, struct cfq_group
*cfqg
)
747 struct rb_node
**node
= &st
->rb
.rb_node
;
748 struct rb_node
*parent
= NULL
;
749 struct cfq_group
*__cfqg
;
750 s64 key
= cfqg_key(st
, cfqg
);
753 while (*node
!= NULL
) {
755 __cfqg
= rb_entry_cfqg(parent
);
757 if (key
< cfqg_key(st
, __cfqg
))
758 node
= &parent
->rb_left
;
760 node
= &parent
->rb_right
;
766 st
->left
= &cfqg
->rb_node
;
768 rb_link_node(&cfqg
->rb_node
, parent
, node
);
769 rb_insert_color(&cfqg
->rb_node
, &st
->rb
);
773 cfq_group_service_tree_add(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
775 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
776 struct cfq_group
*__cfqg
;
784 * Currently put the group at the end. Later implement something
785 * so that groups get lesser vtime based on their weights, so that
786 * if group does not loose all if it was not continously backlogged.
788 n
= rb_last(&st
->rb
);
790 __cfqg
= rb_entry_cfqg(n
);
791 cfqg
->vdisktime
= __cfqg
->vdisktime
+ CFQ_IDLE_DELAY
;
793 cfqg
->vdisktime
= st
->min_vdisktime
;
795 __cfq_group_service_tree_add(st
, cfqg
);
798 st
->total_weight
+= cfqg
->weight
;
802 cfq_group_service_tree_del(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
804 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
806 if (st
->active
== &cfqg
->rb_node
)
809 BUG_ON(cfqg
->nr_cfqq
< 1);
812 /* If there are other cfq queues under this group, don't delete it */
818 st
->total_weight
-= cfqg
->weight
;
819 if (!RB_EMPTY_NODE(&cfqg
->rb_node
))
820 cfq_rb_erase(&cfqg
->rb_node
, st
);
824 * The cfqd->service_trees holds all pending cfq_queue's that have
825 * requests waiting to be processed. It is sorted in the order that
826 * we will service the queues.
828 static void cfq_service_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
831 struct rb_node
**p
, *parent
;
832 struct cfq_queue
*__cfqq
;
833 unsigned long rb_key
;
834 struct cfq_rb_root
*service_tree
;
837 service_tree
= service_tree_for(cfqq
->cfqg
, cfqq_prio(cfqq
),
838 cfqq_type(cfqq
), cfqd
);
839 if (cfq_class_idle(cfqq
)) {
840 rb_key
= CFQ_IDLE_DELAY
;
841 parent
= rb_last(&service_tree
->rb
);
842 if (parent
&& parent
!= &cfqq
->rb_node
) {
843 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
844 rb_key
+= __cfqq
->rb_key
;
847 } else if (!add_front
) {
849 * Get our rb key offset. Subtract any residual slice
850 * value carried from last service. A negative resid
851 * count indicates slice overrun, and this should position
852 * the next service time further away in the tree.
854 rb_key
= cfq_slice_offset(cfqd
, cfqq
) + jiffies
;
855 rb_key
-= cfqq
->slice_resid
;
856 cfqq
->slice_resid
= 0;
859 __cfqq
= cfq_rb_first(service_tree
);
860 rb_key
+= __cfqq
? __cfqq
->rb_key
: jiffies
;
863 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
865 * same position, nothing more to do
867 if (rb_key
== cfqq
->rb_key
&&
868 cfqq
->service_tree
== service_tree
)
871 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
872 cfqq
->service_tree
= NULL
;
877 cfqq
->service_tree
= service_tree
;
878 p
= &service_tree
->rb
.rb_node
;
883 __cfqq
= rb_entry(parent
, struct cfq_queue
, rb_node
);
886 * sort by key, that represents service time.
888 if (time_before(rb_key
, __cfqq
->rb_key
))
899 service_tree
->left
= &cfqq
->rb_node
;
901 cfqq
->rb_key
= rb_key
;
902 rb_link_node(&cfqq
->rb_node
, parent
, p
);
903 rb_insert_color(&cfqq
->rb_node
, &service_tree
->rb
);
904 service_tree
->count
++;
905 cfq_group_service_tree_add(cfqd
, cfqq
->cfqg
);
908 static struct cfq_queue
*
909 cfq_prio_tree_lookup(struct cfq_data
*cfqd
, struct rb_root
*root
,
910 sector_t sector
, struct rb_node
**ret_parent
,
911 struct rb_node
***rb_link
)
913 struct rb_node
**p
, *parent
;
914 struct cfq_queue
*cfqq
= NULL
;
922 cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
925 * Sort strictly based on sector. Smallest to the left,
926 * largest to the right.
928 if (sector
> blk_rq_pos(cfqq
->next_rq
))
930 else if (sector
< blk_rq_pos(cfqq
->next_rq
))
938 *ret_parent
= parent
;
944 static void cfq_prio_tree_add(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
946 struct rb_node
**p
, *parent
;
947 struct cfq_queue
*__cfqq
;
950 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
954 if (cfq_class_idle(cfqq
))
959 cfqq
->p_root
= &cfqd
->prio_trees
[cfqq
->org_ioprio
];
960 __cfqq
= cfq_prio_tree_lookup(cfqd
, cfqq
->p_root
,
961 blk_rq_pos(cfqq
->next_rq
), &parent
, &p
);
963 rb_link_node(&cfqq
->p_node
, parent
, p
);
964 rb_insert_color(&cfqq
->p_node
, cfqq
->p_root
);
970 * Update cfqq's position in the service tree.
972 static void cfq_resort_rr_list(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
975 * Resorting requires the cfqq to be on the RR list already.
977 if (cfq_cfqq_on_rr(cfqq
)) {
978 cfq_service_tree_add(cfqd
, cfqq
, 0);
979 cfq_prio_tree_add(cfqd
, cfqq
);
984 * add to busy list of queues for service, trying to be fair in ordering
985 * the pending list according to last request service
987 static void cfq_add_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
989 cfq_log_cfqq(cfqd
, cfqq
, "add_to_rr");
990 BUG_ON(cfq_cfqq_on_rr(cfqq
));
991 cfq_mark_cfqq_on_rr(cfqq
);
994 cfq_resort_rr_list(cfqd
, cfqq
);
998 * Called when the cfqq no longer has requests pending, remove it from
1001 static void cfq_del_cfqq_rr(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1003 cfq_log_cfqq(cfqd
, cfqq
, "del_from_rr");
1004 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
1005 cfq_clear_cfqq_on_rr(cfqq
);
1007 if (!RB_EMPTY_NODE(&cfqq
->rb_node
)) {
1008 cfq_rb_erase(&cfqq
->rb_node
, cfqq
->service_tree
);
1009 cfqq
->service_tree
= NULL
;
1012 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1013 cfqq
->p_root
= NULL
;
1016 cfq_group_service_tree_del(cfqd
, cfqq
->cfqg
);
1017 BUG_ON(!cfqd
->busy_queues
);
1018 cfqd
->busy_queues
--;
1022 * rb tree support functions
1024 static void cfq_del_rq_rb(struct request
*rq
)
1026 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1027 const int sync
= rq_is_sync(rq
);
1029 BUG_ON(!cfqq
->queued
[sync
]);
1030 cfqq
->queued
[sync
]--;
1032 elv_rb_del(&cfqq
->sort_list
, rq
);
1034 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
)) {
1036 * Queue will be deleted from service tree when we actually
1037 * expire it later. Right now just remove it from prio tree
1041 rb_erase(&cfqq
->p_node
, cfqq
->p_root
);
1042 cfqq
->p_root
= NULL
;
1047 static void cfq_add_rq_rb(struct request
*rq
)
1049 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1050 struct cfq_data
*cfqd
= cfqq
->cfqd
;
1051 struct request
*__alias
, *prev
;
1053 cfqq
->queued
[rq_is_sync(rq
)]++;
1056 * looks a little odd, but the first insert might return an alias.
1057 * if that happens, put the alias on the dispatch list
1059 while ((__alias
= elv_rb_add(&cfqq
->sort_list
, rq
)) != NULL
)
1060 cfq_dispatch_insert(cfqd
->queue
, __alias
);
1062 if (!cfq_cfqq_on_rr(cfqq
))
1063 cfq_add_cfqq_rr(cfqd
, cfqq
);
1066 * check if this request is a better next-serve candidate
1068 prev
= cfqq
->next_rq
;
1069 cfqq
->next_rq
= cfq_choose_req(cfqd
, cfqq
->next_rq
, rq
, cfqd
->last_position
);
1072 * adjust priority tree position, if ->next_rq changes
1074 if (prev
!= cfqq
->next_rq
)
1075 cfq_prio_tree_add(cfqd
, cfqq
);
1077 BUG_ON(!cfqq
->next_rq
);
1080 static void cfq_reposition_rq_rb(struct cfq_queue
*cfqq
, struct request
*rq
)
1082 elv_rb_del(&cfqq
->sort_list
, rq
);
1083 cfqq
->queued
[rq_is_sync(rq
)]--;
1087 static struct request
*
1088 cfq_find_rq_fmerge(struct cfq_data
*cfqd
, struct bio
*bio
)
1090 struct task_struct
*tsk
= current
;
1091 struct cfq_io_context
*cic
;
1092 struct cfq_queue
*cfqq
;
1094 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
1098 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1100 sector_t sector
= bio
->bi_sector
+ bio_sectors(bio
);
1102 return elv_rb_find(&cfqq
->sort_list
, sector
);
1108 static void cfq_activate_request(struct request_queue
*q
, struct request
*rq
)
1110 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1112 cfqd
->rq_in_driver
[rq_is_sync(rq
)]++;
1113 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "activate rq, drv=%d",
1114 rq_in_driver(cfqd
));
1116 cfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
1119 static void cfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
1121 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1122 const int sync
= rq_is_sync(rq
);
1124 WARN_ON(!cfqd
->rq_in_driver
[sync
]);
1125 cfqd
->rq_in_driver
[sync
]--;
1126 cfq_log_cfqq(cfqd
, RQ_CFQQ(rq
), "deactivate rq, drv=%d",
1127 rq_in_driver(cfqd
));
1130 static void cfq_remove_request(struct request
*rq
)
1132 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1134 if (cfqq
->next_rq
== rq
)
1135 cfqq
->next_rq
= cfq_find_next_rq(cfqq
->cfqd
, cfqq
, rq
);
1137 list_del_init(&rq
->queuelist
);
1140 cfqq
->cfqd
->rq_queued
--;
1141 if (rq_is_meta(rq
)) {
1142 WARN_ON(!cfqq
->meta_pending
);
1143 cfqq
->meta_pending
--;
1147 static int cfq_merge(struct request_queue
*q
, struct request
**req
,
1150 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1151 struct request
*__rq
;
1153 __rq
= cfq_find_rq_fmerge(cfqd
, bio
);
1154 if (__rq
&& elv_rq_merge_ok(__rq
, bio
)) {
1156 return ELEVATOR_FRONT_MERGE
;
1159 return ELEVATOR_NO_MERGE
;
1162 static void cfq_merged_request(struct request_queue
*q
, struct request
*req
,
1165 if (type
== ELEVATOR_FRONT_MERGE
) {
1166 struct cfq_queue
*cfqq
= RQ_CFQQ(req
);
1168 cfq_reposition_rq_rb(cfqq
, req
);
1173 cfq_merged_requests(struct request_queue
*q
, struct request
*rq
,
1174 struct request
*next
)
1176 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1178 * reposition in fifo if next is older than rq
1180 if (!list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
1181 time_before(rq_fifo_time(next
), rq_fifo_time(rq
))) {
1182 list_move(&rq
->queuelist
, &next
->queuelist
);
1183 rq_set_fifo_time(rq
, rq_fifo_time(next
));
1186 if (cfqq
->next_rq
== next
)
1188 cfq_remove_request(next
);
1191 static int cfq_allow_merge(struct request_queue
*q
, struct request
*rq
,
1194 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1195 struct cfq_io_context
*cic
;
1196 struct cfq_queue
*cfqq
;
1199 * Disallow merge of a sync bio into an async request.
1201 if (cfq_bio_sync(bio
) && !rq_is_sync(rq
))
1205 * Lookup the cfqq that this bio will be queued with. Allow
1206 * merge only if rq is queued there.
1208 cic
= cfq_cic_lookup(cfqd
, current
->io_context
);
1212 cfqq
= cic_to_cfqq(cic
, cfq_bio_sync(bio
));
1213 return cfqq
== RQ_CFQQ(rq
);
1216 static void __cfq_set_active_queue(struct cfq_data
*cfqd
,
1217 struct cfq_queue
*cfqq
)
1220 cfq_log_cfqq(cfqd
, cfqq
, "set_active");
1221 cfqq
->slice_end
= 0;
1222 cfqq
->slice_dispatch
= 0;
1224 cfq_clear_cfqq_wait_request(cfqq
);
1225 cfq_clear_cfqq_must_dispatch(cfqq
);
1226 cfq_clear_cfqq_must_alloc_slice(cfqq
);
1227 cfq_clear_cfqq_fifo_expire(cfqq
);
1228 cfq_mark_cfqq_slice_new(cfqq
);
1230 del_timer(&cfqd
->idle_slice_timer
);
1233 cfqd
->active_queue
= cfqq
;
1237 * current cfqq expired its slice (or was too idle), select new one
1240 __cfq_slice_expired(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1243 cfq_log_cfqq(cfqd
, cfqq
, "slice expired t=%d", timed_out
);
1245 if (cfq_cfqq_wait_request(cfqq
))
1246 del_timer(&cfqd
->idle_slice_timer
);
1248 cfq_clear_cfqq_wait_request(cfqq
);
1251 * store what was left of this slice, if the queue idled/timed out
1253 if (timed_out
&& !cfq_cfqq_slice_new(cfqq
)) {
1254 cfqq
->slice_resid
= cfqq
->slice_end
- jiffies
;
1255 cfq_log_cfqq(cfqd
, cfqq
, "resid=%ld", cfqq
->slice_resid
);
1258 if (cfq_cfqq_on_rr(cfqq
) && RB_EMPTY_ROOT(&cfqq
->sort_list
))
1259 cfq_del_cfqq_rr(cfqd
, cfqq
);
1261 cfq_resort_rr_list(cfqd
, cfqq
);
1263 if (cfqq
== cfqd
->active_queue
)
1264 cfqd
->active_queue
= NULL
;
1266 if (cfqd
->active_cic
) {
1267 put_io_context(cfqd
->active_cic
->ioc
);
1268 cfqd
->active_cic
= NULL
;
1272 static inline void cfq_slice_expired(struct cfq_data
*cfqd
, bool timed_out
)
1274 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1277 __cfq_slice_expired(cfqd
, cfqq
, timed_out
);
1281 * Get next queue for service. Unless we have a queue preemption,
1282 * we'll simply select the first cfqq in the service tree.
1284 static struct cfq_queue
*cfq_get_next_queue(struct cfq_data
*cfqd
)
1286 struct cfq_rb_root
*service_tree
=
1287 service_tree_for(cfqd
->serving_group
, cfqd
->serving_prio
,
1288 cfqd
->serving_type
, cfqd
);
1290 if (!cfqd
->rq_queued
)
1293 /* There is nothing to dispatch */
1296 if (RB_EMPTY_ROOT(&service_tree
->rb
))
1298 return cfq_rb_first(service_tree
);
1301 static struct cfq_queue
*cfq_get_next_queue_forced(struct cfq_data
*cfqd
)
1303 struct cfq_group
*cfqg
= &cfqd
->root_group
;
1304 struct cfq_queue
*cfqq
;
1306 struct cfq_rb_root
*st
;
1308 if (!cfqd
->rq_queued
)
1311 for_each_cfqg_st(cfqg
, i
, j
, st
)
1312 if ((cfqq
= cfq_rb_first(st
)) != NULL
)
1318 * Get and set a new active queue for service.
1320 static struct cfq_queue
*cfq_set_active_queue(struct cfq_data
*cfqd
,
1321 struct cfq_queue
*cfqq
)
1324 cfqq
= cfq_get_next_queue(cfqd
);
1326 __cfq_set_active_queue(cfqd
, cfqq
);
1330 static inline sector_t
cfq_dist_from_last(struct cfq_data
*cfqd
,
1333 if (blk_rq_pos(rq
) >= cfqd
->last_position
)
1334 return blk_rq_pos(rq
) - cfqd
->last_position
;
1336 return cfqd
->last_position
- blk_rq_pos(rq
);
1339 #define CFQQ_SEEK_THR 8 * 1024
1340 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1342 static inline int cfq_rq_close(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
1345 sector_t sdist
= cfqq
->seek_mean
;
1347 if (!sample_valid(cfqq
->seek_samples
))
1348 sdist
= CFQQ_SEEK_THR
;
1350 return cfq_dist_from_last(cfqd
, rq
) <= sdist
;
1353 static struct cfq_queue
*cfqq_close(struct cfq_data
*cfqd
,
1354 struct cfq_queue
*cur_cfqq
)
1356 struct rb_root
*root
= &cfqd
->prio_trees
[cur_cfqq
->org_ioprio
];
1357 struct rb_node
*parent
, *node
;
1358 struct cfq_queue
*__cfqq
;
1359 sector_t sector
= cfqd
->last_position
;
1361 if (RB_EMPTY_ROOT(root
))
1365 * First, if we find a request starting at the end of the last
1366 * request, choose it.
1368 __cfqq
= cfq_prio_tree_lookup(cfqd
, root
, sector
, &parent
, NULL
);
1373 * If the exact sector wasn't found, the parent of the NULL leaf
1374 * will contain the closest sector.
1376 __cfqq
= rb_entry(parent
, struct cfq_queue
, p_node
);
1377 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1380 if (blk_rq_pos(__cfqq
->next_rq
) < sector
)
1381 node
= rb_next(&__cfqq
->p_node
);
1383 node
= rb_prev(&__cfqq
->p_node
);
1387 __cfqq
= rb_entry(node
, struct cfq_queue
, p_node
);
1388 if (cfq_rq_close(cfqd
, cur_cfqq
, __cfqq
->next_rq
))
1396 * cur_cfqq - passed in so that we don't decide that the current queue is
1397 * closely cooperating with itself.
1399 * So, basically we're assuming that that cur_cfqq has dispatched at least
1400 * one request, and that cfqd->last_position reflects a position on the disk
1401 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1404 static struct cfq_queue
*cfq_close_cooperator(struct cfq_data
*cfqd
,
1405 struct cfq_queue
*cur_cfqq
)
1407 struct cfq_queue
*cfqq
;
1409 if (!cfq_cfqq_sync(cur_cfqq
))
1411 if (CFQQ_SEEKY(cur_cfqq
))
1415 * We should notice if some of the queues are cooperating, eg
1416 * working closely on the same area of the disk. In that case,
1417 * we can group them together and don't waste time idling.
1419 cfqq
= cfqq_close(cfqd
, cur_cfqq
);
1424 * It only makes sense to merge sync queues.
1426 if (!cfq_cfqq_sync(cfqq
))
1428 if (CFQQ_SEEKY(cfqq
))
1432 * Do not merge queues of different priority classes
1434 if (cfq_class_rt(cfqq
) != cfq_class_rt(cur_cfqq
))
1441 * Determine whether we should enforce idle window for this queue.
1444 static bool cfq_should_idle(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1446 enum wl_prio_t prio
= cfqq_prio(cfqq
);
1447 struct cfq_rb_root
*service_tree
= cfqq
->service_tree
;
1449 BUG_ON(!service_tree
);
1450 BUG_ON(!service_tree
->count
);
1452 /* We never do for idle class queues. */
1453 if (prio
== IDLE_WORKLOAD
)
1456 /* We do for queues that were marked with idle window flag. */
1457 if (cfq_cfqq_idle_window(cfqq
))
1461 * Otherwise, we do only if they are the last ones
1462 * in their service tree.
1464 return service_tree
->count
== 1;
1467 static void cfq_arm_slice_timer(struct cfq_data
*cfqd
)
1469 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
1470 struct cfq_io_context
*cic
;
1474 * SSD device without seek penalty, disable idling. But only do so
1475 * for devices that support queuing, otherwise we still have a problem
1476 * with sync vs async workloads.
1478 if (blk_queue_nonrot(cfqd
->queue
) && cfqd
->hw_tag
)
1481 WARN_ON(!RB_EMPTY_ROOT(&cfqq
->sort_list
));
1482 WARN_ON(cfq_cfqq_slice_new(cfqq
));
1485 * idle is disabled, either manually or by past process history
1487 if (!cfqd
->cfq_slice_idle
|| !cfq_should_idle(cfqd
, cfqq
))
1491 * still active requests from this queue, don't idle
1493 if (cfqq
->dispatched
)
1497 * task has exited, don't wait
1499 cic
= cfqd
->active_cic
;
1500 if (!cic
|| !atomic_read(&cic
->ioc
->nr_tasks
))
1504 * If our average think time is larger than the remaining time
1505 * slice, then don't idle. This avoids overrunning the allotted
1508 if (sample_valid(cic
->ttime_samples
) &&
1509 (cfqq
->slice_end
- jiffies
< cic
->ttime_mean
))
1512 cfq_mark_cfqq_wait_request(cfqq
);
1514 sl
= cfqd
->cfq_slice_idle
;
1516 mod_timer(&cfqd
->idle_slice_timer
, jiffies
+ sl
);
1517 cfq_log_cfqq(cfqd
, cfqq
, "arm_idle: %lu", sl
);
1521 * Move request from internal lists to the request queue dispatch list.
1523 static void cfq_dispatch_insert(struct request_queue
*q
, struct request
*rq
)
1525 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1526 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
1528 cfq_log_cfqq(cfqd
, cfqq
, "dispatch_insert");
1530 cfqq
->next_rq
= cfq_find_next_rq(cfqd
, cfqq
, rq
);
1531 cfq_remove_request(rq
);
1533 elv_dispatch_sort(q
, rq
);
1535 if (cfq_cfqq_sync(cfqq
))
1536 cfqd
->sync_flight
++;
1540 * return expired entry, or NULL to just start from scratch in rbtree
1542 static struct request
*cfq_check_fifo(struct cfq_queue
*cfqq
)
1544 struct request
*rq
= NULL
;
1546 if (cfq_cfqq_fifo_expire(cfqq
))
1549 cfq_mark_cfqq_fifo_expire(cfqq
);
1551 if (list_empty(&cfqq
->fifo
))
1554 rq
= rq_entry_fifo(cfqq
->fifo
.next
);
1555 if (time_before(jiffies
, rq_fifo_time(rq
)))
1558 cfq_log_cfqq(cfqq
->cfqd
, cfqq
, "fifo=%p", rq
);
1563 cfq_prio_to_maxrq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1565 const int base_rq
= cfqd
->cfq_slice_async_rq
;
1567 WARN_ON(cfqq
->ioprio
>= IOPRIO_BE_NR
);
1569 return 2 * (base_rq
+ base_rq
* (CFQ_PRIO_LISTS
- 1 - cfqq
->ioprio
));
1573 * Must be called with the queue_lock held.
1575 static int cfqq_process_refs(struct cfq_queue
*cfqq
)
1577 int process_refs
, io_refs
;
1579 io_refs
= cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
];
1580 process_refs
= atomic_read(&cfqq
->ref
) - io_refs
;
1581 BUG_ON(process_refs
< 0);
1582 return process_refs
;
1585 static void cfq_setup_merge(struct cfq_queue
*cfqq
, struct cfq_queue
*new_cfqq
)
1587 int process_refs
, new_process_refs
;
1588 struct cfq_queue
*__cfqq
;
1590 /* Avoid a circular list and skip interim queue merges */
1591 while ((__cfqq
= new_cfqq
->new_cfqq
)) {
1597 process_refs
= cfqq_process_refs(cfqq
);
1599 * If the process for the cfqq has gone away, there is no
1600 * sense in merging the queues.
1602 if (process_refs
== 0)
1606 * Merge in the direction of the lesser amount of work.
1608 new_process_refs
= cfqq_process_refs(new_cfqq
);
1609 if (new_process_refs
>= process_refs
) {
1610 cfqq
->new_cfqq
= new_cfqq
;
1611 atomic_add(process_refs
, &new_cfqq
->ref
);
1613 new_cfqq
->new_cfqq
= cfqq
;
1614 atomic_add(new_process_refs
, &cfqq
->ref
);
1618 static enum wl_type_t
cfq_choose_wl(struct cfq_data
*cfqd
,
1619 struct cfq_group
*cfqg
, enum wl_prio_t prio
,
1622 struct cfq_queue
*queue
;
1624 bool key_valid
= false;
1625 unsigned long lowest_key
= 0;
1626 enum wl_type_t cur_best
= SYNC_NOIDLE_WORKLOAD
;
1630 * When priorities switched, we prefer starting
1631 * from SYNC_NOIDLE (first choice), or just SYNC
1634 if (service_tree_for(cfqg
, prio
, cur_best
, cfqd
)->count
)
1636 cur_best
= SYNC_WORKLOAD
;
1637 if (service_tree_for(cfqg
, prio
, cur_best
, cfqd
)->count
)
1640 return ASYNC_WORKLOAD
;
1643 for (i
= 0; i
< 3; ++i
) {
1644 /* otherwise, select the one with lowest rb_key */
1645 queue
= cfq_rb_first(service_tree_for(cfqg
, prio
, i
, cfqd
));
1647 (!key_valid
|| time_before(queue
->rb_key
, lowest_key
))) {
1648 lowest_key
= queue
->rb_key
;
1657 static void choose_service_tree(struct cfq_data
*cfqd
, struct cfq_group
*cfqg
)
1659 enum wl_prio_t previous_prio
= cfqd
->serving_prio
;
1663 struct cfq_rb_root
*st
;
1664 unsigned group_slice
;
1667 cfqd
->serving_prio
= IDLE_WORKLOAD
;
1668 cfqd
->workload_expires
= jiffies
+ 1;
1672 /* Choose next priority. RT > BE > IDLE */
1673 if (cfq_group_busy_queues_wl(RT_WORKLOAD
, cfqd
, cfqg
))
1674 cfqd
->serving_prio
= RT_WORKLOAD
;
1675 else if (cfq_group_busy_queues_wl(BE_WORKLOAD
, cfqd
, cfqg
))
1676 cfqd
->serving_prio
= BE_WORKLOAD
;
1678 cfqd
->serving_prio
= IDLE_WORKLOAD
;
1679 cfqd
->workload_expires
= jiffies
+ 1;
1684 * For RT and BE, we have to choose also the type
1685 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1688 prio_changed
= (cfqd
->serving_prio
!= previous_prio
);
1689 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
,
1694 * If priority didn't change, check workload expiration,
1695 * and that we still have other queues ready
1697 if (!prio_changed
&& count
&&
1698 !time_after(jiffies
, cfqd
->workload_expires
))
1701 /* otherwise select new workload type */
1702 cfqd
->serving_type
=
1703 cfq_choose_wl(cfqd
, cfqg
, cfqd
->serving_prio
, prio_changed
);
1704 st
= service_tree_for(cfqg
, cfqd
->serving_prio
, cfqd
->serving_type
,
1709 * the workload slice is computed as a fraction of target latency
1710 * proportional to the number of queues in that workload, over
1711 * all the queues in the same priority class
1713 group_slice
= cfq_group_slice(cfqd
, cfqg
);
1715 slice
= group_slice
* count
/
1716 max_t(unsigned, cfqg
->busy_queues_avg
[cfqd
->serving_prio
],
1717 cfq_group_busy_queues_wl(cfqd
->serving_prio
, cfqd
, cfqg
));
1719 if (cfqd
->serving_type
== ASYNC_WORKLOAD
)
1720 /* async workload slice is scaled down according to
1721 * the sync/async slice ratio. */
1722 slice
= slice
* cfqd
->cfq_slice
[0] / cfqd
->cfq_slice
[1];
1724 /* sync workload slice is at least 2 * cfq_slice_idle */
1725 slice
= max(slice
, 2 * cfqd
->cfq_slice_idle
);
1727 slice
= max_t(unsigned, slice
, CFQ_MIN_TT
);
1728 cfqd
->workload_expires
= jiffies
+ slice
;
1729 cfqd
->noidle_tree_requires_idle
= false;
1732 static struct cfq_group
*cfq_get_next_cfqg(struct cfq_data
*cfqd
)
1734 struct cfq_rb_root
*st
= &cfqd
->grp_service_tree
;
1735 struct cfq_group
*cfqg
;
1737 if (RB_EMPTY_ROOT(&st
->rb
))
1739 cfqg
= cfq_rb_first_group(st
);
1740 st
->active
= &cfqg
->rb_node
;
1741 update_min_vdisktime(st
);
1745 static void cfq_choose_cfqg(struct cfq_data
*cfqd
)
1747 struct cfq_group
*cfqg
= cfq_get_next_cfqg(cfqd
);
1749 cfqd
->serving_group
= cfqg
;
1750 choose_service_tree(cfqd
, cfqg
);
1754 * Select a queue for service. If we have a current active queue,
1755 * check whether to continue servicing it, or retrieve and set a new one.
1757 static struct cfq_queue
*cfq_select_queue(struct cfq_data
*cfqd
)
1759 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
1761 cfqq
= cfqd
->active_queue
;
1765 if (!cfqd
->rq_queued
)
1768 * The active queue has run out of time, expire it and select new.
1770 if (cfq_slice_used(cfqq
) && !cfq_cfqq_must_dispatch(cfqq
))
1774 * The active queue has requests and isn't expired, allow it to
1777 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
1781 * If another queue has a request waiting within our mean seek
1782 * distance, let it run. The expire code will check for close
1783 * cooperators and put the close queue at the front of the service
1784 * tree. If possible, merge the expiring queue with the new cfqq.
1786 new_cfqq
= cfq_close_cooperator(cfqd
, cfqq
);
1788 if (!cfqq
->new_cfqq
)
1789 cfq_setup_merge(cfqq
, new_cfqq
);
1794 * No requests pending. If the active queue still has requests in
1795 * flight or is idling for a new request, allow either of these
1796 * conditions to happen (or time out) before selecting a new queue.
1798 if (timer_pending(&cfqd
->idle_slice_timer
) ||
1799 (cfqq
->dispatched
&& cfq_should_idle(cfqd
, cfqq
))) {
1805 cfq_slice_expired(cfqd
, 0);
1808 * Current queue expired. Check if we have to switch to a new
1812 cfq_choose_cfqg(cfqd
);
1814 cfqq
= cfq_set_active_queue(cfqd
, new_cfqq
);
1819 static int __cfq_forced_dispatch_cfqq(struct cfq_queue
*cfqq
)
1823 while (cfqq
->next_rq
) {
1824 cfq_dispatch_insert(cfqq
->cfqd
->queue
, cfqq
->next_rq
);
1828 BUG_ON(!list_empty(&cfqq
->fifo
));
1830 /* By default cfqq is not expired if it is empty. Do it explicitly */
1831 __cfq_slice_expired(cfqq
->cfqd
, cfqq
, 0);
1836 * Drain our current requests. Used for barriers and when switching
1837 * io schedulers on-the-fly.
1839 static int cfq_forced_dispatch(struct cfq_data
*cfqd
)
1841 struct cfq_queue
*cfqq
;
1844 while ((cfqq
= cfq_get_next_queue_forced(cfqd
)) != NULL
)
1845 dispatched
+= __cfq_forced_dispatch_cfqq(cfqq
);
1847 cfq_slice_expired(cfqd
, 0);
1848 BUG_ON(cfqd
->busy_queues
);
1850 cfq_log(cfqd
, "forced_dispatch=%d", dispatched
);
1854 static bool cfq_may_dispatch(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1856 unsigned int max_dispatch
;
1859 * Drain async requests before we start sync IO
1861 if (cfq_should_idle(cfqd
, cfqq
) && cfqd
->rq_in_driver
[BLK_RW_ASYNC
])
1865 * If this is an async queue and we have sync IO in flight, let it wait
1867 if (cfqd
->sync_flight
&& !cfq_cfqq_sync(cfqq
))
1870 max_dispatch
= cfqd
->cfq_quantum
;
1871 if (cfq_class_idle(cfqq
))
1875 * Does this cfqq already have too much IO in flight?
1877 if (cfqq
->dispatched
>= max_dispatch
) {
1879 * idle queue must always only have a single IO in flight
1881 if (cfq_class_idle(cfqq
))
1885 * We have other queues, don't allow more IO from this one
1887 if (cfqd
->busy_queues
> 1)
1891 * Sole queue user, no limit
1897 * Async queues must wait a bit before being allowed dispatch.
1898 * We also ramp up the dispatch depth gradually for async IO,
1899 * based on the last sync IO we serviced
1901 if (!cfq_cfqq_sync(cfqq
) && cfqd
->cfq_latency
) {
1902 unsigned long last_sync
= jiffies
- cfqd
->last_end_sync_rq
;
1905 depth
= last_sync
/ cfqd
->cfq_slice
[1];
1906 if (!depth
&& !cfqq
->dispatched
)
1908 if (depth
< max_dispatch
)
1909 max_dispatch
= depth
;
1913 * If we're below the current max, allow a dispatch
1915 return cfqq
->dispatched
< max_dispatch
;
1919 * Dispatch a request from cfqq, moving them to the request queue
1922 static bool cfq_dispatch_request(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
1926 BUG_ON(RB_EMPTY_ROOT(&cfqq
->sort_list
));
1928 if (!cfq_may_dispatch(cfqd
, cfqq
))
1932 * follow expired path, else get first next available
1934 rq
= cfq_check_fifo(cfqq
);
1939 * insert request into driver dispatch list
1941 cfq_dispatch_insert(cfqd
->queue
, rq
);
1943 if (!cfqd
->active_cic
) {
1944 struct cfq_io_context
*cic
= RQ_CIC(rq
);
1946 atomic_long_inc(&cic
->ioc
->refcount
);
1947 cfqd
->active_cic
= cic
;
1954 * Find the cfqq that we need to service and move a request from that to the
1957 static int cfq_dispatch_requests(struct request_queue
*q
, int force
)
1959 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
1960 struct cfq_queue
*cfqq
;
1962 if (!cfqd
->busy_queues
)
1965 if (unlikely(force
))
1966 return cfq_forced_dispatch(cfqd
);
1968 cfqq
= cfq_select_queue(cfqd
);
1973 * Dispatch a request from this cfqq, if it is allowed
1975 if (!cfq_dispatch_request(cfqd
, cfqq
))
1978 cfqq
->slice_dispatch
++;
1979 cfq_clear_cfqq_must_dispatch(cfqq
);
1982 * expire an async queue immediately if it has used up its slice. idle
1983 * queue always expire after 1 dispatch round.
1985 if (cfqd
->busy_queues
> 1 && ((!cfq_cfqq_sync(cfqq
) &&
1986 cfqq
->slice_dispatch
>= cfq_prio_to_maxrq(cfqd
, cfqq
)) ||
1987 cfq_class_idle(cfqq
))) {
1988 cfqq
->slice_end
= jiffies
+ 1;
1989 cfq_slice_expired(cfqd
, 0);
1992 cfq_log_cfqq(cfqd
, cfqq
, "dispatched a request");
1997 * task holds one reference to the queue, dropped when task exits. each rq
1998 * in-flight on this queue also holds a reference, dropped when rq is freed.
2000 * queue lock must be held here.
2002 static void cfq_put_queue(struct cfq_queue
*cfqq
)
2004 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2006 BUG_ON(atomic_read(&cfqq
->ref
) <= 0);
2008 if (!atomic_dec_and_test(&cfqq
->ref
))
2011 cfq_log_cfqq(cfqd
, cfqq
, "put_queue");
2012 BUG_ON(rb_first(&cfqq
->sort_list
));
2013 BUG_ON(cfqq
->allocated
[READ
] + cfqq
->allocated
[WRITE
]);
2015 if (unlikely(cfqd
->active_queue
== cfqq
)) {
2016 __cfq_slice_expired(cfqd
, cfqq
, 0);
2017 cfq_schedule_dispatch(cfqd
);
2020 BUG_ON(cfq_cfqq_on_rr(cfqq
));
2021 kmem_cache_free(cfq_pool
, cfqq
);
2025 * Must always be called with the rcu_read_lock() held
2028 __call_for_each_cic(struct io_context
*ioc
,
2029 void (*func
)(struct io_context
*, struct cfq_io_context
*))
2031 struct cfq_io_context
*cic
;
2032 struct hlist_node
*n
;
2034 hlist_for_each_entry_rcu(cic
, n
, &ioc
->cic_list
, cic_list
)
2039 * Call func for each cic attached to this ioc.
2042 call_for_each_cic(struct io_context
*ioc
,
2043 void (*func
)(struct io_context
*, struct cfq_io_context
*))
2046 __call_for_each_cic(ioc
, func
);
2050 static void cfq_cic_free_rcu(struct rcu_head
*head
)
2052 struct cfq_io_context
*cic
;
2054 cic
= container_of(head
, struct cfq_io_context
, rcu_head
);
2056 kmem_cache_free(cfq_ioc_pool
, cic
);
2057 elv_ioc_count_dec(cfq_ioc_count
);
2061 * CFQ scheduler is exiting, grab exit lock and check
2062 * the pending io context count. If it hits zero,
2063 * complete ioc_gone and set it back to NULL
2065 spin_lock(&ioc_gone_lock
);
2066 if (ioc_gone
&& !elv_ioc_count_read(cfq_ioc_count
)) {
2070 spin_unlock(&ioc_gone_lock
);
2074 static void cfq_cic_free(struct cfq_io_context
*cic
)
2076 call_rcu(&cic
->rcu_head
, cfq_cic_free_rcu
);
2079 static void cic_free_func(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2081 unsigned long flags
;
2083 BUG_ON(!cic
->dead_key
);
2085 spin_lock_irqsave(&ioc
->lock
, flags
);
2086 radix_tree_delete(&ioc
->radix_root
, cic
->dead_key
);
2087 hlist_del_rcu(&cic
->cic_list
);
2088 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2094 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2095 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2096 * and ->trim() which is called with the task lock held
2098 static void cfq_free_io_context(struct io_context
*ioc
)
2101 * ioc->refcount is zero here, or we are called from elv_unregister(),
2102 * so no more cic's are allowed to be linked into this ioc. So it
2103 * should be ok to iterate over the known list, we will see all cic's
2104 * since no new ones are added.
2106 __call_for_each_cic(ioc
, cic_free_func
);
2109 static void cfq_exit_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2111 struct cfq_queue
*__cfqq
, *next
;
2113 if (unlikely(cfqq
== cfqd
->active_queue
)) {
2114 __cfq_slice_expired(cfqd
, cfqq
, 0);
2115 cfq_schedule_dispatch(cfqd
);
2119 * If this queue was scheduled to merge with another queue, be
2120 * sure to drop the reference taken on that queue (and others in
2121 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2123 __cfqq
= cfqq
->new_cfqq
;
2125 if (__cfqq
== cfqq
) {
2126 WARN(1, "cfqq->new_cfqq loop detected\n");
2129 next
= __cfqq
->new_cfqq
;
2130 cfq_put_queue(__cfqq
);
2134 cfq_put_queue(cfqq
);
2137 static void __cfq_exit_single_io_context(struct cfq_data
*cfqd
,
2138 struct cfq_io_context
*cic
)
2140 struct io_context
*ioc
= cic
->ioc
;
2142 list_del_init(&cic
->queue_list
);
2145 * Make sure key == NULL is seen for dead queues
2148 cic
->dead_key
= (unsigned long) cic
->key
;
2151 if (ioc
->ioc_data
== cic
)
2152 rcu_assign_pointer(ioc
->ioc_data
, NULL
);
2154 if (cic
->cfqq
[BLK_RW_ASYNC
]) {
2155 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_ASYNC
]);
2156 cic
->cfqq
[BLK_RW_ASYNC
] = NULL
;
2159 if (cic
->cfqq
[BLK_RW_SYNC
]) {
2160 cfq_exit_cfqq(cfqd
, cic
->cfqq
[BLK_RW_SYNC
]);
2161 cic
->cfqq
[BLK_RW_SYNC
] = NULL
;
2165 static void cfq_exit_single_io_context(struct io_context
*ioc
,
2166 struct cfq_io_context
*cic
)
2168 struct cfq_data
*cfqd
= cic
->key
;
2171 struct request_queue
*q
= cfqd
->queue
;
2172 unsigned long flags
;
2174 spin_lock_irqsave(q
->queue_lock
, flags
);
2177 * Ensure we get a fresh copy of the ->key to prevent
2178 * race between exiting task and queue
2180 smp_read_barrier_depends();
2182 __cfq_exit_single_io_context(cfqd
, cic
);
2184 spin_unlock_irqrestore(q
->queue_lock
, flags
);
2189 * The process that ioc belongs to has exited, we need to clean up
2190 * and put the internal structures we have that belongs to that process.
2192 static void cfq_exit_io_context(struct io_context
*ioc
)
2194 call_for_each_cic(ioc
, cfq_exit_single_io_context
);
2197 static struct cfq_io_context
*
2198 cfq_alloc_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
2200 struct cfq_io_context
*cic
;
2202 cic
= kmem_cache_alloc_node(cfq_ioc_pool
, gfp_mask
| __GFP_ZERO
,
2205 cic
->last_end_request
= jiffies
;
2206 INIT_LIST_HEAD(&cic
->queue_list
);
2207 INIT_HLIST_NODE(&cic
->cic_list
);
2208 cic
->dtor
= cfq_free_io_context
;
2209 cic
->exit
= cfq_exit_io_context
;
2210 elv_ioc_count_inc(cfq_ioc_count
);
2216 static void cfq_init_prio_data(struct cfq_queue
*cfqq
, struct io_context
*ioc
)
2218 struct task_struct
*tsk
= current
;
2221 if (!cfq_cfqq_prio_changed(cfqq
))
2224 ioprio_class
= IOPRIO_PRIO_CLASS(ioc
->ioprio
);
2225 switch (ioprio_class
) {
2227 printk(KERN_ERR
"cfq: bad prio %x\n", ioprio_class
);
2228 case IOPRIO_CLASS_NONE
:
2230 * no prio set, inherit CPU scheduling settings
2232 cfqq
->ioprio
= task_nice_ioprio(tsk
);
2233 cfqq
->ioprio_class
= task_nice_ioclass(tsk
);
2235 case IOPRIO_CLASS_RT
:
2236 cfqq
->ioprio
= task_ioprio(ioc
);
2237 cfqq
->ioprio_class
= IOPRIO_CLASS_RT
;
2239 case IOPRIO_CLASS_BE
:
2240 cfqq
->ioprio
= task_ioprio(ioc
);
2241 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2243 case IOPRIO_CLASS_IDLE
:
2244 cfqq
->ioprio_class
= IOPRIO_CLASS_IDLE
;
2246 cfq_clear_cfqq_idle_window(cfqq
);
2251 * keep track of original prio settings in case we have to temporarily
2252 * elevate the priority of this queue
2254 cfqq
->org_ioprio
= cfqq
->ioprio
;
2255 cfqq
->org_ioprio_class
= cfqq
->ioprio_class
;
2256 cfq_clear_cfqq_prio_changed(cfqq
);
2259 static void changed_ioprio(struct io_context
*ioc
, struct cfq_io_context
*cic
)
2261 struct cfq_data
*cfqd
= cic
->key
;
2262 struct cfq_queue
*cfqq
;
2263 unsigned long flags
;
2265 if (unlikely(!cfqd
))
2268 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2270 cfqq
= cic
->cfqq
[BLK_RW_ASYNC
];
2272 struct cfq_queue
*new_cfqq
;
2273 new_cfqq
= cfq_get_queue(cfqd
, BLK_RW_ASYNC
, cic
->ioc
,
2276 cic
->cfqq
[BLK_RW_ASYNC
] = new_cfqq
;
2277 cfq_put_queue(cfqq
);
2281 cfqq
= cic
->cfqq
[BLK_RW_SYNC
];
2283 cfq_mark_cfqq_prio_changed(cfqq
);
2285 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2288 static void cfq_ioc_set_ioprio(struct io_context
*ioc
)
2290 call_for_each_cic(ioc
, changed_ioprio
);
2291 ioc
->ioprio_changed
= 0;
2294 static void cfq_init_cfqq(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2295 pid_t pid
, bool is_sync
)
2297 RB_CLEAR_NODE(&cfqq
->rb_node
);
2298 RB_CLEAR_NODE(&cfqq
->p_node
);
2299 INIT_LIST_HEAD(&cfqq
->fifo
);
2301 atomic_set(&cfqq
->ref
, 0);
2304 cfq_mark_cfqq_prio_changed(cfqq
);
2307 if (!cfq_class_idle(cfqq
))
2308 cfq_mark_cfqq_idle_window(cfqq
);
2309 cfq_mark_cfqq_sync(cfqq
);
2314 static void cfq_link_cfqq_cfqg(struct cfq_queue
*cfqq
, struct cfq_group
*cfqg
)
2319 static struct cfq_group
*cfq_get_cfqg(struct cfq_data
*cfqd
, int create
)
2321 return &cfqd
->root_group
;
2324 static struct cfq_queue
*
2325 cfq_find_alloc_queue(struct cfq_data
*cfqd
, bool is_sync
,
2326 struct io_context
*ioc
, gfp_t gfp_mask
)
2328 struct cfq_queue
*cfqq
, *new_cfqq
= NULL
;
2329 struct cfq_io_context
*cic
;
2330 struct cfq_group
*cfqg
;
2333 cfqg
= cfq_get_cfqg(cfqd
, 1);
2334 cic
= cfq_cic_lookup(cfqd
, ioc
);
2335 /* cic always exists here */
2336 cfqq
= cic_to_cfqq(cic
, is_sync
);
2339 * Always try a new alloc if we fell back to the OOM cfqq
2340 * originally, since it should just be a temporary situation.
2342 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
2347 } else if (gfp_mask
& __GFP_WAIT
) {
2348 spin_unlock_irq(cfqd
->queue
->queue_lock
);
2349 new_cfqq
= kmem_cache_alloc_node(cfq_pool
,
2350 gfp_mask
| __GFP_ZERO
,
2352 spin_lock_irq(cfqd
->queue
->queue_lock
);
2356 cfqq
= kmem_cache_alloc_node(cfq_pool
,
2357 gfp_mask
| __GFP_ZERO
,
2362 cfq_init_cfqq(cfqd
, cfqq
, current
->pid
, is_sync
);
2363 cfq_init_prio_data(cfqq
, ioc
);
2364 cfq_link_cfqq_cfqg(cfqq
, cfqg
);
2365 cfq_log_cfqq(cfqd
, cfqq
, "alloced");
2367 cfqq
= &cfqd
->oom_cfqq
;
2371 kmem_cache_free(cfq_pool
, new_cfqq
);
2376 static struct cfq_queue
**
2377 cfq_async_queue_prio(struct cfq_data
*cfqd
, int ioprio_class
, int ioprio
)
2379 switch (ioprio_class
) {
2380 case IOPRIO_CLASS_RT
:
2381 return &cfqd
->async_cfqq
[0][ioprio
];
2382 case IOPRIO_CLASS_BE
:
2383 return &cfqd
->async_cfqq
[1][ioprio
];
2384 case IOPRIO_CLASS_IDLE
:
2385 return &cfqd
->async_idle_cfqq
;
2391 static struct cfq_queue
*
2392 cfq_get_queue(struct cfq_data
*cfqd
, bool is_sync
, struct io_context
*ioc
,
2395 const int ioprio
= task_ioprio(ioc
);
2396 const int ioprio_class
= task_ioprio_class(ioc
);
2397 struct cfq_queue
**async_cfqq
= NULL
;
2398 struct cfq_queue
*cfqq
= NULL
;
2401 async_cfqq
= cfq_async_queue_prio(cfqd
, ioprio_class
, ioprio
);
2406 cfqq
= cfq_find_alloc_queue(cfqd
, is_sync
, ioc
, gfp_mask
);
2409 * pin the queue now that it's allocated, scheduler exit will prune it
2411 if (!is_sync
&& !(*async_cfqq
)) {
2412 atomic_inc(&cfqq
->ref
);
2416 atomic_inc(&cfqq
->ref
);
2421 * We drop cfq io contexts lazily, so we may find a dead one.
2424 cfq_drop_dead_cic(struct cfq_data
*cfqd
, struct io_context
*ioc
,
2425 struct cfq_io_context
*cic
)
2427 unsigned long flags
;
2429 WARN_ON(!list_empty(&cic
->queue_list
));
2431 spin_lock_irqsave(&ioc
->lock
, flags
);
2433 BUG_ON(ioc
->ioc_data
== cic
);
2435 radix_tree_delete(&ioc
->radix_root
, (unsigned long) cfqd
);
2436 hlist_del_rcu(&cic
->cic_list
);
2437 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2442 static struct cfq_io_context
*
2443 cfq_cic_lookup(struct cfq_data
*cfqd
, struct io_context
*ioc
)
2445 struct cfq_io_context
*cic
;
2446 unsigned long flags
;
2455 * we maintain a last-hit cache, to avoid browsing over the tree
2457 cic
= rcu_dereference(ioc
->ioc_data
);
2458 if (cic
&& cic
->key
== cfqd
) {
2464 cic
= radix_tree_lookup(&ioc
->radix_root
, (unsigned long) cfqd
);
2468 /* ->key must be copied to avoid race with cfq_exit_queue() */
2471 cfq_drop_dead_cic(cfqd
, ioc
, cic
);
2476 spin_lock_irqsave(&ioc
->lock
, flags
);
2477 rcu_assign_pointer(ioc
->ioc_data
, cic
);
2478 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2486 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2487 * the process specific cfq io context when entered from the block layer.
2488 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2490 static int cfq_cic_link(struct cfq_data
*cfqd
, struct io_context
*ioc
,
2491 struct cfq_io_context
*cic
, gfp_t gfp_mask
)
2493 unsigned long flags
;
2496 ret
= radix_tree_preload(gfp_mask
);
2501 spin_lock_irqsave(&ioc
->lock
, flags
);
2502 ret
= radix_tree_insert(&ioc
->radix_root
,
2503 (unsigned long) cfqd
, cic
);
2505 hlist_add_head_rcu(&cic
->cic_list
, &ioc
->cic_list
);
2506 spin_unlock_irqrestore(&ioc
->lock
, flags
);
2508 radix_tree_preload_end();
2511 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
2512 list_add(&cic
->queue_list
, &cfqd
->cic_list
);
2513 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
2518 printk(KERN_ERR
"cfq: cic link failed!\n");
2524 * Setup general io context and cfq io context. There can be several cfq
2525 * io contexts per general io context, if this process is doing io to more
2526 * than one device managed by cfq.
2528 static struct cfq_io_context
*
2529 cfq_get_io_context(struct cfq_data
*cfqd
, gfp_t gfp_mask
)
2531 struct io_context
*ioc
= NULL
;
2532 struct cfq_io_context
*cic
;
2534 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2536 ioc
= get_io_context(gfp_mask
, cfqd
->queue
->node
);
2540 cic
= cfq_cic_lookup(cfqd
, ioc
);
2544 cic
= cfq_alloc_io_context(cfqd
, gfp_mask
);
2548 if (cfq_cic_link(cfqd
, ioc
, cic
, gfp_mask
))
2552 smp_read_barrier_depends();
2553 if (unlikely(ioc
->ioprio_changed
))
2554 cfq_ioc_set_ioprio(ioc
);
2560 put_io_context(ioc
);
2565 cfq_update_io_thinktime(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
)
2567 unsigned long elapsed
= jiffies
- cic
->last_end_request
;
2568 unsigned long ttime
= min(elapsed
, 2UL * cfqd
->cfq_slice_idle
);
2570 cic
->ttime_samples
= (7*cic
->ttime_samples
+ 256) / 8;
2571 cic
->ttime_total
= (7*cic
->ttime_total
+ 256*ttime
) / 8;
2572 cic
->ttime_mean
= (cic
->ttime_total
+ 128) / cic
->ttime_samples
;
2576 cfq_update_io_seektime(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2582 if (!cfqq
->last_request_pos
)
2584 else if (cfqq
->last_request_pos
< blk_rq_pos(rq
))
2585 sdist
= blk_rq_pos(rq
) - cfqq
->last_request_pos
;
2587 sdist
= cfqq
->last_request_pos
- blk_rq_pos(rq
);
2590 * Don't allow the seek distance to get too large from the
2591 * odd fragment, pagein, etc
2593 if (cfqq
->seek_samples
<= 60) /* second&third seek */
2594 sdist
= min(sdist
, (cfqq
->seek_mean
* 4) + 2*1024*1024);
2596 sdist
= min(sdist
, (cfqq
->seek_mean
* 4) + 2*1024*64);
2598 cfqq
->seek_samples
= (7*cfqq
->seek_samples
+ 256) / 8;
2599 cfqq
->seek_total
= (7*cfqq
->seek_total
+ (u64
)256*sdist
) / 8;
2600 total
= cfqq
->seek_total
+ (cfqq
->seek_samples
/2);
2601 do_div(total
, cfqq
->seek_samples
);
2602 cfqq
->seek_mean
= (sector_t
)total
;
2605 * If this cfqq is shared between multiple processes, check to
2606 * make sure that those processes are still issuing I/Os within
2607 * the mean seek distance. If not, it may be time to break the
2608 * queues apart again.
2610 if (cfq_cfqq_coop(cfqq
)) {
2611 if (CFQQ_SEEKY(cfqq
) && !cfqq
->seeky_start
)
2612 cfqq
->seeky_start
= jiffies
;
2613 else if (!CFQQ_SEEKY(cfqq
))
2614 cfqq
->seeky_start
= 0;
2619 * Disable idle window if the process thinks too long or seeks so much that
2623 cfq_update_idle_window(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2624 struct cfq_io_context
*cic
)
2626 int old_idle
, enable_idle
;
2629 * Don't idle for async or idle io prio class
2631 if (!cfq_cfqq_sync(cfqq
) || cfq_class_idle(cfqq
))
2634 enable_idle
= old_idle
= cfq_cfqq_idle_window(cfqq
);
2636 if (cfqq
->queued
[0] + cfqq
->queued
[1] >= 4)
2637 cfq_mark_cfqq_deep(cfqq
);
2639 if (!atomic_read(&cic
->ioc
->nr_tasks
) || !cfqd
->cfq_slice_idle
||
2640 (!cfq_cfqq_deep(cfqq
) && sample_valid(cfqq
->seek_samples
)
2641 && CFQQ_SEEKY(cfqq
)))
2643 else if (sample_valid(cic
->ttime_samples
)) {
2644 if (cic
->ttime_mean
> cfqd
->cfq_slice_idle
)
2650 if (old_idle
!= enable_idle
) {
2651 cfq_log_cfqq(cfqd
, cfqq
, "idle=%d", enable_idle
);
2653 cfq_mark_cfqq_idle_window(cfqq
);
2655 cfq_clear_cfqq_idle_window(cfqq
);
2660 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2661 * no or if we aren't sure, a 1 will cause a preempt.
2664 cfq_should_preempt(struct cfq_data
*cfqd
, struct cfq_queue
*new_cfqq
,
2667 struct cfq_queue
*cfqq
;
2669 cfqq
= cfqd
->active_queue
;
2673 if (cfq_slice_used(cfqq
))
2676 if (cfq_class_idle(new_cfqq
))
2679 if (cfq_class_idle(cfqq
))
2682 /* Allow preemption only if we are idling on sync-noidle tree */
2683 if (cfqd
->serving_type
== SYNC_NOIDLE_WORKLOAD
&&
2684 cfqq_type(new_cfqq
) == SYNC_NOIDLE_WORKLOAD
&&
2685 new_cfqq
->service_tree
->count
== 2 &&
2686 RB_EMPTY_ROOT(&cfqq
->sort_list
))
2690 * if the new request is sync, but the currently running queue is
2691 * not, let the sync request have priority.
2693 if (rq_is_sync(rq
) && !cfq_cfqq_sync(cfqq
))
2697 * So both queues are sync. Let the new request get disk time if
2698 * it's a metadata request and the current queue is doing regular IO.
2700 if (rq_is_meta(rq
) && !cfqq
->meta_pending
)
2704 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2706 if (cfq_class_rt(new_cfqq
) && !cfq_class_rt(cfqq
))
2709 if (!cfqd
->active_cic
|| !cfq_cfqq_wait_request(cfqq
))
2713 * if this request is as-good as one we would expect from the
2714 * current cfqq, let it preempt
2716 if (cfq_rq_close(cfqd
, cfqq
, rq
))
2723 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2724 * let it have half of its nominal slice.
2726 static void cfq_preempt_queue(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
)
2728 cfq_log_cfqq(cfqd
, cfqq
, "preempt");
2729 cfq_slice_expired(cfqd
, 1);
2732 * Put the new queue at the front of the of the current list,
2733 * so we know that it will be selected next.
2735 BUG_ON(!cfq_cfqq_on_rr(cfqq
));
2737 cfq_service_tree_add(cfqd
, cfqq
, 1);
2739 cfqq
->slice_end
= 0;
2740 cfq_mark_cfqq_slice_new(cfqq
);
2744 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2745 * something we should do about it
2748 cfq_rq_enqueued(struct cfq_data
*cfqd
, struct cfq_queue
*cfqq
,
2751 struct cfq_io_context
*cic
= RQ_CIC(rq
);
2755 cfqq
->meta_pending
++;
2757 cfq_update_io_thinktime(cfqd
, cic
);
2758 cfq_update_io_seektime(cfqd
, cfqq
, rq
);
2759 cfq_update_idle_window(cfqd
, cfqq
, cic
);
2761 cfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
2763 if (cfqq
== cfqd
->active_queue
) {
2765 * Remember that we saw a request from this process, but
2766 * don't start queuing just yet. Otherwise we risk seeing lots
2767 * of tiny requests, because we disrupt the normal plugging
2768 * and merging. If the request is already larger than a single
2769 * page, let it rip immediately. For that case we assume that
2770 * merging is already done. Ditto for a busy system that
2771 * has other work pending, don't risk delaying until the
2772 * idle timer unplug to continue working.
2774 if (cfq_cfqq_wait_request(cfqq
)) {
2775 if (blk_rq_bytes(rq
) > PAGE_CACHE_SIZE
||
2776 cfqd
->busy_queues
> 1) {
2777 del_timer(&cfqd
->idle_slice_timer
);
2778 __blk_run_queue(cfqd
->queue
);
2780 cfq_mark_cfqq_must_dispatch(cfqq
);
2782 } else if (cfq_should_preempt(cfqd
, cfqq
, rq
)) {
2784 * not the active queue - expire current slice if it is
2785 * idle and has expired it's mean thinktime or this new queue
2786 * has some old slice time left and is of higher priority or
2787 * this new queue is RT and the current one is BE
2789 cfq_preempt_queue(cfqd
, cfqq
);
2790 __blk_run_queue(cfqd
->queue
);
2794 static void cfq_insert_request(struct request_queue
*q
, struct request
*rq
)
2796 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2797 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2799 cfq_log_cfqq(cfqd
, cfqq
, "insert_request");
2800 cfq_init_prio_data(cfqq
, RQ_CIC(rq
)->ioc
);
2802 rq_set_fifo_time(rq
, jiffies
+ cfqd
->cfq_fifo_expire
[rq_is_sync(rq
)]);
2803 list_add_tail(&rq
->queuelist
, &cfqq
->fifo
);
2806 cfq_rq_enqueued(cfqd
, cfqq
, rq
);
2810 * Update hw_tag based on peak queue depth over 50 samples under
2813 static void cfq_update_hw_tag(struct cfq_data
*cfqd
)
2815 struct cfq_queue
*cfqq
= cfqd
->active_queue
;
2817 if (rq_in_driver(cfqd
) > cfqd
->hw_tag_est_depth
)
2818 cfqd
->hw_tag_est_depth
= rq_in_driver(cfqd
);
2820 if (cfqd
->hw_tag
== 1)
2823 if (cfqd
->rq_queued
<= CFQ_HW_QUEUE_MIN
&&
2824 rq_in_driver(cfqd
) <= CFQ_HW_QUEUE_MIN
)
2828 * If active queue hasn't enough requests and can idle, cfq might not
2829 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2832 if (cfqq
&& cfq_cfqq_idle_window(cfqq
) &&
2833 cfqq
->dispatched
+ cfqq
->queued
[0] + cfqq
->queued
[1] <
2834 CFQ_HW_QUEUE_MIN
&& rq_in_driver(cfqd
) < CFQ_HW_QUEUE_MIN
)
2837 if (cfqd
->hw_tag_samples
++ < 50)
2840 if (cfqd
->hw_tag_est_depth
>= CFQ_HW_QUEUE_MIN
)
2846 static void cfq_completed_request(struct request_queue
*q
, struct request
*rq
)
2848 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2849 struct cfq_data
*cfqd
= cfqq
->cfqd
;
2850 const int sync
= rq_is_sync(rq
);
2854 cfq_log_cfqq(cfqd
, cfqq
, "complete");
2856 cfq_update_hw_tag(cfqd
);
2858 WARN_ON(!cfqd
->rq_in_driver
[sync
]);
2859 WARN_ON(!cfqq
->dispatched
);
2860 cfqd
->rq_in_driver
[sync
]--;
2863 if (cfq_cfqq_sync(cfqq
))
2864 cfqd
->sync_flight
--;
2867 RQ_CIC(rq
)->last_end_request
= now
;
2868 cfqd
->last_end_sync_rq
= now
;
2872 * If this is the active queue, check if it needs to be expired,
2873 * or if we want to idle in case it has no pending requests.
2875 if (cfqd
->active_queue
== cfqq
) {
2876 const bool cfqq_empty
= RB_EMPTY_ROOT(&cfqq
->sort_list
);
2878 if (cfq_cfqq_slice_new(cfqq
)) {
2879 cfq_set_prio_slice(cfqd
, cfqq
);
2880 cfq_clear_cfqq_slice_new(cfqq
);
2883 * Idling is not enabled on:
2885 * - idle-priority queues
2887 * - queues with still some requests queued
2888 * - when there is a close cooperator
2890 if (cfq_slice_used(cfqq
) || cfq_class_idle(cfqq
))
2891 cfq_slice_expired(cfqd
, 1);
2892 else if (sync
&& cfqq_empty
&&
2893 !cfq_close_cooperator(cfqd
, cfqq
)) {
2894 cfqd
->noidle_tree_requires_idle
|= !rq_noidle(rq
);
2896 * Idling is enabled for SYNC_WORKLOAD.
2897 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
2898 * only if we processed at least one !rq_noidle request
2900 if (cfqd
->serving_type
== SYNC_WORKLOAD
2901 || cfqd
->noidle_tree_requires_idle
)
2902 cfq_arm_slice_timer(cfqd
);
2906 if (!rq_in_driver(cfqd
))
2907 cfq_schedule_dispatch(cfqd
);
2911 * we temporarily boost lower priority queues if they are holding fs exclusive
2912 * resources. they are boosted to normal prio (CLASS_BE/4)
2914 static void cfq_prio_boost(struct cfq_queue
*cfqq
)
2916 if (has_fs_excl()) {
2918 * boost idle prio on transactions that would lock out other
2919 * users of the filesystem
2921 if (cfq_class_idle(cfqq
))
2922 cfqq
->ioprio_class
= IOPRIO_CLASS_BE
;
2923 if (cfqq
->ioprio
> IOPRIO_NORM
)
2924 cfqq
->ioprio
= IOPRIO_NORM
;
2927 * unboost the queue (if needed)
2929 cfqq
->ioprio_class
= cfqq
->org_ioprio_class
;
2930 cfqq
->ioprio
= cfqq
->org_ioprio
;
2934 static inline int __cfq_may_queue(struct cfq_queue
*cfqq
)
2936 if (cfq_cfqq_wait_request(cfqq
) && !cfq_cfqq_must_alloc_slice(cfqq
)) {
2937 cfq_mark_cfqq_must_alloc_slice(cfqq
);
2938 return ELV_MQUEUE_MUST
;
2941 return ELV_MQUEUE_MAY
;
2944 static int cfq_may_queue(struct request_queue
*q
, int rw
)
2946 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
2947 struct task_struct
*tsk
= current
;
2948 struct cfq_io_context
*cic
;
2949 struct cfq_queue
*cfqq
;
2952 * don't force setup of a queue from here, as a call to may_queue
2953 * does not necessarily imply that a request actually will be queued.
2954 * so just lookup a possibly existing queue, or return 'may queue'
2957 cic
= cfq_cic_lookup(cfqd
, tsk
->io_context
);
2959 return ELV_MQUEUE_MAY
;
2961 cfqq
= cic_to_cfqq(cic
, rw_is_sync(rw
));
2963 cfq_init_prio_data(cfqq
, cic
->ioc
);
2964 cfq_prio_boost(cfqq
);
2966 return __cfq_may_queue(cfqq
);
2969 return ELV_MQUEUE_MAY
;
2973 * queue lock held here
2975 static void cfq_put_request(struct request
*rq
)
2977 struct cfq_queue
*cfqq
= RQ_CFQQ(rq
);
2980 const int rw
= rq_data_dir(rq
);
2982 BUG_ON(!cfqq
->allocated
[rw
]);
2983 cfqq
->allocated
[rw
]--;
2985 put_io_context(RQ_CIC(rq
)->ioc
);
2987 rq
->elevator_private
= NULL
;
2988 rq
->elevator_private2
= NULL
;
2990 cfq_put_queue(cfqq
);
2994 static struct cfq_queue
*
2995 cfq_merge_cfqqs(struct cfq_data
*cfqd
, struct cfq_io_context
*cic
,
2996 struct cfq_queue
*cfqq
)
2998 cfq_log_cfqq(cfqd
, cfqq
, "merging with queue %p", cfqq
->new_cfqq
);
2999 cic_set_cfqq(cic
, cfqq
->new_cfqq
, 1);
3000 cfq_mark_cfqq_coop(cfqq
->new_cfqq
);
3001 cfq_put_queue(cfqq
);
3002 return cic_to_cfqq(cic
, 1);
3005 static int should_split_cfqq(struct cfq_queue
*cfqq
)
3007 if (cfqq
->seeky_start
&&
3008 time_after(jiffies
, cfqq
->seeky_start
+ CFQQ_COOP_TOUT
))
3014 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3015 * was the last process referring to said cfqq.
3017 static struct cfq_queue
*
3018 split_cfqq(struct cfq_io_context
*cic
, struct cfq_queue
*cfqq
)
3020 if (cfqq_process_refs(cfqq
) == 1) {
3021 cfqq
->seeky_start
= 0;
3022 cfqq
->pid
= current
->pid
;
3023 cfq_clear_cfqq_coop(cfqq
);
3027 cic_set_cfqq(cic
, NULL
, 1);
3028 cfq_put_queue(cfqq
);
3032 * Allocate cfq data structures associated with this request.
3035 cfq_set_request(struct request_queue
*q
, struct request
*rq
, gfp_t gfp_mask
)
3037 struct cfq_data
*cfqd
= q
->elevator
->elevator_data
;
3038 struct cfq_io_context
*cic
;
3039 const int rw
= rq_data_dir(rq
);
3040 const bool is_sync
= rq_is_sync(rq
);
3041 struct cfq_queue
*cfqq
;
3042 unsigned long flags
;
3044 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3046 cic
= cfq_get_io_context(cfqd
, gfp_mask
);
3048 spin_lock_irqsave(q
->queue_lock
, flags
);
3054 cfqq
= cic_to_cfqq(cic
, is_sync
);
3055 if (!cfqq
|| cfqq
== &cfqd
->oom_cfqq
) {
3056 cfqq
= cfq_get_queue(cfqd
, is_sync
, cic
->ioc
, gfp_mask
);
3057 cic_set_cfqq(cic
, cfqq
, is_sync
);
3060 * If the queue was seeky for too long, break it apart.
3062 if (cfq_cfqq_coop(cfqq
) && should_split_cfqq(cfqq
)) {
3063 cfq_log_cfqq(cfqd
, cfqq
, "breaking apart cfqq");
3064 cfqq
= split_cfqq(cic
, cfqq
);
3070 * Check to see if this queue is scheduled to merge with
3071 * another, closely cooperating queue. The merging of
3072 * queues happens here as it must be done in process context.
3073 * The reference on new_cfqq was taken in merge_cfqqs.
3076 cfqq
= cfq_merge_cfqqs(cfqd
, cic
, cfqq
);
3079 cfqq
->allocated
[rw
]++;
3080 atomic_inc(&cfqq
->ref
);
3082 spin_unlock_irqrestore(q
->queue_lock
, flags
);
3084 rq
->elevator_private
= cic
;
3085 rq
->elevator_private2
= cfqq
;
3090 put_io_context(cic
->ioc
);
3092 cfq_schedule_dispatch(cfqd
);
3093 spin_unlock_irqrestore(q
->queue_lock
, flags
);
3094 cfq_log(cfqd
, "set_request fail");
3098 static void cfq_kick_queue(struct work_struct
*work
)
3100 struct cfq_data
*cfqd
=
3101 container_of(work
, struct cfq_data
, unplug_work
);
3102 struct request_queue
*q
= cfqd
->queue
;
3104 spin_lock_irq(q
->queue_lock
);
3105 __blk_run_queue(cfqd
->queue
);
3106 spin_unlock_irq(q
->queue_lock
);
3110 * Timer running if the active_queue is currently idling inside its time slice
3112 static void cfq_idle_slice_timer(unsigned long data
)
3114 struct cfq_data
*cfqd
= (struct cfq_data
*) data
;
3115 struct cfq_queue
*cfqq
;
3116 unsigned long flags
;
3119 cfq_log(cfqd
, "idle timer fired");
3121 spin_lock_irqsave(cfqd
->queue
->queue_lock
, flags
);
3123 cfqq
= cfqd
->active_queue
;
3128 * We saw a request before the queue expired, let it through
3130 if (cfq_cfqq_must_dispatch(cfqq
))
3136 if (cfq_slice_used(cfqq
))
3140 * only expire and reinvoke request handler, if there are
3141 * other queues with pending requests
3143 if (!cfqd
->busy_queues
)
3147 * not expired and it has a request pending, let it dispatch
3149 if (!RB_EMPTY_ROOT(&cfqq
->sort_list
))
3153 * Queue depth flag is reset only when the idle didn't succeed
3155 cfq_clear_cfqq_deep(cfqq
);
3158 cfq_slice_expired(cfqd
, timed_out
);
3160 cfq_schedule_dispatch(cfqd
);
3162 spin_unlock_irqrestore(cfqd
->queue
->queue_lock
, flags
);
3165 static void cfq_shutdown_timer_wq(struct cfq_data
*cfqd
)
3167 del_timer_sync(&cfqd
->idle_slice_timer
);
3168 cancel_work_sync(&cfqd
->unplug_work
);
3171 static void cfq_put_async_queues(struct cfq_data
*cfqd
)
3175 for (i
= 0; i
< IOPRIO_BE_NR
; i
++) {
3176 if (cfqd
->async_cfqq
[0][i
])
3177 cfq_put_queue(cfqd
->async_cfqq
[0][i
]);
3178 if (cfqd
->async_cfqq
[1][i
])
3179 cfq_put_queue(cfqd
->async_cfqq
[1][i
]);
3182 if (cfqd
->async_idle_cfqq
)
3183 cfq_put_queue(cfqd
->async_idle_cfqq
);
3186 static void cfq_exit_queue(struct elevator_queue
*e
)
3188 struct cfq_data
*cfqd
= e
->elevator_data
;
3189 struct request_queue
*q
= cfqd
->queue
;
3191 cfq_shutdown_timer_wq(cfqd
);
3193 spin_lock_irq(q
->queue_lock
);
3195 if (cfqd
->active_queue
)
3196 __cfq_slice_expired(cfqd
, cfqd
->active_queue
, 0);
3198 while (!list_empty(&cfqd
->cic_list
)) {
3199 struct cfq_io_context
*cic
= list_entry(cfqd
->cic_list
.next
,
3200 struct cfq_io_context
,
3203 __cfq_exit_single_io_context(cfqd
, cic
);
3206 cfq_put_async_queues(cfqd
);
3208 spin_unlock_irq(q
->queue_lock
);
3210 cfq_shutdown_timer_wq(cfqd
);
3215 static void *cfq_init_queue(struct request_queue
*q
)
3217 struct cfq_data
*cfqd
;
3219 struct cfq_group
*cfqg
;
3220 struct cfq_rb_root
*st
;
3222 cfqd
= kmalloc_node(sizeof(*cfqd
), GFP_KERNEL
| __GFP_ZERO
, q
->node
);
3226 /* Init root service tree */
3227 cfqd
->grp_service_tree
= CFQ_RB_ROOT
;
3229 /* Init root group */
3230 cfqg
= &cfqd
->root_group
;
3231 for_each_cfqg_st(cfqg
, i
, j
, st
)
3233 RB_CLEAR_NODE(&cfqg
->rb_node
);
3235 /* Give preference to root group over other groups */
3236 cfqg
->weight
= 2*BLKIO_WEIGHT_DEFAULT
;
3239 * Not strictly needed (since RB_ROOT just clears the node and we
3240 * zeroed cfqd on alloc), but better be safe in case someone decides
3241 * to add magic to the rb code
3243 for (i
= 0; i
< CFQ_PRIO_LISTS
; i
++)
3244 cfqd
->prio_trees
[i
] = RB_ROOT
;
3247 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3248 * Grab a permanent reference to it, so that the normal code flow
3249 * will not attempt to free it.
3251 cfq_init_cfqq(cfqd
, &cfqd
->oom_cfqq
, 1, 0);
3252 atomic_inc(&cfqd
->oom_cfqq
.ref
);
3253 cfq_link_cfqq_cfqg(&cfqd
->oom_cfqq
, &cfqd
->root_group
);
3255 INIT_LIST_HEAD(&cfqd
->cic_list
);
3259 init_timer(&cfqd
->idle_slice_timer
);
3260 cfqd
->idle_slice_timer
.function
= cfq_idle_slice_timer
;
3261 cfqd
->idle_slice_timer
.data
= (unsigned long) cfqd
;
3263 INIT_WORK(&cfqd
->unplug_work
, cfq_kick_queue
);
3265 cfqd
->cfq_quantum
= cfq_quantum
;
3266 cfqd
->cfq_fifo_expire
[0] = cfq_fifo_expire
[0];
3267 cfqd
->cfq_fifo_expire
[1] = cfq_fifo_expire
[1];
3268 cfqd
->cfq_back_max
= cfq_back_max
;
3269 cfqd
->cfq_back_penalty
= cfq_back_penalty
;
3270 cfqd
->cfq_slice
[0] = cfq_slice_async
;
3271 cfqd
->cfq_slice
[1] = cfq_slice_sync
;
3272 cfqd
->cfq_slice_async_rq
= cfq_slice_async_rq
;
3273 cfqd
->cfq_slice_idle
= cfq_slice_idle
;
3274 cfqd
->cfq_latency
= 1;
3276 cfqd
->last_end_sync_rq
= jiffies
;
3280 static void cfq_slab_kill(void)
3283 * Caller already ensured that pending RCU callbacks are completed,
3284 * so we should have no busy allocations at this point.
3287 kmem_cache_destroy(cfq_pool
);
3289 kmem_cache_destroy(cfq_ioc_pool
);
3292 static int __init
cfq_slab_setup(void)
3294 cfq_pool
= KMEM_CACHE(cfq_queue
, 0);
3298 cfq_ioc_pool
= KMEM_CACHE(cfq_io_context
, 0);
3309 * sysfs parts below -->
3312 cfq_var_show(unsigned int var
, char *page
)
3314 return sprintf(page
, "%d\n", var
);
3318 cfq_var_store(unsigned int *var
, const char *page
, size_t count
)
3320 char *p
= (char *) page
;
3322 *var
= simple_strtoul(p
, &p
, 10);
3326 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3327 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3329 struct cfq_data *cfqd = e->elevator_data; \
3330 unsigned int __data = __VAR; \
3332 __data = jiffies_to_msecs(__data); \
3333 return cfq_var_show(__data, (page)); \
3335 SHOW_FUNCTION(cfq_quantum_show
, cfqd
->cfq_quantum
, 0);
3336 SHOW_FUNCTION(cfq_fifo_expire_sync_show
, cfqd
->cfq_fifo_expire
[1], 1);
3337 SHOW_FUNCTION(cfq_fifo_expire_async_show
, cfqd
->cfq_fifo_expire
[0], 1);
3338 SHOW_FUNCTION(cfq_back_seek_max_show
, cfqd
->cfq_back_max
, 0);
3339 SHOW_FUNCTION(cfq_back_seek_penalty_show
, cfqd
->cfq_back_penalty
, 0);
3340 SHOW_FUNCTION(cfq_slice_idle_show
, cfqd
->cfq_slice_idle
, 1);
3341 SHOW_FUNCTION(cfq_slice_sync_show
, cfqd
->cfq_slice
[1], 1);
3342 SHOW_FUNCTION(cfq_slice_async_show
, cfqd
->cfq_slice
[0], 1);
3343 SHOW_FUNCTION(cfq_slice_async_rq_show
, cfqd
->cfq_slice_async_rq
, 0);
3344 SHOW_FUNCTION(cfq_low_latency_show
, cfqd
->cfq_latency
, 0);
3345 #undef SHOW_FUNCTION
3347 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3348 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3350 struct cfq_data *cfqd = e->elevator_data; \
3351 unsigned int __data; \
3352 int ret = cfq_var_store(&__data, (page), count); \
3353 if (__data < (MIN)) \
3355 else if (__data > (MAX)) \
3358 *(__PTR) = msecs_to_jiffies(__data); \
3360 *(__PTR) = __data; \
3363 STORE_FUNCTION(cfq_quantum_store
, &cfqd
->cfq_quantum
, 1, UINT_MAX
, 0);
3364 STORE_FUNCTION(cfq_fifo_expire_sync_store
, &cfqd
->cfq_fifo_expire
[1], 1,
3366 STORE_FUNCTION(cfq_fifo_expire_async_store
, &cfqd
->cfq_fifo_expire
[0], 1,
3368 STORE_FUNCTION(cfq_back_seek_max_store
, &cfqd
->cfq_back_max
, 0, UINT_MAX
, 0);
3369 STORE_FUNCTION(cfq_back_seek_penalty_store
, &cfqd
->cfq_back_penalty
, 1,
3371 STORE_FUNCTION(cfq_slice_idle_store
, &cfqd
->cfq_slice_idle
, 0, UINT_MAX
, 1);
3372 STORE_FUNCTION(cfq_slice_sync_store
, &cfqd
->cfq_slice
[1], 1, UINT_MAX
, 1);
3373 STORE_FUNCTION(cfq_slice_async_store
, &cfqd
->cfq_slice
[0], 1, UINT_MAX
, 1);
3374 STORE_FUNCTION(cfq_slice_async_rq_store
, &cfqd
->cfq_slice_async_rq
, 1,
3376 STORE_FUNCTION(cfq_low_latency_store
, &cfqd
->cfq_latency
, 0, 1, 0);
3377 #undef STORE_FUNCTION
3379 #define CFQ_ATTR(name) \
3380 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3382 static struct elv_fs_entry cfq_attrs
[] = {
3384 CFQ_ATTR(fifo_expire_sync
),
3385 CFQ_ATTR(fifo_expire_async
),
3386 CFQ_ATTR(back_seek_max
),
3387 CFQ_ATTR(back_seek_penalty
),
3388 CFQ_ATTR(slice_sync
),
3389 CFQ_ATTR(slice_async
),
3390 CFQ_ATTR(slice_async_rq
),
3391 CFQ_ATTR(slice_idle
),
3392 CFQ_ATTR(low_latency
),
3396 static struct elevator_type iosched_cfq
= {
3398 .elevator_merge_fn
= cfq_merge
,
3399 .elevator_merged_fn
= cfq_merged_request
,
3400 .elevator_merge_req_fn
= cfq_merged_requests
,
3401 .elevator_allow_merge_fn
= cfq_allow_merge
,
3402 .elevator_dispatch_fn
= cfq_dispatch_requests
,
3403 .elevator_add_req_fn
= cfq_insert_request
,
3404 .elevator_activate_req_fn
= cfq_activate_request
,
3405 .elevator_deactivate_req_fn
= cfq_deactivate_request
,
3406 .elevator_queue_empty_fn
= cfq_queue_empty
,
3407 .elevator_completed_req_fn
= cfq_completed_request
,
3408 .elevator_former_req_fn
= elv_rb_former_request
,
3409 .elevator_latter_req_fn
= elv_rb_latter_request
,
3410 .elevator_set_req_fn
= cfq_set_request
,
3411 .elevator_put_req_fn
= cfq_put_request
,
3412 .elevator_may_queue_fn
= cfq_may_queue
,
3413 .elevator_init_fn
= cfq_init_queue
,
3414 .elevator_exit_fn
= cfq_exit_queue
,
3415 .trim
= cfq_free_io_context
,
3417 .elevator_attrs
= cfq_attrs
,
3418 .elevator_name
= "cfq",
3419 .elevator_owner
= THIS_MODULE
,
3422 static int __init
cfq_init(void)
3425 * could be 0 on HZ < 1000 setups
3427 if (!cfq_slice_async
)
3428 cfq_slice_async
= 1;
3429 if (!cfq_slice_idle
)
3432 if (cfq_slab_setup())
3435 elv_register(&iosched_cfq
);
3440 static void __exit
cfq_exit(void)
3442 DECLARE_COMPLETION_ONSTACK(all_gone
);
3443 elv_unregister(&iosched_cfq
);
3444 ioc_gone
= &all_gone
;
3445 /* ioc_gone's update must be visible before reading ioc_count */
3449 * this also protects us from entering cfq_slab_kill() with
3450 * pending RCU callbacks
3452 if (elv_ioc_count_read(cfq_ioc_count
))
3453 wait_for_completion(&all_gone
);
3457 module_init(cfq_init
);
3458 module_exit(cfq_exit
);
3460 MODULE_AUTHOR("Jens Axboe");
3461 MODULE_LICENSE("GPL");
3462 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");